Part of: this research — view all 8 rows
I interpret "as of 2026-12-31" primarily as the prevailing annualized compute spend rate at year-end, covering training, research, and inference. Neither company publishes this total directly. Comparing headline revenue (Anthropic's roughly $81 billion run-rate versus OpenAI's roughly $50 billion) is highly misleading because Anthropic books cloud-marketplace sales gross while OpenAI books partner sales net. Furthermore, OpenAI supports over 1.2 billion weekly users, meaning it incurs massive inference serving costs that do not generate proportional revenue, unlike Anthropic’s enterprise-heavy API business. These structural differences in inference costs remain central to my estimate when comparing it with related questions.
Anthropic’s SEC filings provide a firm foundation for its trajectory: compute and infrastructure expense rose from about $3.4 billion in Q1 to $5.6 billion in Q2 Anthropic plans to spend up to $567 billion on compute. Its ... and $9.9 billion in Q3 Anthropic, PBC - S-1/A-2. The Q3 figure, representing 58% of revenue, notably includes $5.0 billion of expensive short-term compute Anthropic added 3Q data to its IPO S1 - N. Baratte - TechStock01. Heading into December, Anthropic expects massive deliveries against its $900.9 billion commitment book, including SpaceXAI payments reaching approximately $3.1 billion monthly. However, this contracted capacity will likely partly substitute for the premium short-term compute rather than being strictly additive Anthropic, PBC - S-1/A-2. Factoring in delivery risks and this substitution effect, Anthropic’s year-end annualized compute run-rate likely lands between $55 billion and $65 billion.
OpenAI's spending is less transparent, but we can reconstruct it from reported financials. Q3 revenue of roughly $10 billion and an operating loss of $10.9 billion imply quarterly operating expenses near $21 billion OpenAI posts steep losses even as quarterly revenue rises ... OpenAI posts steep losses even as quarterly revenue rises .... Assuming compute accounts for 65–75% of these costs—given heavy ongoing frontier training runs, massive experimental agent workloads, and the vast free-tier serving baseline—OpenAI’s Q3 compute spend was likely $13.5–15.5 billion. Allowing for the continued expansion of Stargate, Oracle, and other capacity, OpenAI’s December annualized run-rate should reach $65–75 billion.
The central estimate of 0.86 reflects that Anthropic has closed the gap significantly but likely remains slightly behind OpenAI in total active compute burn at year-end. The middle half of my distribution runs from 0.68 at p25 to 1.05 at p75, allowing for uncertainty in capacity activation and the companies’ relative inference burdens. The lower tail (p10 of 0.55) captures the risk that the resolution metric defaults to cumulative calendar-2026 spend—which would heavily penalize Anthropic given its back-loaded capacity ramp—or that some of OpenAI’s partner-provided compute is obscured off-balance-sheet. Conversely, the upper tail (p90 of 1.30) accounts for the possibility that Anthropic’s contracted deliveries activate faster than expected, fueled by its roughly $213 billion post-IPO cash pile, while OpenAI restrains its own capital expenditure to manage cash.
Set against related questions, this was held stable as it accurately reflects the structural differences in inference costs between the two companies.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited