Compute flips from shortage to glut in 2027, and the AI bubble may not pop
Usable AI data center capacity is about 15 GW at the end of 2026 and races to 45-55 GW by the end of 2027, a 30-40 GW increment. At $50 billion of revenue per gigawatt, the two giants would have to produce $750 billion to $1.5 trillion, and that number does not hold up.
The argument · tap a timestamp to hear it
The window for new AI chips has reopened
Last year the author wrote that ‘the window for new AI chips has closed, except for Google's TPU’, and here he explicitly takes that back and says he was wrong. The trigger was OpenAI unveiling its in-house chip Jalapeno at Hot Chips, a short talk with high information density. Outside the hall a ring of new AI chip startups gathered: Etched, MatX, Posetron, Fractile; Etched gave no talk but built presence by handing out hats and having its founder ask questions all over the Q&A. The author reminds us these are hardware companies, not just chip designers: they have to build their own racks and systems and get their hands dirty in the lab, and recruiting top talent and execution are hard bars. His survival test borrows Nathan Bedford Forrest's controversial line: get there first, and with the most men.
Semicon Taiwan is really TSMC's request for proposals
First-timers are surprised: neither TSMC nor ASML has a booth. But TSMC dominates the whole show — its teams walk the halls visiting new suppliers and scouting new technology, and many information sessions have TSMC executives on stage laying the company's problems out in public. The author's conclusion is that Semicon Taiwan is essentially an RFP that TSMC issues to the entire semiconductor industry. The technical throughline is clear: future manufacturing has to go bigger. AI chips are hitting existing limits, most typically the reticle limit, which determines how large an area a lithography machine can print. TSMC had a slide showing it is preparing to build a ‘battleship-class’ chip spanning 14 reticles, and the industry's response is, as usual, fast.
Glass panel packaging has moved from PPT to the show floor
Two years ago the author made a video on panel-level packaging, which was the theme of Semicon Taiwan 2024, when it was still all talks and slideshows. Now AI systems are so large that the industry wants to use square, glass panel interposers to carry chiplets and memory, replacing silicon interposers. The author's attitude is reserved: can TSMC make them without warping or cracking? ‘I'll believe it when I see the chips.’ But the ecosystem is already moving, and equipment such as handling robots adapted to square panels appeared in the halls. He judges this wave of ‘chipmaxing’ will be one of the most far-reaching technology transitions in semiconductor history, roughly on the scale of the move to 300 mm wafers.
The spot market has vanished, and VCs now help portfolio companies find GPUs
When the author was last in the Bay Area in September 2025, the scene was completely different. Since then long-running agentic AI has swept Silicon Valley, triggered in part by stronger models such as Anthropic's Opus 4.5 around November 2025 and by Open Claw. Prices started rising in February 2026, everyone rushed to lock up compute, and Anthropic led the way signing big deals with SpaceX, Google and others. The author was told that by the time of this trip the spot market had completely disappeared. For Neoclouds (a term coined by semi-analyst Dan Nishbal in April 2024 for pure cloud vendors that resell AI training and inference compute, such as Coreweave and Nebius), this is a huge boon, and they are confident the shortage will last at least another one to two years. The author met several people who did not know each other and each wanted to start a Neocloud. For AI startups this is a business risk: imagine telling investors the product has to wait until mid-2027 because there is not enough compute. The new way VCs help portfolio companies is to go find compute; he heard of people scraping together a thousand GPUs and hunting in unexpected places like Eastern Europe.
Two giants are feasting, and compute is still being wasted
By contrast, Anthropic and OpenAI are feasting internally. Dylan recently said on Dashpod that these two AI giants each have about 5 GW of compute. The author infers that a considerable share of all that compute may be wasted, and anecdotes he heard confirm this ‘rolling in it’ is not being used as efficiently as possible. He also asked every AI insider he talked to one question: will Google return to the frontier? The general consensus is no, though the explanations are vague. Several people pointed out that Google's decision to sell a large amount of compute to Anthropic is itself a sign it does not believe. This is a big reversal from the period when Gemini 3 made some people think OpenAI was finished. What the author heard is that Google suffered severe talent drain, plus a culture unfriendly to frontier AI.
Neolabs are venture capital that treats compute as money
What fascinates the author most is the Neolab: a research startup with basically no revenue, funded to explore an interesting AI idea, whose founders are often elite researchers and whose business plan is little more than ‘discover a completely new architecture’. The most famous and earliest, Safe Superintelligence, raised a billion dollars and has still released no product or model. The label is fuzzy, covering recursive intelligence, Humans, Core Automation, Flapping Airplanes and others; those with products include Japan's Sakana AI, Thinking Machines, Reflection AI, plus a batch in verticals like materials science, world models and robotics. The first reaction is that these assets are dumb: no revenue, no plan for near-term monetization. But the author can understand it: historically breakthrough science came from research universities, government labs or Bell Labs, and now university funding is being cut and talent is being sucked into industry; AI is probably the first major technology in the past 50 years to emerge without government involvement, and in this era research is closed and secretive. So let Nvidia, with its compute and cash, plus a pile of VCs, fund these asymmetric bets. OpenAI and Anthropic are locked into the transformer paradigm and are unlikely to spend on researching these directions. The failure rate is high, but most of the money is given in the form of compute anyway; if someone really strikes gold and forces the giants to pivot (say, to a more data-efficient architecture), the giants will most likely buy the discoverer at a jaw-dropping price, and one such deal pays for everything. The author thinks this is a better use of cash than continuing buybacks, and names Apple.
Tokens are burning, and nobody has measured ROI
Customers really are burning tokens, but are they getting ROI? The author asked across the ecosystem, and the answer remains vague: don't know, nobody has measured it, it's all anecdotal. He believes lower barriers to coding will hit knowledge work — financial firms used to hire a pile of reporting analysts to keep management on top of positions, and now two smart young analysts plus Claude writing a dashboard is enough, which he has seen with his own eyes. But for other AI capabilities, like making spreadsheets and presentations, the things marketing usually pushes, he has seen no obvious impact. Maybe it just needs more time, maybe it is a problem of intelligence itself. OpenAI recently published a blog on how agents accelerate internal research, but still positions AI's capability as intern level: it can execute routine tasks, and is far from a real AI researcher. What he heard is that AI still lacks human research taste, though some believe sufficient scale can achieve it. And one thing holds for everyone in San Francisco: people are working harder than before. If the models are really that capable, why does everyone say they have more work? Maybe AI does not take work off our plate but gives us more work.
Compute flips from shortage to glut in 2027
Everyone is saying shortage, shortage, two more years of shortage, and the author instead suspects things will not go as expected. His arithmetic is not complicated: usable AI data center capacity is estimated at about 15 GW at the end of 2026, with OpenAI and Anthropic each taking about 5 GW; the supply chain will respond, and by the end of 2027 that number is estimated to hit 45-55 GW, an increment of 30-40 GW, with the two giants taking about 30 GW of it. Assume half is for training, leaving 15 GW combined for inference. Inference is currently modeled at about $50 billion of revenue per gigawatt (a figure from Dylan Patel on Dashpod), and it can also be reverse-engineered from Anthropic's deal to buy 300 MW from SpaceX for $1.25 billion a month; Dylan thinks smarter models can eventually push it to $100 billion per gigawatt. On that math, in 2027 the two should produce $750 billion to $1.5 trillion in revenue. The author says that number is absurd: most people agree the two will do $150-200 billion this year, and next year he can barely imagine $400-500 billion, but $750 billion to $1.5 trillion is ‘two Walmarts’ or ‘two Toyotas’, a stretch in credibility, and he does not think it can be reached. He bets revenue per gigawatt is far below $100 billion or even $50 billion, maybe only a small fraction of it. From the infrastructure side it is the same: assume 40 GW lands in 2027, current rent is about $20 million per megawatt, implying a 3-year payback, so those 40 GW would require $800 billion paid to infrastructure providers, in 2027 alone. He does not think that can be collected either, which means rents must fall and someone somewhere has to take a loss. Conclusion: in 2027 the supply chain will bring a lot of compute, most of it in the second half, and he bets this pushes the market from shortage to glut. He stresses he is not saying the AI bubble will pop, only that we will soon be flooded with compute — 2026 is agentic demand matched with 2025 supply, supply will catch up and probably overshoot; but after that maybe the next stage of growth ignites and demand inflates again, for example persistent AI.
Anti-AI sentiment could win elections
The author mentions that in his last Hot Chips video someone predicted AI would solve a Millennium Prize problem within a year, and now there is news that an advanced internal OpenAI model produced a proof of the Navier-Stokes existence and smoothness problem, and his friend called it right. But what should be a remarkable achievement has been tainted by accusations of plagiarism and data theft. This negativity points to something he sees more and more online and in ordinary people: anti-AI sentiment. Many people hate AI, for different reasons — they hate data centers, hate that it represents everything about billionaires, think it is stealing ‘human knowledge’, or think society simply does not need it. Whatever the reason, this sentiment is growing, and he has a bad feeling: this issue can win elections. San Francisco lives in its own bubble, the city itself is wrapped in its own microclimate, but the people inside should pay more attention to the sentiment growing outside the bubble, and things can spread at a terrifying speed.
In their own words · checked verbatim
Last year, I wrote that the window for new AI chips has closed with the exception of Google's TPU. I take that back now. I was wrong.
Semicon Taiwan is TSMC's request for proposal to the whole semiconductor industry.
Can TSMC produce these panels without warping and cracks? I'll believe it when I see the chips
They're often defined by elite researcher founders and little if any business plan other than to discover a brand new structure.
Maybe rather than automating away work, AI just gives us more work to do.
I'm not saying that the AI bubble's popping, just that we will soon be a wash in compute.
And I have a bad feeling that this issue can win elections.
Figures
| Usable AI data center capacity at the end of 2026 | about 15 GW | 15:17 |
| Estimated usable AI data center capacity at the end of 2027 | 45-55 GW | 15:17 |
| Compute of OpenAI and Anthropic each | about 5 GW | 9:12 |
| Anthropic's monthly payment for SpaceX compute | $1.25 billion / 300 MW | 15:17 |
| Revenue model per gigawatt of inference compute | about $50 billion, Dylan thinks it can eventually reach $100 billion | 15:17 |
| Projected 2027 revenue of the two giants | $750 billion to $1.5 trillion | 15:17 |
| Current data center rent | about $20 million per megawatt, implying a 3-year payback | 18:23 |
| Amount to be paid to infrastructure providers for 40 GW in 2027 | $800 billion | 18:23 |
| Largest chip size TSMC is preparing | spanning 14 reticles | 6:08 |
Glossary
- Neocloud
- A pure cloud vendor that resells AI training and inference compute, such as Coreweave and Nebius.
- Neolab
- A research startup with basically no revenue, funded to explore a completely new AI architecture.
- reticle
- The maximum area a lithography exposure can print, which sets the upper limit on a single chip's size.
- panel level packaging
- Using square glass panels instead of silicon interposers to carry chiplets and memory.
- RSI
- recursive self-improvement, the ability of AI to autonomously improve its own capabilities.
How to listen
Founders and investors watching AI compute supply and demand and data center investment; hardware engineers who want to understand where TSMC and semiconductor packaging are heading.
The opening Hot Chips venue atmosphere and booth observations can be fast-forwarded; go straight to the compute section after 08:12.