Gavin Baker: Frontier labs will pace AI by spending compute on alignment, making safety a compute demand signal

On Sept 15, investor Gavin Baker argued that Anthropic and OpenAI will control the pace of the frontier not by slowing research itself but by dedicating more time and compute to alignment, monitoring, and evals—paying extra compute cost for safety, i.e., "margin for safety." The next day, citing OpenAI executives' statements, he pushed back on market fears that a slower AI pace would drag down infrastructure demand, arguing that alignment also requires massive compute; however, he noted that not all scenarios are bullish for compute—an OpenAI–Anthropic duopoly and open-source losing could both suppress demand. This remains an investor's analytical framing; OpenAI executives themselves did not use the "alignment compute" framework.

Confirmed

• This is Baker's own publicly stated and repeatedly elaborated view: labs wanting to "brake" will simply spend more time and compute on alignment, monitoring, and evals, accepting lower margins.
• Baker added in a reposted discussion that incremental alignment compute spend is one of many factors affecting frontier-lab margins, a judgment he attributed to a senior OpenAI employee.
• The original poster tszzl made a "pace"-related rebuttal aimed at skeptics inside and outside government; Baker's view extends that discussion, with Justin Halford adding that labs choosing to slow down will pay a price.
• On Sept 16, Baker compiled statements from two OpenAI executives: CFO Sarah Friar said in a CNBC interview that from her perspective there are abundant growth opportunities, with the focus still on acquiring more compute to fuel the growth flywheel; VP of compute strategy Sachin Katti said the way to ensure frontier models are safely aligned is to invest compute. These were used to argue that safety and alignment themselves require substantial compute, so infrastructure demand will not weaken if the pace slows. Poster eyishazyer commented on the reversal: "safety," once a reason to slow down, has become the hardest compute-demand signal.
• Also on Sept 16, Baker responded to the "AI is bullish for compute in every scenario" meme, calling it "funny and almost right," but identified two exceptions that would actually hurt compute demand: an OpenAI–Anthropic duopoly—where compute demand worsens whether the frontier speeds up or slows down—and open-source models losing.

Unconfirmed

• The identity of the "senior OpenAI employee" and the original context of their claim lack a primary source.
• No data supports the actual scale of labs' alignment compute spend or its quantified impact on margins.
• The direct mapping between the executives' statements and the "alignment compute" framework is an interpretation by Baker and posters; Friar and Katti did not systematically frame it this way.

Why it matters

• This framing translates "AI safety" from an abstract promise into a concrete economics question: alignment, monitoring, and evals are real compute expenses that directly compress frontier-lab margins.
• If the judgment holds, investors evaluating AI labs must fold alignment costs into unit economics, and labs' "slow down or not" choices will show up as margin differences rather than slogans.
• For compute and infrastructure markets, it implies post-training alignment and eval compute demand provides support even if the frontier pace slows—"safety" becomes a hard compute-demand signal; but Baker himself draws the boundary—under a duopoly or open-source defeat, "always bullish for compute" does not hold.

2026-09-15 ~ 2026-09-16 · 8 related posts

• [source] Frontier labs will pace AI progress by spending more compute on alignment and evals, says investor — GavinSBaker · 2026-09-15
• Alignment compute spend will compress frontier lab margins, investor argues — soumitrashukla9 · 2026-09-15
• How labs would actually "pace the frontier": more compute on alignment at lower margins — AccBalanced · 2026-09-15
• How frontier labs would 'pace' AI: more compute on alignment and evals, at lower margins — nbaschez · 2026-09-15
• [source] OpenAI execs: safety alignment itself requires more compute, demand stays hot — GavinSBaker · 2026-09-16
• [source] 'Everything is bullish for compute' gets pushback: duopoly and open source loss aren't — GavinSBaker · 2026-09-16
• 'Safety' quietly became the strongest demand signal for compute, as OpenAI execs say alignment requires spending more — eyishazyer · 2026-09-16

1 near-duplicate retellings: Justin_Halford_