Post-training and fine-tuning plausibly absorbs about $400B of 2026 AI spend — anywhere from $273B to $816B depending on scope
Asked:
“What percent of AI spend today is dedicated to post-training and fine-tuning? How much is that in dollar terms globally?”
No authority publishes a global line item for post-training spend. The best public allocation estimate is SemiAnalysis’ late-2026 figure for combined OpenAI and Anthropic compute — 55% post-training/RL, 37–38% inference, 7% pre-training — and that 55% is single-sourced. Applying it to four independent 2026 denominators (USD, worldwide) yields the four modeled scenarios below; all shares are measured against Gartner’s $2.67T total worldwide AI-spending forecast.
Four scopes, one 55% assumption: implied post-training dollars, 2026
Grey bar = denominator (total scope, USD billions). Dark bar = 55% of it, the implied post-training spend. Hover a bar for scope detail.
The 55% post-training/RL compute share describes OpenAI plus Anthropic capacity only — a frontier-lab allocation, not 55% of the whole AI economy, and it is single-sourced.
officechai.comThe practical midpoint is $407B — 55% of the $740B hyperscaler AI-capex denominator — equal to 15.2% of Gartner’s $2.67T worldwide AI-spend forecast.
elfysworld.comScope alone moves the answer threefold: a narrow server-and-storage scope implies $273.35B (10.2% of worldwide spend), while Gartner’s broad AI-infrastructure scopes imply $814–816B (30.5–30.6%).
softwarestrategiesblog.comCompute share may not equal spend share: pre-training, post-training, and inference run on different hardware with different utilization economics, so every dollar figure here is a modeled estimate, not a reported market total.
gartner.com
Scenario detail
All amounts USD billions, 2026. Implied post-training = denominator × 55%.
| Scenario scope | Denominator ($B) | Implied post-training ($B) | Share of $2.67T worldwide AI spend | Source |
|---|
Method: 4 denominator scenarios compiled from public 2026 forecasts (IDC-scope narrow infrastructure, hyperscaler AI capex, and two Gartner-scope broad definitions), each multiplied by SemiAnalysis’ single-sourced 55% post-training/RL compute share for combined OpenAI and Anthropic capacity; shares expressed against Gartner’s $2,670.46B worldwide AI-spending forecast for 2026. Dollar figures are modeled estimates, not reported totals; compute share may differ from spend share. Research current through 2026-10-27.