The US–China Capability Lag Under a Compute Deficit

Sources#

Summary#

The Open-Weight Frontier Gap measures how far the best Chinese models trail the US frontier. This page is about why the number is so small. JS Denain of Epoch AI states the puzzle in his September 2026 conversation with Nathan Lambert (Debating RSI, the US-China Gap, and Jaggedness with JS Denain of Epoch AI, Interconnects podcast, practitioner-opinion): Chinese labs "have way less capital and particularly way less compute than US labs. And that delay is huge." Whether the capability lag is four months or eight, it is "a much shorter delay in capabilities than in capital. And so, this requires some explanation." Neither speaker measures any of the explanations below. The page records who argues for each, how much weight they give it, and what would test it.

How big the lag is, by ruler#

Who	Estimate	Ruler	Tier
Denain (Epoch)	6–8 months, "four to eight seems reasonable"; 2 months "seems a bit more overfitting-y"	Epoch Capabilities Index, release date to release date	`practitioner-opinion` reading Epoch's own data insight (not in `raw/`)
Lambert, briefing (The current balance of power in open models)	2–5 months, decreasing over three years	AAII snapshot and his own dashboards	`practitioner-opinion`
Lambert, on air (2026-09-22)	Kimi K3 and GLM 5.2 "closer to like two to three or four", then: "I think I intuitively have landed something closer to you"	AAII	`practitioner-opinion`
Epoch real-work tasks	Kimi K3 ~7 points below Grok 4.6 at the same ECI of 158	Epoch's own work, one grader	`empirical`, n = 1 per task

Three details from Denain bear on reading the table:

Contamination does not explain the ECI lag. Epoch "did internally look into" an ECI built only from private benchmarks, or from benchmarks published after a model's release, and found the overfitting effect "basically not statistically significant." He expects model-card numbers would show an effect where ECI's own runs do not. This is an unpublished internal analysis, relayed in conversation.
Both indices may understate the lag, for different reasons. AAII "might understate the gap for curation reasons." ECI may understate it because OpenAI and Anthropic optimize capabilities that "appear less in benchmarks" and serve a broader set of users. The second point is the same direction as Epoch's own real-work finding in the table.
The interval depends on the lag method. Forward or backward matching, staircase or interpolation: Denain says most of his 4–8 range is method uncertainty. Lambert's release-cadence argument (a year ago Chinese labs shipped within days of training, US labs within months, which "would artificially squeeze the gap") is in Denain's view already absorbed by that range.

Reconciliation. The two estimates do not share a ruler, so neither supersedes the other. Lambert's own on-air concession moves his figure toward Denain's a day after his briefing was published, and Epoch's empirical report says the index Denain uses errs on the small side. So the corpus's best reading is that 2–5 months is a lower bound on the AAII and 6–8 months a central estimate on ECI.

The candidate explanations#

Explanation	Mechanism	Argued by	Weight they give it
Distillation	API outputs from frontier models used as training data; turns money into high-quality data "you would otherwise need compute for" (Lambert)	Denain, against Lambert	Denain: "updated towards distillation is actually a really big factor"; Lambert: "without distillation, the gap would be fairly similar." See below
Router traffic	The relay services in Anthropic's September 2026 threat report give realistic prompt distributions, not only completions	Denain	Useful for real-world capability; weak for the ECI lag, since benchmark prompt distributions are public anyway (Lambert's objection, which Denain accepts)
Bought RL environments	A US data vendor iterates with a frontier lab, which spends R&D compute finding which environments work; the vendor then builds similar ones and sells them later, including in China	Denain proposed; Lambert has heard it "multiple times, including from people in China"	"Implicit R&D compute that was spent and that it gets saved." Lambert's multi-hop rumour: "we just have to wait a certain amount of time and then we buy the RL environments that Anthropic bought for a 10th of the price"
Labor and data lags shorter than compute lag	Two years behind on compute, six months on data, level or ahead on researchers	Denain	Named, not sized
Proof-of-concept ("four-minute mile") effects	Knowing a result is possible saves the exploration budget	Lambert, named by Denain	Denain asks for examples and offers a counterexample: DeepSeek Math preceded o1, so reasoning models were not this. Lambert relocates it to org structure: Chinese labs spend on next-generation compute, US labs also on models two or three generations out
Personnel leakage	Architecture innovations circulate among US labs "very quickly through personnel turnover" and then reach Chinese labs	Lambert	Enough, with the others, that distillation is not needed
Benchmark focus	Some Chinese labs are public companies, and close benchmark scores help them financially	Lambert	"A month or two" at most
Grind on mundane data work	Talent organized for hill-climbing data work	Lambert	"Not a gigantic lead cause"
Model spillovers short of distillation	Using Claude as a reward model or data cleaner; inferring training choices by playing with a model	Denain	"Some spillovers"
Idea leakage	Ideas leak	Denain	"I don't suspect that's a huge effect"

The decisive split is between distillation and everything else. Denain summarizes Lambert's position as "with the other explanations, you can just get pretty far, such that you don't really need distillation to explain a large fraction of the gap", and Lambert agrees. He adds a reason to doubt distillation's reach. If a distinctive reasoning-trace style must be kept stable through mid-training and SFT, a lab cannot simply swap in a teacher's traces each cycle.

How much distillation buys: two estimates, one shape#

Asked where the gap would sit in 6–12 months if distillation were effectively stopped, "without RSI being super crazy," Denain answers: "currently if I'm saying six months gap, then in six months, it's probably not literally 12 months… you go from like six to nine maybe," i.e. +2–3 months. His reasons for updating toward distillation are, in his order: "a lame reason… more people seem to be saying it"; the Stolen Thoughts paper suggesting traces were "relatively easy to get" (not in raw/); the router channel; Beren Millidge's claim that mid-training gets "80% of the way there", which Denain reads as evidence that SFT on trajectories matters more than he thought; and the weakness of the alternative explanations.

Lambert's briefing estimate is +1–2 months (Illicit Distillation).

The rhetorical disagreement is larger than the numerical one. On a 6-month base Denain's +2–3 months is roughly a third to a half. On Lambert's 2–5 month base, +1–2 months is roughly a quarter to a half, a ratio Lambert's own page already computes. Both say stopping distillation widens the lag by a large fraction and leaves the Chinese labs months, not years, behind. They differ on the base and on the adjective. Both are judgements with no method. The falsifier is the same for both: a release cycle under KYC-gated API access, which Illicit Distillation records as its open test.

Lambert adds a scenario condition: if RSI goes "super crazy", "the labs pull ahead by a lot more". On that reading the gap is a stationary lag only while the frontier's own rate is unchanged.

What the job postings show#

Epoch's study of 1,604 Chinese AI job postings (Cheryl Wu and Denain, June 2026, not in raw/) is the one piece of evidence either speaker brings from inside the Chinese industry. Denain calls it a source of "interesting anecdotes or interesting facts" rather than trend lines:

Own data centers. Data-center roles posted in particular regions suggest that "at least some of these labs are basically constructing their own data centers as opposed to renting compute from Alibaba."
Strategy differences. Z.ai (Zhipu, GLM (Z.AI)) posts far more B2B sales roles than MiniMax or Moonshot (Kimi (Moonshot AI)). MiniMax expands internationally more than Z.ai. DeepSeek shows little of either.
In-housing and interns. More in-housing in general, including data work, and heavy hiring of interns, "some of that might be for data purposes."

This bears on the "labor and data" explanation without sizing it. Lambert's request for a follow-up — hiring whose job is buying data — is the measurement the bought-environments explanation would need. Both note that data is the hardest input to track: "it's quite secretive", and "if the data companies were public, it would probably be easier to tease some things apart."

Connections#

The Open-Weight Frontier Gap — the measurement this page explains; its axes all read a gap larger than the indices show, and its Lambert section is the source of the 2–5 month figure
Illicit Distillation — the campaigns behind the first explanation, the 1–2 month estimate, and the KYC test both estimates share
Open Weights as Competitive Strategy — Ng's "vastly overstated" verdict on distillation, which Lambert's "fairly similar" supports and Denain's update leans against
Epoch Automation Reports — the empirical reason to read ECI-based lags as a floor: Kimi K3 level with Grok 4.6 on ECI and ~7 points below it on Epoch's own work
Artificial Analysis — the index behind Lambert's months; Denain suspects it understates the gap for curation reasons
Kimi (Moonshot AI) / GLM (Z.AI) — the two Chinese labs Lambert names as closest, and two of the companies in Epoch's job-postings comparison
Epoch AI — Denain's organization, the owner of the ECI, and the author of the job-postings study
Nathan Lambert — the other side of the distillation disagreement, and the source of the bought-environments rumour from Chinese contacts

Open Questions#

Denain's +2–3 months and Lambert's +1–2 months agree in ratio and differ in base. Does a KYC-gated release cycle widen the ECI lag by about a third to a half, or by much less? The test is shared with Illicit Distillation's distillation-size question; track it there.
The bought-RL-environment channel is a multi-hop rumour from both speakers. Does any vendor disclosure, job-postings study or reporting show US-built environments being resold to Chinese labs, with a lag and a price ratio?
If the capability lag is set by explanations other than compute, it should not widen when the compute gap widens. Does the ECI lag stay at 4–8 months through 2027 while the US–China compute ratio grows? Trigger: Epoch's next US-vs-China ECI update.

Sources#

Debating RSI, the US-China Gap, and Jaggedness with JS Denain of Epoch AI — Nathan Lambert with JS Denain (Epoch AI, leads its Insights Team), Interconnects podcast #19, 2026-09-22, practitioner-opinion (published transcript, 95 speaker turns). §24:20–48:13: the gap estimates, the internal contamination analysis, the distillation counterfactual and Denain's reasons for updating, the explanation list, the bought-environments rumour, and the job-postings findings. Every Epoch analysis cited here is relayed in conversation; the data insight, the internal analysis and the job-postings post are not in raw/. COI: Lambert is an open-model advocate; Denain reports on his own organization's index
The current balance of power in open models — Nathan Lambert, Interconnects, 2026-09-21, practitioner-opinion. The 2–5 month gap and the 1–2 month distillation estimate. Full treatment on The Open-Weight Frontier Gap and Illicit Distillation
Can AI automate Epoch? — Hong & Burnham (Epoch AI), 2026-10-08, empirical. Kimi K3 at ECI 158 ≈ Grok 4.6, ~7 points below it on real work. Full treatment on Epoch Automation Reports