In Silicon Valley in the fall of 2026, the celebratory champagne has barely stopped flowing. OpenAI launched its new flagship model, GPT-6 Astra, which President Greg Brockman called a historic turning point. Nvidia CEO Jensen Huang also tweeted that general AI has arrived. Judging by what the top executives are saying, the AI arms race is nearing its finish line, and it is now time for commercial harvesting.
Offstage, however, the story is completely different. Over the past two years, a wave of top researchers who directly built today’s mainstream large models have left to strike out on their own. Former OpenAI Chief Scientist Ilya Sutskever, former Meta Chief AI Scientist Yann LeCun, and former Google Chief Scientist Jeff Dean, along with Stanford professor Fei-Fei Li and a cohort of technical leads who left major tech companies around the same time, have founded a total of 45 new companies. According to a recent report by German tech media outlet THE DECODER, confirmed funding across these startups has reached $53.8 billion. Counting only the companies that have disclosed their valuations, their combined worth is already approaching $245 billion.
These veteran model builders know the inner workings of Big Tech better than anyone. The fact that they are lining up to leave suggests the current technical paradigm has hit a real wall. Yet a closer look at where the money is going is enough to make anyone sweat: more than half of the funding has gone to empty-handed companies that have not released a single line of public code. Where do they think today’s models fall short? And with all that capital deployed, who actually has substance and who is just selling hype?
When we use ChatGPT, Claude, or Gemini to write code or revise proposals, the models seem to get noticeably smarter every few months, making it easy to assume that scaling up compute along the current trajectory will eventually solve everything. In the eyes of the scientists who design the underlying architectures, however, today’s chat-based large language models are born with three fatal flaws that cannot be fixed by feeding them more text. Only by understanding these three root problems can we make sense of what this new wave of founders is trying to build.
First, they are frozen at birth. Whether it is GPT, Gemini, or Claude, once a model finishes training and ships, its neural network connections are locked in place. No matter how much reference material or how large a retrieval database you give it, the model is merely cribbing from a cheat sheet within a single conversation. Close the window and start over, and it will step into the exact same traps today that it fell into yesterday. It cannot review its mistakes on the job like a human intern, wire those lessons into its brain, and get sharper with practice.
The second shortcoming is armchair theorizing: memorizing every recipe in the book without ever cracking an egg by hand. Large models have read every physics textbook on the internet cover to cover, but web text is ultimately just a post-hoc human summary of reality. How much force to ease off the instant your fingertips touch an eggshell, how friction shifts when you grip a wet cup at an angle, or how to cleanly catch a sliding part on a high-speed conveyor belt: none of that precise 3D spatial scaling or tactile force feedback is written into web pages. Put a pure text model directly in charge of high-frequency physical control, and it falls apart the moment it meets real-world gravity and changing friction.
The third practical bottleneck is that Big Tech’s web crawlers have nearly exhausted the public text on the internet. For writing articles, coding, or handling customer support, humans have left behind massive troves of examples online over the years, allowing models to learn by imitation and perform convincingly. But step into frontier sciences like novel small-molecule drugs, advanced solid-state materials, extreme aerodynamics, or deep-sea microbial genomics, and nature has not posted the answers online. Chewing through old papers, no matter how much compute you throw at them, will never conjure up real physical, chemical, or biological experimental results.
Most of the 45 new labs are searching for remedies to these exact bottlenecks. Some want models to build lasting memory as they work, some are teaching models to understand 3D space and control machines, others are doing the hard work of collecting first-hand scientific experimental data or having AI run its own research verification, and still others are building model-customization tools or new hardware. Seeing the flaws of the old path is one thing, though; actually delivering a working cure is another.
When most practitioners and investors look at this startup wave, their first instinct is to check which founder has the biggest name and raised the most money. But if you line up the product progress of these 45 new companies against their fundraising totals, a jaw-dropping inversion emerges: the 32 teams that have earnestly built and released public products, APIs, or open-weight models have raised only about $20.9 billion in total. Meanwhile, 8 companies that have yet to release a single public model, API, or research paper have walked away with nearly $30 billion on promises alone, taking nearly 56% of all funding. The remaining $2.9 billion went to 5 companies still in private beta.
Whose hands did the lion’s share of that nearly $30 billion land in? The biggest fundraising magnet, Prometheus, focuses on industrial manufacturing and scientific automation. Although it is led by Amazon founder Jeff Bezos and the former head of Google’s life sciences unit, its R&D roster is packed with core researchers poached from major AI labs; that single company has raised $18.2 billion in total investment, yet the outside world has not seen a trace of a product.
The heaviest concentration of paper-only fundraising is in continual learning, or teaching models to build lasting memory on the job, which is the hottest direction in AI right now. Companies in this category have raised $10.85 billion in total, 99% of which flowed into 5 teams with neither a product nor a paper. The poster child is Safe Superintelligence, or SSI, co-founded by former OpenAI Chief Scientist Ilya Sutskever; refusing to build a conventional chatbot first and without publishing a single paper to date, it has already raised about $8 billion. The handful of smaller teams that actually have public code or APIs were left splitting the scraps.
Meanwhile, the companies that actually put their models to the test were immediately blindsided by a general-purpose foundation model. In a simulation benchmark run by Nvidia’s Seattle Robotics Lab, GPT-6 Astra, a general-purpose large model never specifically trained to control robots, borrowed a coding-assistant harness to issue step-by-step action commands to a robot and took first place with a 74.8% success rate. By contrast, π0.5, a specialized robotic action model from multibillion-dollar robotics unicorn Physical Intelligence, scored just 28.0% on the exact same benchmark. In a separate real-world robotic arm evaluation released the same day, where models controlled a robotic arm on a static tabletop via a coding-agent harness to complete tasks like handwriting and block stacking, Astra took first place yet again.
How did a chat model that never practiced driving a robot beat the specialists? Because on a stationary tabletop, the real test is understanding complex instructions and sequencing multi-step plans. With its broad world knowledge and strong reasoning, a general-purpose large model naturally holds a massive advantage there. Yet its fatal flaw is just as obvious: it is far too slow and far too expensive. During testing, Astra paused for more than 13 seconds to think before every single move, taking an average of 30 minutes of deliberation to finish one task on a static desk. Running a single task through the two benchmarks also racked up $16 to $36 in API fees. Spending half an hour carefully trial-and-erroring on a fixed tabletop can win a benchmark, but it is useless on a moving factory assembly line or in a real home. A machine that zones out for a dozen seconds before every movement will never catch a slipping part, and nobody is going to pay tens of dollars just to get the dishes washed.
By comparison, a specialized small robotics model makes an action decision in about 0.05 seconds, roughly 300 times faster than a general-purpose model. Big Tech’s flagship models excel at interpreting complex scenes and planning multi-step tasks, but they cannot perform millisecond-level real-time motor adjustments. If robotics startups want to survive Big Tech’s squeeze, their way out is not competing over who can plan more elaborate task sequences, but accumulating the real-world physical contact data that Big Tech lacks to train a rock-solid sensorimotor cerebellum capable of split-second hand-eye reflexes.
Another group of founders who set out to train models from scratch to simulate realistic 3D physical environments quickly did the math as well: a startup’s billion-dollar war chest is nowhere near enough to fill the bottomless pit of long-term GPU purchases and low-level hardware optimization. Spatial intelligence startup World Labs, co-founded by Fei-Fei Li, raised a massive round early this year, only to fold into a hardware giant a few months later in order to secure dedicated silicon support. Chipmaker AMD recently issued an official acquisition filing announcing it would buy the company for roughly $8.2 billion in an all-stock deal; once the transaction closes, Fei-Fei Li will become AMD’s Chief Scientist, reporting directly to AMD CEO Lisa Su.
Among these 45 new companies, what kind of team can genuinely withstand the pressure from OpenAI and Google? The startups that have delivered solid results and established a real foothold all hold one of two trump cards. The first is real-world physical and biochemical experimental data that Big Tech’s web crawlers cannot scrape, much like the real-machine contact data in robotics.
In frontier science, the companies building truly deep moats are the ones willing to go out into the physical world to collect samples and generate proprietary data. Biotech startup Basecamp Research didn’t sit in an office scraping existing databases online; instead, it signed benefit-sharing agreements compliant with international conventions across 26 countries and sent teams to polar regions, the deep sea, and volcanic vents to collect samples, assembling an exclusive database of 9.8 billion novel genes, equivalent to 10 times the size of public gene repositories. Today, pharmaceutical researchers can simply click a dedicated connector plugin right inside Claude to call Basecamp’s biological models and design new genes.
Another breed of data builder relies on self-built automated labs, exemplified by Lila Sciences, which has raised $2.55 billion. The company built fully automated chemical and materials synthesis facilities where models propose candidates and robotic arms synthesize and test them on the spot, closing the loop between computer simulation and physical experimentation. Negotiating cross-border sampling treaties and building wet labs may look slow and grueling, but it erects a concrete barrier against software giants accustomed to making fast money online. No matter how many GPUs OpenAI and Google hoard, writing code alone cannot conjure up microbes from deep-sea hydrothermal vents or chemical reactions in test tubes.
The second trump card is pairing the model with an uncompromising computer referee that can instantly judge right from wrong. Large models love to hallucinate when writing prose because no one can automatically verify every step at low cost. In black-and-white domains like mathematical proofs or code optimization, however, a machine checker can act as a ruthless examiner. Take the classification of finite simple groups, a famous theorem in mathematics whose original proof is scattered across roughly 15,000 pages of journal papers. According to an official technical report from startup Axiom Math, they coordinated multiple AI agents to rewrite the entire theorem into 43.1 million lines of Lean code in just 24 days. Lean is a programming language built specifically to verify mathematical proofs, where every deductive step is automatically checked by computer.
Not only did the system translate legacy literature, but as the computer referee verified the proof line by line, it also caught 34 substantive errors in the human mathematicians’ original texts and even patched a proof gap that had lurked in an old paper for 40 years. The exact same logic applies to tuning model-training code, where a single test run after each edit immediately tells you whether the change helped or hurt. Mirendil, founded by a group of former Big Tech researchers, built its workflow around this idea. Every time the AI modifies the training code, the system immediately runs a test to see if the model’s score improves. Guided by that automated scoring feedback, the AI ran unattended for 50 straight hours and tuned the model to outperform its baseline on its own.
This also draws a clear boundary around where the computer-referee playbook works: checking a single attempt must be fast and cheap. Verifying a step in a math proof or tweaking a few lines of training code gives a computer program rapid feedback; but validating a brand-new foundational model architecture from scratch cannot be done with quick, small-scale benchmarks, and instead requires massive compute and long, large-scale training runs. A computer referee cannot give you cheap, rapid trial-and-error there. Discovery Loop, the new company founded by former Google Chief Scientist Jeff Dean and legendary engineer Sanjay Ghemawat, is betting on having AI autonomously invent next-generation algorithms. Although the company is already discussing a fundraising valuation between $10 billion and $50 billion right out of the gate, it has yet to release a single public paper or product.
At this point, a natural business question arises: if so much capital is pouring into companies without even a public product, why aren’t the backers writing tens of billions of dollars in checks worried about losing their shirts? In fact, the most frequent check-writer behind the scenes is none other than GPU maker Nvidia. A look through the public portfolio of Nvidia’s venture arm, combined with publicly disclosed funding rounds, shows that at least 26 of the 45 companies, or more than half, have received direct investment or compute partnerships from Nvidia, spanning every single technical direction discussed above.
A traditional venture fund that backs the wrong zero-product startup can lose everything, but for Nvidia it is a rain-or-shine compute recycling loop. By chipping into the funding rounds of dozens of new labs, Nvidia catalyzes tens of billions of dollars in follow-on capital from other VCs. And the very first thing startups do with that fresh cash is rent or buy high-margin Nvidia GPUs. Even when Nvidia writes a massive check itself, it uses the deal to pull giant compute orders back into its own camp. For example, after Nvidia struck a partnership and invested $5 billion in Sutskever’s SSI, SSI immediately shifted its primary training workloads away from rented Google custom chips and onto Nvidia GPUs. Regardless of which technical bet ultimately pays off, and even if most of these labs fold within two or three years, Nvidia comes out ahead: the tens of billions of venture dollars flooding into this startup wave have already been booked in advance as Nvidia GPU and cloud revenue.
Meanwhile, the frenzy of commanding astronomical valuations on Big Tech star power alone is already starting to cool. Former OpenAI CTO Mira Murati founded Thinking Machines Lab to build model-customization tools. Even though the company already has a public product in hand, when it recently moved to raise a new round amid co-founder departures and a tighter market, its pre-money valuation expectations dropped from an initial $50 billion ask to $40 billion. Two or three years from now, only a tiny handful of these 45 new labs, specifically those holding exclusive real-world data or an uncompromising computer referee, will grow into independent survivors. The majority that neither deliver breakthrough products nor withstand the crushing cost of compute will either limp along via down-round financing or get absorbed into tech giants the way World Labs sold to AMD, sending their talent and compute right back into the hands of the incumbents.
The next time a former Big Tech luminary strikes out on their own and touts a sky-high valuation, you only need to ask two questions to tell whether there is real substance behind the pitch. Do they have real-world physical or biochemical experimental data that Big Tech’s web crawlers cannot reach? And does their business have an uncompromising computer referee that can verify right from wrong in an instant? Companies that have neither, and rely purely on grand algorithmic visions to hold the stage, are mostly just helping Jensen Huang pad next quarter’s GPU sales.