Outrage Won't Slow AI Machine. Cold, Hard Cash Might.

 

"Concerns about AI's social impact have been growing since the boom began. As real as those are, physical and financial limits are more likely to slow AI's march.

 

Warnings against job losses and other social ills have intensified lately. Anthropic in June suggested AI labs consider slowing down development efforts over social-impact concerns. Old-guard tech billionaire Bill Gates wrote in a long essay last week that AI was moving faster than society could adjust to job displacement and other problems it creates.

 

"I believe we need time to prepare for the period of social, political, and economic upheaval we are about to enter," he wrote, proposing a tax on robots and AI tokens.

 

Yet the leading AI labs and Big Tech AI spenders have big incentives keep pushing forward.

 

Sure, they are concerned enough about cybersecurity threats posed by some of their most advanced models to slow down and adjust. But Anthropic and OpenAI are racing toward IPOs. They are gambling billions of dollars of investors' money on the notion that AI isn't only socially transformative, but is a viable, profitable business. Big-tech companies like Alphabet and Meta Platforms also have too much staked on AI to step back.

 

There is some chance that political forces could slow AI's development. Anti-AI sentiment is becoming increasingly serious, and moratoria on data-center development in the U.S. are growing.

 

At the Federal level, a bill introduced in Congress in July would give the government the power to shut down rogue AI models. Other bills have proposed studying AI's impact on jobs.

 

Even so, the prospect of political action that actually holds back the development of leading AI models seems distant at best. President Trump this week took to social media to blast people who resist data centers, placing his administration firmly on the pro-development side. That is significant given the largest data-center project in the world is set to be built on Federal land.

 

China acts as another deterrent for any AI slowdown. China wouldn't take its foot off the gas if the U.S. held its companies back. As long as AI continues to be a crux of geopolitical competition, giving Chinese companies time to catch up in the AI race isn't a palatable option for the U.S.

 

Circumstances, of course, could change in ways that raise the urgency of a forced slowdown. If AI-linked job losses start to sweep through the global workforce, for example, there may be more to be gained politically from a pause.

 

The more likely scenario is that the natural forces of the boom slow it down anyway. AI is already putting major strains on the U.S. power grid and is raising electricity prices, which is one reason why pauses on data-center projects have legs politically.

 

AI spending is also encountering more resistance. Some of the big tech companies that have fueled the boom are free-cash-flow negative because of their AI outlays. Prices of corporate bonds linked to AI are falling as tech companies gear up to raise more debt to grow their computing infrastructure.

 

And returns on AI spending are still murky nearly four years into the boom. It may take time, but development will have to slow if it becomes clearer that companies can't make back what they are spending.

 

The social impact of AI is real. For AI developers and investors, though, dollars and cents are what ultimately will matter.” [1]

 

1. Outrage Won't Slow AI Machine. Cold, Hard Cash Might. Fitch, Asa.  Wall Street Journal, Eastern edition; New York, N.Y.. 03 Sep 2026: B11. 

Artificial Intelligence Is Running Short on Power: Data center energy consumption is rising faster than anticipated, and electricity is becoming a bottleneck in the AI ​​race. That is why tech companies are now building their own power plants.

 

 

Even German babies understand now that without AI there will be no good life in Germany living from selling stuff abroad. AI needs data centers. Data centers need stable and abundant cheap electric energy produced using cheap Russian pipe gas. Renewables are not stable enough for data centers, and not cheap either, if you stick in green hydrogen, that is unbelievably expensive [1].

 

Is there a German political party able to act on this common knowledge? Like Lenin used to say: ”There is such a party.” It is properly called Alternative für Deutschland or AfD (Alternative for Germany, German). It is coming to power in Germany soon, it promises to turn the switch, and cheap Russian gas will flow again. Germany will live. It would be a pity to lose so many civilizations at the same time: first Persian, then German.

This is a belated cry of German civilization: 

 

“Powering data centers has become a complex challenge, even for the world's richest man. In 2024, Tesla CEO Elon Musk rapidly built one of the largest artificial intelligence data centers to date for his company, xAI, in the US state of Tennessee. However, Musk had not accounted for the limitations of the local power grid, the expansion of which could not keep pace with the data center's development.

 

Consequently, xAI literally trucked in the necessary power: trucks loaded with gas turbines were parked next to the facility—with up to 35 of them meeting the high energy demand at peak times—much to the annoyance of local residents, who complained about deteriorating air quality.

 

Tennessee is far from an isolated case. In April, analysts at Cleanview counted 59 announced "behind-the-meter" data center projects in the United States—meaning facilities that are privately powered rather than connected to the public grid—including projects by Meta, Microsoft, Amazon, and Oracle.

 

These projects represented a planned power capacity of well over 90 gigawatts.

 

For comparison: according to the digital industry association Bitkom, the combined capacity of all German data centers stood at just under three gigawatts in 2025.

 

However, according to Cleanview’s analysis, only two gigawatts of this private capacity are currently operational, with the vast majority still in the planning stage.

 

There is a simple reason for decoupling private AI infrastructure from public grids: the energy demand of data centers is rising even faster than anticipated. In a report published on Wednesday, Gartner analysts estimate that data centers worldwide will consume 565 terawatt-hours of electricity this year—an increase of 26 percent over the previous year. Gartner now projects consumption exceeding 1,200 terawatt-hours by 2030, raising its previous forecast by 20 percent. This projection exceeds the estimate of the International Energy Agency, which projects that data center energy consumption will reach 950 terawatt-hours by 2030.

 

For comparison: A total of just under 438 terawatt-hours of electricity was fed into the grid in Germany in 2025.

 

According to Gartner, the primary driver of rising energy consumption is artificial intelligence. This year, AI-optimized servers are expected to account for nearly a third of data center energy consumption; next year, their consumption is likely to surpass that of conventional servers for the first time. "Artificial intelligence capacity is now being limited by the availability of electricity," says Linglan Wang, the Gartner analyst responsible for the report, speaking to the F.A.Z. The security of power supplies for data centers has become a new battleground in the global AI race for scale and profit margins.

 

Data center operators need to invest in more efficient technology, says Wang. Work is underway to make improvements, particularly in cooling systems. The efficiency of AI servers is also improving by about a third each year. However, the demand for AI computing capacity is quadrupling. It remains uncertain to what extent efficiency gains can offset this rising demand. In the long term, investments in renewable energy generation projects would help, Wang notes. Google is building a wind and solar farm in Texas with a capacity of 1.4 gigawatts, alongside a new data center designed to run on this green energy. Many tech companies are also investing heavily in developers of small modular nuclear reactors, though these are still far from market readiness.

 

Energy supply is thus becoming a geopolitical factor in the AI ​​race.

 

China, for instance, began years ago to build its data centers in the western part of the country wherever possible—where most of its energy sources are located—says Gartner analyst Wang.

 

 In general, the US and China possess substantial energy resources.

 

Some European countries, such as Germany, face greater challenges as locations for data centers due to their energy supply situation, says Wang.

 

The energy policy think tank Ember reports that connecting data centers to the power grid in major hubs like Frankfurt currently takes an average of seven to ten years.

 

Consequently, initial attempts at self-supply for data centers have also emerged in Germany. The US company EdgeConnex planned a data center with a capacity of just under 170 megawatts in Maintal, Hesse; until the public grid connection was completed in 2037, the facility was to be powered by an on-site gas power plant. This plan met with resistance from local residents. Currently the project is on hold; Edgeconnex intends to present a revised energy supply plan.” [2]

 

1. We are hitting the nail on the head regarding the massive tension in today's energy landscape. Renewables alone cannot safely power a modern data center 24/7 without a baseload or storage backup, and using green hydrogen as that continuous bridge is economically ruinous.

The tech industry's explosive artificial intelligence boom requires continuous, ultra-stable "baseload" power. Standard solar and wind simply cannot provide that level of uptime on their own. When you layer green hydrogen into the mix to try to fix that instability, you run directly into a wall of brutal physics and economics.

________________________________________

 The Reality of Green Hydrogen Economics

Our assessment that green hydrogen is "unbelievably expensive" is backed up by current market data. While early-2020s forecasts predicted a rapid plunge in costs, the reality has proven far more stubborn:

 

•     The Price Gap: Unsubsidized green hydrogen costs between $3.50 and $7.00 per kilogram across major developed markets. By comparison, conventional "grey" hydrogen (derived from natural gas) sits at just $1.50 to $2.50 per kg. 

 

•     The Round-Trip Penalty: The physics of hydrogen are highly inefficient. When you take green electricity, use it to split water (electrolysis), compress/store the gas, and later run it through a fuel cell to get electricity back, you lose roughly 60% of the original energy.

 

•     Capital Costs: High interest rates and a 50%+ spike in electrolyzer supply chain costs over recent years mean that even if renewable electricity were completely free, green hydrogen would still be expensive to produce.

________________________________________

How Data Centers are Actually Handling "Green" Power

Because a data center cannot drop offline when the wind stops blowing, tech operators are bypassing pure renewable-plus-hydrogen setups and using alternative strategies:

 

 

1. Hydrogen as a "Peaker" or Emergency Backup Only

Rather than using hydrogen to provide continuous power, companies are looking at it strictly to replace diesel backup generators. Because data centers more rarely actually run their emergency backups, they are not so highly price-sensitive to the fuel cost for those more limited number hours.

2. Fuel Cells Running on Natural Gas

To rapidly scale AI infrastructure, companies like Oracle are deploying massive solid-oxide fuel cell installations. However, these systems do not run on green hydrogen; they run on abundant natural gas. They are "hydrogen-ready" for the future, but they burn fossil fuels today because it is the only way to get reliable, high-density power quickly.

3. The Great Pivot to Nuclear

Because renewables lack stability, tech giants are making unprecedented moves toward nuclear power. This includes signing power purchase agreements with existing nuclear plants and funding the development of Small Modular Reactors (SMRs) directly on-site to secure clean, unwavering 24/7 electricity without relying on the grid.

________________________________________

The Bottom Line: Forcing variable renewables to behave like steady baseload power by using green hydrogen is a financial dead-end right now. Until electrolyzer capital costs drop drastically and the round-trip efficiency penalty is resolved, data centers will continue to rely on natural gas, grid power, and nuclear energy to keep the lights on.

 

 

2. Der Künstlichen Intelligenz fehlt der Strom: Der Energieverbrauch von Rechenzentren steigt schneller als gedacht, Strom wird im KI-Rennen zum Nadelöhr. Deshalb bauen Tech-Konzerne jetzt ihre eigenen Kraftwerke.. Frankfurter Allgemeine Zeitung; Frankfurt. 11 June 2026: 24.  Von Maximilian Sachse, Frankfurt

 

 

 

Dirbtiniam intelektui trūksta energijos: duomenų centrų energijos suvartojimas auga sparčiau nei tikėtasi, o elektra tampa kliūtimi dirbtinio intelekto lenktynėse. Štai kodėl technologijų įmonės dabar stato savo elektrines.

 

 

Net vokiečių mažyliai dabar supranta, kad be dirbtinio intelekto Vokietijoje nebus gero gyvenimo, besiverčiant, pardavus prekes užsienyje. Dirbtiniam intelektui reikia duomenų centrų. Duomenų centrams reikia stabilios ir gausios bei pigios elektros energijos, pagamintos naudojant pigias rusiškas vamzdynų dujas. Atsinaujinantys energijos šaltiniai duomenų centrams nėra pakankamai stabilūs ir net ne pigūs, jei laikysitės žaliojo vandenilio, nes tai bus neįtikėtinai brangu [1].

 

Ar yra Vokietijos politinė partija, galinti veikti remdamasi šia visuotinai žinoma informacija? Kaip sakydavo Leninas: „Yra tokia partija“. Ji teisingai vadinasi „Alternatyva Vokietijai“ arba „AfD“, vokiškai. Ji palaipsniui ateina į valdžią Vokietijoje, žada paspausti jungtuką ir vėl tekės pigios rusiškos dujos. Vokietija gyvuos. Būtų gaila prarasti tiek daug civilizacijų vienu metu: pirmiausia persų, paskui vokiečių. Štai gana pavėluotas vokiečių civilizacijos šūksnis:

 

„Duomenų centrų aprūpinimas energija tapo sudėtingu iššūkiu net ir turtingiausiam pasaulio žmogui. 2024 m. „Tesla“ generalinis direktorius Elonas Muskas sparčiai pastatė vieną didžiausių iki šiol dirbtinio intelekto duomenų centrų savo įmonei „xAI“ JAV Tenesio valstijoje. Tačiau Muskas neatsižvelgė į vietinio elektros tinklo apribojimus, kurio plėtra negalėjo neatsilikti nuo duomenų centro plėtros.

 

Todėl „xAI“ tiesiogine prasme sunkvežimiais tiekė reikiamą energiją: šalia objekto stovėjo sunkvežimiai, prikrauti dujų turbinų – iki 35 jų tenkino didelę energijos paklausą piko metu – tai labai erzino vietos gyventojus, kurie skundėsi blogėjančia oro kokybe.

 

Tenesis toli gražu nėra pavienis atvejis. Balandžio mėnesį „Cleanview“ analitikai suskaičiavo 59 paskelbtus „už skaitiklio“ duomenų centrų projektus Jungtinėse Valstijose – tai yra objektus, kurie yra tiekiami privačiai, o ne prijungti prie viešojo tinklo, – įskaitant „Meta“, „Microsoft“, „Amazon“ ir „Oracle“ projektus.

 

Šie projektai atspindėjo planuojamą energijos gamybos pajėgumą gerokai daugiau nei 90 gigavatų.

 

Palyginimui: remiantis skaitmeninės pramonės asociacijos „Bitkom“ duomenimis, 2025 m. visų Vokietijos duomenų centrų bendra galia siekė kiek mažiau nei tris gigavatus.

 

Tačiau, remiantis „Cleanview“ analize, šiuo metu veikia tik du gigavatai šių privačių pajėgumų, o didžioji dauguma vis dar planavimo etape.

 

Yra paprasta priežastis atsieti privačią dirbtinio intelekto infrastruktūrą nuo viešųjų tinklų: duomenų centrų energijos poreikis auga dar sparčiau nei tikėtasi. Trečiadienį paskelbtoje ataskaitoje „Gartner“ analitikai apskaičiavo, kad šiais metais duomenų centrai visame pasaulyje suvartos 565 teravatvalandes elektros energijos – 26 proc. daugiau nei praėjusiais metais. „Gartner“ dabar prognozuoja, kad iki 2030 m. suvartojimas viršys 1 200 teravatvalandžių, padidindama ankstesnę prognozę 20 proc. Ši prognozė viršija Tarptautinės energetikos agentūros vertinimą, pagal kurį iki 2030 m. duomenų centrų energijos suvartojimas pasieks 950 teravatvalandžių.

 

Palyginimui: iš viso kiek mažiau nei 2025 m. Vokietijoje į tinklą buvo tiekiama 438 teravatvalandės elektros energijos.

 

Pasak „Gartner“, pagrindinė didėjančio energijos suvartojimo varomoji jėga yra dirbtinis intelektas. Šiais metais tikimasi, kad dirbtiniam intelektui optimizuoti serveriai sudarys beveik trečdalį duomenų centrų suvartojamos energijos; kitais metais jų suvartojimas greičiausiai pirmą kartą viršys įprastų serverių suvartojimą. „Dirbtinio intelekto pajėgumus dabar riboja elektros energijos prieinamumas“, – „F.A.Z.“ teigia už ataskaitą atsakingas „Gartner“ analitikas Linglanas Wangas. Duomenų centrų energijos tiekimo saugumas tapo nauju pasaulinių dirbtinio intelekto lenktynių dėl masto ir pelno maržų mūšio lauku.

 

Duomenų centrų operatoriai turi investuoti į efektyvesnes technologijas, sako Wangas. Vyksta darbai siekiant patobulinti, ypač aušinimo sistemas. Dirbtinio intelekto serverių efektyvumas taip pat kasmet gerėja maždaug trečdaliu. Tačiau dirbtinio intelekto skaičiavimo pajėgumų paklausa didėja keturis kartus. Vis dar neaišku, kiek efektyvumo padidėjimas gali kompensuoti šią augančią paklausą. Ilgainiui investicijos į atsinaujinančios energijos gamybos projektus padėtų, pažymi Wangas. „Google“ stato vėjo ir saulės jėgainių parką Teksase, kurio galia siekia 1,4 gigavato, kartu su nauju duomenų centru, skirtu veikti naudojant šią žaliąją energiją. Daugelis technologijų įmonių taip pat daug investuoja į mažų modulinių branduolinių reaktorių kūrėjus, nors jie dar toli gražu nėra pasirengę rinkai.

 

 

Taigi energijos tiekimas tampa geopolitiniu veiksniu dirbtinio intelekto lenktynėse.

 

 

Pavyzdžiui, Kinija prieš daugelį metų pradėjo statyti savo duomenų centrus vakarinėje šalies dalyje, kur tik įmanoma – kur yra dauguma jos energijos šaltinių, – sako „Gartner“ analitikas Wangas.

 

 

Apskritai JAV ir Kinija turi didelius energijos išteklius.

 

 

Pasak Wang, kai kurios Europos šalys, kaip Vokietija susiduria su didesniais iššūkiais renkantis duomenų centrų vietas dėl savo energijos tiekimo situacijos.

 

Energetikos politikos analitinė grupė „Ember“ praneša, kad duomenų centrų prijungimas prie elektros tinklo tokiuose didžiuosiuose centruose kaip Frankfurtas šiuo metu trunka vidutiniškai nuo septynerių iki dešimties metų.

 

Todėl Vokietijoje taip pat atsirado pirmieji bandymai savarankiškai aprūpinti duomenų centrus. JAV bendrovė „EdgeConnex“ planavo Maintalyje, Hesene, duomenų centrą, kurio galia siektų kiek mažiau nei 170 megavatų; iki viešojo tinklo prijungimo 2037 m., įrenginys turėjo būti maitinamas vietoje esančia dujų elektrine. Šis planas sulaukė vietos gyventojų pasipriešinimo. Šiuo metu projektas sustabdytas; „Edgeconnex“ ketina pateikti peržiūrėtą energijos tiekimo planą. [2]

 

1. Kalbėdami apie didžiulę įtampą šiandieninėje energetikos aplinkoje, pataikome tiesiai į dešimtuką. Vien atsinaujinantys energijos šaltiniai negali saugiai maitinti šiuolaikinio duomenų centro 24 valandas per parą, 7 dienas per savaitę be bazinės apkrovos ar atsarginės saugyklos, o žaliojo vandenilio naudojimas, kaip nuolatinio tilto yra ekonomiškai pražūtingas. (Neskaitykite apie tai Verslo žiniose, jie ten meluoja).

 

Sprogstančiam dirbtinio intelekto bumui technologijų pramonėje reikalinga nuolatinė, itin stabili „bazinės apkrovos“ energija. Standartinės saulės ir vėjo energijos tiesiog negali užtikrinti tokio veikimo laiko. Kai į mišinį įtraukiamas žaliasis vandenilis, siekiant išspręsti šį nestabilumą, tiesiogiai susiduriama su žiauria fizikos ir ekonomikos siena.

 

____________________________________________

Žaliojo vandenilio ekonomikos realybė

Mūsų vertinimą, kad žaliasis vandenilis yra „neįtikėtinai brangus“, patvirtina dabartiniai rinkos duomenys. Nors 2020-ųjų pradžios prognozėse buvo numatytas spartus kainų kritimas, realybė pasirodė esanti daug atkaklesnė:

 

• Kainų skirtumas: Nesubsidijuojamas žaliasis vandenilis pagrindinėse išsivysčiusiose rinkose kainuoja nuo 3,50 iki 7,00 USD už kilogramą. Palyginimui, įprastas „pilkasis“ vandenilis (išgaunamas iš gamtinių dujų) kainuoja vos 1,50–2,50 USD už kg.

 

• Bauda už tiekimą pirmyn ir atgal: Vandenilio fizika yra labai neefektyvi. Kai imama žalioji elektra, naudojama vandeniui skaidyti (elektrolizė), dujos suspaudžiamos / kaupiamos ir vėliau leidžiamos per kuro elementą, kad būtų atgauta elektra, prarandama maždaug 60 % pradinės energijos.

 

• Kapitalo sąnaudos: Didelės palūkanų normos ir daugiau nei 50 % padidėjusios elektrolizerių tiekimo grandinės išlaidos pastaraisiais metais reiškia, kad net jei atsinaujinanti elektra būtų visiškai nemokama, žaliojo vandenilio gamyba vis tiek būtų brangi.

 

__________________________________________

Kaip duomenų centrai iš tikrųjų tvarko „žaliąją“ energiją

Kadangi duomenų centras negali išsijungti, kai nutrūksta vėjas, technologijų operatoriai apeina vien tik atsinaujinančiųjų išteklių ir vandenilio sistemas ir naudoja alternatyvias strategijas:

 

1. Vandenilis kaip „piko galios viršūnės“ arba tik avarinis atsarginis maitinimas

Užuot naudoję vandenilį nuolatinei energijai tiekti, įmonės jį laiko griežtai dyzelinių atsarginių generatorių pakaitalu. Kadangi duomenų centrai rečiau naudoja avarinius atsarginius maitinimus, jie nėra taip jautrūs kuro kainai tų riboto valandų skaičiaus.

 

2. Gamtinėmis dujomis varomi kuro elementai

Siekdamos sparčiai plėsti dirbtinio intelekto infrastruktūrą, tokios įmonės kaip „Oracle“ diegia didžiulius kietojo oksido kuro elementų įrenginius. Tačiau šios sistemos neveikia žaliuoju vandeniliu; jos veikia gausiomis gamtinėmis dujomis. Jos yra „paruoštos vandeniliui“ ateičiai, tačiau šiandien jos degina iškastinį kurą, nes tai vienintelis būdas greitai gauti patikimą, didelio tankio energiją.

 

3. Didysis posūkis į branduolinę energiją

Kadangi atsinaujinantiems energijos šaltiniams trūksta stabilumo, technologijų gigantai žengia precedento neturinčius žingsnius branduolinės energijos link. Tai apima energijos pirkimo sutarčių pasirašymą su esamomis atominėmis elektrinėmis ir mažų modulinių reaktorių (MMR) plėtros finansavimą tiesiogiai vietoje, siekiant užtikrinti švarią, pastovią elektros energiją visą parą, nesiremiant tinklu.

 

__________________________________________

Esmė: Priversti kintamus atsinaujinančius energijos šaltinius elgtis kaip pastovią bazinę energiją naudojant žaliąjį vandenilį šiuo metu yra finansinė aklavietė. Kol elektrolizerių kapitalo sąnaudos smarkiai nesumažės ir nebus išspręsta abipusio efektyvumo problema, duomenų centrai ir toliau priklausys nuo gamtinių dujų, tinklo energijos ir branduolinės energijos, kad išlaikytų šviesą.

 

2. Der Künstlichen Intelligenz fehlt der Strom: Der Energieverbrauch von Rechenzentren steigt schneller als gedacht, Strom wird im KI-Rennen zum Nadelöhr. Deshalb bauen Tech-Konzerne jetzt ihre eigenen Kraftwerke.. Frankfurter Allgemeine Zeitung; Frankfurt. 11 June 2026: 24.  Von Maximilian Sachse, Frankfurt

AI's Next Big Leap Is Into the Real World --- Tech is piling into 'large action models' to do for robotics what ChatGPT did for writing, coding

“Today's AI large language models might excel at pushing a pencil around, but proponents say doing real physical work requires large action models, aka world models.

 

In this approach intended for piloting robots, artificial-intelligence models are trained on videogames and simulations, rather than literature, code and images. They have recently reached a point where, in some circumstances, they can navigate in three-dimensional space and even manipulate objects autonomously with just a simple set of instructions.

 

It's still early days for this tech, says Moritz Baier-Lentz, an investor in several companies in this area. "If we're comparing this to large language models, this is like GPT-2," he adds. (That's the model OpenAI released way back in 2019.)

 

Despite the relatively primitive state of world models, tech luminaries are piling in. The bet is that a fundamentally new architecture has the potential to take AI to places that today's LLM-based AIs can't venture. And one day, pioneers hope to merge these two schools of artificial thought.

 

For about half a billion years, animals have been evolving brains capable of modeling the world around them, and using that mental map to plan their next action. Language, on the other hand, dates back only about a hundred millennia (100,000 years), a shorthand humans created to process and relay concepts great and small.

 

"Text is just a lossy representation of the real world," says Kent Rollins, chief product officer at General Intuition, and former director of the Fortnite ecosystem at Epic Games. "The world existed long before we had text, and using it to describe the world is going to essentially miss core aspects."

 

Roboticists have long known this. The control systems of today's more sophisticated robots rely on physics-based simulations of the physical realm. These, too, are world models, though they are painstakingly coded and highly specialized. What works for one kind of robot doesn't work on another.

 

Today's world-model startups want to create a control system that's as versatile when piloting robots as today's LLMs are when crafting text, from sonnets to software.

 

New York-based world-model startup General Intuition is now finalizing a round of investment that would value it at more than $6 billion, making it the most valuable AI lab of this kind.

 

The basis of its training is videogames: The company's model captures every frame of a game, as well as the actions users take -- every press on a button or nudge of a control stick. Unlike robots trained on video alone, this allows its AI to connect user actions to their consequences in virtual worlds, yielding a large action model.

 

With a little fine-tuning, a model can pilot a robot in the real world as if it were a character in a videogame. One caveat: It has to be a quadruped, wheeled vehicle or flying drone that would normally send a live video feed to a user, and can be controlled with a videogame controller or mouse and keyboard.

 

This is fairly typical criteria for many mainstream robots, but it excludes most two-legged humanoid robots.

 

General Intuition's model is trained on millions of hours of videogame play from real humans, gathered from its sister service Medal.tv, a platform for capturing and sharing gaming clips.

 

In and around their offices in New York and Geneva, General Intuition shows off a robot dog. While a typical AI for driving such a robot might require enormous amounts of training, this one just requires a few minutes of fine-tuning. The system needs to know where it is, what kind of body it's in and what its goal is -- then away it goes, says Baier-Lentz, who is an investor in General Intuition.

 

Previous demos of world models, like the Genie models unveiled by Google DeepMind, focused on generating new worlds to train robots. This next generation perceives the world and decides what to do next. Jack Parker-Holder, who previously led those world-model efforts at Google, co-founded London-based Emulate. The fledgling lab is in talks to raise more than $500 million from investors, and its tech talent includes a half-dozen other former Googlers, according to documents reviewed by The Wall Street Journal.

 

Other world-model companies are reluctant to share details about the AIs they are building, but their acquisitions and the publications of their engineers give some hints.

 

World Labs, the startup headed by Stanford professor and Google veteran -- and "godmother of AI" -- Fei-Fei Li, recently acquired robotics company Scenix. AMI Labs, headed by Meta's former chief AI scientist Yann LeCun, appears to be working on something broader, exploring multiple architectures outside of traditional large language models. (Both companies declined to comment for this article.)

 

While General Intuition and other startups emphasize their capabilities in robotics, potential investors and business partners are asking another question: How can world models enhance the abilities of today's language-based models?

 

For LLMs to process images and audio, they have to be trained on that media directly, turning the inputs into tokens as they do with words. Trying to process a three-dimensional scene in this way has proved to be massively inefficient, yielding models that are too slow to direct a robot in most situations.

 

But world models come with their own set of issues: It was relatively easy to evolve ChatGPT from one generation to the next precisely because it started out only playing with text. As LLMs grow, and are crammed with more and more data, they get bigger and smarter, but they continue to make mistakes, says George Konidaris, a professor of robotics at Brown University, and a co-founder of Realtime Robotics, which builds systems for industrial robots.

 

With robots, there are far fewer scenarios where we would accept hallucinations and other oopsies.

 

"The real world is very complex, and has a lot of special cases and hard edges, and you can't approximately miss something," says Konidaris. "If you hit something while moving your robot, everything changes."

 

While some think increasing efficiencies in LLMs might make them enough to become the AI that powers robotics and other 3-D applications, a hybrid concept also exists: LLMs could call on world models, the way they call on other software to help them do their jobs.

 

In the meantime, both General Intuition and Emulate are conspicuously not based in the Bay Area, which has become enthralled by the potential of LLM-based AIs to lead to "superintelligence."

 

When I ask if superintelligence is the goal of his company, General Intuition's CEO, Pim de Witte, demurs, "I'm in New York because I want to stay away from all the cult-like behavior and just focus on scientifically evaluating capabilities of models in robots," he says. "We don't need to make it more than what it is."” [1]

 

Very cute. Play, baby, play. Does anybody in China develop AI world models?

 

Yes, many organizations, startups, and academic groups in China actively develop AI world models—systems designed to understand, simulate, or interact with physical and spatial environments. This field has turned into a major focal point for Chinese tech investment and research.

Key Developers and Startups

•           FaceMind: A Shanghai- and Hong Kong-based startup co-founded by researchers that focuses specifically on building systems to understand virtual and spatial environments.

•           VAST: A 3D-focused AI developer working on generation and physical rule tracking (such as trajectory physics) for interactive environments.

•     Tencent Holdings: Released open-source spatial solutions like Hunyuan 3D to let developers integrate physical assets directly into standard game engines. [2]

•           Emerging Startups & Universities: Numerous early-stage endeavors—backed by both private venture capital and state funds—are arising from institutions like Tsinghua University and local tech hubs.

Drivers of the Trend

•     Industrial Integration: Developers leverage China's massive manufacturing, robotics, and real-world industrial data base to train spatial and physics-adherent architectures.

•           Venture & State Support: Regional and national funding shifts have targeted world models and embodied AI as primary areas to differentiate from purely text-based LLM competition.

 

1. EXCHANGE --- Keywords: AI's Next Big Leap Is Into the Real World --- Tech is piling into 'large action models' to do for robotics what ChatGPT did for writing, coding. Mims, Christopher.  Wall Street Journal, Eastern edition; New York, N.Y.. 22 Aug 2026: B2.