明摆着 means plainly visible. 暗摆着 , After Dark still starts with data, and uses analytical frameworks assuming rational actors to make inferences about what isn't. The possible dark matter hidden from plain sight.
The starting gun nobody heard
OpenAI kicked off America's AI race in late 2022. The GPU frenzy, the Nvidia stock run, the permeation of the zeitgeist. Two years later DeepSeek dropped, and the Western narrative absorbed it neatly: China has joined the race.
The chronology is backward. China has been running full-throttle since the spring of 2016, 10 years ago. The starting gun wasn't fired in San Francisco. It came from London.
By the time ChatGPT was launching to breathless commentary and a Mag7 capex arms race, China had already implemented and iterated a national AI program across two Five-Year Plans. It was drafted swiftly, in response to a single figure almost no one talks about in connection to the Chinese AI Cambrian explosion, even though he's the one who lit the flame.
If you think Chinese AI started with DeepSeek, you cannot understand Chinese AI.
The day Demis shocked China
In March 2016, Demis Hassabis and the DeepMind team defeated Lee Sedol at Go, a game thousands of years old and a point of civilizational pride in China. An estimated 280 million Chinese watched those matches. The single biggest Baidu spike for 人工智能 artificial intelligence) landed during the Lee Sedol matches..
The American response to DeepMind was mostly polite applause and a Netflix documentary. The gushing compliments came from one of the creators of Deep Blue — which of course famously beat Kasparov at chess. But that was the emphasis: it was seen as having solved board games. The following year, President Trump proposed cutting National Science Foundation funding for AI.
In China, they had experienced their equivalent of the Sputnik moment. Of course AI research was already underway well before AlphaGo; but it was not a national priority as such, and seen more as a niche computer science project than civilizational infrastructure.
That all changed when Lee Sedol conceded match 3. Less than a month after AlphaGo's triumph, the People's Liberation Army (PLA) convened a top-level workshop titled A Summary of the Workshop on the Game between AlphaGo and Lee Sedol and the Intelligentization of Military Command and Decision-Making. Yes, it was serious.
On July 20, 2017, the State Council issued the 新一代人工智能发展规划 , the New Generation Artificial Intelligence Development Plan. A staged 2020 → 2025 → 2030 blueprint covering basic research, industrial applications, talent pipelines, and ethical norms.
That's a top to bottom national strategy, five years before ChatGPT.
One country treated AlphaGo as a tech demo that beat the hardest board game ever; the other as if it had seen fire for the first time.
The West saw a game; Beijing saw a method
What landed in Beijing wasn't just a board-game victory. It was a demonstration of a specific kind of intelligence — one that taught itself by self-play, learned a value function over positions, and produced moves no human had ever seriously considered. Small data, big task. A crow, not a parrot. Hassabis didn't build a machine that mimicked; he built one that seemed to reason.
And he kept going. Each result was the same method — search, self-play, a learned sense of what counts as good — pointed at a new scientific problem:
- AlphaFold (2020) collapsed a fifty-year problem in structural biology and became public scientific infrastructure within two years.
- GraphCast outperformed the world's leading numerical weather models on the large majority of measures.
- GNoME (2023) turned up 2.2 million new crystal structures, 380,000 of them predicted stable — roughly an order-of-magnitude jump in the known catalogue of stable inorganic materials.
- The Nobel Prize in Chemistry, 2024, for the AlphaFold work.
This was not AI as a chatbot. This was a scientist delivering civilizational infrastructure on a first-principles basis, paper after paper.
This is the asymmetry that matters. The West looked at AlphaGo and saw a solved game. Beijing looked at the same machine and saw an existence proof — a general method for original discovery in any domain you can simulate or score — and acted on it, realigning capital, institutional mandates, and engineering talent, and writing it into a national plan inside sixteen months. One capital read the result correctly. The other filed it under games.
So the lineage that matters for Chinese AI runs through DeepMind, not OpenAI — and not as a matter of taste, but of method. The frontier the West narrates runs through OpenAI: scale up next-token prediction until capability emerges. The one China has been building toward runs through DeepMind: reinforcement learning, search, and self-play, aimed at problems that have a right answer. Two different bets about what intelligence is for — and when DeepSeek's R1 landed, it was the second one.
Two Engines, One Brain
This story has already happened in modern Chinese history.
In 1955, after a five-year fight with the FBI, Qian Xuesen flew home to China carrying American rocketry and a decade at JPL in his head. The state built an institution around him. What followed was 两弹一星 Two Bombs, One Satellite — an impossibly ambitious program for an impoverished country with a destroyed industrial base. It was widely considered impossible after the Soviets withdrew their support. And for good reason — China at that time was far poorer than most of sub-Saharan Africa and much of the rocket program was done in the rural west in places that started without running water or electricity.
But then it produced exactly what it set out to: the atom bomb in 1964, hydrogen bomb in 1967, satellite in 1970. One phrase, one program, one national push that turned the People's Republic into a strategic power despite simultaneous US and Soviet pressure.
It is taught proudly in Chinese schools to this day. It is the canonical Chinese name for generational state-plus-science executed against a great-power deadline. Imported catalyst, indigenous system, civilisational soul.
Zhu Songchun, China's patriarch of AGI research and director of BIGAI, has formally proposed to China's top political advisory body that AI research be elevated to the same level as Two Bombs One Satellite. He is very influential.
Which leaves a question: if AGI is the successor to Two Bombs, One Satellite, what is its institutional vehicle? My analytical conjecture: China doesn't need to create one, because it already has a home.
两机 Two Engines is an existing top-level state program. It covers aero engines and heavy-duty gas turbines, written into the 13th and 14th Five-Year Plans, executed through the consolidation of forty-six fragmented engine units into the Aero Engine Corporation of China (AECC) in 2016. It's the closest thing China has to a strategic-tier, state-mobilised, generationally compounding scientific program.
What is Two Engines, and why does it matter?
Aero engines and gas turbines are the "Two Engines". They convert chemical and nuclear energy into electricity, into thrust, into industrial work and they do it at ultra-high temperatures that surpass the melting point of most viable materials.
For seventy years, 心脏病 literally heart disease, has been the standard Chinese figurative term for the country's inability to build a world-class jet engine. A chronic condition you simply live with. China could build the bullet train, the third-generation nuclear reactor, the lunar sample-return mission — and still had to buy the engines that hang under the wings of the C919, its own flagship airliner. A body with chronic heart disease cannot run, however strong the legs.
The reason it's hard isn't exotic engineering. It's materials. A modern turbine blade runs in gas hotter than the melting point of the metal it's made from, and survives only through a stack of metallurgical tricks refined over decades. Each new generation of alloy took the West roughly fifteen years to reach. Consensus in 2020 put China thirty-plus years behind on the hot-section — the searing core of the engine, where the metal lives closest to failure. Most Western analysis still says that today.
My analysis leads me to append cognition to the program name for completeness: 两机一智 Two Engines, One Brain. The brain isn't a chatbot. It's the AI that searches for materials, optimises industrial cycles, runs the closed loops, and ships the next-generation turbine hot-section the body can't ship without. Two Engines in plain sight, One Brain in the dark — and my belief is, executed together, written into the same plan.
The DeepSeek moment, revisited
On January 20, 2025, DeepSeek released R1. Western coverage absorbed it as a cost story: China caught up to OpenAI for $5 million. Scary. Recalibrate. The real story isn't just about cost, it's about method.
OpenAI's o1 had launched four months earlier — the first publicly released LLM trained with large-scale reinforcement learning on reasoning. But o1 was closed: the blog post described outcomes, the method was a trade secret, the system card was unreproducible. R1 was the same kind of model, but the entire training pipeline was disclosed in a peer-reviewable paper: algorithm pseudocode, reward functions, training dynamics, failure modes all directly into the public domain.
DeepSeek was first to publish, and in doing so, launched the post-training era. Within sixty days, every frontier lab on Earth had reproduced GRPO or a close variant. Within six months it was the field standard.
The method was the secret sauce: train a model not to mimic correct text, but to be rewarded for being right. Replace the supervised teacher with a value function over outcomes. The humans designed the reward, and then got out of the loop.
That isn't an LLM technique, that's the AlphaGo technique, applied to language. It was reinforcement learning over a search space, with a verifier replacing the human grader, producing reasoning traces no human had taught the model to write.
After ChatGPT launched, Google panicked. Their first response was Bard, which flubbed its first public demo in February 2023 and wiped $100 billion off Alphabet's market cap in a single day. So the panic worsened. Two months later DeepMind was merged with Google Brain, and Hassabis became CEO of the combined entity. Gemini became his day job. Hassabis had been derailed by his own employer, whereas the AI community in China had no Sundar Pichai.
The Chinese arrived late to the LLM battle, but when they did, they brought the crow with them.
The Dark Matter of Chinese AI is materials
Why materials? Because that's where the AlphaFold-and-GNoME template fits cleanest, and because materials is where every twenty-first-century industrial bottleneck actually lives. The gas turbine hot-section is a high-temperature-materials problem. Supercritical CO₂ power cycles are an alloy-corrosion problem. Fusion is a plasma-facing-materials problem. Battery density is a chemistry-search problem. Post-silicon compute is a 2D-materials problem. Every one of these is a high-dimensional combinatorial search over composition and processing — exactly the shape AlphaFold solved for proteins and GNoME extended to crystals. The lack of progress on these axes is exactly the cause of China's chronic heart disease.
AI for Materials(AI4Mat) by itself isn't enough. A model that proposes a million candidate alloys is useless if you can't make them and test them. That's the actual bottleneck — and it's why Western superalloys have walked from second to third generation at roughly one step every fifteen years. Most of that time wasn't science. It was institutional friction: lab to foundry, foundry to OEM, OEM to certification, each handoff a contract renegotiation.
China built something different. AECC's consolidation behind the Two Engines program collapsed the OEM layer into a single state-directed counterparty. Baowu and Ansteel supply the upstream metallurgy at national-champion scale; the Wuxi precision-casting cluster and AECC's own foundry network handle the hot-section parts. The whole stack is purpose-engineered to compress the lab-to-qualified-part handoff from years into months.
And CAS Shenzhen's 19-agent autonomous-discovery system isn't a standalone candidate generator — it's wired directly into that stack, routing proposals through pre-negotiated qualification protocols that would take a Western lab years just to set up.
A foundry like Wuxi Jinye doesn't wait for peer review. It waits for simulation convergence, then pours metal.
This is the half even DeepMind structurally cannot buy. The Hassabis template plus a continental-scale validation surface, under one coordination umbrella. China is the only country currently holding both halves.
And the payoff has started to arrive.
Heart disease, on the mend
Five years of preparation. Then, beginning in 2020, a phase transition.
The CJ-2000 core engine achieves first ignition, reaching 100.6% of design speed. The C929's future power plant clears its first wall.
China's first fully domestic 300 MW F-class heavy-duty gas turbine ignites in Shanghai — eight years of development, 50,000 components, sharing its hot-section metallurgy with the aero-engine program.
The CJ-2000 completes its full-state high-altitude test bench campaign. Peak thrust 35.2 metric tons — the Trent 1000 / GEnx class. Cumulative operation over 3,000 hours, with a claimed 15% fuel-efficiency improvement over the GEnx. The blade scrap rate — how many cast blades fail inspection — has fallen from 30% to 7.8%, a manufacturing-maturity signal as telling as the thrust.
A paper lands in npj Computational Materials from Central South University: CSU-S1, a low-cost nickel-based single-crystal superalloy (the same class of metal turbine blades are cast from) designed end-to-end by an AI pipeline that screened 340,000 virtual compositions down to one. It survives 224 hours under load at 1100°C, performance comparable to Western third-generation alloys that took fifteen years of iterative metallurgy to develop, and at lower cost. The first peer-reviewed confirmation that AI4Mat in China moved beyond proposing candidate materials to producing turbine-grade alloys.
Two engine programs — wide-body commercial and heavy-duty industrial — moving from far behind to possible parity in five years, when the consensus said three decades. The 3,000-hour CJ-2000 figure is the signal that matters most. The historical Chinese failure mode was creep fatigue — metal slowly deforming when it is held at high heat under load — the wall every prior program hit and died at. Three thousand hours on the bench means that wall has been cleared, at least at the test stand.
To be clear, no one is claiming these specific engines were designed by AI. The CJ-2000 has been in development for many years, and brute-force engineering could account for the milestones on its own. The papers are careful, and so am I. What we can say is that the tooling is now in hand, the validation surface is hot, and the cycle time is collapsing. The next generation — the one arriving in the 2030s — is almost certainly already moving through AI4Mat pipelines at every Chinese lab that matters.
If it works, what breaks
If half of this thesis is what it looks like, the following things break before 2035:
- The single-crystal superalloy moat collapses. GE and Rolls-Royce lose their last structural advantage in commercial wide-body propulsion
- Supercritical CO₂ cycles go from product category to platform, wrapping every high-temperature thermal source on Earth in compact efficient conversion.
- Fusion compresses on the materials side.
- Battery chemistry iteration cements into a permanent lead.
- Hualong, Linglong, and HTR-PM nuclear plants quietly become best-in-class via metallurgy.
Every one of these would arrive in Western press as an engineering announcement; nobody covers as an "AI story" because it's never announced that way. The AI layer that produced them stays invisible dark matter.
Close
The LLM race is public and loud. And it's enormous business. But just like in our universe, what we can't see is more massive than what's visible.
The heavier bet is the crow Hassabis showed them how to build — the one his own company had to set partly aside. It was placed in 2016, alongside AECC and the Two Engines plan, inside the same Five-Year Plan. It has been running for ten years. DeepSeek R1 was a visible flame. The wildfire is still burning in places with less visibility.
The payoff will not arrive as a chatbot. It will arrive (perhaps has started arriving) as alloys, blades, reactor vessels, battery chemistries, et cetera, delivered through a COMAC aircraft, AECC turbine, a CNNC reactor, or a CATL pack. And when it arrives, they'll call it an engine, they'll call it a reactor or they'll call it a car. Nobody is going to call it Chinese AI.
And seventy years after the diagnosis, China no longer has to hang a Russian engine off its most advanced fighter. The heart disease is lifting — and nobody will call that AI either.
Two signals will tell you, over the next few years, whether this thesis is right — both mostly invisible unless you know to look for them.
The rhenium ratchet
Start with rhenium, because it is the cleanest thing to watch. The best turbine blades survive temperatures past the melting point of their own metal partly because of a few rare elements blended in, and rhenium is the rarest and most important of them. A few percent buys a large jump in how long a blade lasts, or how much hotter the engine can run — and that margin is roughly the difference between one generation of Western alloy and the next.
The catch is supply. The world produces only 50–60 tonnes of rhenium a year — for scale, gold runs about 3,000 — and it comes out only as a byproduct of a byproduct of copper mining, so you cannot simply decide to make more. China produces almost none, and roughly 80% of global supply was already committed to Western jet engines. After the US cut China off from the LEAP-1C engine in 2025, "use less rhenium without losing performance" became the stated priority across every Chinese superalloy roadmap.
A high-end superalloy blends eight to twelve elements, and the number of possible recipes is astronomically large — far too large to search by trial and error. That is exactly the kind of problem machine learning is built for. The CSU-S1 paper is the first public sign it is working: 340,000 candidate recipes screened down to one, third-generation performance at lower cost. Lower cost almost certainly means less rhenium, since rhenium runs about $4,000 a kilogram against nickel's $15. If that holds, the rhenium bottleneck stops being a bottleneck within a decade — not because China found more of it, but because it engineered around needing it.
The pilot cycle speed
Inventing an alloy in a lab is the easy part. Qualifying it for production turbine blades in a flight engine takes, in the West, about fifteen years — and almost none of that time is chemistry. It's institutional friction.
An alloy has to walk through three stages: lab scale (grams of material, ideal conditions), pilot scale (the Chinese term is 中试 — kilograms to tonnes, real foundry equipment, realistic process variation), and finally production (industrial volumes, certified suppliers, regulatory sign-off). The hard part isn't any single stage. It's the handoffs between them, because in the Western model the people running each stage barely talk to each other. Lab researchers publish; foundry engineers wait for the paper; OEMs wait for the foundry; certification authorities arrive last. Each interface gets rebuilt from scratch every time a new alloy generation comes through, usually inside a project-specific consortium — GE plus Pratt plus Rolls plus a national lab plus a university — that didn't exist the year before and won't exist the year after. That's why third-generation alloys took thirty years to walk from lab demonstration to qualified flight hardware. Most of those thirty years were spent waiting for the next institution to pick up the phone.
What China built after 2016 is purpose-engineered to compress that handoff into a standing pipeline. 揭榜挂帅 — "post the challenge, claim the command" — lets any team bid on state-funded materials projects, including the qualification work, so the consortium is permanent rather than rebuilt each cycle. MOST's AI for Science program funds the algorithmic layer. AECC's consolidation collapsed the OEM side into a single state-directed counterparty. And Baowu, Ansteel, the Wuxi precision-casting cluster, and AECC's own foundry network supply the 中试 (pilot-scale) infrastructure the alloy has to pass through to reach production. The whole stack is a single coordinated pipeline. The CAS Shenzhen 19-agent autonomous-discovery system doesn't just propose alloys — it routes them straight into that pipeline, through pre-negotiated qualification protocols that would take a Western consortium years just to negotiate before any metal got poured.