Compute Scarcity & IPO Talk
A recent 20VC episode argued that frontier AI is now governed by compute allocation, not just model quality, and floated big claims — Anthropic’s revenue surpassing OpenAI and SpaceX/other frontier IPOs re-opening the public market. The episode framed vendors’ capacity constraints as a strategic filter that forces pricing, customer prioritisation and more selective access to training/inference resources. For investors, that reframes diligence: model access is necessary but not sufficient; who gets priority compute matters for revenue durability. (youtube.com)
The strange part of frontier artificial intelligence in 2026 is that the bottleneck is no longer just who has the smartest model. It is who can actually get chips, power, and cloud capacity when demand spikes. (anthropic.com) Anthropic said on April 6 that it signed a new deal with Google and Broadcom for multiple gigawatts of next-generation tensor processing unit capacity starting in 2027. Anthropic also said its run-rate revenue had passed $30 billion, up from about $9 billion at the end of 2025. (anthropic.com) OpenAI made the same point from the other side on March 31, when it raised $122 billion and said “durable access to compute” is the strategic advantage that compounds across research, products, and delivery cost. OpenAI also said it is now generating $2 billion in revenue per month. (openai.com) Compute is the rented factory floor of artificial intelligence. If your customers want answers every second and your training runs need thousands of chips for weeks, a better model on paper does not help much if someone else got the machines first. (openai.com) That changes how these companies sell. Anthropic said more than 1,000 business customers are now spending over $1 million each on an annualized basis, and it highlighted that Claude is available through Amazon Web Services, Google Cloud, and Microsoft Azure, which spreads demand across three giant cloud pipes instead of one. (anthropic.com) The shortage is not only about chips. Bloom Energy’s 2026 Data Center Power Report said power availability has become the defining boundary on data center growth, with capital moving toward places that can deliver electricity faster and campuses pushing toward gigawatt scale. (bloomenergy.com) So “who has the best model” is turning into “who gets first call on the grid.” If a lab has priority access to power, cloud contracts, and custom silicon, it can keep response times steady and keep big customers from being rationed during peak demand. (bloomenergy.com) That is why the revenue race suddenly looks different. Anthropic’s own April 6 post put its run-rate revenue above $30 billion, while OpenAI’s March 31 post put its revenue at $2 billion per month, which annualizes to about $24 billion. (anthropic.com) (openai.com) The public-market angle is getting louder at the same time. Reuters reported on April 7 that SpaceX is preparing a $75 billion initial public offering that could value the company at as much as $1.75 trillion, with a roadshow planned for the week of June 8. (cnbc.com) But Reuters also reported that a deal that large could soak up investor attention instead of neatly reopening the market for everyone else. In the same report, Renaissance Capital data showed 35 initial public offerings had priced so far in 2026, down 37.5% from a year earlier. (usnews.com) So the cleaner read is not “the public market is back.” The cleaner read is that 2026 is rewarding the companies that control scarce things first, whether that scarce thing is compute for artificial intelligence or investor attention for a giant listing. (openai.com) (usnews.com)
