China’s AI Advantage Is Bigger Than You Think

Summary

• AccelXR opened this research as a distillation bear and flipped: “distillation just can’t explain a lot of the things that they’ve shipped.” The tell was Chinese labs stabilizing mixture-of-experts at multi-trillion-parameter scale and open-sourcing training recipes (DeepSeek’s GRPO/RLVR spawning DAPO and GSPO variants within months) — and the flow now runs backwards, with Thinking Machines distilling on top of Kimi and Hugging Face reportedly using a Chinese model to fight off an attack during the OpenAI agent incident.
• The core macro thesis: open-weight price wars push tokens toward “an electricity derivative” — and on electricity, “China is just absolutely trouncing the US right now.” Electricity is ~10-20% of token cost today; China’s grid generates 2x US electricity and is adding capacity ~6x faster, with AI needing only 1-5% of capacity added over five years versus 50-70% in the US. He hedges the endpoint — “I’m not entirely sure that it commoditizes to this extent” — which is why policy likely intervenes.
• His hardware call: “By 2028, we have a mass-produced, cost-competitive Chinese AI stack.” AccelXR’s confirming signal is the first multi-trillion-parameter model trained on Ascend chips; DeepSeek tried, with Huawei engineers on site, and reverted to Nvidia. Tommy notes that China could still be only ~10-15% of US total FLOPs output by 2028, but argues that the shift from training to serving could make those efficiency gains decisive.
• The “light switch” scenario: China dumps cheap open weights to suppress US lab margins until it reaches training sovereignty, then cuts off exports. Evidence it’s live: Xi’s viral pro-open-source speech came about three days before the FT revealed that MOFCOM was consulting labs on restricting open weights. AccelXR’s pessimistic default is “bridge demolition” within a couple of years; his ideal — the US competing through openness — “feels a little optimistic.”
• Lab rankings: DeepSeek takes “four of the top five” innovations (MLA, MoE, FP8 quantization in pre-training), Moonshot second with 1:56 sparsity, MuonClip, and hand-built below-CUDA kernels for Kimi K3, then Qwen/Alibaba and Z.AI, whose Cambricon/Ascend co-design he calls the coming training-side breakthrough. Baidu and Tencent are “fast followers.”
• The contrarian policy take: the best US move is getting China “back on the Nvidia bridge,” because export tightening paradoxically accelerates their sovereignty push. China already sees this — regulators are slow-walking H200 imports and pushing state-funded data centers to swap foreign chips for domestic ones, even as the Trump administration resumed H20 sales with revenue sharing in January of this year and H200 sales.
• The gap and the race: US labs hold roughly an eight-month capability lead, and because “the student can’t surpass the teacher,” distillation keeps China behind until training sovereignty — but if recursive self-improvement kicks in, eight months “feels like a good enough gap” for a US takeoff. Meanwhile the talent pipeline is reversing: Chinese AI researcher returnee rates doubled from 12% (2019) to 28% (2025), China produces 2x the AI PhDs, and China-educated top researchers rose from 27% to 38% — while Chinese lab valuations run 40-120x revenue. Tommy considered buying GLM but hesitated at roughly 1,000x price-to-sales despite its claimed ARR ramp from ~$100M to over $1B in five months.

Deep dive

1. A distillation bear walks in, an innovation bull walks out

• AccelXR’s opening confession: “When I first started this report, I was convinced that China was just doing benchmark maxing… but the deeper I dove in, the more I really came around to the other side that distillation just can’t explain a lot of the things that they’ve shipped.” Distillation explains a chunk of the capability and benchmark side; it can’t explain what they’ve built under hardware constraints.
• The flip moment: MoE training was “relatively unstable” before Chinese labs, across a handful of papers, “showed that you can actually stabilize multi-trillion parameter sized models” — mixing US-originated techniques and porting them to frontier training scale. That, plus an AEI think-tank report on China’s grid buildout, turned him.
• The research now flows both ways: Thinking Machines distilling on top of Kimi, and Hugging Face — during the OpenAI agent attack — reportedly resorting to a Chinese model to thwart attackers because of US frontier-model restrictions. Tommy’s reaction: “So the US is now benefiting from China’s AI research… that’s crazy.”

2. The league table: DeepSeek four of the top five, and why open source compounds faster

• His rankings: DeepSeek clear #1 (“MLA, MoE stuff… the first ones to do FP8 quantization during the pre-training pipeline”); Moonshot #2 — aggressively pushing sparsity to roughly 1:56, introducing MuonClip (now adopted or adapted by the other Chinese labs), KDA linear attention, and “good distillation” via MoPD; then Qwen/Alibaba; then Z.AI, less architecturally novel but doing heavy hardware co-design. Baidu and Tencent: fast followers.
• The velocity argument: DeepSeek R1’s GRPO and RLVR spawned variants like DAPO and GSPO “within a few months” because the training recipes are open. “I know for a fact our labs aren’t communicating to that degree… because it’s proprietary.”

3. Below CUDA: the co-design work that makes domestic training viable

• The mechanism, kept high-level: while most labs program GPUs through CUDA and its ecosystem, Chinese labs are handwriting low-level GPU instructions “at the assembly level” — Moonshot hand-built core GPU programs for a Chinese chip for Kimi K3, and Chinese labs also use RL to train a model to self-write and optimize kernels; DeepSeek handwrote low-level code with DeepEP.
• Why it matters: “That’s not necessary in the US… we can just build on top of Nvidia.” Optimizing for weaker chips means China can eventually “serve inference significantly cheaper than the US in the long term” — and, combined with quantization work and Cambricon/Ascend supply chains, become viable on the training side in the next few years. He thinks some people still get this wrong in assuming China can never train on its own chips.

4. US labs: capability lead intact, sane multiples, and a pivot from training to serving

• His US read: “It’s hard to see the US labs slow down in capability frontier leadership” — Chinese labs stay fast followers until training sovereignty. And on valuation, US labs look “still relatively cheap” versus Chinese peers at 40-120x revenue; Tommy considered buying GLM but hesitated at roughly 1,000x price-to-sales, even as GLM’s ARR ramped from ~$100M to reportedly over $1B in five months.
• The regime change: “We’re in a scenario now where it’s starting to transition from being able to train these models at frontier capabilities to actually being able to serve them at scale.” Expect US labs to adopt Chinese-style efficiency research and shift capability gains toward post-training and RL through the Chinese labs’ STR strategy — where “the scaffold or the harness for an agent can actually explain a large amount of capability gained” — rather than ever-larger pre-training runs.

5. Tokens as an electricity derivative — China’s structural trump card

• The chain: efficiency gains get passed through as price cuts, not margin; with open weights, “theoretically there’s no equilibrium where your per-token margin recovers.” Electricity is already ~10-20% of token cost, so as chip capex, software margin, and R&D amortization are squeezed, “tokens converge towards almost an electricity derivative” — and then China’s grid (2x US generation, adding capacity ~6x faster, AI needing 1-5% of five-year added capacity versus 50-70% in the US) dominates. His hedge, exactly as given: “I’m not entirely sure that it commoditizes to this extent.”
• China’s leadership sees it — he cites Ren Zhengfei of Huawei: China’s power generation and grid transmission are very good, and AI development “requires this power guarantee.” The chip answer is aggregation: a superpod of over 500,000 Ascend chips, 2.5x xAI’s Colossus by chip count and ~1.3x its total compute — weaker chips, more of them, until “it comes down purely to electricity bottleneck.”
• AccelXR says US misinformation about data centers is slowing deployments, while China runs “East Data, West Computing” — siting compute on cheap renewables stacked with tax and energy subsidies. One constraint he concedes: per the AEI report, Chinese chip production stays supply-constrained; he believes the ratio may be roughly 5:1, with production still insufficient for domestic demand. That is exactly why the co-design optimizations matter.

6. Monetization, the price war, and three endgame scenarios

• The price war is real — Tommy’s Silicon LLM Index, a weighted cost-per-million-tokens measure, fell from 2.2 on June 2 to ~1 — but monetization is dispersing: Moonshot now requires model-as-a-service providers above $20M revenue to sign separate agreements, Alibaba is adding a revenue threshold, Z.AI is looking at on-premises sovereign deployments, and DeepSeek added peak pricing and raised its rates. The open question: can they justify 6x-plus the revenue multiples of US peers? Meanwhile agentic workloads are “extremely cost sensitive,” driving a composition shift toward “cheapest but still adequately capable models” — exactly where Chinese labs are optimizing.
• His three scenarios: today’s “functioning bridge” (US frontier APIs own regulated workloads, Chinese open weights own the cost-sensitive mid-market, western enterprises hedge by running both); “bridge demolition” via US cloud/distillation restrictions or Beijing cutting exports — his pessimistic default; or the US “competing through openness” with cheap frontier-adjacent open weights — his ideal, but “it feels a little optimistic.”
• The light-switch evidence: Xi’s speech encouraging open source came about three days before the Financial Times revealed that MOFCOM was consulting labs on restricting open weights. The good strategy, in his telling: “dump cheap open weights on the market to slow down the US… until China can catch up itself, and then you flip.”

7. Reverse distillation, the eight-month gap, and reversing talent flows

• Tommy’s spicy question — odds that China trains a superior frontier model and US labs distill it — gets a structural answer: “A US lab’s best distillation teacher legally is the Chinese open-weight models.” AccelXR would not be surprised if some smaller US labs lean in; regarding Anthropic, he says he is not sure it would. Tommy’s rejoinder lands: “They are creating god though.”
• The gap math: consensus is roughly an eight-month capability lead, and since “the student can’t surpass the teacher model,” distillation freezes the status quo until Chinese training sovereignty. Is eight months enough for a US recursive-self-improvement takeoff — Anthropic’s “million geniuses in a data center”? “If it gets self-recursive, I think so. It feels like a good enough gap.”
• The talent underpinning is shifting: foreign-PhD Chinese AI researchers returning home doubled from 12% (2019) to 28% (2025), China produces 2x the US AI PhD pipeline, and the share of top AI researchers educated in China rose from 27% (2017) to 38% (2024). Tommy’s summary: “It’s just so hard to be bearish on China when you talk about them in any domain.”

8. What to watch — and the paradoxical case for selling China Nvidia chips

• His flipping-point indicator: “clear disclosure of actual full Ascend usage for pre-training at trillion-parameter scale.” People conflate serving on domestic chips with training on them — DeepSeek tried training on Ascend with Huawei engineers on site, hit instability and slow interconnect, and reverted to Nvidia. A confirmed multi-trillion-parameter Ascend-trained model would be “the biggest confirmation” of a coming policy shift; secondary signal: whether Chinese AI cloud pricing normalizes in the second half of this year after Alibaba, Tencent, Baidu, and Z.AI all raised prices.
• The policy paradox (he notes Sacks has made a version of this): loosening exports and getting China back on Nvidia “should slow down their development… it slows their sovereignty push, and any kind of tightening actually accelerates their push.” China gets it — regulators are slow-walking H200 imports and encouraging state-funded data centers to swap foreign chips for domestic alternatives, even after the US resumed H20 sales with revenue sharing in January of this year and H200 sales.
• On the US side he expects labs to lobby to wall off the ecosystem (“it just makes fundamental sense from a business perspective”), possible US-government ownership stakes in frontier labs, and possible government positions in discoveries those labs produce. The closing security paradox — if companies cannot access frontier models, they lose the ability to defend against them when released — is why Tommy stakes his flag: “That’s why I’m bullish open source… it’s the only thing that makes sense to me.”