Graphic · AI Compute · 2025–2026

AI Company Costs Breakdown: Compute vs. Staff (2025–2026)

The Compute Ruler: line up the filed accounts of Anthropic, MiniMax and Z.ai and compute lands in the same place every time, at 54.5% to 63.8% of operating costs. Staff, the famous AI salaries included, took 14.1% to 24.0%.

Fig. 1

The Compute Ruler

AI company expenses broken down by what the money bought, from each company's own filings. Compute is drawn flush left, so its share lines up like a mark on a ruler: more than half of operating costs at every company, every period.

ComputeStaffMarketing & partnershipsOther
0%25%50%75%100%Compute share: 54.5%–63.8%Anthropic, 2025US$ thousandsFull year 2025Compute 57.9%Staff 14.5%4.5%Other 23%Anthropic, 2026 YTDUS$ thousandsJan–Sep 2026Compute 54.5%Staff 14.1%16%Other 15.4%MiniMaxUS$ thousandsJan–Sep 2025Compute 63.8%Staff 17.9%12.8%5.5%Z.ai (Zhipu)RMB thousandsJan–Jun 2025Compute 57.7%Staff 24%4.6%Other 13.8%

Operating costs are cost of revenue (or sales) plus operating expenses, including stock-based pay. Non-operating items such as fair-value changes on convertible notes and preferred shares are excluded. Anthropic reports its costs by nature; the MiniMax and Z.ai splits are our sums of the line items in each prospectus. Values are in thousands; Anthropic reports in millions. Periods differ, and Z.ai is in RMB, so compare shares, not amounts.

Anthropic's compute and infrastructure costs, 2025
$7,326M
57.9% of its $12,645M operating expenses (derived)
Compute's share of operating costs at Anthropic, MiniMax and Z.ai
54.5%–63.8%
Latest disclosed periods, 2025–2026 (derived from filings)
Anthropic's compute costs as a share of revenue
251% → 58%
Q1 2025 to Q3 2026, as reported in the S-1
Anthropic's committed future spend, mostly compute
$900.9B
$451.6B of it expected before the end of 2030
Fig. 2

Compute alone cost Anthropic about 1.6 times its 2025 revenue

Anthropic revenue and operating expenses, 2025, US$ millions

Revenue$4,585MCompute & infra$7,326MPersonnel$1,839MBartz settlement$1,500MOther & admin$1,412MSales & marketing$568M

Full S-1 line names: compute and infrastructure; personnel; Bartz settlement; other operating and administrative; sales, marketing, and partnerships. Compute and infrastructure covers cloud compute fees, storage, networking, colocation, equipment leases and committed power. The multiple is $7,326M ÷ $4,585M = 1.598. Total operating expenses were $12,645M, for an operating loss of $8,060M.

Fig. 3

Anthropic's revenue overtook its compute bill in early 2026

Anthropic quarterly revenue vs. compute and infrastructure costs, US$ millions

0$6B$12B$18BQ1 2025Q3 2025Q1 2026Q3 2026Revenue$16,956M in Q3 2026Compute$9,863M in Q3 2026

Quarterly figures are unaudited. Compute and infrastructure costs rose in every quarter, but revenue rose faster: the S-1 puts compute at 251% of revenue in Q1 2025 and 58% in Q3 2026.

Fig. 4

Half of Anthropic's 2025 costs went to R&D, not to serving customers

Anthropic operating expenses by function, 2025, US$ millions

50.1%21.9%11.9%8.9%
  • Research and development$6,338M50.1%
  • Cost of revenue$2,773M21.9%
  • Bartz settlement$1,500M11.9%
  • General and administrative$1,123M8.9%
  • Sales and marketing$911M7.2%
  • Total operating expenses$12,645M100%

From Note 8 of the S-1. R&D is mainly compute for model development, research and internal use, plus research staff. Cost of revenue is substantially all compute for customer-serving workloads. Sales and marketing includes compute for free access to Claude. R&D's share is $6,338M ÷ $12,645M = 50.1%.

Fig. 5

At MiniMax and Z.ai, most compute went to training, not to serving users

AI compute costs by use, training vs. inference, latest disclosed period

Training and researchServing customersMarketing and free use
0%25%50%75%100%MiniMaxUS$ thousands · Jan–Sep 2025Training and research 78.9%21.1%Z.ai (Zhipu)RMB thousands · Jan–Jun 2025Training and research 95.3%

Anthropic does not split its compute by use. Its cost of revenue ($2,773M in 2025), which is substantially all customer-serving compute plus some staff and overhead, was 37.9% of its $7,326M compute bill (derived), so at least about 62% went elsewhere: R&D, internal use and free access.

Fig. 6

Final training runs were a small slice of R&D compute

Estimated share of R&D compute spending on the final training runs of released models: OpenAI in 2024, MiniMax Q4 2024–Q3 2025, Z.ai H2 2024–H1 2025

MiniMax22.6%Z.ai12.3%OpenAI9.6%

Epoch AI estimates. R&D compute is from the IPO filings (MiniMax $141M, Z.ai $216M at 7.2 RMB per US$) and Epoch's earlier estimate for OpenAI ($5B). Final-run costs are Epoch's estimates for models released one quarter after each window. The rest went to experiments, synthetic data and research. Shares are of money spent, not of computing operations.

Fig. 7

Pre-IPO estimates missed $2,945M of Anthropic's 2025 costs

Anthropic 2025: Epoch AI estimate (Feb 2026) vs. S-1 actuals, US$ millions

Epoch AI estimate S-1 actual Gap (S-1 actual − Epoch AI estimate)
0$8B$16BGapTotal operating costs$9,700M$12,645M+$2,945MStaff and other costs$2,900M$5,319M+$2,419MCompute$6,800M$7,326M+$526MRevenue$4,500M$4,585M+$85M

Epoch built its estimate from reporting by The Information before Anthropic filed. S-1 staff and other = personnel $1,839M + Bartz settlement $1,500M + sales, marketing and partnerships $568M + other $1,412M. Epoch's compute estimate is R&D compute ($4,100M) plus inference ($2,700M). The estimate got compute nearly right; most of the miss was in everything else.