OpenAI Projects $50 Billion in Compute Spending for 2026

In a San Francisco courtroom Tuesday, OpenAI President Greg Brockman dropped a number that reshapes how we understand the AI race: $50 billion. Under oath in the Musk‑Altman trial, Brockman revealed that OpenAI projects spending that staggering sum on computing power in 2026 alone, according to Reuters and Bloomberg. It is the first time OpenAI has publicly quantified its infrastructure ambitions, and the number is breathtaking.

For context: $50 billion is more than the annual GDP of many countries. It rivals the full‑year capital expenditures of Google, Microsoft, and Meta individually. It is, by any measure, a staggering commitment to compute for a single company — let alone one that has never turned a profit.

The disclosure came during cross‑examination in the Musk‑Altman trial, now in its second week. Brockman's testimony has been central to the case — he recounted a 2017 confrontation where Musk allegedly became physically threatening after being denied control of OpenAI, and described Musk's demand to raise $80 billion for Mars colonization, per Reuters.

The $50 billion figure emerged not as a boast but as a data point in describing the financial reality of frontier AI. Brockman was testifying about the resources OpenAI requires to compete, and the number slipped into the public record under oath — giving it more weight than any press release could.

At current pricing, $50 billion could acquire roughly 1.25 million NVIDIA B200 GPUs at bulk rates, or 2‑3 million H100 equivalents. The power draw alone would exceed 1.5 gigawatts — enough to power a small city. Add networking (InfiniBand or Spectrum‑X at $3‑5B), data center leases, liquid cooling, and energy contracts, and the physical footprint becomes staggering.

This is not a training budget. It is a total compute envelope — spanning GPU clusters for training frontier models, inference serving for ChatGPT's hundreds of millions of users, API compute for enterprise customers, and research experimentation. Per Bloomberg, the spending supports OpenAI's entire AI services operation.

Here's the problem: OpenAI's revenue, while growing fast, does not come close to covering this spend. Revenue was estimated at roughly $3.7 billion in 2024 and projected around $11.6 billion for 2025. Even with aggressive growth to $25‑30 billion in 2026, compute alone would consume nearly double that revenue. Add $5‑10 billion for headcount (4,000+ employees), research, and operations, and OpenAI faces a roughly $25‑30 billion annual loss.

The company has raised approximately $20 billion cumulatively, including Microsoft's $13 billion and a $6.6 billion October 2024 round at a $157 billion valuation. At this burn rate, OpenAI would deplete every dollar ever raised in under a year — which is why the $50 billion figure forces a reckoning about how AI labs actually fund the frontier.

The math only works because of Stargate — the $500 billion infrastructure joint venture announced in January 2025 with SoftBank, Oracle, MGX, and others. Stargate is effectively OpenAI's off‑balance‑sheet compute vehicle: SoftBank and partners provide the capital expenditure, and OpenAI "rents" the resulting capacity.

Without Stargate, OpenAI would need to put $50 billion annually on its own balance sheet — an impossibility for a company with no profits and limited equity. With Stargate, the compute spend flows through a separate entity, keeping OpenAI's financials (relatively) cleaner and enabling continued fundraising. As Reuters framed it, the disclosure reveals Stargate not as an ambitious side project but as an existential necessity.

At $50 billion per year for one company, the barrier to competing at the AI frontier becomes effectively insurmountable for startups. No VC‑backed AI company can raise even 10% of that. This entrenches a compute oligarchy of perhaps five players: OpenAI/Microsoft, Google DeepMind, Anthropic/Amazon, Meta, and xAI.

The compute cost floor for training a frontier model may already be $5‑10 billion, and inference at ChatGPT scale costs billions more. The message is clear: if you want to build frontier AI, you need a hyperscaler partner — or a sovereign wealth fund. The era of garage‑to‑AGI is over.

If $50 billion is 2026, and the trajectory has been roughly doubling each year, 2027 could see $75‑100 billion and 2028 could approach $150 billion. Three futures branch from here: OpenAI achieves AGI that generates commensurate economic returns; OpenAI is effectively absorbed by Microsoft, which already provides most of its compute; or investor appetite wanes before the payoff arrives, creating an existential funding cliff. The Musk‑Altman trial itself adds uncertainty — if Musk prevails, OpenAI's for‑profit conversion could be blocked, cutting off its primary funding mechanism at the worst possible moment.