OpenAI's computer costs were aboiut $30 million in 2017. What are they now?

I know what your thinking. Snitz, bring on the analysis. Your wish is my command. Notes to graph below the article.

This is a tremendous amount of money, on the other hand is it? Considering the impact it's having on our economy and GDP maybe it's not? Plus it's going to be heavily used in the next decade by the Federal Gov and in particular the Dept of Defense. Compared to the Federales burn rate of using money, it's a drop in the bucket.

OpenAI is planning to spend $50 billion on computing in 2026, its president says in court

Greg Brockman testified that the company's computing costs have climbed from about $30 million in 2017 to tens of billions this year

By Colleen Cabili. Quartz Media

Published 15 hours ago

Testimony from OpenAI president Greg Brockman on Tuesday placed the company's projected 2026 computing expenditure at $50 billion — a figure that reflects dramatic growth from the roughly $30 million the company spent on computing in 2017, which has since ballooned into the tens of billions annually.

The disclosure adds detail to a spending trajectory OpenAI has outlined elsewhere. Those figures sit alongside earlier financial commitments OpenAI has made public: a February pledge to investors of roughly $600 billion in spending through 2030, and a separate commitment totaling more than $1.4 trillion earmarked for AI infrastructure.

OpenAI's ongoing court battle with Musk has drawn attention to the company's finances and governance. On his first day of testimony Monday, Brockman disclosed that his equity stake in the company is valued at almost $30 billion — a figure he reached after initially acknowledging a valuation above $20 billion. He said he invested none of his own money to acquire that stake, and argued that OpenAI has "created the most well-resourced nonprofit in history."

Central to Musk's case are allegations that Brockman and CEO Sam Altman abandoned OpenAI's charitable mission for personal financial gain. Among the remedies Musk is seeking is the removal of both men from their positions. Altman has said publicly that he holds no ownership stake in the company. Private investors now value OpenAI at over $850 billion.

The case is being heard in the U.S. District Court for the Northern District of California by Judge Yvonne Gonzalez Rogers. Musk filed it in 2024 on four main claims: unjust enrichment, fraud, constructive fraud, and breach of charitable trust. OpenAI has maintained that the lawsuit's real purpose is to harm a competitor.

OpenAI did not immediately respond to a request for comment.

Claude AI notes

Before I build this, an important note on the data: meaningful "compute spending" estimates for most of these companies don't exist for 2017. Anthropic was founded in 2021, xAI in 2023, and DeepSeek in 2023 — so for them the line simply starts when they did. For Google, "Gemini" wasn't a product until late 2023, so I'll use Alphabet's total annual capex (overwhelmingly AI/data-center driven in recent years) as the closest available proxy. OpenAI's pre-2020 compute spend is also essentially a rounding error — they only had ~$28M in total revenue in 2021, so compute was tiny.

Numbers below are best public estimates from Epoch AI, company disclosures, The Information, WSJ, and SEC filings — they're the order of magnitude rather than precise figures.

A few things worth flagging on what you're looking at:

The Alphabet line is the most apples-to-oranges of the five — it's the company's total annual capex, which is a useful upper bound but includes all data centers, networking, and offices, not just Gemini. Google doesn't disclose Gemini-specific compute. In 2017–2022 most of that capex was traditional infrastructure; only from roughly 2023 onward does the curve really reflect AI.

OpenAI's pre-2022 numbers are rough back-of-envelope from total operating spend (the company only had ~$28M in 2021 revenue, so compute was tiny). The 2024 figure of ~$7B comes from Epoch AI's breakdown ($5B R&D + ~$2B inference), and 2025 at ~$15B is also Epoch's estimate.

Anthropic's 2025 figure of $6.8B is Epoch AI's published estimate. The 2026 budget — not on the chart — is reported at roughly $19B ($12B training + $7B inference).

xAI's curve is steep because Colossus 1 (100K H100s, mid-2024) and Colossus 2 (the ~2GW Memphis/Southaven buildout reaching ~555K GPUs and ~$18B by end of 2025) compressed what would normally be five years of ramp into eighteen months.

DeepSeek looks tiny because it largely is, at least for disclosed model training. The widely-cited $5.6M figure for V3's final training run is real but only a sliver — total annual opex is more plausibly in the few-hundred-million range, and they're using H800s constrained by US export controls.

If you'd like, I can redo this with just the four pure-play labs (dropping Alphabet) to make the comparison cleaner, or zoom into 2023–2025 where the real action happens.

\

Opus 4.7

Adaptive