How AI labs split spending: training, inference and products, sourced

Frontier AI labs spend in three buckets: training and research compute, inference compute to serve customers, and everything else, mainly people and products. The best public numbers say training and research take the largest share while revenue is small, and inference grows as customers arrive. SpaceX's quarterly filings on xAI now add primary detail.

This post looks at xAI after its merger into SpaceX and SpaceX's public listing, at what an independent research group found in the accounts of Anthropic and two Chinese labs, and at what has leaked about OpenAI. It also explains the terms, because the same words (training, inference, research compute) are used loosely.

Private company figures are unaudited. Several xAI numbers for 2025 and early 2026 come from news summaries of SpaceX's stock offering document, and the later numbers come from SpaceX's own quarterly filing. Each figure below says which. Nothing here is investment advice.

Key numbers at a glance

• SpaceX AI segment revenue, Q2 2026 (April to June): $2.561 billion, against $0.737 billion a year earlier. Source: SpaceX Q2 2026 earnings release, filed with the SEC on 4 August 2026.
• SpaceX AI segment capex, Q2 2026: $15.828 billion, against $0.749 billion a year earlier and $7.723 billion in Q1 2026. Source: same release.
• SpaceX AI segment loss from operations, Q2 2026: $1.257 billion; adjusted EBITDA a profit of $1.146 billion. Source: same release.
• Cloud Services Agreements, Q2 2026: $14.1 billion of contracted sales, and $1.6 billion of incremental AI infrastructure revenue in the quarter. Source: same release.
• AI nameplate compute: 1.4 GW at 30 June 2026, against 1.0 GW in Q1 2026. Source: same release.
• SpaceX AI segment, Q1 2026: revenue $818 million, operating loss $2.469 billion. Source: SpaceX offering document, as reported by Storage Newsletter, 11 June 2026.
• xAI, full year 2025: revenue about $3.2 billion and an operating loss of $6.4 billion. Source: Storage Newsletter and Revista Mercado, both summarising the offering document.
• SpaceX listing: 12 June 2026 on Nasdaq under SPCX, 555.56 million shares at $135, raising $75 billion. Source: Business Today, 12 June 2026.
• Anthropic lease of SpaceX capacity: $1.25 billion a month through May 2029. Source: Storage Newsletter and Revista Mercado.
• Anthropic 2025 spending estimate: $9.7 billion total, of which $4.1 billion training and research compute and $2.7 billion inference. Source: Epoch AI, 4 February 2026, from reporting by The Information.
• Z.ai, first half of 2025: $312 million of expenses on $27 million of revenue. Source: Epoch AI, from Hong Kong listing filings.
• OpenAI training compute: $32 billion in 2026 and about $65 billion in 2027, per The Information via The Decoder, 21 February 2026.

Three buckets, defined

• Training and research compute. The computing used to build models: the final big training run, plus the many smaller experiments that never ship. Epoch AI, an independent research group, calls this R&D compute.
• Inference compute. The computing used each time a model answers a request. This grows with usage, and it is the part that earns revenue.
• Staff, products and other. People, offices, sales, free tiers, data licences. It is usually the hardest to separate from outside.

A note on the word "compute". It means computing capacity, mostly graphics chips and the data centres around them. Prices are often quoted per GPU hour, meaning one chip for one hour.

Two more terms appear below. Capital expenditure, or capex, is money spent on long lived assets such as buildings, power equipment and chips; it is not an expense in the year it is paid but is spread over the asset's life through depreciation. Adjusted EBITDA is earnings before interest, tax, depreciation and amortisation, with further adjustments chosen by the company. For a lab that is buying chips at scale, adjusted EBITDA can look healthy while the operating line, which includes depreciation, shows a loss.

xAI inside SpaceX

In early 2026 SpaceX acquired xAI in an all stock deal that valued xAI at $250 billion and the combined company at $1.25 trillion, according to Storage Newsletter's reading of the offering document. SpaceX then listed on 12 June 2026. Business Today reports the offering priced at $135 a share, sold 555.56 million shares for $75 billion, valued the company at $1.77 trillion at pricing and opened at $150. The share price has moved since and is not tracked here.

What the offering document showed

The offering document gave outsiders the first detailed look at xAI's finances. Two news summaries, Storage Newsletter and Revista Mercado, both rely on it:

• The AI segment, which includes xAI and the X social network, had revenue of $818 million and an operating loss of $2.469 billion in the first quarter of 2026, with adjusted EBITDA a loss of $609 million (Storage Newsletter).
• For 2025, xAI had an operating loss of $6.4 billion on $3.2 billion of revenue, against a $1.56 billion loss in 2024 (Storage Newsletter). Revista Mercado reports the same revenue and loss.
• Data centre capex was about $13 billion in 2025 (Storage Newsletter) or $12.7 billion (Revista Mercado). Storage Newsletter says AI capex now exceeds rocket related capex.
• Usage figures given are 550 million monthly users on X (March 2026) and 117 million monthly users of the Grok assistant, per Storage Newsletter.

What the second quarter filing shows

SpaceX's earnings release for the quarter ended 30 June 2026 is a primary source and settles some of the differences between those summaries. Its capex table gives AI segment capex of $7.723 billion for Q1 2026, which matches Revista Mercado's $7.7 billion and is a little above Storage Newsletter's $7.5 billion.

For the second quarter itself, the release reports:

• AI segment revenue of $2.561 billion, which the release describes as up 213 percent on the previous quarter and 247 percent on a year earlier. Working back from the 213 percent gives roughly the $818 million Q1 figure in the offering summaries.
• AI segment capex of $15.828 billion, out of $18.369 billion for the whole company. For the first half of 2026, AI capex was $23.551 billion.
• AI segment loss from operations of $1.257 billion, which the release says is a 49 percent cut from Q1, and adjusted EBITDA of positive $1.146 billion.
• Nameplate compute of 1.4 GW, up from 1.0 GW in Q1 and 0.4 GW a year earlier. Nameplate means the rated capacity of the equipment, not how much power it actually draws. One sentence in the release labels the 0.4 GW figure as Q2 2026, but its table places it in the June 2025 column, so the year earlier reading is the consistent one.

In spending split terms, the second quarter is a picture of a lab whose capex is about six times its revenue in a single quarter. Most of that capex buys chips and buildings that will serve both training and inference over several years, which is why capex alone cannot be divided between the two buckets. The gap between the positive adjusted EBITDA and the operating loss is mostly depreciation on equipment already bought. For a longer walk through these filings, see our post on xAI's SpaceX filings.

A training cluster becomes a landlord

The most interesting finding is that xAI's big cluster is being rented to other labs.

• Anthropic. Both summaries say Anthropic pays $1.25 billion a month through May 2029. Storage Newsletter says this covers Colossus 1, about 220,000 Nvidia GPUs and 300 megawatts, with a total of roughly $40 billion, and that either party can end it on 90 days' notice. Revista Mercado estimates about $15 billion a year, or roughly $45 billion over the term, and says the arrangement uses Colossus and Colossus II. The two reports disagree on the total and on which cluster. The headline monthly rate and the termination right appear in both.
• Google. Storage Newsletter reports a cloud services agreement signed on 5 June 2026 at $920 million a month from October 2026 to June 2029, about $30 billion, for roughly 110,000 Nvidia GPUs, with a reduced fee while capacity ramps.
• Cursor. Both reports describe a compute and model collaboration with the coding tool maker Cursor, and an option for SpaceX to acquire it at an implied $60 billion, with a $10 billion fee payable if not exercised or on breach.

The Q2 release confirms the direction without naming customers. It says SpaceX closed several Cloud Services Agreements totalling $14.1 billion of contracted sales, and that these added $1.6 billion of AI infrastructure revenue in the quarter. That is well over half of the AI segment's $2.561 billion of revenue for the quarter.

What this tells us about the spending split is that a lab that builds training capacity can also earn from it by renting it out. Seen from the tenant's side, Anthropic's payments count as compute spending, whether used for training or for serving. Seen from SpaceX's side, the same money is revenue that offsets the cost of the build. The offering document also mentions possible space based compute from 2028, which is an aspiration and not a product. A 90 day exit right is worth noting too: these leases are less firm than a long take or pay contract, where the customer pays whether or not it uses the capacity. That kind of contract term is covered in our Anthropic compute explainer.

Mistral, briefly

Mistral, the French lab, raised a Series D in September 2026 and has said it will build European compute capacity of up to 1 GW by 2030. It has not published revenue or a spending split, so it cannot be placed in the three buckets. The round, the compute plan, the company's revenue projection and its Large 4 model are covered in our Mistral explainer.

How labs actually split spending

Public splits exist for only a few labs. Epoch AI's analysis of 4 February 2026 covers Anthropic, from reporting by The Information, and two Chinese labs, Z.ai and Minimax, whose January 2026 initial public offerings in Hong Kong required detailed financial filings. Other well funded labs have not published comparable figures, so they are not covered here.

• Anthropic, 2025: total expenses $9.7 billion. R&D compute $4.1 billion (42 percent), inference compute $2.7 billion (28 percent), and $2.9 billion (30 percent) for staff and everything else. Epoch says the totals come from projections as of December 2025 and calls them "somewhat speculative." Inference was estimated from a reported 40 percent gross margin on $4.5 billion of revenue. Gross margin is revenue minus the direct cost of serving customers, divided by revenue. For more on that number, read our post on Anthropic revenue and margin.
• Z.ai, first half of 2025: expenses of $312 million on revenue of $27 million. R&D compute $183 million (about 59 percent), inference $9 million (about 3 percent), staff $72 million (about 23 percent) and other $48 million (about 15 percent).
• Minimax, first three quarters of 2025: expenses of $314 million on revenue of $53 million. R&D compute $142 million (about 45 percent), inference $38 million (about 12 percent), staff $48 million (about 15 percent) and other $86 million (about 27 percent).

The percentages are calculated from Epoch's dollar figures. Epoch's totals include non-cash costs such as stock compensation and leave out financing costs.

Two caveats on Epoch's own summary. It says spending at each lab was two to three times revenue, but its table implies about 2.2 times for Anthropic, about 6 times for Minimax and about 12 times for Z.ai. And its headline range for compute as a share of costs (54 to 62 percent) does not match its table for Anthropic, where R&D plus inference is about 70 percent. Use the table and note the discrepancy.

The pattern is still clear. Labs with little revenue spend almost all their compute on training and research: Z.ai spent about twenty times as much on R&D compute as on inference. Inference grows as revenue grows, and at Anthropic it was already two thirds the size of R&D compute in 2025.

That fits OpenAI's reported numbers. The Decoder, citing The Information, reports that OpenAI expects to spend $32 billion on training compute in 2026 and about $65 billion in 2027, nearly $440 billion through 2030. It reports total spending on training and operating models of $665 billion through 2030, that day to day inference costs quadrupled in 2025, and that adjusted gross margin was 33 percent in 2025 against 40 percent in 2024. Revenue in 2025 was $13.1 billion, with $30 billion forecast for 2026. These are reported leaks of internal forecasts, labelled as such by The Decoder, and they change from quarter to quarter. Our OpenAI compute guide covers how the deals behind them are structured.

The falling gross margin is the inference bucket showing up in the numbers. When more people use a model, inference cost rises with them; unless prices or efficiency keep pace, margin shrinks even as revenue grows.

The hyperscalers split differently, because they report what they buy, not what the models do. Alphabet said about 60 percent of its latest quarter's infrastructure spending went to servers and 40 percent to data centres and networking, according to MarketBeat's summary of the Q2 2026 call. Microsoft said about two thirds of its quarter was short lived equipment, per the Motley Fool transcript summary. Neither company separates training from inference. For the chips behind these splits, see our AI chip market guide.

How to read a lab's financial claim

A few habits prevent most misreadings.

• Check the status of the source. A quarterly filing with a securities regulator is a legal document. A leak to a news outlet is not, and neither is an analyst model. The Decoder says outright that its OpenAI numbers come from reporting by The Information on internal documents.
• Check what is counted. Epoch counts non-cash costs such as stock compensation in its totals and leaves out financing costs. Another analyst may count differently.
• Separate capex from expense. A quarter's capex is cash out the door for assets that last years. Comparing one quarter's capex with one quarter's revenue overstates the yearly cost of the build.
• Ask whether a number is a rate or a total. A monthly lease fee is a rate. A total over three years depends on the termination terms.
• Look at the date. Labs revise forecasts every quarter. The Decoder's table shows OpenAI's 2026 cash burn forecast rising from about negative $8 billion in early 2025 to negative $17 billion in late 2025 and negative $25 billion in early 2026.
• Look for a second source. The Q1 2026 capex figure differs between the two offering summaries, and the Q2 release settles it. That is typical, and it is why the filing beats the summary.

What is firm and what is uncertain

Firm, from SpaceX's Q2 2026 earnings release filed with the SEC: AI segment revenue, capex, operating loss, adjusted EBITDA, nameplate compute and contracted Cloud Services Agreement sales. These are unaudited quarterly figures, but they are a company's legal filing.

Reported through summaries of the offering document, so check the document: xAI's 2025 results, the Q1 2026 operating loss and the lease terms with Anthropic, Google and Cursor.

Estimated: Anthropic's split (Epoch from The Information) and OpenAI's forecasts (The Decoder from The Information). Both come from outside the companies.

Reported from filings by an independent researcher: Z.ai and Minimax splits, from Hong Kong listing documents covering 2025 periods.

Unknown: the true training versus inference split inside SpaceX's AI segment, the split for any private lab after 2025, and Mistral's actual revenue.

What to watch next

• SpaceX's third quarter 2026 report, for AI capex, nameplate compute, contracted sales and how much revenue comes from leasing capacity.
• Whether the Anthropic and Google leases appear by name in later filings, and whether any 90 day exit right is used.
• Epoch's and others' updates on lab spending, which tend to follow new filings.
• Z.ai's and Minimax's first annual reports as listed companies, which would give full year 2025 splits.
• Physical limits. Epoch's data insights list says the largest AI data centre by power capacity has been doubling about every ten months since mid 2024 (4 September 2026), and that global AI computing capacity doubles about every seven months (9 January 2026). Power and chips, not money alone, set the pace.

Quick check

If your products depend on any of these models, Safeguard offers SBOM and dependency scanning for the software around them, with a free scan at https://gold.safeguard.sh/scan.

Sources

• SpaceX, second quarter 2026 earnings release, SEC Exhibit 99.1, 4 August 2026: https://www.sec.gov/Archives/edgar/data/1181412/000162828026052515/earningsreleaseq22608042.htm
• Storage Newsletter, SpaceX IPO coverage, 11 June 2026: https://www.storagenewsletter.com/2026/06/11/spacex-ipo-is-coming-very-soon-now/
• Revista Mercado, SpaceX, AI losses and capex before the IPO: https://mercado.com.ar/para-entender/la-compania-detras-de-la-compania-6-400-millones-de-perdida-y-30-000-millones-de-capex-anualizado
• Business Today, SpaceX Nasdaq debut, 12 June 2026: https://www.businesstoday.in/amp/markets/stocks/story/spacex-shares-make-nasdaq-debut-after-historic-ipo-check-listing-price-marketcap-536680-2026-06-12
• Epoch AI, company spending breakdown, 4 February 2026: https://epoch.ai/data-insights/company-spending-breakdown
• Epoch AI, data insights list: https://epoch.ai/data-insights
• The Decoder, OpenAI cash burn forecast, 21 February 2026: https://the-decoder.com/openai-adds-111-billion-to-its-cash-burn-forecast-as-ai-costs-spiral-beyond-projections/
• MarketBeat, Alphabet Q2 2026 earnings call highlights: https://www.marketbeat.com/instant-alerts/alphabet-q2-earnings-call-highlights-2026-07-22/
• Motley Fool, Microsoft Q4 2026 earnings call transcript summary: https://www.fool.com/earnings/call-transcripts/2026/08/07/microsoft-msft-q4-2026-earnings-call-transcript/

Last reviewed: 19 October 2026.