GPU spending signals
Industry estimates floating this week include OpenAI projecting up to $600 billion of GPU spend through 2030 and Anthropic at roughly $50 billion — figures that frame the scale of long‑term procurement bets. (x.com) Practically, startups still report burning tens of thousands of dollars on idle GPU capacity over three months, and buyers note spot/reserved discounts in the range of 30–70% for committed capacity. (x.com)
The numbers around artificial intelligence chips now run into the hundreds of billions, even as many companies still pay for graphics processors that sit unused. (cnbc.com) OpenAI told investors in February 2026 that it is targeting roughly $600 billion in total compute spending by 2030, down from the $1.4 trillion figure Sam Altman had discussed earlier. Anthropic said on November 12, 2025 that it would invest $50 billion in U.S. computing infrastructure, starting with data centers in Texas and New York built with Fluidstack. (cnbc.com) (anthropic.com) A graphics processing unit, or GPU, is the chip that does the matrix math behind training and running large language models. The spending plans matter because frontier model companies are no longer talking only about software budgets; they are planning power, buildings, networking gear, and long-term chip supply. (openai.com) (anthropic.com) OpenAI’s January 21, 2025 Stargate announcement put that shift in public view with a plan to invest $500 billion over four years in U.S. artificial intelligence infrastructure, with $100 billion slated for immediate deployment. Anthropic’s 2025 build-out made the same move in a smaller but still enormous range: secure capacity directly instead of relying only on rented cloud supply. (openai.com) (anthropic.com) That top-down race coexists with a bottom-up cost problem: unused capacity. AWS says Spot Instances can be priced at up to a 90% discount to on-demand, Google says Spot virtual machines are discounted 60% to 91%, and AWS says one- to three-year commitments through Reserved Instances or Savings Plans lower standard prices as well. (aws.amazon.com) (cloud.google.com) (aws.amazon.com) Microsoft makes a similar pitch, saying Azure Spot Virtual Machines can offer discounts of up to 90% versus pay-as-you-go pricing. Those discounts exist because cloud providers are trying to sell spare capacity that would otherwise go idle. (azure.microsoft.com) The result is a split market. The biggest labs are trying to lock in future supply years ahead, while smaller buyers are still arbitraging hourly, spot, and reserved markets to avoid paying full price for every workload. (cnbc.com) (aws.amazon.com) (cloud.google.com) That mismatch shows up inside clusters too. GPU operators and infrastructure vendors now openly market tools for “idle” or “stranded” capacity, and some say typical fleets run far below full utilization because training jobs, inference traffic, and developer schedules rarely line up perfectly. (getlilac.com) (businesswire.com) (aptlytech.com) So the signal from this week’s spending figures is not that GPU scarcity has disappeared. It is that the market now has two truths at once: long-term chip access is scarce enough to justify $50 billion and $600 billion plans, and day-to-day usage is uneven enough to keep producing discounts on the capacity companies already bought. (cnbc.com) (anthropic.com) (aws.amazon.com) (cloud.google.com)
