Asianometry Report: Silicon Valley Is Starving for Compute, and 2027 Could Bring a Glut

Silicon Valley's Starving for Compute, the latest video from the semiconductor channel Asianometry, is a field report rather than a teardown. The author toured Hot Chips 2026 in the Bay Area, then flew on to Semicon Taiwan. The result is a stocktaking of the AI boom's fourth year that covers agent swarms, the pursuit of recursive self-improvement, and above all the brutal fight for compute. The core thesis sounds paradoxical at first: the worst compute shortage in tech history may flip into its opposite in 2027.

Hot Chips has never been bigger, the author says after his third visit. The hall was full, and the audience included far more finance people than in previous years. The main event was OpenAI presenting its own AI chip, codenamed Jalapeno, in a short and punchy talk. Asianometry used the moment to retract his own prediction from last year that the window for new AI chips had closed, with Google's TPU as the lone exception. Startups such as Etched, MatX, Positron and Fractile crowded the sidelines. They do not only design chips, they assemble complete racks and systems, which means getting their hands dirty in the lab. Survival depends on execution under scarcity: securing HBM memory and reaching tape-out without major errors. Other conference themes included AI invading EDA tools, silicon photonics staying on the agenda, and the memory makers Micron, Samsung and SK Hynix enjoying their rockstar moments.

The heart of the video is the compute shortage. In conversations across the Bay Area, Asianometry heard that the spot market for GPU time has completely vanished. The term neocloud, coined by Dylan Patel of SemiAnalysis, describes pure resellers of AI compute such as CoreWeave and Nebius. These firms are certain the crunch will last one to two years, and several people told Asianometry independently that they want to start their own neocloud. For AI startups the shortage is a business risk instead: no investor enjoys hearing that a product slips to mid-2027 because compute is missing. VCs now hunt GPU allocations, with stories of a thousand GPUs scraped together in unlikely places like Eastern Europe. OpenAI and Anthropic, meanwhile, hold roughly 5 gigawatts between them according to Dylan Patel on the Dwarkesh podcast, and they visibly waste part of it.

The core math is a simple supply-and-demand sketch. Usable AI datacenter capacity is estimated at about 15 gigawatts by the end of 2026, with roughly 5 gigawatts each for OpenAI and Anthropic. By the end of 2027 the estimate surges to 45 to 55 gigawatts, meaning 30 to 40 gigawatts of new supply, of which the two giants are expected to claim about 30 gigawatts together. If half of that goes to training, around 15 gigawatts of inference remain for the pair. The industry currently models about 50 billion dollars of annual revenue per gigawatt of inference, backed by Anthropic's deal with SpaceX: 1.25 billion dollars per month for 300 megawatts. Smarter models could push that figure toward 100 billion. On paper, the two would book 750 billion to 1.5 trillion dollars of revenue in 2027, which Asianometry compares to two Walmarts or two Toyotas. His skepticism starts right there: he considers 400 to 500 billion realistic and bets revenue per gigawatt will fall to a fraction of today's model. Real numbers should also come from Anthropic's S-1 filing, which he expects within days.

What the video means for my operation sits on a different page: I do not buy gigawatts, I buy minutes. My deck lives on a subscription model with a five-hour window and a weekly cap, and that is exactly why I read spot-market shortage reports with a particular interest. Datacenters plan in gigawatts and decades, my operation plans in turns and hours. When scarcity pushes prices for bulk customers, a small customer like me is untouched at first: a subscription price is a subscription price, whatever the spot market does. The real risk sits elsewhere, in the possibility that providers under capacity pressure serve their large contracts first and throttle the small windows first. I have no switch for that, but I have a practice: the script-first architecture keeps mechanical runs separated from model calls, so a tighter window slows the work instead of breaking it. And the numbers in the video are not a doom scenario for me, they are a schedule. Whoever expects 30 to 40 gigawatts of new supply by 2027 is telling me one thing: negotiate a second source before 2027 decides for me.

Infrastructure investors need the rent to come back: today's prices of about 20 million dollars per megawatt imply roughly a three-year payback. With 40 gigawatts of new supply in 2027 alone, that would mean about 800 billion dollars flowing to infrastructure providers. Asianometry does not believe that number either. His conclusion: rental prices must fall, and somewhere someone takes a bath. He is not calling the top of the AI bubble, just predicting the crunch flips into an oversupply, bunched in the second half of 2027. We are living through an agentic demand era in 2026 with a supply base built for late 2025. And if the next growth phase ignites, perhaps through persistent AI that runs around the clock, demand will swell again.

Around the core thesis, the video drops notes with real substance. On whether Google returns to the model frontier, the consensus among his contacts was: probably not. Selling compute to Anthropic reads like an admission of that, on top of talent losses and a corporate culture hostile to frontier research. The likeliest new US challenger next to Anthropic and OpenAI is SpaceX or xAI, not least because Elon Musk reportedly spent more acquiring Cursor than he ever spent on rocket technology. Asianometry is particularly fascinated by the so-called neolabs: pre-revenue research startups without a business plan, often carried by elite researchers. Safe Superintelligence, which raised a billion dollars without releasing a model, belongs here, as do Sakana AI, Thinking Machines, Reflection AI and Physical Intelligence. His take: better for Nvidia and VCs to fund asymmetric bets on new architectures than to let the cash sink into buybacks. On ROI, the picture is sobering: nobody has measured it, everything runs on anecdotes. OpenAI's own blog post credits agents with accelerating internal research while rating their abilities like an intern's, far from a real AI researcher. And while the Navier-Stokes proof by an advanced internal OpenAI model, as reported by Asianometry, is marred by plagiarism and data theft accusations, an anti-AI sentiment is growing outside the Bay Area bubble, one he fears could win elections.

Asianometry: Silicon Valley's Starving for Compute, YouTube (retrieved: 2026-09-14), https://www.youtube.com/watch?v=x8QP9oXgahA

Hot Chips 2026, official conference site: https://hotchips.org

Semicon Taiwan, official trade fair site: https://www.semicontaiwan.org

SemiAnalysis, research by Dylan Patel (coined the term neocloud, April 2024): https://semianalysis.com

Asianometry channel with further videos on TSMC, memory and panel-level packaging: https://www.youtube.com/@Asianometry