The Open/Close← All briefs
Discussion · 2026-09-15

Pacing the frontier means more compute, not less — at the labs' margins' expense

Gavin Baker reframes the 'pace the frontier' push: pacing does not mean building less — it means Anthropic…

AI LabsSemis & Chips

Follow-up

Follow-up to Does “pacing the frontier” break the AI capex complex? — incremental development in the same debate.

The Debate

For vs against
Essence
Gavin Baker reframes the 'pace the frontier' push: pacing does not mean building less — it means Anthropic and OpenAI will spend *more* compute on alignment, monitoring and evals, accepting slightly lower margins. The thread splits between those who see this as capex-bullish (a new marginal buyer of GPUs and memory) and skeptics who argue alignment-as-compute is contested science, the safety metrics are undefined, and the policy fight may be theater. A follow-up to the Sep 13 debate on whether pacing breaks the AI capex complex — the new turn is that pacing may be the thing that *saves* it.
For
  • Gavin Baker (@GavinSBaker): pacing the frontier does not mean slowing down — it means frontier labs spend more time and more compute on alignment, monitoring and evals, spending slightly more money on compute at the cost of lower margins. Pacing is compute-bullish, not capex-bearish.
  • A senior OpenAI employee (roon @tszzl, quoted by Baker): 'pacing the frontier' will compress the margins of the frontier labs — 'a heavy cost imposed asymmetrically on model developers with the strongest AIs in America. by its nature, it would be a terrible regulatory capture tactic.'
  • Dan Druckenmiller (@xEBITDA): the memory-infrastructure bill for training, aligning, evaluating, monitoring and safely deploying each generation of frontier model is growing faster than the capability spend itself — a direct capex implication for memory and compute suppliers.
  • Rob Schoening (@sentientvector): use cybersecurity spend as a share of IT spend as the floor — if safety/alignment is ~10% of all training and inference spend, that is 'a really big number' of GPUs absorbed.
  • Almost Tomorrow (@almosttomorrowe): spending more compute on alignment changes what the marginal GPU is optimizing — the test is whether safety evidence improves faster than the systems being evaluated.
Against
  • Anthony Stafford, CFA (@stafant): the premise that alignment is just a function of compute is 'highly contested' — a lack of alignment more likely needs a fundamental reconsideration of how reinforcement learning works, so rewards are a function of outcome and behavior, not just more GPU-hours thrown at evals.
  • Vijay Walunj (@walunjlab): higher compute budgets may accelerate alignment work, but nobody has quantified the trade-off — what metrics or benchmarks decide when extra spend hits diminishing returns?
  • Shako Janashvili (@shakojanashvil2): the margin hit is the easy part to model; the harder part is whether extra compute actually changes the release bar — 'Pace' only means something if that bar is explicit.
  • Andrew Wilkinson (@StartupsILike): if pacing really just means burning more compute at a slightly worse margin, the whole policy fight reduces to labs asking permission to spend money they already have — theater dressed as restraint.

View the thread on X

The Tweets

4 posts
GBGavin Baker@GavinSBaker · 2:30 PM · Sep 14, 2026X Post

The way that Anthropic and OpenAI are going to "pace" the frontier is by spending more time and more *compute* on alignment, monitoring and evals. The frontier labs that choose to "pace" likely spend slightly more money on compute at the cost of lower margins. That's it.

239 Reposts · 2300 Likes · View on X →
GBGavin Baker@GavinSBaker · 2:30 PM · Sep 14, 2026X Post

Many factors would go into the lower margins, but incremental compute spend on alignment would be one. Don't take it from me, take it from a senior OpenAI employee:

11 Reposts · 212 Likes · View on X →
ASAnthony Stafford, CFA@stafant · 9:30 PM · Sep 14, 2026X Post

Assuming that "alignment" is just a function of compute - something that is highly contested. More likely a lack of alignment needs fundamental reconsideration of how reinforcement learning works to ensure rewards are a function of outcome and behaviour of the models.

View on X →
AWAndrew Wilkinson@StartupsILike · 11:30 PM · Sep 14, 2026X Post

If pacing the frontier really just means burning more compute on evals and alignment work at a slightly worse margin, the entire policy fight of the last few days boils down to labs asking permission to spend money they already have.

View on X →
On this page
01 / Follow-up02 / The Debate03 / The Tweets