• Semiautonomous drone deployment: Ukrainian forces now rely on drones that can switch from pilot control to AI-guided terminal strikes, exemplified by Bumblebee and other systems achieving over 1,000 combat flights.
• Underdog module: NORDA Dynamics created an aftermarket module that enables F.P.V. drones to lock onto targets and complete terminal attacks autonomously, extending range to roughly 2 kilometers.
• Swarm coordination: Sine Engineering’s Pasika app lets a single operator autonomously launch, loiter, and transfer control among dozens of drones, creating rapid-sequence swarm strikes.
• Counter-jamming navigation: Vermeer’s Visual Positioning System equips deep-strike drones with camera-based navigation that bypasses GPS denial, guiding missions across contested airspace.
• Schmidt-backed platforms: Eric Schmidt’s venture supplies Bumblebee, Hornet, Merops, and other AI-enabled drones to elite Ukrainian brigades, combining target recognition, terminal guidance, and jamming resistance.
• Human oversight emphasis: Developers and operators maintain humans designate targets, and Ukrainian doctrine aims to keep people in the loop despite growing autonomous capabilities.
• Ethical concerns: Technologists and activists warn that AI-enabled targeting raises moral questions about dehumanized kill decisions, especially once algorithms pursue people independently.
• Strategic implications: Ukrainian leaders treat AI-enhanced drones as essential to national defense, while Western observers worry existing militaries are unprepared for the emerging autonomous warfare landscape.

---

In the past year, drone warfare in Ukraine has undergone a chilling transformation.

Most drones require a human pilot. But some new Ukrainian drones, once locked on a target, can use A.I. to chase and strike it — with no further human involvement.

This is the story of how the battlefield became the birthplace of a powerful new weapon.

On a warm morning a few months ago, Lipa, a Ukrainian drone pilot, flew a small gray quadcopter over the ravaged fields near Borysivka, a tiny occupied village abutting the Russian border. A surveillance drone had spotted signs that an enemy drone team had moved into abandoned warehouses at the village’s edge. Lipa and his navigator, Bober, intended to kill the team or drive it off.

Another pilot had twice tried hitting the place with standard kamikaze quadcopters, which are susceptible to radio-wave jamming that can disrupt the communication link between pilot and drone, causing weapons to crash. Russian jammers stopped them. Lipa had been assigned the third try but this time with a Bumblebee, an unusual drone provided by a secretive venture led by Eric Schmidt, the former chief executive of Google and one of the world’s wealthiest men.

Bober sat beside Lipa as he oriented for an attack run. From high over Borysivka, one of the Bumblebee’s two airborne cameras focused on a particular building’s eastern side. Bober checked the imagery, then a digital map, and agreed: They had found the target. “Locked in,” Lipa said.

With his right hand, Lipa toggled a switch, unleashing the drone from human control. Powered by artificial intelligence, the Bumblebee swept down without further external guidance. As it descended, it lost signal connection with Lipa and Bober. This did not matter: It continued its attack free of their command. Its sensors and software remained focused on the building and adjusted heading and speed independently.

Another drone livestreamed the result: The Bumblebee smacked into an exterior wall and exploded. Whether Russian soldiers were harmed was unclear, but a semiautonomous drone had hit where human-piloted drones missed, rendering the position untenable. “They will change their location now,” Lipa said. (Per Ukrainian security rules, soldiers are referred to by their first name or call sign.)

Throughout 2025 in the war between Russia and Ukraine, in largely unseen and unheralded moments like the warehouse strike in Borysivka, the era of killer robots has begun to take shape on the battlefield. Across the roughly 800-mile front and over the airspace of both nations, drones with newly developed autonomous features are now in daily combat use. By last spring, Bumblebees launched from Ukrainian positions had carried out more than 1,000 combat flights against Russian targets, according to a manufacturer’s pamphlet extolling the weapon’s capabilities. Pilots say they have flown thousands more since.

Bumblebee’s introduction raised immediate alarms in the Kremlin’s military circles, according to two Russian technical intelligence reports. One, based on dissection of a damaged Bumblebee collected along the front, described a mystery drone with chipsets and a motherboard “of the highest quality, matching the level of the world’s leading microelectronics manufacturers.” The report noted the sort of deficiencies expected of a prototype but ended with an ominous forecast: “Despite current limitations,” it declared, “the technology will demonstrate its effectiveness” and its range of uses “will continue to expand.”

That conclusion was prescient but understated, for the simple reason that Bumblebees hardly fly alone. Under the pressures of invasion, Ukraine has become a fast-feedback, live-fire test range in which arms manufacturers, governments, venture capitalists, frontline units and coders and engineers from around the West collaborate to produce weapons that automate parts of the conventional military kill chain. Equipped with onboard proprietary software trained on large data sets, and often run on off-the-shelf microcomputers like Raspberry Pi, drones with autonomous capabilities are now part of the war’s bloody and destructive routine.

In repeated visits to arms manufacturers, test ranges and frontline units over 18 months, I observed their development firsthand. Functions now performed autonomously include: pilotless takeoff or hovering, geolocation, navigation to areas of attack, as well as target recognition, tracking and pursuit — up to and including terminal strike, the lethal endpoint of the journey. Software designers have also networked multiple drones into a shared app that allows for flight control to be passed between human pilots or for drones to be organized into tightly sequenced attacks — a step toward computer-managed swarms. Weapons with these capabilities are in the hands of ground brigades as well as air defense, intelligence and deep-strike units.

Image

Bumblebee attack drones at a combat testing range outside Kharkiv, Ukraine.Credit...Finbarr O'Reilly for The New York Times

Drones under full human control remain far more abundant than their semiautonomous siblings. They cause most battlefield wounds. But unmanned weapons are crossing into a new frontier. And while no publicly known drone in the war automates all steps of a combat mission into a single weapon, some designers have put sequential steps under the control of artificial intelligence. “Any tactical equation that has a person in it could have A.I.,” said the founder of X-Drone, a Ukrainian company that has trained software for drones to hunt for and identify a stationary target, like an oil-storage tank, and then hit it without a pilot at the controls. (The founder asked that his name be withheld for security reasons.)

The Kremlin’s forces are also adopting A.I.-enhanced weapons, according to examinations of downed Russian drones by Conflict Armament Research, a private arms-investigation firm. With both sides investing, Mykhailo Fedorov, Ukraine’s first deputy prime minister, said A.I.-powered drones are at the center of a new arms race. Ukraine’s defenders must field them in large numbers quickly, he said, or risk defeat. “We are trying to stimulate development of every stage of autonomy,” he said. “We need to develop and buy more autonomous drones.”

To be sure, the familiar weapons of modern battlefields, all under human control, have caused immeasurable harm to generations of soldiers and civilians. Even weapons celebrated by generals and pundits as astonishingly precise, like GPS-guided missiles or laser-guided bombs, have often struck the wrong places, killing innocents, often without accountability. No golden age is being left behind. Rather, semiautonomous drones compound existing perils and present new threats. Peter Asaro, vice chair of the Stop Killer Robots Campaign and a philosopher and an associate professor at the New School, warned of rising dangers as weapons enter unmapped practical and ethical terrain. “The development of increasing autonomy in drones raises serious questions about human rights and the protection of civilians in armed conflict,” he said. “The capacity to autonomously select targets is a moral line that should not be crossed.”

The concept of a killer robot is vague and prone to hype, invoking T-800 of “The Terminator,” an adaptive mobile killing machine deployed by an artificial superintelligence, Skynet, that perceives humanity as a threat. Nothing close exists in Ukraine. “Everybody thinks, Oh, you are making Skynet,” said a captain responsible for integrating new technology into the 13th Khartia Brigade of Ukraine’s National Guard, one of the country’s most sophisticated units, in which Lipa and Bober serve. “No, the technology is interesting. But it’s a first step and there are many more steps.”

The captain and other technicians working with A.I.-enhanced weapons said they tend to be brittle, limited in function and less accurate than weapons under skilled human control. Many have a short battery life and brief flight times. Autonomous weapons with sustained endurance, high flexibility and the ability to discern, identify, rank and pursue multiple categories of targets independent of human action have yet to appear, and they would require, the captain said, “a waterfall of money” plus much imagination and time. “It’s like the staircase of the Empire State Building,” he said. “That’s how many steps there are, and we are inside the building but only on the first floor.”

As a safeguard against A.I.-powered weapons slipping the leash, humanitarians and many technologists have long advocated keeping “humans in the loop,” shorthand for preventing weapons from making homicidal decisions alone. By this thinking, a trained human must assess and approve all targets, as Lipa and Bober did, ideally with the power to abort an individual strike and a kill switch to shut an entire system down. Strong guardrails, the argument goes, are necessary for accountability, compliance with laws of armed conflict, legitimacy around military action and, ultimately, for human security.

Schmidt has emphasized the necessity of human oversight. But at the end of a flight, some semiautonomous weapons in Ukraine can already identify targets without human involvement, and many Ukrainian-made systems with human override are inexpensive and could be copied and modified by talented coders anywhere. Some of those designing A.I.-enhanced weapons, who consider their development necessary for Ukraine’s defense, confess to unease about the technology that they themselves conjure to form. “I think we created the monster,” said Nazar Bigun, a young physicist writing terminal-attack software. “And I’m not sure where it’s going to go.”

The Dawn of Autonomous Attack

Bigun’s own journey exemplifies how the duration and particulars of the war incentivized the creation of semiautonomous weapons. When Russia rolled mechanized divisions over Ukraine’s border in 2022, Bigun was leading a team of software engineers at a German tech start-up. In early 2023, he founded a volunteer initiative for the military that eventually manufactured 200 first-person-view (F.P.V.) quadcopters a month. It was a significant contribution to Ukraine’s war effort at a time when low-cost and explosive-laden hobbyist drones, not yet widely recognized as the transformative weapons they are, remained in short supply. His focus might have remained there. But as he and his peers heard from frontline drone pilots, they became concerned about declining success rates of kamikaze drones in the face of defensive adaptations, and they joined the search for solutions.

The problems were many. As more drones took flight, both sides developed physical and electronic countermeasures. Soldiers erected poles and strung mesh to snag drones from the air, and they covered the turrets and hulls of military vehicles with protective netting, grates or welded cages. Among the most frustrating countermeasures were jammers that flooded the operating frequencies used for flight control and video links, generating electronic noise that reduced signal clarity in pilot-to-drone connections. The systems became standard around high-value targets, including command bunkers and artillery positions. They also appeared on expensive mobile equipment, like air-defense systems, multiple-rocket launchers and tanks.

Image

A Ukrainian military vehicle traveling under a web of netting meant to protect against drone attacks near the frontline city Pokrovsk.Credit...Finbarr O'Reilly for The New York Times

This complex puzzle led to the creation of drones that fly on fiber-optic cables, one solution that has appeared on the battlefield. It also fueled Bigun’s interest in a form of computer-enabled attack, known as last-mile autonomous targeting, in which computer vision and autonomous flight control would guide drones through the final stage of attack without radio-signal inputs from a pilot. Such systems promised another benefit as well: They would increase the efficacy of strikes at longer range and over the radio horizon, when terrain or the Earth’s curvature interfere with a steady radio signal.

In theory, the technical remedy was simple. When pilots anticipated a break in communications, they could pass flight control to an automated substitute — a powerful chipset and extensively trained software — that would complete the mission. With this tech coupled to onboard sensors and a camera, the pilot could lock the mini-aircraft on a target and release the drone to strike alone. The company Bigun co-founded in 2024, NORDA Dynamics, did not manufacture drones, so it set to work creating an aftermarket component to attach to other manufacturers’ weapons. With it, a pilot would still fly the drone from launch until it neared a target. Then the pilot would have the option of autonomous attack.

Boosted by funding from the Ukrainian government and venture capital firms, NORDA spent much of 2024 testing a prototype that evolved into its flagship offering, Underdog, a small module that fastens to a combat drone. When flying an Underdog-equipped quadcopter, a pilot with F.P.V. goggles still controls the weapon from takeoff almost to destination. But in a flight’s final phase, the pilot has the choice — via an onscreen window that zooms in on objects of interest, like a building or car — of approving an autonomous attack in a process called pixel lock. At that moment, Underdog takes over.

Underdog began with tests on stationary objects, but after repeated updates, its software chased moving quarry. Range extended, too. Early modules allowed 400 meters of terminal attack; by summer 2025, with the fifth version of NORDA’s software, pixel lock reached 2,000 meters — about 1.25 miles. By then the modules had been distributed to collaborating F.P.V. teams at the front. “We have some very good feedback,” Bigun said. A company bulletin board listed early hits, among them Russian artillery pieces, trucks, mobile radar units and a tank.

Image

Nazar Bigun (center), co-founder and chief executive of the Ukrainian defense-tech start-up NORDA Dynamics, with colleagues on lunch break from a nearby arms expo.Credit...Finbarr O'Reilly for The New York Times

One summer afternoon in western Ukraine, Bigun and several employees arrived at a tree line of wild pear, apple and plum dividing agriculture fields. Cows meandered past, swishing tails to shoo flies. Two white storks glided to the ground, alighted and picked their way through the furrows, hunting. NORDA’s technicians sent a black S.U.V. with a driver and two-way radio to drive along the fields.

A test pilot, Janusz, a Polish citizen who had volunteered as a combat medic in Ukraine before joining the company, sat in the van wearing goggles and holding a hand-held radio controller. Once the S.U.V. drove away, he commanded an unarmed F.P.V. drone through liftoff. “I’m flying,” he said over the radio.

The video feed showed golden fields and green windbreaks, overlaid with dirt roads. The drone climbed to about 200 feet. Its camera revealed the black S.U.V. less than a mile away. Onscreen, Janusz slipped a square-shaped white cursor over the image of the vehicle. A pop-up window appeared in the upper-left corner containing a stabilized close-up of the S.U.V. With his left hand, Janusz selected pixel lock. The word “ENGAGE” appeared within a red banner onscreen. Thin black cross-hairs settled on the center of the S.U.V.

Janusz lifted his hands from the controller. From an altitude of about 215 feet, the drone entered a slow dive. Within seconds it had flown almost to the moving S.U.V.’s windshield. Janusz switched back to human piloting and banked the quadcopter away, sparing the vehicle damage from impact.

At his command, the drone climbed, spun around and resumed pursuit, this time from more than 500 feet up. Its prey bounded along a road. Janusz lined up the cursor and engaged pixel lock again. The drone entered a second independent dive, accelerating toward the fleeing car. Once more Janusz overrode the software at the last moment. The quadcopter buzzed so closely that the whine of its engines was picked up by the vehicle’s two-way radio and broadcast inside the pilot’s van. Janusz smiled.

He swung the drone around, showing the van he sat in. The cursor briefly presented the possibility of pixel lock on himself. Janusz chuckled and steered the weapon away, back toward the S.U.V. The driver’s voice crackled over the radio. “We will right now make a turn,” he said.

For the next half-hour, the driver’s maneuvers made no difference. No matter what he did, the drone, once pixel-locked, closed the distance autonomously, harassing the moving vehicle with the tenacity of an obsessed bird.

Compared with conditions common in war, the field exercise was simple. Groundspeeds were slow, flights were by daylight, no power lines or tree branches blocked the way. The drone maintained a constant line of sight with the S.U.V., and the software had to lock on a lone vehicle, not on a target weaving through traffic or passing parked cars. But with more training and computational power, the software could be improved to discern and prioritize military targets based on replacement cost or degree of menace, or fine-tuned to strike armored vehicles in vulnerable places, like exhaust grates or where turrets meet hulls. It might be trained to hunt most anything at all — a bus, a parked aircraft, a lectern where a speaker is addressing an audience, a step-down electrical transformer distributing power to a grid.

Image

An employee from NORDA Dynamics launching a Dart-2 fixed-wing strike drone for engineers to fine-tune their automated flight software. Credit...Finbarr O'Reilly for The New York Times

For Bigun, the natural worry that such technology could be turned against civilians has been overridden by the imperatives of survival. Beyond coding, his work involves interacting with arms designers from Ukraine and the West, including at weapons expositions, where he seeks partners and clients. But he often visits the Field of Mars, a cemetery in Lviv that is a repository of solemn memory and raw pain.

Bigun’s great-uncle was a Ukrainian nationalist during the totalitarian rule of Stalin. For this, Bigun’s grandfather was deemed an enemy of the state by association, and shipped to Siberia at age 16. Both men are buried on the grounds, where they have been joined by a procession of soldiers killed since the full invasion. On an evening following one of Bigun’s arms-show appearances, mourners at the field sanded tall wooden crucifixes by hand, then reapplied lacquer; a widow sat beside a grave talking to her lost husband as if he were sipping tea in an opposite chair; a family formed a semicircle around a plot covered in flowers with each member taking turns catching up a deceased soldier on household news. The field reached full capacity in December — almost 1,300 graves — prompting Lviv to open a second cemetery for its continuing flow of war dead. Just before Christmas, the second field held 14 fresh mounds.

Bigun abhors the need for these places. But beside commemorations of friends snatched early from life by the war, he said, he finds inspiration to continue his work. “This is where I feel the price we pay,” he said, “and it motivates me to move forward.” By the end of the year, NORDA Dynamics had provided frontline units fighting in the East more than 50,000 Underdog modules.

The Rise of the Swarm

The Ukrainian military’s hard pivot to drone warfare helped save the nation. For almost four years, while fielding the world’s first armed force to reorganize around unmanned weapons, it blunted the ground assaults of Russia’s far larger military. It continues to do so even as the Kremlin replenishes its thinned divisions, drawing from oil-state revenue and a population at least four times the size of Ukraine’s.

But the weapon has a limit. Almost all short-range kamikaze drones — a primary means of stopping advancing Russian soldiers — are flown by individual pilots, one at a time. Each is a vicious aerial acrobat: From airspeeds up to 70 miles an hour, small multicopters can stop, hover, turn and fly off in new directions for minutes on end, traits empowering pilots to find, chase and kill their human victims with chilling efficacy. And yet during sustained Russian attacks, typical frontline conditions can force drone teams to fight slowly. The pace is set by the duration between each drone’s launch and final approach, which at common standoff distances often stretches past 20 minutes. When Russian soldiers infiltrate in large numbers, single-drone strike sequences can feel slow and insufficient. Between sorties, enemies escape.

Given the enduring challenge of massing drone firepower, designers of autonomous combat-drone technology have sought to assemble drones into swarms, the allure of which is obvious to a nation under attack. Even small swarms would allow pilots to concentrate multiple weapons in punishing rapid-fire strikes, stiffening defenses and raising the prospect of overwhelming machine-only attacks.

Not long after Janusz’s terminal-attack flights, technicians from another Ukrainian company, Sine Engineering, gathered near a rural village to train drone pilots on its entrant to the swarm-tech field, called Pasika, the Ukrainian word for apiary. The heart of Pasika’s hardware is its radio modems — small frequency-hopping transceivers that act as beacons for flying drones. In flight, each quadcopter’s altitude and location update several times a second by measuring the differences in arrival times of radio signals from several known positions. Pasika software also provides automated flight control. At its current stage of development, a sole pilot can manage dozens of drones through autonomous launch, navigation and hovering — a pre-attack phase during which massed drones loiter pending instructions.

During a multiday training session for quadcopter teams fresh from frontline duty, Pavlo, a former infantryman who serves as Sine Engineering’s liaison to combat brigades, coached pilots as they practiced. The tech was futuristic but the scene was characteristically rural Ukrainian. The test range, hectares of hayfield and sunflower, was not secured behind fences or watched over by control towers. The students, including pilots from Kraken 1654 Unmanned Systems Regiment of the Third Army Corps and Samosud Team, a drone unit of the 11th National Guard Brigade, worked in casual clothes and colorful T-shirts, eating artisanal pizzas as they tinkered. A small horse stood tethered to a stake beside a wooden cart, chewing grass into a flat circle.

Via an app, the pilots took turns ordering drones to launch and navigate to a point on the map. Unaided by human hand, quadcopters rose in the air and sped off over the countryside to loiter together about a mile away. Tablet screens revealed their progress. A video feed from each quadcopter showed rolling cropland below. The pilots kept their hands off the controls. A rusty tractor puttered past.

Image

Soldiers attending training with Sine Engineering, a Ukrainian defense-tech company. Sine’s product, Pasika, lets a single operator manage several drones at a time, hitting targets in quick-succession “swarm attacks.”Credit...Finbarr O'Reilly for The New York Times

For a final exercise, a pilot with the call sign Kara directed her team to gather two drones autonomously over an opposite field. Another team flew three more. Once the drones reached their loitering point, Kara said, pilots would take control and fly them manually toward targets to practice a massed attack.

Pasika also allows pilots on the app to exchange control of individual drones among themselves. In this way, any pilot could use them to attack across a short distance with brief intervals between strikes. The concept could be extended further. With Pasika or a similar system, quadcopters stockpiled in boxes near a front could be commanded by a sole operator, whether A.I. or human, creating a dense swarm of drones to face ground attacks without delay.

Multiplying a sole soldier’s combat power in these ways made sense to Kara, whose brother, husband and husband’s twin brother all rose to resist the Russian invasion. Ukraine grants humanitarian discharges to siblings of service members killed in action, and when her brother-in-law was killed, her husband transferred to the reserve. Upon his return home, Kara enlisted and became a drone pilot. Her call sign means retribution.

Pavlo, too, had his motivations. In summer 2022, he was almost killed by conventional military incompetence when his commander gathered more than 300 soldiers in buildings in Apostolove, a city north of Crimea. Dense public garrisoning offers rich targets for long-range weapons, which arrived as a barrage of S-300 missiles. Pavlo was inside a middle school when the first missile exploded in the yard. He huddled with others under a stairwell. The next missile hit the school squarely, leaving him pinned under rubble as yet another roared in and exploded. At least four peers in the stairwell died. Pavlo suffered burns on his head, torso and arms. After convalescing, he trained as a drone pilot and began flying reconnaissance missions behind Russian lines. Experience told him Ukraine needed high-tech solutions to secure its future. “It woke something in me,” he said, of nearly dying because of an old-school infantry commander’s lazy mistake. “An instinct for survival on a more sophisticated level.”

Can A.I. Replace GPS?

In late 2023, Brian Streem, the founder of a niche drone-cinematography business, was meeting with a Latvian company about putting a visual navigation system on long-range drones when one of his hosts suggested he bring his ideas to Ukraine. Deep-strike drones, essentially slow-flying cruise missiles that can travel hundreds of miles, require precise navigation to move through foreign airspace for hours, and to follow zigzag routes that change altitude frequently to evade air defenses. Ukraine had high hopes for its growing deep-strike arsenal to target Russian fuel and arms depots. But Russia had shut down GPS over the front and its western territory. Vaunted GPS-reliant systems were failing. The Latvian company thought Streem had a solution.

Streem was new to the arms business, and his journey had been anything but direct. A native of Bayside, Queens, he graduated from New York University’s Tisch School of the Arts in 2010 and started producing independent films and commercials. His work indulged a fascination for difficult photographic challenges, which led him to drones and the creation of a company, Aerobo, that shot aerial footage for music videos and Hollywood productions, including “A Quiet Place,” “Spider-Man: Homecoming” and “True Detective.”

Image

Brian Streem, the founder of Vermeer, a company that develops visual positioning systems (V.P.S.) that enable drones to navigate without GPS assistance.Credit...Finbarr O'Reilly for The New York Times

Aerobo’s services were in demand. Streem might easily have settled in, but he found working in Hollywood to be stressful and sometimes stultifying. Every shoot had to be both technically perfect and aesthetically pleasing, even beautiful. “If you think military drone missions are hard,” he says, “try making Steven Spielberg happy.” Moreover, the available tech, though expensive, felt janky and resistant to graceful use. Quadcopters were just beginning to enter markets, and the larger drones Aerobo flew required at least two people — a pilot controlling the airframe and an operator moving a camera system on a gimbal, “in this kind of dance,” Streem said. Frustrated with the equipment and unsure how to scale up his business, he shuttered Aerobo in 2018, renamed the company Vermeer and started developing software to make drone cinematography more intuitive.

Revenue dried up. In 2019, feeling desperate, he attended an investor speed-dating seminar in Buffalo and sat across from Warren Katz, an entrepreneur who directed a U.S. Air Force start-up accelerator that funded military tech. Streem explained what he was working on. Katz suggested that the Air Force “would be very interested in what you’re doing.” Streem was astonished. He was not a weapons designer; one of his last professional collaborations was a music video by the rapper Cardi B. “If you’re coming to me for help,” he said, “we’re in a lot of trouble.”

“Well,” Katz said, “you might be surprised.”

Katz urged him to apply to a program for start-ups of Air Force interest. Streem hastily filled out forms. A month later, Vermeer was selected. Streem moved to Boston, began meeting officials from the Pentagon and was at once startled and pulled in. “I realized, OK, I don’t know much about A.I.,” he said. “But as I am talking to these people, I’m kind of thinking to myself, I don’t think they know much about A.I., either.”

Streem is amiable, persistent and energized by a seemingly instinctual inclination for sales. When Covid closed offices, he retreated to a lakeside cabin and created an internet-scraping tool that yielded the email addresses of more than 50,000 military officers. From social isolation, Streem started writing them to discuss what they saw as the military’s most pressing technical challenges. Answers clustered around reliance on GPS. Entire arsenals of American military equipment, he heard, depended on a satellite navigation system that a sophisticated enemy could disable. The more he learned, the more his ideas clicked into place: I may know how to solve these people’s problems, he thought.

His idea was straightforward. He would program autopilot software to process visual information from multiple cameras, compare it with onboard three-dimensional terrain maps, then triangulate to fix a weapon’s location. Called a visual positioning system, or V.P.S., the software could be loaded onto any number of flying platforms, equipping them to navigate over terrain with no satellite link at all. It could not be jammed or spoofed, because it would neither emit nor receive a signal.

Image

V.P.S. units designed by Vermeer sitting in storage. The system uses cameras to analyze an environment and match it to 3-D maps, determining precise locations even when satellite signals are jammed or unavailable.Credit...Finbarr O'Reilly for The New York Times

In early 2024, Streem rode a train from Warsaw to Kyiv and messaged Ukrainian officials on LinkedIn, introducing himself and his work. Soon he was invited to the Cabinet of Ministers building, where he made his way past the sandbagged entrance, attended an impromptu birthday party for a government employee and was led into a room with an official who wanted to put Vermeer’s modules on deep-strike drones. The building’s power was out. The office was dark. Ukraine was short on money, soldiers and time. The official did not mince words. “He essentially had a big map of Russia and Ukraine behind him, and immediately he started telling me about targets we’re going to hit,” Streem said.

Streem had talked his way into Ukraine’s war. Over the next year, he met drone manufactures around the country, tested prototypes on various drones and released the VPS-212, a roughly one-pound box with two cameras and a minicomputer. Assembled in an office beside a bagel shop in Brooklyn, the module can fix its location at speeds up to 218 m.p.h. — not fast enough for a proper cruise missile, but sufficient for most deep-strike drones. By summer 2025, Vermeer’s technicians were helping soldiers attach them to drones flown by several units that attack strategic targets in Russia. For security reasons, Vermeer does not publicly discuss specific strikes, but Streem and a Vermeer employee in Ukraine said V.P.S. modules had guided drones to verified hits.

With these results, Vermeer emerged as a winner in the scrum for contracts and funding, raising $12 million in a recent round that included the venture capital firm Draper Associates. In 2025, it attracted renewed attention from the Pentagon, which proposed attaching Vermeer’s V.P.S. to new fleets of deep-strike drones of its own. The U.S. Air Force also contracted the company to develop a similar system that would mount atop drones to look skyward and navigate celestially, like an A.I.-enabled sextant for precision flight above clouds or over water.

By then, Streem was immersed in his new line of work. Late in 2025, he sent me a tongue-in-cheek text about his passage from filmmaker to manufacturer of self-navigating camera modules that steer high-explosive payloads into Russia. “Stream reclined in his chair, vapor from his vape pen curling lazily between his fingers, as if the war beyond the walls were a distant rumor rather than the air he breathed,” the text read. The prose, he said, had been generated by A.I. It misspelled his name.

The Great Convergence

Few people could be as conversant in the promises and the perils of A.I.-enhanced weapons as Eric Schmidt, who served as chairman of the U.S. Defense Innovation Board from 2016 to 2020 and led the bipartisan National Security Commission on Artificial Intelligence from 2018 to 2021. Since working with Ukraine, he has been mostly tight-lipped about his wartime ventures, which have operated under multiple names, including White Stork, Project Eagle and Swift Beat. Through a public-relations firm, he declined to respond to multiple requests for comment for this article. But in a talk at Stanford in 2024 he called himself “a licensed arms dealer.” And by that fall, his operation had hired former Ukrainian soldiers to meet with active drone teams near Kyiv and train them on semiautonomous weapons to take to the front.

Ukrainian units vary widely in quality, and Schmidt’s team appears to have chosen carefully. Those collaborating with his operation are among Ukraine’s best managed, with reputations for innovation, battlefield savvy and sustained success against much larger Russian forces. Among them are the Khartia Brigade, which formed in 2022 to defend Kharkiv, Ukraine’s second largest city, and grew into a disciplined, tech-centric battlefield force. With an extensive fleet of aerial and ground drones, modern command posts and an ethos of data-driven decision-making, Khartia operates from hiding in the city and countryside. Its officers track Russian actions with networked ground cameras and sensors on the battlefield and reconnaissance drones in the air. The raw information is run through software producing outputs that resemble, its analysts say, “Moneyball” goes to war.

“We collect a lot of statistics and data,” said Col. Daniel Kitone, the brigade commander. “In conditions where resources are limited, we are providing for efficiency, and sometimes statistics give interesting answers that drive operations.” Rapid processing of multiple forms of surveillance data, the brigade’s officers say, can illuminate patterns, like recurring times when Russian troops resupply positions. They then use these insights to synchronize strike-drone flights with anticipated Russian movement, hoping to catch targets in the open.

Image

Khartia Brigade soldiers on shift in an underground command center in northeastern Ukraine.Credit...Finbarr O'Reilly for The New York Times

Bumblebee’s combat trials began with multiple brigades last winter. One test pilot, who flew the drone near Kupiansk in December 2024, said his team put the new quadcopters through progressively harder tests against Russian positions and vehicles, with the manufacturer’s engineers on call for technical assistance. “During this whole time, the product constantly improved,” the pilot said. “We constantly provided the developers with feedback: what works, what really does not work, which features are useful and which are not.” As with most A.I. products, more data can lead to smarter software. During Bumblebee’s quiet rollout, mission data from combat flights was logged and analyzed, several pilots said. Developers then pushed software updates to brigades that drone teams could download remotely.

One early attack hit the entrance to a root cellar where Russian soldiers sheltered. As Bumblebees evolved, pilots flew them further distances and against moving targets. In January 2025, an autonomous Bumblebee attack stopped a Russian logistics truck as it drove behind enemy lines, the test pilot said. In April, after more updates, Bumblebees flew autonomously against a Russian armored vehicle driving with a jammer. The vehicle was so covered with anti-drone protective measures that it resembled a porcupine. It absorbed the first strike and kept moving. A second Bumblebee hit immobilized it. This amounted to a milestone: A vehicle with all the protective steps Russia could muster had been removed from action.

Bumblebee’s entrance to the war had been shrouded in secrecy. But with strikes like that the hush could not last. Russia took note. Its soldiers outside Kupiansk and Kharkiv, where early Bumblebee sorties occurred, reported that these strange new drones seemed impervious to interference: They flew smoothly through jamming and kept racking up hits. The only sure way to stop them was to shoot them down.

Two days after the strike on the armored vehicle, the Center for Integrated Unmanned Solutions, a Russian drone-manufacturing business outside Moscow, issued findings from its analysis of a downed Bumblebee recovered near the front. It nicknamed the quadcopter Marsianin, Russian for “Martian,” based on the assumption that the prototypes descended from NASA’s Ingenuity program, which developed a small autonomous helicopter that flew on Mars. The report’s author declared that Ukraine had fielded an A.I.-enhanced drone capable of “operating in total radio silence” while flying complex routes and maneuvers “completely independent of navigation systems” — or even a human pilot.

The following month, the Novorossiya Assistance Coordination Center, a Russian ultranationalist organization that provides training and equipment to Russian soldiers, published a second analysis, a 49-page report loaded with warnings. Through open-source sleuthing, its author had found a photo of a Bumblebee posted by a Reddit user who in late 2024 obtained a broken specimen from garbage discarded at a Michigan National Guard facility. With that evidence, the author, Aleksandr Lyubimov, who organizes combat-drone exhibitions in Russia, echoed the first analyst’s suspicion that the Bumblebee had some connection to the United States. He noted that it “poses a serious threat” and asserted, accurately, that “its current use does not yet reveal its full capabilities and is most likely of a combat testing nature.” Nonetheless, he added, “there are no effective countermeasures against it, and none are expected in the near future.”

The Bumblebee’s resistance to jamming, according to a sales pamphlet in limited circulation in Ukraine, was tied in part to “redundant comms,” radio-wave frequency hopping and navigation by visual inertial odometry — technical solutions to signal jamming. But what made the weapon remarkable was not any one autonomous feature. It was the convergence of several.

According to the pamphlet, which claimed an “over 70-percent direct-hit rate via autonomous terminal guidance,” Schmidt’s quadcopters are also capable of autonomous target recognition. By overlaying bright green squares on items of interest appearing in video feeds, pilots said, Bumblebee’s software highlights potential targets, including foot soldiers, bunkers, vehicles and other aerial drones, often before human pilots can spot them. The combination of features, they said, result in an autonomous attack capability more robust and reliable than others available to date.

Bumblebees can also be controlled over the internet, which Lipa did in the strike in Borysivka. This keeps pilots away from the front, and from many weapons that might counterattack. In theory, as long as a Bumblebee’s ground station maintains a stable Wi-Fi or broadband connection, a pilot can operate it from almost anywhere — a capability demonstrated this past summer when Schmidt visited Kyiv and observed a Khartia team flying a Bumblebee released by a ground crew outside Kharkiv. According to a review of footage and people familiar with the mission, the drone passed over the lines and hit a four-wheel-drive Russian military van, known as a bukhanka — from roughly 300 miles away.

A Billionaire’s Growing Fleet

The Bumblebee is not a one-off project. It’s part of an experimental pack. Schmidt’s operation has also supplied Ukrainian units with a medium-range fixed-wing strike drone with a two-meter wingspan, marketed under the name Hornet, according to another sales pamphlet from early 2025. Like Bumblebee, it has A.I.-powered target recognition and terminal-attack guidance, along with jam-resistant communication and navigation systems. The pamphlet advertised an 11-pound payload, a cruise speed of 62 m.p.h. and range exceeding 90 miles. “Our A.I.-powered platform processes battlefield data in real time, adapting to changing conditions without human intervention,” the pamphlet says. “Neutralize more targets at a fraction of legacy system costs. Deploy at scale to achieve overwhelming force multiplication against sophisticated threats.” The pamphlet claimed a “future monthly production” of more than 6,000 units.

Schmidt has also become an ally in Ukraine’s defense against Shaheds, the Iranian-designed long-range drones that pummel Ukrainian cities almost nightly. In July 2025, he appeared with Ukraine’s president, Volodymyr Zelensky, to announce a strategic partnership to provide Ukraine with A.I.-powered drones, with an emphasis on an interceptor system known as Merops. Ukraine was developing its own human-piloted anti-Shahed drones, and had some early successes. But Schmidt’s weapons, Ukrainian officials said, were more effective. Mykhailo Fedorov, Ukraine’s first deputy prime minister, showed videos of them striking Shaheds at high speed. Merops had a hit rate as high as 95 percent, he said. (After its combat trials in Ukraine, Merops is now being deployed on NATO’s eastern flank.)

Ukrainian pilots and officers say Schmidt’s products have shown more promise than most, though reviews for Bumblebee have not been universally glowing. The weapon has no night cameras, limiting flights to daytime. Lyubimov, the Russian technical analyst, described the drone as inadequately weather resistant. “The design exhibits many ‘childhood flaws,’” he wrote. But Schmidt’s technicians are available on the Signal app and responsive to suggestions, said Serhii, a Bumblebee combat pilot and Khartia’s chief technical consultant. The first Bumblebees were difficult to operate, but through cycles of feedback and updates they became better. “In the beginning it didn’t fly without a professional pilot,” he said. “Now it can fly with a newbie.” Serhii said he had tested 15 semiautonomous terminal-attack drones from different manufacturers, and Bumblebee was the best. A new generation of Bumblebees is in the works, he added, with a stronger airframe and night optics.

Hardware upgrades would be welcome. But the captain who integrates new tech into Khartia’s operations said hardware was not the secret ingredient behind Bumblebee’s performance. It is the firmware and the flight software, BeeQGroundControl, that separates the drone from others. “Eric Schmidt made a very innovative drone,” he said, adding that Bumblebee is one of the only drones in Ukraine that “is ready out of the box.” Teams simply add an explosive charge and begin work.

In one coordinated kamikaze attack, Serhii said, three Bumblebees and a standard F.P.V. drone destroyed a 152-millimeter howitzer inside Russia that was protected under a bunker roofed with logs. The first Bumblebee dove into camouflage netting and set it afire; the second breached the roof. Then the standard F.P.V. plunged inside and the final Bumblebee hovered overhead and scattered 10 small anti-personnel land mines around the site. The strikes were timed two minutes or less apart. Khartia has repeated the tactic. “There have been many similar attacks,” Serhii said.

Bumblebees are so valuable, he added, that teams flying them are assigned only to important missions — principally hunts behind Russian lines for artillery and logistics vehicles, and to carry relay transmitters that extend other drones’ ranges. Other units deliver packets of blood to bunkers that medics use to stabilize wounded soldiers awaiting evacuation, an officer supervising strike-drone teams said.

For all the ways that Bumblebees have brought together multiple autonomous features, Schmidt’s engineers, Ukrainians said, have not programmed weapons for full autonomy. Like NORDA Dynamics’s Underdog, Bumblebees require a human to designate targets before attack. “My opinion is that we need to leave the final judgment to the human being,” one Bumblebee test pilot said. Schmidt has agreed. “There’s always this worry about the ‘Dr. Strangelove’ situation, where you have an automatic weapon which makes the decision on its own,” he said in a televised 2024 appearance. “That would be terrible.”

Image

A soldier known by the call sign Lipa (left) from Ukraine’s Khartia Brigade, updating the software on Bumblebee attack drones.Credit...Finbarr O'Reilly for The New York Times

In “Dr. Strangelove,” the 1964 Stanley Kubrick movie, a Soviet doomsday device detects a rogue American attack and automatically launches a nuclear response, effectively exterminating humankind. Schmidt’s caution aligns with Pentagon policy, which embraces keeping “humans in the loop,” at least in an aspirational spirit. Its 2023 guidance for autonomy in weapons systems, known as Directive 3000.09, requires “appropriate levels of human judgment over the use of force,” and senior-level official review of all autonomous systems in development or to be deployed. But the directive offers no clarity about what, exactly, constitutes “appropriate levels of human judgment.”

Further, no global consensus or convention exists for these ideas or other forms of design constraint. The arms race is afoot without mutually accepted guardrails. Schmidt has prefaced his support for keeping people in the loop by pointing out that “Russia and China do not have this doctrine,” suggesting that weapons that kill outside of human supervision could find their way to the battlefield no matter anyone’s positions or desires.

At the Edge of Total Autonomy

The compete-or-die paradigm has brought semiautonomous weapons into new territory fast. X-Drone, for example, merges multiple forms of autonomous tech onto long-range drones. Its software helps navigate the weapons to a distant area, like a seaport, then uses computer vision to identify and attack specific targets — warships, fuel-storage tanks, parked aircraft. “You fly 500 kilometers and you miss a target by a little bit, and your mission is wasted,” the company’s founder said. “Now we train on an oil tank and it hits.”

Drones with X-Drone’s software have also hit trains carrying fuel and expensive Russian air-defense radar systems, clearing routes for more drones to follow, according to the founder. Andrii, a pilot of medium-range strike drones, said he flew more than 100 A.I.-enhanced missions in 2025 with the company’s software. His work involved flying to areas where reconnaissance flights detected a valuable target, then passing control to the software for terminal attack. On a sortie this fall, he said, the drone struck a mobile air-defense system.

By late 2025, the founder said, X-Drone had provided Ukrainian units with more than 30,000 A.I.-enhanced weapons. The company is experimenting with more complex capabilities, including loading facial recognition technology into drones that could identify then kill specific people, and coupling flight-control and navigation software with large language models, or L.L.M.s, “so the drone becomes an agent,” he said. “You can literally speak to the drone, like: ‘Fly to right, 100 meters. What do we see? Do you see a window? Fly inside the window.’”

Using armed quadcopters to peer into windows or enter structures is not new. The practice is common in shattered neighborhoods at the front. Videos posted on Telegram by Ukrainian units show piloted quadcopters performing exactly such feats, slipping into occupied buildings to kill Russian soldiers within. Adding a role for A.I. on such flights could allow this particular form of violence to expand.

With A.I.-enhanced weapons, the ethical distinction between two broadly different types of strike — a drone selecting a large inanimate object for attack and a drone autonomously hunting human beings — is large. But the technical difference is smaller, and X-Drone has already crept from the inanimate to the human. X-Drone has developed A.I.-enhanced quadcopters that, its founder says, can attack Russian soldiers with or without a human in the loop. He said the software allows remote human pilots to abort auto-selected attacks. But when communications fail, human control can become impossible. In those cases, he said, the drones could hunt alone. Whether this is occurring yet is not clear.

As semiautonomous weapons train to pursue people, Asaro, of the Stop Killer Robots Campaign, warned that entering this uncharted moral frontier was deeply worrisome, because computer programs applying rules to patterns of sensor data should not determine people’s fates. “These things are going to decide who lives and who dies without any sort of access to morality,” he said, and were the essence of digital dehumanization. “They are amoral. Machines cannot fathom the distinction between inanimate objects and people.” Ukrainians involved often agree. But whether fighting in brigades or coding in company offices, as members of a population unwillingly greased in blood, they speak of ample motivation to continue their work, and of little time for regret.

X-Drone’s founder had not intended to become an arms manufacturer; it is a role he neither sought nor foresaw. Born in Soviet Russia, he worked a long career in the United States and was arranging tech deals in Kyiv when the Kremlin’s forces invaded in 2022. A physicist by training, he joined a neighborhood defense unit, which gave him a rifle. When the Russian vanguard breached the capital near his home, he turned out to meet it, then recorded videos of the bloodied bodies of soldiers his group killed. Over a meal in Kyiv in 2025, he showed one of the videos almost incredulously. “All of my career I worked in Silicon Valley and on Wall Street,” he said, “and one day I am shooting Russians near my house?”

Now the front was only a few hours’ drive away, and Russia’s missiles and long-range drones could kill anyone in Ukraine in their sleep on any night. He nodded toward a peaceful daytime street scene. “It feels normal, but it’s just not,” he said. “This is an illusion of normality.”

Wars can push everyday people to extreme positions, which for a physicist can take the form of pragmatic epiphany: With Ukraine’s defenses slowly yielding ground, and defeat meaning a return to life under Moscow’s boot, developing A.I.-enhanced drones was logical, obvious and necessary. Western militaries were far behind, he said, and Ukraine’s resistance was buying them time. Its innovations might save them, too. “Drones with A.I. are the big game-changer,” he said. “The whole military infrastructure previously is obsolete.”

‘They Should Have Stopped the War Early On’

On the range where Pavlo trained pilots on Sine Engineering’s tech, one student, Yurii, who commands an F.P.V. platoon in a frontline brigade, brushed aside philosophical discussion. He had participated in some of the war’s most prominent battles, including the incursion in 2024 into Kursk. When the full invasion began, he was a medical doctor in Western Europe. Now he killed Russian soldiers, a career deviation he insisted contained no breach of his Hippocratic oath. While practicing medicine, he prescribed antibiotics to kill microbes to save patients and stop contagions’ spread. Strikes on Russian soldiers, he offered, amounted to a similar public service. “Now we are killing bugs, too,” he said. “They’re just larger bugs.”

A front-row participant in Ukraine’s adoption of drone warfare, Yurii had seen his share of new weapons. In his view, A.I.-powered drones were inevitable. “Any large-scale war, it delivers demons,” he said. “It unleashes something powerful and it accelerates developments which otherwise would have taken decades.” World War I saw rollouts of combat aviation, tanks and artillery, alongside widespread use of chlorine, phosgene and sulfur mustard. World War II ended after the United States destroyed Hiroshima and Nagasaki with nuclear bombs. “Who knows what this war is going to unleash,” Yurii said. “If the international community is concerned about this, then they should have stopped the war early on.”

Weapons as transformative as the combat drones proliferating in Ukraine are historically uncommon. They enforce tectonic shifts in military tactics, budgets, doctrines and cultures. Organizations that adapt to the new capabilities and dangers can thrive; those that do not suffer battlefield humiliations and miseries for their rank and file. The rapid evolution of drones, now accelerating through integration with autonomy, is a moment potentially analogous to the rise of the machine gun during the Russo-Japanese War.

That new weapon’s power was demonstrated during the 1904 siege of Port Arthur. Tsarist forces were thick with peasants and drunkards. Japan’s imperial ranks were motivated and well trained. But when they attacked fortified Russian positions in massed infantry assaults, they rushed into machine guns, previously unseen in state-on-state conventional war, that shattered their lines with sweeping bursts of fire. Western military attachés were present to observe events that darkened dirt red. And yet most nations failed to take notice. European armies, oblivious to what machine guns would mean for their soldiers’ fates, continued to feed cavalry horses and to preach the glories of the open-ground charge. A decade later, hapless generals poured away young lives on the Western Front, out of step with the technology of their time.

The state of Western military readiness today galvanizes Deborah Fairlamb, a founding partner of Green Flag Ventures, a Delaware-registered venture capital fund investing in Ukrainian defense-tech start-ups. Even before autonomous drones appeared, she said, the extraordinary proliferation of unmanned weapons outran nations’ defensive abilities. “Most people in the West do not understand what is happening here,” she said, and the gap could mean stinging defeat and enormous loss of life.

Image

The military section at Cemetery No. 18 in Kharkiv, one of many burial grounds for Ukrainian soldiers killed in its war with Russia.Credit...Finbarr O'Reilly for The New York Times

Fairlamb lives in Kyiv. Her alarm sounded after American veterans of Afghanistan and Iraq visited Ukraine to prospect for business or fight as volunteers. They entered a war in which new tech has caused almost unfathomable carnage. Russia and Ukraine have suffered well over a million combined casualties in less than four years, with most wounds caused by drones. “They come back from the front, like, shaken,” she said, and they share a refrain: “My team would not last for 48 hours out there.” With A.I.-enhanced drones joining the action, Fairlamb described the need to boost A.I.-arms development as no less than existential, prompting her to approach embassies and arms manufacturers with urgency. “It really and truly is about making people understand how dramatically different this technology is,” she said. “And how unbelievably unprepared the United States is.”

For Schmidt, multiple motivations appear to overlap. At Stanford in 2024, he said he entered drone manufacturing after seeing Russian tanks destroy apartments with elderly women and children inside. His entrance to the war earned him genuine gratitude from Ukrainians, whether they fly Bumblebees or are at decreased risk every time a Merops interceptor hits a Shahed. With a year-plus of combat shock-testing of its products, his operation is also well positioned for potential profits as nations reassess and update their arsenals in light of lessons learned from Ukraine.

He has framed his movement into the A.I. arms sector as implicitly humanitarian. “Now you sit there and you go, Why would a good liberal like me do that?” he said at Stanford. “The answer is that the whole theory of armies is tanks, artilleries and mortars, and we can eliminate all of them and we can make the penalty for invading a country, at least by land, essentially be impossible.” A.I.-powered weapons, he suggested, could end this kind of warfare.

This is a prediction with precedent from when machines guns were poised to upend ground combat as people knew it. In 1877, Richard Gatling, inventor of the Gatling gun, a prominent forerunner of automatic fire, proposed that as an efficient multiplier of lethal violence his weapon might spare people the horrors of war. “It occurred to me,” he wrote, that “if I could invent a machine — a gun — which could by rapidity of fire, enable one man to do as much battle duty as a hundred, that it would, to a great extent, supersede the necessity of large armies.”

Maybe the future will prove Eric Schmidt’s vision right. Whatever is coming will reveal itself in time. History shows Gatling was spectacularly wrong.

---

Yurii Shyvala contributed reporting.

• Republican decline: New York GOP influence in Albany fell sharply after the 2018 collapse of the Independent Democratic Conference, leaving them unable to prevent Democratic supermajorities.
• Union alignment issue: Some state Republicans are cozying up to public-sector unions, making them harder to distinguish from Democrats and undermining reform efforts.
• Messaging vacuum: Unlike candidates with clear platforms, today’s state GOP lacks a coherent identity, leading to voter confusion about what the party stands for.
• Fiscal contradiction: Albany Republicans repeatedly backed large budget increases, resulting in roughly $20 billion more annual spending compared to a 2018 baseline adjusted for inflation.
• Medicaid spending pressure: GOP lawmakers joined unions like 1199 SEIU in demanding higher Medicaid payments, despite New York spending nearly $100 billion annually—about 80 percent above the national per capita average.
• Train staffing stance: Bruce Blakeman and lawmakers supported union-backed legislation requiring two-person subway crews, even though one-person operations are standard elsewhere and already used selectively in New York.
• Union resistance to reform: Long outdated MTA work rules hinder innovation and efficiency, but the TWU and similar unions strongly resist change to protect their interests.
• Opportunity for contrast: Republicans could gain ground by confronting union-driven costs and focusing on taxpayer value, yet accommodating those interests keeps them stuck without a compelling message.

---

[

](https://substackcdn.com/image/fetch/$s_!j1QU!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fffd4356c-10d7-4a04-bb41-3f7e753fd734_3000x2000.jpeg)

Courtesy Lev Radin/Pacific Press/LightRocket/Getty.

It is no secret that New York State Republicans are struggling. While there have been occasional glimmers of hope—most notably former Rep. Lee Zeldin’s relatively close loss to Gov. Kathy Hochul in 2022—these have been the exception rather than the rule.

The party’s fortunes turned decisively in 2018, when the collapse of the Independent Democratic Conference—a group of moderate Democrats in the state legislature—ended a governing arrangement that had allowed Republicans to partner with the center Left to keep the far Left sidelined. In the aftermath, Republicans have seen their influence in Albany sharply curtailed. They are now barely able to prevent Democrats from securing a supermajority in the state senate.

The state GOP could respond to this decline by making themselves a serious alternative to the Democrats, one capable of addressing New York’s most pressing problems. Instead, some state Republicans are aligning themselves with the same interest groups that have long driven those problems in the first place, most notably public-sector unions.

These Republicans may believe that cozying up to unions will make them more viable. In reality, it’ll just make them less distinguishable from Democrats and less able to actually make things better for everyday New Yorkers.

One of the New York Republican party’s core weaknesses is the absence of a well-defined platform. Regardless of the merits of his proposals, voters knew exactly what Zohran Mamdani stood for. He returned to affordability relentlessly, no matter the question, and in doing so established a coherent political identity. It didn’t matter that large portions of his program are fiscally or legally implausible—when voters went to the polls, they knew what was on offer.

By contrast, it is difficult to say what today’s New York Republican Party is actually for. Zeldin had a singular focus—crime—that showed the power of staying on message. Three years on, however, if you were to ask voters on the street what the state GOP is about, you’d be unlikely to hear a consistent answer, much less a positive one.

That confusion is not an accident. It reflects the party’s voting record and behavior in Albany, and specifically its abandonment of fiscal responsibility.

As my Manhattan Institute colleague Ken Girardin has repeatedly pointed out, Republican lawmakers have backed massive budget increases over the past several years. The state now spends roughly $20 billion more per year, in inflation-adjusted terms, than it would have if the 2018 budget had simply kept pace with inflation. Blame Medicaid and school aid for that growth, both relentlessly pushed upward by powerful unions like 1199 SEIU and the United Federation of Teachers.

A party that routinely votes for the largest drivers of spending growth cannot credibly claim to be the party of fiscal restraint. Nor can it say that it is delivering better value for taxpayers when it largely goes along with Democrats’ unprincipled spending.

New York’s GOP has increasingly been coopted by the same union forces that prevent Democrats from pursuing meaningful reforms. Girardin notes, for example, that last year Republican Senator Patrick Gallivan joined a chorus of Democrats, conducted by the 1199 SEIU healthcare workers union, to demand “Medicaid equity.” In reality, the ask is for higher payments for hospitals and other medical providers—despite the state’s spending about $100 billion a year on Medicaid, nearly 80 percent above the national average per capita, and the most of any state.

More recently, presumptive GOP gubernatorial nominee Bruce Blakeman voiced his disapproval of Hochul’s vetoing legislation that would have required two-person subway operation. The union-backed bill passed unanimously in the Assembly, and all but two Republican senators voted for it.

For context, most of the country and the world operate trains with one person: a driver, sans conductor. In a Marron Institute analysis of 400 train lines around the world, only 6 percent operated with two-person crews. The MTA utilizes one-person operations for subway shuttles and shorter lines on off hours.

If signed, this legislation would have effectively taken the issue off the table when the MTA negotiates its next collective bargaining agreement with the Transport Workers Union. Blakeman wrote that he stood “shoulder to shoulder” with TWU Local 100, arguing that conductors are essential for public safety.

Though the TWU and other advocates frequently claim that conductors are necessary for public safety, conductors aren’t required to assist in threatening situations. That makes sense—train operators shouldn’t need to put themselves potentially in harm’s way. But it seriously undermines the claim that conductors play a vital public-safety role.

A TWU vice president recently acknowledged as much on X, posting that conductors are only required to report incidents to the MTA and NYPD. “When we take action we get written up for dismissal or other negative actions by the MTA,” he wrote.

It’s not unreasonable to oppose one-person train operation, which—despite being the global and national standard—must be considered carefully given the immense complexity and intensive use of the MTA’s transit system. But it’s another thing to support banning the MTA from using it at all. One-person operations work well enough in the few instances the MTA uses it today—why upset that apple cart?

It’s also understandable that Republicans like Blakeman would want to peel away union rank-and-file members to their corner. But if they’re serious about governing differently than Democrats, they can’t cozy up to the forces that are making New York expensive and slowing down innovation.

Long outdated work rules are endemic in the MTA, for example. They stand in the way of managerial reforms that would improve fiscal efficiency and the quality of riders’ experience. Don’t expect the TWU to give those up without a big fight.

Friendliness to unions also makes little sense politically. Even if Blakeman and co. manage to peel off some rank-and-file support, New York’s union leadership (putting aside police unions) is unlikely to break with Democrats—particularly given the current national political climate.

With New York’s Democrats firmly in league with reform-thwarting unions, Republicans have an opportunity to side with the public’s well-being against the special interests. A serious focus on greater affordability through results for tax dollars could form the basis of a winning message. But that’s only if the party is willing to confront, not accommodate, those interests.

If Republicans are unwilling to do so, the party will remain stuck where it is now: unable to articulate what it stands for, and even less able to convince voters to give them a chance.

Share

• Leadership transition: Warren Buffett, after six decades, has passed Berkshire Hathaway CEO duties to Greg Abel, marking a significant succession.
• Abel’s background: The 63-year-old vice-chair, a low-profile executive from Berkshire’s utilities division, now leads the conglomerate.
• Buffett’s legacy: Buffett transformed Berkshire from a textile mill into a $1.1tn conglomerate spanning railroads, utilities, and insurance.
• Capital position: Berkshire holds more than $350bn in cash and short-term Treasuries plus $283bn in publicly traded stock while generating almost $900mn weekly from its businesses.
• Investment philosophy: Abel has pledged to maintain Buffett’s focus on businesses with significant cash flows and a long-term horizon, including thorough economic forecasts.
• Communication expectations: Investors seek clarity on whether Abel will introduce quarterly earnings calls, additional qualitative unit performance insights, or reshape the annual shareholder letter.
• Technology bets: Market watchers wonder if Abel influenced the $4.3bn Alphabet investment, which could signal Berkshire’s stance on fast-growing tech opportunities.
• Organizational moves: Abel is appointing a new CFO, first general counsel, and promoting NetJets’ CEO to oversee hundreds of consumer and service businesses to bolster a lean Omaha office.

---

Warren Buffett has handed over Berkshire Hathaway’s reins after six decades at the helm, in one of the most consequential corporate successions of a generation.

Berkshire vice-chair Greg Abel, 63, took charge as chief executive on Thursday, leaving a low-profile protégé of one of history’s savviest investors to lead the sprawling company.

Buffett, 95, transformed Berkshire from a struggling New England textile mill into a $1.1tn financial behemoth that spans railroads, utilities and insurance operations.

With Buffett promising to go “quiet”, Wall Street is watching to see how Abel will deploy Berkshire’s vast portfolio.

The company has more than $350bn of cash and short-term Treasuries and $283bn in publicly traded stock. Investors will also scrutinise how Abel allocates the almost $900mn of cash that flows in from its businesses each week.

“He’s inheriting the most privileged place in American business,” said Christopher Davis, a partner at Berkshire investor Hudson Value Partners. “Buffett was not only a great investor but someone people . . . looked up to for doing the right thing and dealing fairly and that gave Berkshire some pretty broad latitude.”

Investors have more questions than answers as they wait to see how Abel, a longtime Berkshire executive, begins to leave his mark. They are keen to see whether he will hew to Buffett’s value investing philosophy, which in recent years has meant Berkshire has passed on several big-ticket deals and avoided many flashy tech investments.

Some analysts and investors have also pressed Buffett to launch quarterly earnings calls or provide better qualitative insight into how individual units are performing — a decision that now rests with Abel.

The new CEO has only seldom made himself available to the press and investors.

Shareholders have instead drawn on his comments at Berkshire’s annual meetings in Omaha, Nebraska, for clues on how he will size up investment opportunities and the business attributes he will be looking for when going elephant hunting — Buffett’s term for the group’s mammoth corporate takeovers.

Abel, a Canadian who rose up through Berkshire’s utilities division, has signalled the company’s investment philosophy will not change when he takes over.

Last year he told investors he would continue to target businesses that generate significant cash flows and the company’s long-term investment horizon would remain intact.

Investors are waiting to see if Greg Abel, centre, will continue to use Warren Buffett’s annual shareholder letter in February to lay out his vision for Berkshire © Brendan McDermid/Reuters

He added Berkshire would still need to have a view on the economic prospects of a company in 10 or 20 years before investing, whether buying a business outright or when purchasing a minority stake.

“It is really the investment philosophy and how Warren and the team have allocated capital for the past 60 years,” Abel said last May. “It will not change and it’s the approach we’ll take as we go forward.”

He declined to comment for this story.

Each February investors and corporate executives pore over Buffett’s annual shareholder letter. Buffett has said he will not be writing the next missive, and investors are waiting to see if Abel will use the annual letter to lay out his vision for Berkshire.

Investors are particularly focused on whether Abel was involved in Berkshire’s $4.3bn investment in Google owner Alphabet in late 2025, or if it was Buffett who signed off on the wager.

If Abel was a key driver of the investment, then investors could view that as a sign that Berkshire is open to making big bets on fast-growing technology companies.

“What is it in the capital allocation model that just made them decide to buy Alphabet now?” asked Christopher Rossbach, chief investment officer of J Stern & Co, a longtime Berkshire investor.

He added: “What makes Berkshire special is the public equity portfolio. And the question is: is it really going to be Greg managing that in the way that Warren did?”

Abel has begun to leave his stamp on some parts of Berkshire’s sprawling empire. He will be joined by a new chief financial officer next year as well as the company’s first general counsel.

He has also promoted the CEO of fractional jet ownership company NetJets as president of 32 of Berkshire’s consumer, retail and service businesses — a division that generated more than $40bn of revenues in the first nine months of 2025.

Investors such as Darren Pollock, a portfolio manager at Cheviot, said they saw the appointments as a step by Abel to fill in gaps at Berkshire’s slim home office in Omaha, which at the end of 2024 employed just 27 people.

“This paves the way for Greg’s larger priority — to win the confidence of shareholders who can’t help but be less comfortable with Abel than they are [with] the legendary Buffett,” Pollock said.

“Now, with the spotlight on him, Abel needs to show what people within Berkshire know him to be: a steady and razor-sharp leader.”

• Regulatory Pressure: Japanese regulators targeted a surge of scam ads on Meta’s platforms, prompting Meta to fear mandatory advertiser identity verification.
• Ad Library Management: Meta analyzed keyword searches by Japanese regulators and deleted problematic ads to improve perceived outcomes and reduce scam discoverability.
• Global Playbook: The tactic of mimicking regulator searches was expanded into a global strategy to delay or weaken advertiser verification in markets worldwide.
• Revenue Concerns: Documents show Meta estimates universal verification would cost about $2 billion to implement and could cut up to 4.8% of total revenue by blocking unverified advertisers.
• Reactive Stance: Meta officially maintains a “reactive only” approach, resisting verification unless law requires it and prioritizing voluntary programs instead.
• Verification Impact: Taiwan’s mandatory advertiser verification reduced scam ads significantly, but Meta’s algorithms reroute disallowed ads to other countries, shifting harm.
• Industry Comparison: Meta notes Google’s universal verification reset regulatory expectations, yet refuses comparable investments due to costs and potential revenue loss.
• Risk Ranking: Fraudulent advertising earned the highest internal risk rating, with potential liability in Europe and Britain estimated at up to $9.3 billion if users’ losses became Meta’s responsibility.

---

SAN FRANCISCO - Japanese regulators last year were upset by a flood of ads for obvious scams on Facebook and Instagram. The scams ranged from fraudulent investment schemes to fake celebrity product endorsements created by artificial intelligence.

Meta, owner of the two social media platforms, feared Japan would soon force it to verify the identity of all its advertisers, internal documents reviewed by Reuters show. The step would likely reduce fraud but also cost the company revenue.

To head off that threat, Meta launched an enforcement blitz to reduce the volume of offending ads. But it also sought to make problematic ads less “discoverable” for Japanese regulators, the documents show.

The documents are part of an internal cache of materials from the past four years in which Meta employees assessed the fast-growing level of fraudulent advertising across its platforms worldwide. Drawn from multiple sources and authored by employees in departments including finance, legal, public policy and safety, the documents also reveal ways that Meta, to protect billions of dollars in ad revenue, has resisted efforts by governments to crack down.

In this case, Meta’s remedy hinged on its “Ad Library,” a publicly searchable database where users can look up Facebook and Instagram ads using keywords. Meta built the library as a transparency tool, and the company realized Japanese regulators were searching it as a “simple test” of “Meta’s effectiveness at tackling scams,” one document noted.

To perform better on that test, Meta staffers found a way to manage what they called the “prevalence perception” of scam ads returned by Ad Library searches, the documents show. First, they identified the top keywords and celebrity names that Japanese Ad Library users employed to find the fraud ads. Then they ran identical searches repeatedly, deleting ads that appeared fraudulent from the library and Meta’s platforms.

Instead of telling me an accurate story about ads on Meta’s platforms, it now just tells me a story about Meta trying to give itself a good grade for regulators.

Sandeep Abraham, former Meta fraud investigator

The tactic successfully removed some fraudulent advertising of the sort that regulators would want to weed out. But it also served to make the search results that Meta believed regulators were viewing appear cleaner than they otherwise would have. The scrubbing, Meta teams explained in documents regarding their efforts to reduce scam discoverability, sought to make problematic content “not findable” for “regulators, investigators and journalists.”

Within a few months, they said in one memo after the effort, “we discovered less than 100 ads in the last week, hitting 0 for the last 4 days of the sprint.” The Japanese government also took note, the document added, citing an interview in which a prominent legislator lauded the improvement.

Meta has studied searches of its Ad Library and worked to reduce the "discoverability" of problematic advertising. Documents reviewed by Reuters, and highlighted here by the news agency, show internal discussions about the effort. REUTERS

“Fraudulent ads are already decreasing,” Takayuki Kobayashi, of the ruling Liberal Democratic Party, told a local media outlet. Kobayashi didn’t respond to a Reuters request for comment about the interview.

Japan didn’t mandate the verification and transparency rules Meta feared. The country’s Ministry of Internal Affairs and Communications declined to comment.

So successful was the search-result cleanup that Meta, the documents show, added the tactic to a “general global playbook” it has deployed against regulatory scrutiny in other markets, including the United States, Europe, India, Australia, Brazil and Thailand. The playbook, as it’s referred to in some of the documents, lays out Meta’s strategy to stall regulators and put off advertiser verification unless new laws leave them no choice.

The search scrubbing, said Sandeep Abraham, a former Meta fraud investigator who now co-runs a cybersecurity consultancy called Risky Business Solutions, amounts to “regulatory theater,” distorting the very transparency the Ad Library purports to provide. “Instead of telling me an accurate story about ads on Meta’s platforms, it now just tells me a story about Meta trying to give itself a good grade for regulators,” said Abraham, who left the company in 2023.

Meta spokesperson Andy Stone in a statement told Reuters there is nothing misleading about removing scam ads from the library. “To suggest otherwise is disingenuous,” Stone said.

By cleaning those ads from search results, the company is also removing them from its systems overall. “Meta teams regularly check the Ad Library to identify scam ads because when fewer scam ads show up there that means there are fewer scam ads on the platform,” Stone wrote.

Advertiser verification, he said, is only one among many measures the company uses to prevent scams. Verification is “not a silver bullet,” Stone wrote, adding that it “works best in concert with other, higher-impact tools.” He disputed that Meta has sought to stall or weaken regulations, and said that the company’s work with regulators is just part of its broader efforts to reduce scams.

Those efforts, Stone continued, have been successful, particularly considering the continuous maneuvers by scammers to get around measures to block them. “The job of chasing them down never ends,” he wrote. The company has set global scam reduction targets, Stone said, and in the past year has seen a 50% decline in user reports of scams. “We set a global baseline and aggressive targets to drive down scam activity in countries where it was greatest, all of which has led to an overall reduction in scams on platform.”

Meta’s internal documents cast new light on the central role played by fraudulent advertising in the social media giant’s business model – and the steps the company takes to safeguard that revenue. Reuters reported in November that scam ads Meta considers “high risk” generate as much as $7 billion in revenue for the company each year. This month, the news agency found that Meta tolerates rampant fraud from advertisers in China.

In response to Reuters’ coverage, two U.S. senators urged regulators at the Securities and Exchange Commission and the Federal Trade Commission to investigate and “pursue vigorous enforcement action where appropriate.” Citing Reuters reporting, the attorney general of the U.S. Virgin Islands also sued Meta this month for allegedly “knowingly and intentionally” exposing users of its platforms to “fraud and harm” and “profiting from scams.” Stone said Meta strongly disagrees with the lawsuit’s allegations.

In Brussels, where European authorities have also been focused on scams, a spokesperson for the European Commission told Reuters its regulators had recently asked Meta for details about its handling of fraudulent advertising. “The Commission has sent a formal request for information to Meta relating to scam ads and risks related to scam ads and how Meta manages these risks,” spokesperson Thomas Regnier wrote. “There are doubts about compliance.” He didn’t elaborate.

The documents reviewed by Reuters show that Meta assigned its handling of scams the top possible score in an internal ranking of regulatory, legal, reputational and financial risks in 2025. One internal analysis calculated that possible regulation in Europe and Britain that would make Meta liable for its users’ scam losses could cost the company as much as $9.3 billion.

EMPLOY A “REACTIVE ONLY” STANCE

One big push among regulators is to get Meta and other social media companies to adopt what is known as universal advertiser verification. The step requires all advertisers to pass an identity check by social media platforms before the platforms will accept their ads. Often, regulators request that some of an advertiser’s identity information also be viewable, allowing users to see whether an ad was posted locally or from the other side of the world.

Google in 2020 announced that it would gradually adopt universal verification, and said earlier this year it has now verified more than 90% of advertisers. Along with requiring verification in jurisdictions where it’s legally mandated, Meta offers to voluntarily verify some large advertisers and sells “Meta Verified” badges to others, combining identity checks with access to customer support staff.

Documents reviewed by Reuters say that 55% of Meta’s advertising revenue came from verified sources last year. Stone, the spokesperson, added that 70% of the company’s revenue now comes from advertisers it considers verified.

The internal company documents show that unverified advertisers are disproportionately responsible for harm on Meta’s platforms. One analysis from 2022 found that 70% of its newly active advertisers were promoting scams, illicit goods or “low quality” products. Stone said that Meta routinely disables such new accounts, “some on the very day that they’re created.”

Meta’s documents also show the company recognizes that universal verification would reduce scam activity. They indicate that Meta could implement the measure in any of the countries where it operates in less than six weeks, should it choose to do so.

But Meta has balked at the cost.

Despite reaping revenue of $164.5 billion last year, almost all of which came from advertising, Meta has decided not to spend the roughly $2 billion it estimates universal verification would cost, the documents show. In addition to that cost of implementation, staffers noted, Meta could ultimately lose up to 4.8% of its total revenue by blocking unverified advertisers.

I expected that the company would have continued to do more verification, and personally felt that was something that all major platforms should be doing.

Rob Leathern, a former senior director of product management at Facebook

Instead of adopting verification, Meta has decided to employ a “reactive only” stance, according to the documents. That means resisting efforts at regulation – through lobbying but also through measures like the scrubbing of Ad Library searches in Japan last year. The reactive stance also means accepting universal verification only if lawmakers mandate it.

So far, just a few markets, including Taiwan and Singapore, have done so.

Even then, the documents show, the financial costs to Meta have remained small. Meta’s own tests showed verification immediately reduced scam ads in those countries by as much as 29%. But much of the lost revenue was recouped because the same blocked ads continued to run in other markets.

If an unverified advertiser is blocked from showing ads in Taiwan, for example, Meta will show those ads more frequently to users elsewhere, creating a whack-a-mole dynamic in which scam ads prohibited in one jurisdiction pop up in another. In the case of blocked ads in Taiwan, “revenue was redistributed/rerouted to the remaining target countries,” one March 2025 document said, adding that consumer injury gets displaced, too. “This would go for harm as well,” the document noted.

Meta analyses found that even when verification blocked ads in one market, those same ads would still generate revenues for the company in other markets. Highlighting of internal document reviewed by Reuters. REUTERS

Meta’s documents show the company believes its efforts to defeat regulation are succeeding. In mid-2024, one strategy document called the prospect of being “required to verify all advertisers” worldwide a “black swan,” a term used to describe an improbable but catastrophic event. In the months afterwards, policy staffers boasted about stalling regulations in Europe, Singapore, Britain and elsewhere.

In July, one Meta lobbyist wrote colleagues after they thwarted stricter measures considered by financial regulators in Hong Kong against financial scams. To get ahead of the effort, staffers helped regulators draft a voluntary “anti-scam charter.” They coordinated with Google, which also signed the charter, to present a “united front,” the document says. “Through skillful negotiations with regulators,” the Meta lobbyist wrote, Hong Kong relaxed rules that would have forced verification of financial advertisers. “The finalised language does not introduce new commitments or require additional product development.”

Hong Kong regulators, the lobbyist added, “have shown huge appreciation for Meta’s leading participation.”

Meta staffers boasted about success slowing the push by authorities for advertiser verification. In one document, highlighted here by Reuters, Meta employees say their lobbying in Hong Kong thwarted "new commitments" in local regulations. REUTERS

A Google spokesperson said the company signed onto the charter because it believed it would benefit customers. Google participated, he said, of its own accord and as the result of direct engagement with Hong Kong regulators.

In a statement, Hong Kong financial regulators said that “advertiser verification is one of many ways social media platforms can protect the investment public.” They declined to respond to Reuters’ questions about Meta and noted that the regulators involved with the charter don't themselves have the authority to impose advertiser verification requirements.

“All social media platforms should strengthen their efforts to detect and remove fraudulent and unlawful materials,” they added.

“INDUSTRY AND REGULATORY EXPECTATIONS”

Fraud across social media platforms has surged in recent years, fueled by the rise of untraceable cryptocurrency payments, AI ad-generation tools and organized crime syndicates. Mob rings have found the business so lucrative that they employ forced labor to staff well-documented “scam compounds” that generate waves of fraudulent content from southeast Asia. Internally, Meta has cited estimates that such compounds are responsible for $63 billion in annual damage to consumers worldwide.

In some countries, regulators have determined that Meta platforms host more fraudulent content than its online competitors. In February 2024, Singapore police reported that more than 90% of social media fraud victims in the city state had been scammed through Facebook or Instagram. In a statement to Reuters, a spokesperson for Singapore’s Ministry of Home Affairs wrote that “Meta products have persistently been the most common platforms used by scammers.”

“We have repeatedly highlighted our deep concern over the continued prevalence of scams on Meta’s platforms,” the statement continued. After Reuters’ inquiries for this report, it added, Singapore authorities have asked Meta for more information and will broaden existing verification measures, including some mandating the use of facial recognition technology to prevent the impersonation of public figures. “We have reiterated that more needs to be done to secure Meta’s products and protect users from scams, instead of prioritising its profits. We have requested for a formal explanation from Meta and will take enforcement action if Meta is found to be in violation of legal requirements.”

A known weakness in Meta’s defenses is the ease of advertising on its platforms.

To purchase most advertisements, all a client needs is a user account – easily created with an email or phone number and a user-supplied name and birthdate. If Meta doesn’t verify those details, it can’t know who it’s doing business with. Even if an advertiser gets banned, there is nothing to stop it from returning with a new account. A fraudster can merely sign up again.

Meta has known about the problem for years, documents and interviews with former staffers show.

In the 2016 U.S. presidential election, fake political ads flooded Facebook with disinformation. In response, the company took steps to reduce chances that could happen again. Back then, foreign actors seeking to influence the election easily placed ads masquerading as Americans. Some Russian advertisers pretending to be American political activists even paid for such ads in rubles, Meta has said.

Starting in 2018, the company began requiring a valid identity document and a confirmed U.S. address before clients could place political ads. In addition to providing verification for the company itself, the general details, including the name and location of the advertiser, could be viewed by users, too.

Rob Leathern, a former senior director of product management at Facebook who oversaw the effort to verify political advertisers, said the added transparency and accountability led some staffers to believe that Meta would broaden it to all advertisers. “I expected that the company would have continued to do more verification, and personally felt that was something that all major platforms should be doing,” said Leathern, who left the company at the end of 2020.

Meta in 2018 also introduced its Ad Library, an easily searchable database of all ads that run on its platforms. The company, the documents show, expected to generate goodwill with the library, particularly with regards to political advertisements. Competitors, including Google, soon launched ad libraries of their own.

In the years that followed, Meta continued to acknowledge the effectiveness of both transparency and verification. So-called “know your customer policies,” Meta staffers wrote in a November 2024 document, are “commonly understood to be effective at reducing scam-risks.” They noted a competitive component, too, citing Google’s move at the start of the decade to adopt universal verification: “Google’s approach to verify all advertisers is recalibrating industry and regulatory expectations.”

Meta, however, has been reluctant to pay for it.

The internal documents show that last year Meta consulted with a company that works with Google to verify advertisers. Meta officials, according to the documents, wanted to know how much it would cost to follow suit. But the answer – at least $20 per advertiser – proved too costly for their liking, one document said.

The Meta spokesperson said that the company, regardless of cost, didn’t work with the vendor because its verification process took too long.

The potential for lost revenue has also given the company pause.

In addition to lost income from advertisers culled by verification, stricter measures could also cannibalize a paid program through which Meta already charges advertisers for similar status. The program, known as “Verified for Business,” costs clients as much as $349.99 per month and allows businesses to display a badge assuring users that Meta has authenticated their profile. Meta describes the program as more than just basic verification, offering advertisers better customer support and protections against impersonation.

Still, the documents show, Meta managers fear those revenues could shrivel if the company adopts verification for all advertisers.

“WE HAVE AN OPPORTUNITY”

In 2023, because of a sharp rise in ads for investment scams, Taiwan passed legislation ordering social media platforms to begin verifying advertisers of financial products. The self-governing island, population 23 million, is small compared to Meta’s major markets, but the company’s response there helps illustrate how resistant Meta has been to growing regulatory scrutiny worldwide.

In private conversations, the documents show, Taiwanese regulators told Meta it needed to demonstrate it was taking concrete steps to help reduce financial scam ads. When it came to financial fraud, the regulators said, Meta needed to verify the identity of those advertising financial services and respond to reports of fraud within 24 hours.

Meta, according to the documents, told Taiwan it needed more time to comply. Regulators agreed. But Meta, the documents show, in the months that followed didn’t address the problem to the government’s satisfaction.

Frustrated, the Taiwanese regulators last year issued new demands. Now, the new regulations stated, Meta and the owners of other major platforms would have to verify all advertisers. Regulators told Meta it would be fined $180,000 for every unverified scam ad it ran, Meta staffers wrote.

If it didn’t comply, the staffers calculated, the resulting fines would exceed Meta’s total profits in Taiwan. It would be cheaper to abandon the market than to disobey, they concluded.

Meta complied, rushing to verify advertisers ahead of regulators’ deadlines.

In a statement to Reuters, Taiwan’s Ministry of Digital Affairs said stricter regulations over the past year brought down rates of scam ads involving investments by 96% and identity impersonation by 94%. In addition to requiring major social media platforms to verify advertisers, Taiwan has developed its own AI system to scan ads on Meta’s platform, set up a portal for citizens to report fraudulent ads, and established public-private partnerships to detect scams, the ministry added.

Over the course of 2025, the statement said, Taiwan has fined Meta about $590,000 for four violations of the law. The ministry said it “will maintain a close watch on shifting fraud risks.”

The new rules gave Meta the opportunity to study the impact that full verification would have on its business. Before the new regulation, according to internal calculations, about 18% of all Meta advertising in Taiwan, or about $342 million of its annual ad business there, broke at least one of the company’s rules against false advertising or the sale of banned products. Unverified advertisers, one analysis found, produced twice as much problematic advertising as those who submitted verification details.

Their analyses also revealed the whack-a-mole dynamic.

Because scamming is a global business – and Meta’s algorithms allow clients to choose multiple markets in which to advertise – many advertisers seeking to place fraudulent posts do so in more than one geography. Meta experiments showed that while fraudulent ads decreased in Taiwan after the rule change, its algorithms simply rerouted them to users in other markets.

“The implication here is that violating actors that only require verification in one country, will shift their harm to other countries,” one analysis spelled out. Unless advertiser verification was “enforced globally,” staffers wrote, Meta wouldn’t so much be fighting scams as relocating them.

The documents included briefing notes prepared for Chief Executive Mark Zuckerberg about the dynamic. Reuters couldn’t determine whether the Meta boss ever saw the notes or was briefed on their contents. But the message delivered a similar conclusion. It also warned of a complication: If enforcement in one jurisdiction worsened the problem of fraud in others, regulators in the newly impacted markets were likely to crack down, too.

Meta spokesperson Stone said he couldn’t determine whether Zuckerberg received the briefing described in the document reviewed by Reuters.

Faced with the prospect of ever-expanding scrutiny, Meta considered embracing full verification voluntarily, the documents show. The goal, staffers wrote, could enable the company to appear proactive but also set terms and a timeline on its own. “We have an opportunity to set a goal of verifying all advertisers (and communicate our intention to do so externally, in order to better negotiate with lawmakers),” a November 2024 strategy document noted. Meta could “stage the rollout over time and set our own definitions of verification.”

Policy staff even planned to announce the decision during the first half of 2025, the documents show. But for reasons not specified in the documents, they postponed an announcement until the second half of the year and then cancelled it altogether. Leadership had changed its mind, a document noted, without saying why.

“MIMIC WHAT REGULATORS MAY SEARCH FOR”

Instead, Meta began to apply some of the lessons it learned in Japan.

That experience helped the company realize that Tokyo wasn’t the only government using Ad Library searches as a means of tracking online fraud. “Regulators will open up the ads library and show us multiple similar scam ads,” public policy staffers lamented in one 2024 document. Staffers also noted authorities were employing one feature that was proving especially useful: a keyword search. Unlike Google’s version, the Meta library made it easy to find scam ads through searches with terms like “free gift” or “guaranteed profit.”

Managers overseeing a revamp of the Ad Library proposed eventually killing the keyword feature entirely, the documents show. Wary of blowback from regulators, however, Meta decided not to. The Meta spokesperson said Meta is not considering it.

The company did, however, change the library so that searches returned fewer objectionable ads.

One adjustment made searches default to active ads, reducing the number of search results by eliminating content that Meta had already blocked through prior screening. The change made fraudulent ads from the past absent from new search results.

Staffers also made Meta’s systems rerun enforcement measures on all ads that appeared during new Ad Library searches, the documents show. That adjustment gave Meta a second chance to scrap violators that had previously evaded fraud filters.

One of the most useful tactics it learned in Japan was Meta’s mimicry of searches performed by regulators. After repeating the same queries, and deleting problematic results, staffers could eventually go days without finding scam ads, one document shows.

As a result, Meta decided to take the tactic global, performing similar analyses to assess “scam discoverability” in other countries. “We have built a vast keyword list by country that is meant to mimic what regulators may search for,” one document states. Another described the work as changing the “prevalence perception” of scams on Facebook and Instagram.

Meta’s perception-management tools are now part of what the company has referred to as its “general global playbook” for dealing with regulators. The documents reviewed by Reuters repeatedly reference the “playbook” as steps the company should follow in order to slow the push toward verification in any given jurisdiction.

Beginning one year ahead of expected regulation, the playbook advises, Meta should tell the local regulators it will create a voluntary verification process. When doing so, the documents add, Meta should ask those authorities for time to let the voluntary measures play out. To buy yet more time, and further gauge reactions from regulators, Meta after six months should force verification upon “new and risky” advertisers, the playbook continues.

Meta has devised a “global playbook,” summarized in the document here, to delay and weaken the push by regulators to mandate advertiser verification. Internal documents reviewed by Reuters show that verification reduces scam ads, but also costs Meta revenue. REUTERS

If ultimately regulators force mandatory verification for all, the playbook states, Meta should once again stall. “Keep engaging with regulator on extension,” one document advises.

The documents show Meta staffers celebrating the success of their efforts to change some perceptions.

In March, industry officials and regulators met for a conference in London organized by the Global Anti-Scam Alliance, a group that organizes regular gatherings to address online fraud. Meta staffers in one document celebrated the lack of scorn heaped on the company compared with previous events.

“There was a drastic shift in tone,” a project manager noted. “Meta was rarely called out whereas previously we were explicitly and repeatedly shamed for lack of action in countering fraud.”

Reporting by Jeff Horwitz. Additional reporting by Anton Bridge, Kentaro Okasaka, Yoshifumi Takemoto in Japan; Xinghui Kok in Singapore; Clare Jim, Selena Li and James Pomfret in Hong Kong; Yimou Lee and Emily Cha in Taiwan; Brad Haynes in Brazil; Phoebe Seers in UK; Panu Wongcha-um and Poppy McPherson in Thailand and Philip Blenkinsop in EU. Editing by Steve Stecklow and Paulo Prada.

Our Standards: The Thomson Reuters Trust Principles., opens new tab

• Yearly reading count: The author usually averages over 27 books per year but counted only six full-length books in 2025, excluding repeated short children’s books.
• Survey participation: YouGov found 59% of Americans read at least one book in 2025, consistent with 2023 and 2024 survey results.
• Reading distribution: 40% read no books, 27% read one to four, and 19% read ten or more, with 9% reaching 10-19, 6% reaching 20-49, and 4% reading 50+.
• Inequality impact: The top 19% of readers accounted for 82% of all books read, while the bottom 40% read none.
• Demographics: Median books read is two across age groups, but those 65+ averaged 12.1 books, and higher education correlated with higher averages.
• Formats and preferences: 46% read physical books, 24% digital, 23% audiobooks; 29% used multiple formats, and digital-preferring readers were overrepresented among heavy readers.
• Genres and education: Mystery/crime topped at 35%, while genre preferences varied by gender and education, with literary and academic genres favored by college graduates.
• Library engagement: 51% have library cards, 24% use them monthly, and 19% use other library services monthly, mostly checking out physical books.

---

Featuring data on how many books Americans read in 2025.

---

It’s the end of the year, which means it’s time for me to count up my reading stats and see how I did. I’m normally a pretty heavy reader, and have averaged more than 27 books per year over the past decade.

2025 was either my weakest reading year in more than a dozen years — or by far my best. It all depends on whether you count the short children’s books I’ve read over and over again to my infant daughter. Hand, Hand, Fingers, Thumb is hardly Moby Dick, but one of those is sitting on my shelf with a bookmark a quarter of the way through, and the other’s got a read count in the dozens.

Personally I don’t count short children’s books as part of my reading. (Should I? Leave a comment.) I do count full-length books I read aloud while trying to help her sleep, so Seamus Heaney’s Beowulf translation counts (and worked great, I might add). But even with those, I’m only at six full-length books read this year, with an outside chance at a seventh before the ball drops on the 31st.

The good news is that that still makes me well above average in book count, as our new survey shows:

Six in 10 Americans (59%) say they read at least one book in 2025, a new YouGov survey finds. That’s in line with similar YouGov surveys in 2024 and 2023. Most Americans who did read books only finished a handful of books, while a minority of Americans were plowing through the pages. Here’s what YouGov found about Americans’ 2025 book-reading habits:

Reading inequality

---

Besides the 40% of Americans who didn’t read any books in 2025, another 27% read one to four books. And 13% read five to nine books. That leaves 19% of Americans who read 10 or more books, including 9% who read 10 to 19 books, 6% who read 20 to 49 books, and 4% who say they read 50 or more books.

The median American read two books in 2025. On average, Americans read eight books. The average books read is higher than the median books read because the small number of heavy readers increases the total number of books read by the same number of people.

Americans with more education are more likely to read more books: Those with postgraduate degrees read a median of five books and an average of 13.6 books, for example, while those with a high school education or less read a median of zero books and an average of 4.6.

On average, Americans 65 and older read significantly more books (12.1) than those 45 to 64 (6.4), 30 to 44 (8.2), and 18 to 29 (5.8). But all four age groups have the same median number of books read. Those 65 and older are no more likely to read any books than younger Americans, but they are more likely to be heavy readers: 24% of those 65 and older have read at least 10 books in 2025, compared to 17% of those 45 to 64, 17% of those 30 to 44, and 19% of those 18 to 29.

Democrats read more books on average than Republicans, though this effect is driven more by heavy readers than the reading habits of typical Democrats and Republicans — both parties have a median books read of two. Independents read less than members of either party, both on average and as a median.

Americans who pay more attention to what’s going on in government and public affairs are more likely to read more books than those who pay less attention.

With most Americans reading no books or just a few books, and a minority reading lots of books, that means that the distribution of the total books read in the U.S. is very unequal. The 4% of Americans who say they read 50 or more books alone account for 46% of all books read. Add in the 6% of Americans who read between 20 and 49 books, and the 9% who read between 10 and 19 books, and the top 19% of U.S. adult citizens account for 82% of all books read in 2025.

The middle 40% read 18% of all books, and the bottom 40% read no books.

[

](https://substackcdn.com/image/fetch/$s_!RP8S!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F250bd685-2ccc-4274-901e-a68af13bdff1_1200x869.png)

Someone who read just three books in 2025 did more book reading than 57% of Americans. Reading five books is more than two-thirds of Americans, and 10 books is more than 81% of Americans read.

What and how Americans read

---

Among Americans who did read in 2025, about half read only in one type of book: physical, digital, or audiobooks. The other half of readers read books in multiple formats.

Overall in 2025, 46% of Americans read at least one physical book, 24% read at least one digital book, and 23% listened to at least one audiobook. 29% of Americans consumed books in multiple formats, while 23% say they read only physical books, 5% read only digital books, and 2% consumed only audiobooks.

Only 14% of Americans say they prefer to read digital books, but these are some of the country’s heaviest readers. 13% of them say they read 50 or more books in 2025, compared to 4% of those who prefer physical books and 5% of those who prefer audiobooks.

The most popular genre of books that Americans read in 2025 was mystery and crime: 35% of Americans who read at least one book read a mystery or crime novel. Other top genres are history (30%), biography and memoir (27%), thrillers (23%), fantasy (23%), and romance (23%).

Some genres, such as biographies, thrillers, and fantasy, are read by similar shares of men and women. Other genres have sizable gender skews: 42% of male readers read a history book in 2025, compared to 20% of female readers. 32% of women who read at least one book read a romance, compared to 12% of men.

Another factor associated with what types of books Americans read is education. Several genres, including literary fiction, academic books, politics, historical fiction, and science fiction, are substantially more likely to be read by people with college degrees than by those without degrees.

In contrast, books on religion and spirituality, young adult novels, and dramas are slightly more likely to be read by those without degrees. Fantasy, graphic novels, and romance novels are read by similar shares of those with and without college degrees.

About half (51%) of Americans have a library card, while 46% don’t. Around one-quarter (24%) of Americans say they are heavy library users — they check out books at least once a month. Another 20% of Americans have library cards but say they rarely check out books, while 7% have library cards and never check out books. Physical books are the most common choice for library users: Only 9% of Americans have library cards but never check out physical books, while 29% have cards but never check out digital books and 31% have cards but never check out audiobooks.

About 19% of Americans say they use other library services — such as programs and events for children or adults, accessing computers, or just spending time there — at least once a month.

---

Learn more and get in touch

---

Sign up to participate in our polls and share your opinions on the people, companies, and issues that shape your life and the world

For more polling data and insights, visit our website or follow us on X, Threads, and Bluesky.

Have questions or feedback?

• If you have a press or media inquiry, email uspress@yougov.com

• If you have an idea for a future survey, share it here

• If you are a panelist and have questions, contact support here

• For information on privacy, please read our privacy notice

Carl Bialik contributed to this newsletter.

• Forecast Validation: The 2023 report accurately predicted AI datacenter demand rising from ~3GW to over 28GW in the US by 2026, triggering concerns about grid strain.
• Grid Saturation: Texas alone receives tens of gigawatts of monthly datacenter load requests while barely a gigawatt is approved annually, signaling a sold-out grid.
• BYOG Momentum: AI labs and hyperscalers increasingly deploy onsite gas generation to circumvent multi-year grid upgrades, with xAI pioneering truck-mounted turbines to bring Colossus online rapidly.
• Equipment Diversity: Datacenters select from aeroderivative/industrial gas turbines, reciprocating engines, and Bloom Energy fuel cells, balancing cost, lead time, ramp rate, and redundancy.
• Deployment Strategies: Bridge power, Energy-as-a-Service, and mixed fleets (aeros + RICE) exemplify approaches to match grid-level uptime demands, often involving N+1/N+1+1 redundancy and batteries or flywheels.
• Permitting & Costs: Onsite generation faces higher power costs, permitting delays, and complex operations, though fuel cells ease permitting while turbines and engines require specialized alloys and maintenance.
• Supply Constraints: Manufacturers struggle with lead times (18–36 months), blade metallurgy bottlenecks, and logistic challenges for heavy-duty turbines amid limited factory expansion commitments.
• Market Expansion: New entrants and adjacent industries (Doosan, Wärtsilä, Boom Supersonic, Boom’s Superpower) push triple-digit growth in onsite gas, while existing OEMs cautiously scale production.

---

The Grid is Old and Tired

Nearly two years ago, we were the first to predict a looming power crunch. In our report AI Datacenter Energy Dilemma - Race for AI Datacenter Space, we forecasted AI Power Demand in the US to grow from ~3GW in 2023 to over 28GW by 2026 – a pressure that would overwhelm America’s supply chains. Our prediction proved very accurate.

AI Datacenter Energy Dilemma - Race for AI Datacenter Space

Dylan Patel, Daniel Nishball, and Jeremie Eliahou Ontiveros

·

March 13, 2024

Read full story

The chart below tells the story: in Texas alone, tens of gigawatts of datacenter load requests pour in each month. Yet in the past 12 months, barely more than a gigawatt has been approved. The grid is sold out.

Source: ERCOT 2024 Large Flexible Load Task Force (LFLTF)

However, AI infrastructure cannot wait for the grid’s multiyear transmission upgrades. An AI cloud can generate revenue of $10-12 billion dollars per gigawatt, annually. Getting a 400 MW datacenter online even six months earlier is worth billions. Economic need dwarfs problems like an overloaded electric grid. The industry is already searching for new solutions.

Eighteen months ago, Elon Musk shocked the datacenter industry by building a 100,000-GPU cluster in four months. Multiple innovations enabled this incredible achievement, but the energy strategy was the most impressive. xAI entirely bypassed the grid and generated power onsite, using truck-mounted gas turbines and engines. As shown below, xAI has already deployed over 500MW of turbines near its datacenters. In a world where AI Labs are racing to be first with a Gigawatt datacenter, speed is the moat.

Source: SemiAnalysis Datacenter Industry Model

One by one, hyperscalers and AI Labs are following suit and temporarily abandoning the grid to build their own onsite power plant. As we discussed months ago in the Datacenter Model, in October 2025, OpenAI and Oracle placed the largest order ever for onsite gas generation, with a 2.3GW plant in Texas. The market for onsite gas generation is entering an era of triple-digit growth annual growth.

The beneficiaries extend far beyond the usual suspects. Yes, GE Vernova and Siemens Energy have seen their stocks surge. But we’re witnessing an unprecedented wave of new entrants, such as:

• Doosan Enerbility, the Korean industrial giant, timing its H-class turbine launch perfectly and already booked 1.9GW of US datacenter orders.

• Wärtsilä, historically a ship engine manufacturer, realized the same engines that power cruise ships can power large AI clusters. It has already signed 800MW of contracts.

• Boom Supersonic—yes, the supersonic jet company—announced a 1.2 GW turbine contract with Crusoe, treating the margin from datacenter power generation as another round of funding for their Mach 2 passenger jets.

To understand growth and market share by supplier, we built a building-by-building tracker of sites deploying onsite gas in our Datacenter Model. The results surprised us: 12 different suppliers have now secured >400 MW of datacenter orders each in the US alone, for onsite gas generation.

Source: SemiAnalysis Datacenter Industry Model

Upgrade to paid

However, onsite power generation brings its own set of challenges. Power costs are often (much) more expensive than via the grid, as detailed below. Permitting can be a lengthy and complicated process. And it’s already causing some datacenter delays - most notably one of the Oracle/Stargate GW-scale facilities, which our Datacenter Industry Model predicted three weeks prior to the Bloomberg headlines by analyzing the whole permitting process.

Again, clever firms like xAI have found remedies. Elon's AI Lab even pioneered a new site selection process - building at the border of two states to maximize the odds of getting a permit early! While Tennessee couldn't deliver on time, Mississippi happily enabled Elon to build a GW-scale power plant.

Source: SemiAnalysis Datacenter Industry Model

This report is a deep dive into Bring Your Own Generation (BYOG). We begin with why the grid can’t keep up, then provide a technical breakdown of every generation technology available to datacenters—GE Vernova’s aeroderivatives, Siemens’ industrial turbines, Jenbacher’s high-speed engines, Wärtsilä’s medium-speed engines, Bloom Energy’s fuel cells, and much more.

Then we examine deployment configurations and operational challenges: fully islanded datacenters, gas + battery hybrids, Energy-as-a-Service models, and the economics that determine which solutions win. Behind the paywall, we share our views on manufacturer positioning, d and the future of onsite generation.

Is the Electric Grid Dead in the AI Era?

Before we dive into solutions, we need to understand why the grid is failing. To be fair, America’s electrical system has been the primary enabler of AI infrastructure so far. Elon aside, every major GPU & XPU clusters today runs on grid power. We’ve covered many of them in prior SemiAnalysis deep dives:

• Microsoft’s AI Strategy showing the massive grid-connected facilities for OpenAI in Wisconsin, Georgia and Arizona.

• Our Multi-Datacenter Training report, digging into Google’s massive grid-powered clusters in Ohio and Iowa/Nebraska, as well as OpenAI’s gigawatt cluster in Abilene, TX with Oracle, Crusoe and Lancium.

• Our Meta Superintelligence article laying out their AI large plans, which include some onsite gas generation, but remain primarily served by AEP’s system in Ohio and Entergy in Louisiana.

• Our Amazon’s AI Resurgence thesis, discussing AWS’ massive Trainium clusters for Anthropic, connected as well to AEP and Entergy’s infrastructure.

These insights appeared in our Datacenter Industry Model months or years before official announcements. Our model tracks dozens more large-scale clusters under construction for 2026 delivery and beyond—including their exact start dates, full capacity, end-users, and energy strategies.

But we’ve hit a tipping point. The large datacenters coming online in 2024-25 secured their power in 2022-23, before the gold rush. Since then, the scramble has been relentless. We estimate roughly a terawatt of load requests have been submitted to US utilities and grid operators.

Source: SemiAnalysis Datacenter Industry Model

The result is gridlock - literally. As we explained in AI Training Load Fluctuations at Gigawatt-Scale, the grid is slow by design:

1. Real-time balancing: Electricity supply and demand must match nearly perfectly, every second. A mismatch risks blackouts for millions, as we saw with the Iberian Peninsula blackout in April 2025.

2. System studies: Every large new load (datacenter) or supply (power plant) triggers deep engineering studies to ensure it won’t destabilize the network. And in some places, grid topology changes so quickly that load studies go obsolete before they’re completed.

Source: 2025 ITP portfolio

When hundreds of developers simultaneously submit interconnection requests, the system seizes up. It becomes a prisoner’s dilemma:

• If everyone coordinated, the grid could handle more requests faster.

  • FERC Order 2023 has pushed grid operators to adopt cluster studies for this purpose, but those reforms were solidified only in 2025.

• In practice, “gold rush” behavior means developers submit multiple speculative requests to different utilities simultaneously

  • For example as of mid-2024, AEP Ohio had 35 GW of load requests—and 68% didn’t even have land control

• Speculative requests clog the queue for everyone, encouraging more speculative requests elsewhere

• The vicious cycle accelerates

Source: PJM Load Analysis Subcommittee

The supply side is equally constrained. The timeline from interconnection request to commercial operation now stretches to five years for most generation types.

Souce: Lawrence Berkeley National Lab

AI infrastructure developers cannot wait five years. In many cases, they cannot wait six months, because waiting six months costs billions of dollars of lost opportunities.

Enter BYOG - Bring Your Own Generation

The core value proposition of BYOG is simple: start operating without waiting for the grid. A datacenter can run indefinitely on local generation, then convert that equipment to backup power once grid service eventually arrives.

That’s exactly xAI’s strategy. They built Colossus using mobile gas turbines, bringing the facility online in months rather than years. Now everyone is following the playbook.

Let’s examine how.

How to Bring Your Own Generation

The Old World vs The New World

BYOG involves a complete re-thinking of the way we build power plants. Traditionally, we deliver power via large, centralized GW-scale baseload generators – accompanied by smaller peaker plants to handle spikes in grid-wide load. Heavy-duty gas turbines in combined cycle mode are the most common modern deployment. Their unmatched fuel efficiency (>60%) provides the backbone of our modern civilization. However, their main issue is deployment speed:

• There is typically a multi-year lead time to get large turbines, and current lead times are at an all-time high.

• Once delivered, construction and commissioning of a large combined-cycle power plant takes ~2 years - an eternity in the AI era.

A combined cycle gas turbine (CCGT). Source: Knoxville News Sentinel

AI Datacenter “BYOG” power plants re-shape the playbook, and xAI led the way for the industry. To deploy faster, Elon’s AI Lab relied on small modular 16MW turbines from Solar Turbines, a CAT subsidiary. The turbines are small enough to be transported by standard long-haul trucks. They’re deployed in a matter of weeks. Elon didn’t even buy them – he rented from Solaris Energy Infrastructure to bypass the equipment lead time. He also leveraged VoltaGrid’s fleet of mobile truck-mounted gas engines to deliver faster!

Solar SMT130 (rated for 16 MW). Truck for scale. Source: CAT (Solar Turbines)

Source: Tom’s Hardware

Other hyperscalers quickly followed suit. Meta’s deployment in Ohio, with Williams, is illustrative – with their power plant comprising five different types of turbines & engines, clearly the design pattern was “I’ll deploy whatever I can get on time!”

Socrates South Satellite Image (Nov 11, 2025)

Let’s now dig into the different types of equipment available to datacenter operators.

Equipment Landscape Overview

Among gas generators available to datacenter developers, there are three broad categories:

1. Gas Turbines (GTs) - low-temp, slow-to-ramp industrial gas turbines (IGTs); high-temp, fast-to-ramp aeroderivative gas turbines (Aeros); very large heavy-duty gas turbines.

2. Reciprocating Internal Combustion Engines (RICEs) - both smaller, 3-7 MW high-speed engines; and larger, 10-20 MW medium-speed engines. Sometimes called “recips” for short.

3. Solid-oxide fuel cells (SOFCs) - the main option available so far is from Bloom Energy.

There are additional onsite power options such as co-locating with an existing nuclear power plant, building onsite SMRs, Geothermal, and many more, but we won’t discuss them in this report. For the most part, these other solutions are not driving net new power generation in the next ~3 years.

Understanding which solutions are the best fit for certain use-cases requires digging into the core tradeoffs. We see the following as most relevant:

• Cost: Usually listed as $/kW. These cost estimates vary wildly and are consistently rising across every generator category. Note that maintenance expenses are also relevant: certain systems have lower useful life, i.e. higher annual maintenance costs.

• Lead Time (shipment and installation): Usually listed in months or years. Lead times are increasing across every generator category as demand growth outstrips supply.

  • Note that other factors outside generator availability can affect time-to-power. Most notably, air permitting for onsite generation can take a year or more, even in fast-to-permit states like Texas.

  • In addition, installation time varies widely across systems. Some can take barely a few weeks from delivery onsite to power generation, such as small truck-mounted turbines or engines, as well as fuel cells. Large CCGTs can take over 24 months to assemble.

• Redundancy & uptime: the expected availability of the generator, expressed in % of uptime over a year, or in “nines” of uptime. The US Electric grid averages 99.93% (3 nines) over the last ten years, with some areas even higher. For an onsite power plant, redundancy can be managed by adding hot spares and cold spares, or by having additional backup power. The larger the individual turbines, the more difficult managing spares & backup is.

• Ramp Rate: Measured as minutes between cold start and maximum output. A ramp rate of less than 10 minutes makes a generator eligible as reserve generation for an electric grid or backup power. A slow ramp-rate means that the unit is primarily focused on baseload power.

• Land Use: Measured as MW/acre. This matters more in space-constrained areas. Water use for small generation systems is insubstantial, even as a fleet. However, very large turbines do require significant water use for cooling.

• Heat Rate and Fuel Efficiency: Measured as BTU of natural gas per kWh. A higher heat rate means lower efficiency—more fuel in, same electricity out, more waste left behind. Nameplate heat rate assumes “peak” operating conditions, typically maximum output. Efficiency drops substantially below 50% output.

  • Many of these onsite gas systems can be configured as combined heat and power (CHP) systems. For datacenters, this would entail using the waste heat from a gas generator for an absorption cooling system, allowing for reduced electricity use in cooling the datacenter.

In reality, we observe that whoever has an open orderbook and can provide good timelines tends to win deals, regardless of most other specs!

Having said that, let’s now deep dive into the different types of gas power plants.

Aeroderivatives and IGTs – highly attractive for datacenters

Gas turbines run on a Brayton Cycle: compress air, burn fuel in it, and route the hot gas through a turbine. Turbines are differentiated by inlet temperatures. Lower temperatures correspond to lower installation costs, lower maintenance costs, lower peak efficiency, and slower ramp rates.

An aeroderivative gas turbine is simply a jet engine bolted to the ground. GE Vernova’s aeros derive from GE jet engines; Mitsubishi Power’s from Pratt & Whitney; Siemens Energy’s from Rolls-Royce. Because jet engines are designed to deliver massive power in a compact, flight-worthy package, they are relatively easy to adapt for stationary power. Extend the turbine shaft, bolt a generator coil to the end, add intake and exhaust mufflers, and feed fuel from tanks or a pipeline. This is, in part, why Boom Supersonic could pivot so quickly into aeroderivative gas turbines: most of their engineering and manufacturing is carryover.

Mitsbishi Heavy FT8 MOBILEPAC (rated for 30 MW). Source: Mitsubishi Heavy Industries

We show below a view of the Martin Drake power plant, w/ 6x GE Vernova LM2500XPRESS units. This is how electric utilities deploy aeroderivatives, as “peaker plants” for sudden supply shortages in the grid.

The core manufacturers for aeroderivative gas turbines are similar to those of heavy-duty gas turbines: GE Vernova, Mitsubishi Power, and Siemens Energy dominate the market, selling both aeros and lower-temp industrial gas turbines (IGTs). Additionally Caterpillar also produces IGTs under the Solar brand name, as does Everllence (formerly MAN Energy Systems).

Two GE Vernova designs dominate the aeroderivative market:

• LM2500 – ~34 MW, optimized for fast deployment, especially as LM2500XPRESS.

• LM6000 – ~57 MW, now available in fast-deploy LM6000VELOX configurations.

Aeros are reasonably efficient with fuel but extremely efficient with respect to space and weight. They can fit in tight footprints, and in some configurations can be transported on a pair of tractor trailers. Simple-cycle aeros typically come in 30-60 MW packages and can ramp from cold to full output in 5-10 minutes. However, efficiency suffers if they are at less than full steady load. Aeros can also be configured as small combined-cycle plants:

• 1x1 (one combustion turbine feeding one steam turbine), or

• 2x1 (two combustion turbines feeding one steam turbine).

These combined-cycle setups deliver higher efficiency and more output at the cost of ramp speed. Startup times lengthen to 30–60 minutes.

At current rates, aeros cost $1,700-2,000/kW in all-in capital expenditure, and based on recent orders, they have lead times of 18-36 months and rising. Smaller turbines can have lead times as short as 12 months, and larger aeros (~50 MW) can take up to 36 months. These systems are quick to install (2-4 weeks usually), but the factories are heavily booked. One workaround is truck-mounted turbines, which can be rented and deployed quickly, if available. xAI used this exact strategy, partnering with Solaris Energy Infrastructure to shrink their time-to-power for Colossus 1 and 2.

Industrial Gas Turbines (IGTs)

Industrial gas turbines work on the same cycle as aeros and share benefits like compact footprints, modularity, and relatively fast lead times. But they are designed from scratch for stationary use rather than adapted from aviation. They typically run at lower inlet temperatures and use simpler designs, which lowers service costs at the expense of efficiency and ramp speed.

Cutaway of SMT130 IGT. Source: Solar Turbines

Simple-cycle IGTs span roughly 5–50 MW and ramp from cold to full output in ~20 minutes. That makes them too slow, on their own, to serve as peaker plants or emergency backup without help from batteries or diesel units. Like aeros, IGTs can be upgraded to combined-cycle configurations, improving efficiency while further slowing ramp rate.

The most common dedicated industrial gas turbines are the Siemens Energy SGT-800 and Solar Titan Series. However, smaller heavy-duty gas turbines like the GE Vernova 6B also sometimes take on similar use cases.

Solar SMT130 (rated for 60 MW). Truck for Scale. Source: CAT (Solar Turbine)

At current rates, IGTs cost $1,500-1,800/kW in all-in capital expenditure, with lead times of approximately 12-36 months, similar to aeros. However, procuring a used or refurbished IGT can shrink lead times to under 12 months, which is how Fermi America is procuring power.

Overall, we believe that aeroderivatives and IGTs are a very attractive solution for onsite power generation, because:

• They are the “right” size: small enough to facilitate redundancy, large enough to avoid having too many units onsite and complexifying maintenance.

• They have a fast ramp-rate: while they aren’t as energy-efficient as others, they can more easily be repurposed for backup power.

• They are quick to deploy, normal trucks and construction crews can transport and install them, instead of the insane heavy-lift infrastructure necessary for heavy-duty turbines.

We’ll discuss these concepts later in the report when discussing deployment considerations. The main issue with aeros and IGTs is, increasingly, lead times.

The most supply-constrained component in gas turbines are the turbine blades and cores, which must handle high temperatures and speeds. These blades use exotic monocrystalline nickel alloys that include rare-earth metals like rhenium, cobalt, tantalum, tungsten, and yttrium. Notably, yttrium is among the rare earths under export control from the Chinese government. The cores, meanwhile, require high-temperature ceramics that are in short supply.

Reciprocating Engines (RICE)

Reciprocating engines function like automotive engines, but at a much larger scale, an 11MW engine can be more than 45 feet (14 m) long. They use four-stroke combustion cycles and are divided by rotation speed:

• High-speed engines – ~1,500 rpm; smaller in footprint and output.

• Medium-speed engines – ~750 rpm; generally lower maintenance costs due to lower mechanical stress.

RICEs can ramp from cold to full output in 10 minutes, similar to aeros in practice. This lets RICEs work as peaker plants or as backup generators, eliminating the need for diesel backups. On paper, RICE O&M looks higher than for turbines because there are more moving parts. In practice, they handle fuel impurities, dust, and high ambient temperatures better than many turbines and suffer less de-rating in hot climates.

Medium-speed engine manufacturing is fairly consolidated, with the primary manufacturers being Wärtsilä, Bergen Engines, and Everllence (formerly MAN Energy).

Bergen B36:45V20AG (rated for 11.3 MW). Person for scale. Source: Bergen Engines

High-speed engine manufacturing is not as consolidated as turbines. Outside the prominent players in Jenbacher, CAT, Cummins, and Rolls Royce subsidiary MTU, there are a wide range of manufacturers, because high-speed gas engines are functionally equivalent to the diesel engine designs currently used for backup power at many datacenters. The most consequential reciprocating engine is the Jenbacher J624, a 4.5MW turbocharged gas engine that can be containerized for easier logistics. This system is the preferred generator for VoltaGrid’s energy integration services.

Source: VoltaGrid

RICE systems typically generate less power per unit than equivalent turbines. Medium-speed engines run between 7 MW and 20 MW, with the higher power outputs enabled by turbocharging. High-speed engines are even smaller, with per-unit outputs between 3 MW and 5 MW. However, RICE generators are more efficient than turbines when running at partial loads between 50% and 80%.

Reciprocating engines operate at much lower temperatures than gas turbines, closer to 600°-700°C. This dramatically reduces their need for high-performance alloys. Only the high-temperature components in the pistons, combustion chambers, and turbochargers still need rare nickel and cobalt alloys, and the rest can be manufactured with simple cast iron, steel, and aluminum. However, RICEs overall are less dependent on critical minerals, especially if emissions controls are relaxed during a materials supply crunch.

At current rates, reciprocating engines cost $1,700-2,000/kW in all-in capital expenditure and have lead times of 15-24 months. Compared to turbines, these systems are less delayed in manufacturing; the manufacturing timeline is closer to 12-18 months. However, medium-speed RICEs are considerably heavier than turbines, and installation and commissioning can take up to ~10 months.

High-speed engines can be much faster to deploy. For example, at the initial Colossus 1 deployment, xAI leveraged 34 VoltaGrid truck-mounted systems, incorporating high-speed engines from Jembacher. High-speed engines, in particular, are popular with energy procurement vendors (described later). Their wide availability and small unit size offer faster time-to-power. We show below a VoltaGrid 50MW deployment in San Antonio, with twenty Jembacher J620 (rated 3.36kW per unit).

Source: Voltagrid

The tradeoff is scale: to build a 2 GW onsite gas system with 5 MW engines, you need 500 units! That has major operational consequences. If each engine needs minor servicing every 2,000 hours, the maintenance staff would perform more than 2,000 services per year, almost 40 per week. These costs are more predictable than turbine overhauls (which can involve swapping entire cores), but they add up, especially for fleets with many small units. Space and spares inventories grow similarly, although vertical stacking of small generators can mitigate land use, a trick not available for medium-speed engines.

Fuel cells and Bloom Energy’s ascent

A fairly niche solution is now taking an increasingly large share of the pie: fuel cells. Often associated with hydrogen, Bloom Energy’s SOFC fuel cells can run on natural gas too and are pitched as baseload generation. We first called out Bloom Energy as a big winner in last 2024 in the datacenter model. Since then the orders have skyrocketed.

Source: Power Engineering

Bloom’s “Energy Server” is made up multiple ~1kW stacks, assembled into ~65kW modules, and packaged into a 325kW power generator. To date, the largest operational SOFC-based power plants are in the tens of MW, mostly in the US and Korea.

Source: Bloom Energy

The way they generate energy is very different from that of traditional generators. There is no combustion process. Instead, oxygen is electrochemically reduced to oxide ions, which flows through a ceramic electrolyte. At the other end of the fuel cell, these ions combine with hydrogen atoms stripped from methane natural gas. This combination releases water, CO2, and electricity.

This fundamental difference provides Bloom’s fuel cells with a key advantage: they do NOT generate material air pollution, besides CO2. The permitting at the EPA level is significantly smoother and easier than that of combustion generators. That’s why we often see them in population centers, such as near offices.

Bloom’s killer feature is the speed of deployment. It barely requires precast pads and a simple installation of modules. Once factoring-in the electrical work, installation & commissioning can be done in a matter of weeks, matching the speed of aeroderivatives and high-speed RICE.

In the AI era where speed is the moat, that advantage alone is enough to place Bloom on the map.

Source: Bloom Energy Installation [YouTube]

Bloom’s main challenge is cost. Fuel cell efficiency is quite good, with an equivalent heat rate of 6,000-7,000 BTU/kWh, which is on-par with CCGTs. However, the costs for fuel cell systems are notably higher than turbines or RICE systems, at a capex cost between $3,000-$4,000/kW. Bloom does not advertise ramp rates, suggesting these units are too slow to function as peakers or emergency backup.

Maintenance has historically also been notably higher than other solutions. Individual fuel cell stacks last roughly 5-6 years, then must be replaced and refurbished. This per-cell replacement makes up ~65% of service costs, although specific numbers are kept close to vest. Bloom discloses little about its materials beyond the use of ceramics in the cell core, but claim that their fuel cells have no critical mineral dependence on China or other contested regions.

Source: Bloom Energy

We provide TCO estimates for Bloom fuel cells behind the paywall.

Heavy-duty gas turbines: the future of BYOG?

Before ChatGPT, only utilities and independent power producers (IPPs) had any reason to buy a gas turbine larger than 250 MW, because turbines above that threshold are simply too large to use for most industrial applications. As explained above, speed of deployment is an issue, however, we’re increasingly seeing developers provide “bridge power” via smaller aeroderivatives/RICE then shift them as backup/redundancy once the big CCGT is operational.

Large turbines are grouped into classes based on combustion (turbine inlet) temperature and technology stack:

E-Class and F-Class – Older, lower-temperature, lower-efficiency designs. Some F-class units are still sold, usually into developing markets, because they offer decent efficiency at lower capex. The line between “industrial” turbines and small E/F-class frames is fuzzy, with the below famous models straddling that boundary:

• GE Vernova 6B

• GE Vernova 7E

• Siemens Energy SGT6-2000E

H-Class and equivalents – Modern, high-temperature designs. These run firing temperatures comparable to modern aeros and jet engines, but with roughly 10x the per-unit power. The most prominent examples are:

• GE Vernova HA series (e.g., HA.02)

• Siemens Energy H/HL

• Mitsubishi Heavy Industries J series (e.g., H510J)

• More recently, Korean firm Doosan Enerbility has started production of a new H-class turbine, the DGT6. It’s rare to see new entrants in a decade-old market, but Doosan has deep experience in steam turbine production and a track record of building Mitsubishi-designed F-class turbines.

As shown below, these systems are both very large and heavy. The installation and commissioning process can take a while.

A view of the Three Rivers CCGT in Grundy County, IL. Satellite Image.

Combined-Cycle Gas Turbines (CCGTs)

Combined-cycle gas turbines (CCGTs) exploit the fact that simple-cycle exhaust is still very hot, hot enough to boil water into steam. Routing exhaust through a heat recovery steam generator (HRSG) produces steam for a separate steam turbine and generator. The result is a second round of power from the same fuel. By turning one turbine’s trash into another turbine’s treasure, CCGTs can run 50-80% more efficiently than a simple-cycle turbine.

The CCGTs most vaunted for large loads are heavy-duty CCGTs, which can reach gigawatt-scale outputs. However, even small aeroderivative or industrial gas turbines can be sold with an integrated steam turbine, which can dramatically increase power output with near-identical fuel inputs. Common configurations are:

• 1x1 – One gas turbine feeding one steam turbine

• 2x1 – Two gas turbines feeding one steam turbine

In theory, more gas turbines can feed a single steam turbine but returns diminish. The primary drawback of a CCGT system is the ramp rate: the addition of the steam turbine slows the time from cold start to full output to 30 minutes or more.

The other major drawback is the lead time. Installation & commissioning is even longer than for a simple cycle deployment.

From equipment to execution: deployment, challenges, economics

Understanding the equipment landscape is necessary but not sufficient. The real complexity in onsite gas isn’t choosing between an LM2500 and a Jenbacher J624—it’s figuring out how to configure, deploy, and operate these systems to meet datacenter uptime requirements.

The electric grid is a marvel of systems engineering: thousands of generators, hundreds of transmission lines, and sophisticated market mechanisms that together deliver 99.93% average uptime. When you go off-grid, you’re taking on that complexity yourself—with a single plant that has to match grid-level reliability. Redundancy and uptime are the key reason why onsite gas power costs are, in most cases, structurally much more expensive than power delivered by the grid.

The next section examines how leading deployments are solving this challenge, and what it means for equipment selection.

Crusoe and xAI: bridge power deployment

One of the most popular onsite gas strategies so far has been “bridge power”. The datacenter campuses have an active discussion with the grid to get electrical service, but begin operations before via onsite power.

Bridge power clears electricity as a bottleneck to operation, allowing a datacenter to start training models or generating revenue several months earlier. This speedup is significant! AI cloud revenue can net $10-12M per MW annually, meaning that getting 200 MW of datacenter powered and online even six months earlier can net $1-1.2 billion in revenue.

Bridge power brings two advantages:

1. The uptime requirements can be matched to the workload. For example, in Abilene TX and Memphis TN, both xAI and Crusoe/OpenAI are deploying large training clusters. Training jobs don’t need particularly elevated uptime, given the inherent unreliability of large GPU clusters. As such, “overbuilding” the power plant for redundancy can be avoided. Once a grid connection is secured, the campus can be more fungible and also used for inference.

2. Favorable economics via removal of diesel generator backup. In both Memphis and Abilene, the absence of backup reduces datacenter capex/MW. Once a grid connection is secured, the turbines can act as backup – as such, fast ramp-rate systems are preferred, e.g. aeroderivatives.

To ensure reasonable uptime, xAI paired the turbines with MegaPacks. That also enables to smooth out load fluctuations – an issue we’ll discuss below.

Satellite image over xAI Memphis

Staying Off-Grid Forever: redundancy challenges, Energy-as-a-Service

Many generator vendors suggest that datacenter owners should never bother interconnecting with the broader electric grid; instead, they argue that their datacenter customers should stay off-grid forever. Firms like VoltaGrid offer a full “Energy-as-a-Service” package managing all aspects of electric service:

• Electric energy – MW of capacity and MWh of energy

• Power quality – Voltage and frequency tolerances

• Reliability – Targeted “nines” of uptime

• Time-to-power – Months from contract to operation

They typically sign long-term PPAs with customers who pay for electric service – the EaaS vendors essentially acts as a utility. They procure equipment, design the deployment, sometimes assemble the BoM, and maintain & operate the power plant.

A key challenge when deploying off-grid generation is managing redundancy. For example, the 1.4GW Vantage DC campus in Shackelford County, TX will deploy 2.3GW of VoltaGrid systems. These systems being small facilitates redundancy – but if you were to deploy onsite power with large heavy duty turbines, redundancy might be to simply have two power plants, if not more.

Generation vendors will suggest at minimum an N+1 configuration, if not an N+1+1 configuration. An N+1 configuration maintains full generation capacity even if one generator unexpectedly shuts down, whereas an N+1+1 configuration enables this flexibility while also keeping another generator on standby to enable maintenance cycles. It’s the equivalent of driving a car with a spare tire and a tire repair kit. Note that N+1 or N+1+1 does not necessarily refer to a literal count of generators, given that datacenter loads are typically much larger than individual onsite gas generators. For example, consider a datacenter with an all-in (IT + non-IT) power demand of 200 MW:

Example 1: 11-MW RICEs

• Generation fleet: 26 × 11 MW RICE units

• Total capacity: 286 MW

Under normal operation:

• 23 engines run at ~80% load to produce 200+ MW.

• One generator failure: 22 engines ramp modestly to ~82% load.

• 3 spare engines remain for maintenance or as cold standby.

Running engines below full load reduces O&M, and the extra units provide a buffer for maintenance scheduling.

Nexus Datacenter is using a similar approach: they have recently applied for an air permit for a fleet of thirty Everllence 18V51/60G gas engines, each good for 20.4 MW, for a total of 613 MW of generation. This site will also include 152 MW of diesel backup generation, which likely fulfills the N+1 redundancy requirements for the total site.

Example 2: 30-MW Aeroderivatives

• Generation fleet: 9 × 30 MW aeros

• Total capacity: 270 MW

Under normal operation:

• 7 turbines run at ~95% load for best efficiency.

• One turbine failure: the 8th turbine starts, maintaining output.

• The 9th turbine remains in reserve for maintenance.

Because turbine overhauls are more disruptive than engine maintenance, some vendors offer hot-swap programs: a turbine due for major service is swapped out for a replacement core.

In hot climates, such as the American Southwest, derating may require 10–11 aeros to maintain N+1+1 redundancy.

Crusoe’s Abilene site for Oracle and OpenAI uses a version of this setup, with a deployed fleet of ten turbines, with five GE Vernova LM2500XPRESS aeroderivative gas turbines and five Titan 350, good for 360MW of nameplate generation.

Source: Citrini Research

Example 3: Meta + Williams Socrates South

Meta and Williams are building a pair of 200 MW behind-the-meter gas plants to power Meta’s New Albany Hub, which we have covered in this article: Meta’s new ultra-fast “tent” datacenters in Ohio – SemiAnalysis

Socrates South Satellite Image (Nov 11, 2025)

The Socrates South project is a hybrid fleet:

• 3 × Solar Titan 250 IGTs (23 MW)

• 9 × Solar Titan 130 IGTs (16.5 MW)

• 3 × Siemens SGT-400 IGTs (14.3 MW)

• 15 × Caterpillar 3520 fast-start engines (3.1 MW)

Nameplate capacity inside the fence is 306 MW: roughly 260 MW from turbines and 46 MW from engines. Under normal conditions, a subset of IGTs runs steadily to deliver 200 MW. If one or two IGTs trip, the RICE fleet can ramp quickly to cover the gap. Additional IGTs remain available for maintenance switchover. This supports an N+1+1 behind-the-meter design.

However, this is a patchwork implementation compared to the first two examples. The turbines don’t match, and the engines used are smaller, 1800-rpm high-speed gas engines. This suggests that Williams prioritized time-to-power over standardized maintenance schedules.

Match the grid’s uptime: Overbuild, Grid-as-backup, Batteries

To match the “three nines” of uptime provided by the grid, an onsite power plant must be “overbuilt” for redundancy. This is typically the key reason for higher onsite generation power costs, relative to the grid.

Redundancy introduces a new headache for operators: there is a tradeoff between the size of a system and the “overbuild” ratio. While H class and F class turbines are more energy-efficient than aeros, the higher redundancy needs means than, if poorly designed, an islanded system based on heavy duty turbines can yield higher power costs than aeros. Other solutions than a simple “overbuild” must be considered, such as using smaller turbines as “backup”, batteries, or even a grid connection.

To understand the overbuild ratio, we can use a practical example. In Shackelford County, TX, VoltaGrid is poweing a 1.4GW datacenter (IT capacity) with 2.3GW of Jembacher systems (64% overbuild). We can break this down in the following way:

• Peak PUE overbuild: as is typical for a grid-connected sites in Texas, there is a 1.4x - 1.5x over provisioning, largely related to cooling.

• There is an additional 10-17% overbuild related to redundancy.

For H/F class systems, a simple overbuild is often not the most economical path. Some operators are considering a grid connection solely for backup purposes - but that introduces interconnection timeline challenges, and complicates the site selection process (need access to high-voltage lines). A huge battery plant can also be built - as we illustrate below with xAI’s Colossus 2 deployment - but that’s both expensive and impractical, given 2-4hrs of typical storage duration. Lastly, a combination of different sizes of turbines and engines can be used, with H-class in combined-cycle mode operating as baseload, and IGTs/aeros/RICE as backup—but that’s typically more expensive than a grid connection or a 2-4hr BESS.

Managing Load Surges

AI compute load, particularly training, is highly variable, including megawatt-scale power surges and dips on a sub-second basis. The more inertia a power system has, the better it can manage short-term power fluctuations while maintaining power frequency. If frequency deviates too far from the 50 Hz or 60 Hz baseline, the power fluctuations can trip breakers or cause malfunctions. All thermal generators have some inertia, because they are generating electricity with a spinning heavy object. However, a developer can increase inertia with auxiliary systems:

• Synchronous condensers – These are essentially generators spun up as motors, with no mechanical load. Once synchronized to the grid, they consume only small losses. During sudden load changes, they absorb or supply reactive power, stabilizing voltage and adding short-duration inertia. Their energy capacity is small, so they help for seconds, not minutes.

Source: Baldor.com

• Flywheels – These add a real rotational energy buffer. A motor-generator is coupled to a large flywheel and connected between generation and load. Flywheels can inject or absorb real power (not just reactive) for 5–30 seconds, smoothing transients, generator trips, and voltage dips. Bergen, for example, packages flywheels alongside its engines via an affiliate vendor.

Source: Piller Power

• Battery energy storage systems (BESS) – Batteries can ramp as quickly as the load changes, providing “synthetic inertia” through high-speed control, as described in an earlier article. They excel at frequency regulation, but because inverters are current-limited, they contribute less to reactive power and fault currents than synchronous machines.

VoltaGrid combines RICE fleets with synchronous condensers. Bergen Engines has sold engines with flywheels from a vendor under the same parent company. Engine manufacturer Wärtsilä has a battery energy storage vertical that they may bundle with datacenter projects. Bloom claims that their fuel cell systems don’t need any equipment to manage load fluctuations. The exact system used depends a bit on local constraints and mostly on what the vendor prefers to use. xAI prefers to use Tesla’s Megapacks for backup and handling load fluctuations.

Megapacks + MACROHARD

Can we even build enough gas power plants to power AI?

Current lead times for gas generation systems are unprecedented. Historically, gas turbine manufacturers have only taken orders on average 20 months in advance of shipment from factories, but now the Big Three of manufacturers, GE Vernova, Siemens Energy, and Mitsubishi Power, are accepting orders into 2028 and 2029, with nonrefundable reservation slots beyond that Every public manufacturer of gas systems reports rising datacenter demand, but most are responding with caution, not a full-send buildout.

• GE Vernova has promised to increase production to 24 GW/year, but that only returns them to its 2007–2016 levels. They are investing in new staff in machinery, but do not intend to increase factory footprint.

• Siemens Energy also plans to invest in production without increasing factory footprint. They are instead prioritizing price increases, leaning on service revenue, and prioritizing investments with short payback periods. They plan to scale annual capacity from ~20GW to >30GW by 2028-30.

• Mitsubishi Heavy Industries has guided to increase gas turbine & combined-cycle production by 30% in recent earnings calls, contrary to Bloomberg reporting about plans to double capacity by 2027.

• Caterpillar plans to double engine production and 2.5x turbine production between 2024 and 2030, but their Solar-branded turbine production averaged ~600 MW/year between 2020–2024, with a 2022 peak production of 1.2 GW.

• Wärtsilä has promised only incremental expansion, preferring to “wait and see” on datacenter demand and preserve relationships with marine customers.

Of the major gas generation manufacturers, only Bloom Energy, Caterpillar, and newcomer Boom Supersonic have announced ambitious expansion plans. Bloom Energy has claimed they can reach 2 GW/year of production capacity by end of 2026, and Boom Supersonic plans to reach 2 GW/year by end of 2028. At first glance, few manufacturers appear fully “AGI-pilled” despite surging demand. Some of that hesitation reflects real manufacturing limits; much of it reflects PTSD from 30 years of boom-bust cycles in gas generation. Notably, the worst bottlenecks are in heavy-duty turbines. Aeros, IGTs, and RICE systems are less constrained.

The Two Boom-Bust Cycles of Gas Turbines

Since the mid-‘90s, the gas turbine industry has seen two boom-bust cycles rock the industry. The first boom, between 1997 and 2002, was driven by electric power deregulation in parts of the United States, which pulled in new companies as independent power producers, as well as (ironically enough) high expectations of electric demand growth coming from the dotcom bubble, as popularized by the Huber and Mills paper “The Internet Begins with Coal.” Large players like Calpine, Duke, Williams, and NRG placed block orders for turbines, sending GE Vernova (then GE Power) and Siemens Energy (then Siemens AG’s power segment) into lunar order volumes. GE shipped more than 60 GW of gas turbines in 2001; Siemens peaked at 20+ GW in 2002.

Source: Energy Information Administration

The crash came fast. The dot-com bubble burst, the Enron scandal shook the power trading business, and orders dried up, leaving GE and Siemens in a manufacturing winter for the next few years. The second “boom” in the gas turbine industry was less a boom than a stabilization of orders. Between 2006 and 2016, GE averaged about 20 GW/year of turbine shipments, and Siemens about 15 GW/year. Then, between 2017 and 2022, the bottom fell out on the market, with both GE and Siemens seeing production lows under 10 GW/year.

These two large companies have both institutional memory of the Y2K gas turbine boom and recent memory of generationally low sales. Notably, Mitsubishi Heavy Industries has largely escaped these boom-bust cycles. Until extremely recently, MHI has sold a fraction of the hardware of GE Vernova and Siemens Energy. It has only become part of a “Big Three” because the larger companies have shrunk to its sales volume and other players like Alstom Energy and Westinghouse have shuttered or been acquired. This may in part explain MHI’s interest in expansion, although its supposed doubling plan has not been corroborated in earnings calls.

Supply Chain Bottlenecks

However, within gas turbines, even a guarantee of high future demand may not push forward increased production, because of internal bottlenecks in the production and logistics of gas turbine cores.

Gas turbine blades and vanes are among the high watermarks for civilizational technological competence, requiring an insane quality of metallurgy and machining to manufacture properly.

Machining a single turbine blade. Source: Reliable Turbine Services, LLC

Turbine blades and vanes are among the most demanding components modern industry makes. Manufacturing them requires extraordinary metallurgical and machining precision. As a result, Western production is concentrated in four firms:

• Precision Castparts Corporation (PCC)

• Howmet Aerospace

• Consolidated Precision Products (CPP)

• Doncasters

These companies supply not only industrial and electrical gas turbines but also civilian and military jet engines as well. All except CPP have vertically-integrated metals supply, but they are a fraction of the size of their customers, and thus much more vulnerable to market shocks. The second gas turbine bust coincided with a COVID-driven slump in aerospace orders, meaning these companies have recently been hit quite hard. An increase in demand would require these companies not only to hire more specialized staff, but also to reckon with their supply chain for materials like yttrium, rhenium, single-crystal nickel, and cobalt. More importantly, they are likely reluctant to make these investments because they stand to lose the most if they follow an AI bubble off a cliff.

Additionally, heavy-duty gas turbine production is constrained by logistics. The turbine cores alone are 300-500-ton systems that need specialized barges, rail cars, and truck trailers to transport. Even after permitting, heavy-duty gas turbines need 24-30 months to build, install, and test before they are ready to run. Aftermarket OEMs can build new plants around refurbished cores, but moving and integrating those cores remains a major challenge. These constraints are less severe for aeros and IGTs, which are small enough to ship on standard containers or conventional trailers.

Siemens SGT5-800H H-Class turbine on a self-propelled modular transport (SPMT). Source: Siemens China

New entrants to the rescue: from jets to ships?

As often, in times of constraints, many smart firms are exploring solutions. ProEnergy was one of the first to come with innovations. Its PE6000 program retrofits CF6-80C2 engine cores from Boeing 747 and delivers operational aeroderivative gas turbines with near-identical specs and packaging to the GE Vernova LM6000.

ProEnergy PE6000. Source: Datacenter Dynamics

More recently, Boom Supersonic has announced the development of the Superpower aeroderivative gas turbine, based on their supersonic jet engine design. Its proposed form factor looks remarkably similar to the GE Vernova LM2500, and it operates on the same principle: a small jet engine that can fit in one shipping container (with auxiliary intake, controls, and exhaust equipment fitting in 1-2 more shipping containers). Testing for this engine is still underway, but preliminary advertised specs suggest the Superpower will produce 42 MW per unit, even at high ambient air temperatures.

Source: @bscholl, Twitter

The first 1.2 GW of production has already been booked for Crusoe, with a targeted 200 MW of production in 2027 and 1 GW in 2028, and 2 GW in 2029. The initial order price suggests a hardware cost of $1,000/kW, but that figure does not include balance of plant, shipping, or commissioning, and should not be directly compared against all-in cost figures. Boom Supersonic have vertically integrated production for blade and vane production, but rely on external vendors for metallurgy, which may remain a supply chain bottleneck.

We haven’t yet seen other firms jump on the retrofit wagon. However, medium-speed engines are largely manufactured by firms with a long experience building ship engines – such as Wärtsila. In fact, they are largely the same engines and can be manufactured in the same facility. When will we see old ship engines retrofitted to power datacenters?

Let’s now turn our attention to comparing the different solutions and manufacturers. We’ll also analyze the economics and TCO of onsite power generation, and compare it to the electric grid in the US.

Onsite gas TCO analysis and leading manufacturers...