The efficiency counter-thesis — DeepSeek, Jevons, and China's domestic AI stack

Web-verified 2026-06-11. Structured + sources: fin-ai-efficiency-counter-thesis.json. Overlay — evidence-graded, excluded from the proofs. The honest steelman of the strongest challenge to this project's own thesis. Connects to fin-microsoft-openai (the $1.4T), fin-ai-depreciation-debttrap, geopolitics-chip-chokepoint-war, macro-ai-power-grid-bottleneck.

A zero-trust project has to doubt its own thesis too. The strongest argument against "the AI build is an over-leveraged bubble" is efficiency: if intelligence gets radically cheaper, maybe the brute-force compute build is either unnecessary (bearish for the spenders) or about to be swamped by demand (bullish). We hold both — and flag the tell.

The DeepSeek shock

DeepSeek R1 (late Jan 2025) delivered frontier-comparable performance at a small fraction of the claimed training cost. Markets read it as "compute is over-valued" — Nvidia lost ~$600B of market cap in a single day, the largest one-day loss in history (fact). Two readings followed, both fact-based:

• Bear (efficiency obviates brute force): if a model this good costs ~1/10–1/50 the compute, the $1.4T of OpenAI compute commitments + ~$600B/yr hyperscaler capex looks over-built — supporting the "largest misallocation of capital in history" view (fin-ai-depreciation-debttrap).
• Bull (Jevons paradox): cheaper intelligence → demand explodes. Satya Nadella tweeted "Jevons paradox strikes again!"; Goldman's 2026 forecast says at ~$0.02/kWh, AI inference is cheaper than human labor for ~23% of service jobs.

The tell

What happened next is the diagnostic: despite DeepSeek, 2026 hyperscaler capex grew to ~$602B (+36% YoY). The narrative resolved bullish either way — when compute was scarce you "needed to build more"; when DeepSeek made it cheap, Jevons said you "needed to build more." A thesis that justifies the same action under opposite facts is unfalsifiable — the structural signature of a bubble narrative, not an analysis. (This doesn't prove over-building; it flags that "efficiency = bullish" is being asserted, not demonstrated.)

China routes around the chokepoint

China is building an AI stack separate from Nvidia/CUDA, eroding the export-control moat (geopolitics-chip-chokepoint-war) from two sides at once:

• Efficiency: DeepSeek does more with less (a Huawei-led team post-trained DeepSeek V4-Pro, a 1.6-trillion-parameter model, on ~1,000 Ascend 910C chips).
• Brute force with subsidized power: Huawei plans ~600,000 Ascend 910C chips in 2026 (~1.6M Ascend dies total; customers Alibaba/Tencent/DeepSeek). The CloudMatrix 384 rivals Nvidia's NVL72 on inference but draws ~560kW vs ~145kW (4.1× the power for ~1.7× the compute) — inefficient, but China subsidizes energy for priority industries, so it's viable.

CSIS/RAND debate whether DeepSeek undermines or reinforces the case for export controls; a US lawmaker alleged (Jan 2026) that Nvidia co-designed the DeepSeek model (contested). Either way, smarter algorithms need fewer chips, and a domestic (power-hungry) stack supplies the rest.

The honest assessment (where it leaves the bubble thesis)

Crucially, the project's machine-proven results don't hinge on this. The circular-funding, self-marked-value, and depreciation proofs are about how the build is financed and accounted for — defective regardless of end-demand. The efficiency debate bears on a different question — "is the capex justified by real demand?" — and there, intellectual honesty requires saying: it is genuinely uncertain, the Jevons case is real, and the project does NOT claim the demand is fake. What it can say:

1. The unfalsifiable "build-more-either-way" narrative is bubble-shaped.
2. Efficiency makes the depreciation trap worse — today's expensive brute-force GPUs are obsoleted faster by tomorrow's efficiency (fin-ai-depreciation-debttrap).
3. China's efficiency + domestic stack undercut the chip-chokepoint moat the bull case leans on.

The distillation-by-proxy wrinkle (2025-2026)

A live wrinkle in the efficiency story: US labs allege part of the Chinese cost gap comes from distilling their models via proxied access. OpenAI flagged DeepSeek (2025); Anthropic (24 Feb 2026) named DeepSeek, Moonshot AI, and MiniMax - estimating ~16M exchanges from ~24,000 fraudulently-created accounts (MiniMax the largest, ~13M) - and later added Alibaba/Qwen. Because Anthropic and OpenAI sell no API in China, the requests were allegedly routed to the US providers through commercial proxies + overseas shell accounts - i.e. Chinese AI pipelines quietly pointing at American models. Fact that the allegations were made; contested per-firm culpability - the named firms have not confirmed.

Why it matters here. It partially undercuts the cleanest reading of DeepSeek: some efficiency may be free-riding on US models' training signal, not purely novel algorithm. But it does not collapse the counter-thesis - DeepSeek's architectural gains (MoE, MLA, FP8, Huawei-Ascend post-training) are independently documented, and distillation is a standard, cheap technique regardless of source. Both hold: real efficiency AND unauthorized distillation-by-proxy.

Honesty guards. (1) Composition fallacy - "Chinese labs" is not one mind; the evidence differs per firm. (2) The dispute is narrow - using a competitor's model + circumventing ToS/geo-limits, not distillation itself (which Anthropic/OpenAI/Google all do to their own models). (3) Anthropic's own detection drew backlash: per the Washington Post (Mar 2026) it quietly deployed software to unmask China-based users, then pulled it after privacy criticism - a surveilling-its-own-customers episode that cuts against a clean-hands framing. See spec-china-ai-stack-censorship for the graph edges + fuller account.

Sources: CNBC - Anthropic joins OpenAI in flagging distillation by Chinese AI firms (2026-02-24); CNBC - Anthropic's distillation battle turns to the dark web (2026-09-03).

← Research index · structured data: fin-ai-efficiency-counter-thesis.json · fin-ai-efficiency-counter-thesis.md