Fifty billion dollars. That is the amount OpenAI president Greg Brockman told a court the company expects to spend on compute this year.
The setting was not a product launch. It was the witness stand in Musk v. Altman. On his second day of testimony, Brockman was asked about OpenAI's costs and gave the number.
A 1,666-fold jump in nine years
In 2017, OpenAI spent about $30 million on compute.
By 2026, that figure has become $50 billion. Over nine years, the bill has risen 1,666 times.
The curve is not mainly about compute becoming more expensive. It is about OpenAI trying to do something different. Brockman's argument in court was direct: building AGI requires the ability to attract investment partners at the scale of tens of billions of dollars, such as Microsoft. A nonprofit shell could not hold that much capital, which is why OpenAI created a capped-profit subsidiary.
That was OpenAI's explanation to the court, and a direct answer to Musk's claim that the company should never have taken the for-profit step.
$80 billion of the $110 billion is conditional
The $50 billion figure only matters if investors can see where the money is coming from.
In February 2026, OpenAI secured $110 billion in combined commitments from Amazon, Nvidia and SoftBank. The headline sounds overwhelming, but the details are more constrained:
| Backer | Amount | Condition |
|---|---|---|
| Amazon | $35 billion conditional portion | OpenAI must rent 2 gigawatts of Amazon Trainium capacity |
| Nvidia | $30 billion | OpenAI must deploy 5 gigawatts of Nvidia compute, at an implied valuation near $300 billion |
| Microsoft | $25 billion in 2026 AI spending | Primarily routed through Azure |
Roughly $80 billion of the $110 billion is conditional. In plain terms, the investors are putting in money while requiring OpenAI to buy capacity back from them.
The industry already knows this pattern of investment flowing back to suppliers. Amazon's investment in Anthropic and Google's investment in Anthropic follow the same logic: capital leaves one hand and returns to the other. On paper it is financing; in practice it also works like a large customer prepayment.
The other side of the testimony
Brockman was not only talking about money. One of his emails was read in court, including his warning that, if AGI is built, the conversation over who controls it "will be the highest-stakes conversation the world has ever seen."
Musk's lawyers pressed Brockman on OpenAI's equity structure and on his own stake, reportedly valued at about $30 billion. Brockman did not take that bait directly, but he described a different detail: Musk, he said, had agreed to a for-profit subsidiary years earlier, on the condition that he would have full control and a far larger stake than the other co-founders.
That was Brockman's rebuttal to Musk's original version of the lawsuit: the issue, he argued, was not simply that OpenAI had abandoned its mission, but that Musk had once wanted a bigger share and stronger control for himself.
What $50 billion means
At current market prices, $50 billion roughly maps to:
- The purchase-equivalent cost of about 160,000 to 180,000 H100-class GPUs
- Or a full year of normal operating expenses for a 4-5 gigawatt data-center footprint
- For comparison, Nvidia's planned 5-gigawatt Stargate capacity for OpenAI carries an estimated total cost of about $300 billion
So the $50 billion figure looks less like a single hardware order and more like a 2026 operating total: some owned buildout, some operating cost, and some rented Amazon and Nvidia compute. Microsoft's spending sits on a separate track.
The signal to capital markets
Three years ago, the industry was still arguing over how much it cost to train GPT-4. Now one company's annual compute budget exceeds the combined 2024 AI spending of the world's large technology companies.
OpenAI is betting on scale: if it can pile up enough compute, it expects the capability frontier to keep moving. But for that thesis to work, revenue has to catch up. Sam Altman has recently talked about a trillion-dollar IPO target; to make a $50 billion annual compute plan make sense, the revenue side needs to support even more aggressive numbers before 2027.
The $50 billion line in court was a warning shot to capital markets. OpenAI's cash needs over the next two years will be severe. Microsoft, Amazon and Nvidia have all received the signal. They have also already signed contracts that put OpenAI's spending back on their own order books.
Sources: CocoLoop, OpenAI exec says it will burn $50B on compute this year (The Register); Greg Brockman testifies in Musk v. OpenAI trial (multiple media reports, 2026-05-05)