"Heaven and earth are the furnace; creation is the smith; yin and yang are the charcoal; all things are the bronze."
— Jia Yi, Owl Rhapsody, 2nd century BC

Autumn, 1793. Jehol, north of the Great Wall.

Lord Macartney's embassy came ashore at Guangzhou with six hundred men and crates containing a working steam engine, field guns, astronomical instruments, and looms. He wanted an audience with the Emperor. He wanted to open trade.

The Emperor's reply ran to eight characters: "The Celestial Empire is abundant in all things, and lacks nothing."

Forty-seven years later, the gunboats arrived at the Pearl River.

This ledger starts there.

I. Three Absences, Three Different Prices

Four industrial tides have risen in four centuries. The first in Britain. The second in Germany and America. The third in America. The fourth — on both shores of the Pacific.

The first three times, there was no Chinese flag on the mountain.

But "absence" is too gentle a word. Three absences were three very different tuition bills.

The first: not "didn't sit at the table" — the table was on top of them

Before 1840, China still held roughly a third of the world's output. But per capita, it was a fraction of Britain's.

This was not non-participation. It was participation on someone else's terms — tariffs, ports, indemnities, all written by others. The tuition was not technology. It was sovereignty.

The second: they did try — but only bought the furnace

This is the most bitter one, because China genuinely tried.

The Self-Strengthening Movement ran from 1861 to 1895: the Jiangnan Arsenal, the Fuzhou shipyard, the Hanyang ironworks. Real money, real machines. In the same window, Japan was rewriting its own house rules — the Meiji Restoration, 1868.

Thirty years later, the First Sino-Japanese War settled it. Buying the furnace without changing the household does not work. The furnace survived. The mountain did not.

The third: not late — wrong layer

In the information tide, China's application layer was more aggressive than America's: e-commerce, mobile payments, short video, delivery. It went global.

But the blade was forged from someone else's iron: chips, operating systems, EDA tools, instruction sets. The foundation was almost entirely held elsewhere.

ZTE in 2018 and Huawei in 2019 demonstrated this in the most painful way possible. One sentence — "we won't sell" — and you discover: however ornate your technique, the sword is in someone else's hand.

String the three together and a pattern emerges: the application layer catches up extremely fast; the foundation layer is repeatedly missing.

That pattern has not broken. The chip is simply its latest manifestation.

So what is genuinely new the fourth time is not that China joined. It is that for the first time it climbed with its own furnace, its own iron, and its own internal art.

The internal art is still three parts short of mastery. But "three parts short" and "has none" are different things. The first is a question of winning. The second is a question of survival.

II. Five Passes

On December 3, 2025, Jensen Huang sat down with John Hamre of CSIS in Washington and drew a diagram dividing the whole craft into five layers, inviting a layer-by-layer comparison.

Three months later he fixed it in writing on Nvidia's site: energy, chips, AI infrastructure, models, applications.

In jianghu terms — the martial world of sects and masters — I would render them as earth-fire, dark iron, the training hall, the internal art, and technique.

The nine layers: five passes as the base, four beams as the frame

Pass one, earth-fire: the only short-term irreversible gap

In 2025, China's total electricity consumption crossed 10 trillion kWh for the first time — more than twice America's. Since late 2021, China's additions to generating capacity have exceeded the entire installed base of the US grid. In 2025 alone, its capacity growth was roughly seven times America's.

Price tells the same story. Industrial power in China's western energy provinces, with local subsidies, runs as low as 0.4 yuan per kWh. The US industrial average is about 9.1 cents. Nearly double.

But America's problem is not a shortage of electricity. It is when the power can be connected — interconnection queues, transformer lead times, skilled labor. Which produces a striking image: Meta and xAI are building their own power plants next to their data centers.

One of the richest companies on earth, burning gas on its own campus, because the grid has no slot.

Pass two, dark iron: the only generational gap

In 2025, US-headquartered semiconductor firms booked $425 billion in sales — 53.4% of the global market, the highest share since 1984. R&D spending: $76.8 billion.

Chinese semiconductor firms hold roughly 7.2% of global sales. R&D intensity: 10.66%, against 17.6%.

But something important shifted on the Chinese side: in 2025, domestic AI accelerators reached 41% of the Chinese market (Nvidia 55%, AMD 4%), with about 1.65 million units shipped — roughly 812,000 of them from Huawei.

That is volume substitution, not performance substitution. Per-chip capability, HBM, advanced process, the CUDA ecosystem — the gaps remain.

So the Chinese approach is explicit: if you cannot win on one chip, win on the array. Huawei's Ascend Atlas 950 SuperPoD links 1,024 chips into a single super-node at 1 EFLOPS (FP8). If one blade is not enough, bring a thousand.

That recovers some training efficiency. It does not recover per-chip energy efficiency — and per-chip efficiency determines how many tokens you get per kilowatt-hour.

Pass three, the training hall: stock versus slope

By the end of 2026, US data center capacity is estimated at about 56 GW (Goldman Sachs says 64 GW); China at just over 24 GW. Roughly 2.3x. China also has ~20 GW with firm timelines and ~30 GW merely announced.

But China builds faster and cheaper: a 100 MW facility can be delivered in about 12 months, at roughly $7.1 million per MW versus a global average near $11.3 million.

Two different disciplines. America practices scale economics — a handful of gigawatt campuses, extreme concentration, brute compute. China practices network economics — national compute corridors, one grid, unified scheduling.

One optimizes the peak. The other optimizes the whole.

Pass four, the internal art: from contest to price war

In March 2026, the best US model scored 1503 on Arena; the best Chinese model, 1464. A gap of 2.7%. In May 2023, that gap was 17.5% to 31.6%.

Stanford's 2026 AI Index used the phrase "effectively closed."

But 2.7% on a benchmark is not 2.7% in aggregate capability. The ceiling still differs by an order of magnitude: roughly 10 trillion parameters against 1.6 trillion.

China's real edge is efficiency: per dollar of private investment, it produces 12.7 times as many notable models. Open weights, brutal cost compression, flexible deployment — that is the differentiation.

Pass five, technique: the most fractured layer

America wins monetization. Private AI investment in 2025: $285.9 billion in the US against $12.4 billion in China — 23x. 1,953 US AI companies raised new funding, against 161 in China.

China wins volume and cost. In the first week of July 2026, Chinese models served 23.45 trillion tokens a week; American models, 4.28 trillion. 5.5x. API prices differ by one to two orders of magnitude: DeepSeek-V4-Pro outputs at $0.87 per million tokens; GPT-5.5 at $30.

The sharpest number is in the developer market: on OpenRouter, the US share of model requests fell from 72% to 33% in a single year. More than 30% of US enterprises' weekly model calls now go to Chinese models, peaking at 46%.

Two years ago, that sentence would have been unthinkable.

Pass	Who leads	Key figure
Earth-fire	China	2x consumption, 7x annual growth
Dark iron	United States	53.4% vs 7.2% global share
Training hall	US stock, Chinese slope	56 GW vs 24 GW
Internal art	US ceiling, Chinese volume	2.7% benchmark gap
Technique	Structurally split	23x investment, 5.5x usage

Read across the five passes and it is not 5–0. It is not 3–2 either.

It is two parties sharing one table, neither able to reach across it.

III. Two Leaks

That table has a trap. It invites you to count squares — America wins two, China wins one, two are draws. As if it were five to five.

It is not. These five passes are not five tables standing side by side. They are one chain.

Fire feeds the furnace → the furnace makes compute → compute trains the art → the art becomes technique → technique earns silver → silver buys more iron and more fire.

Five seams. Break one, and the whole chain leaks.

A chain of five: break one link, and every layer falls

America leaks at electricity

Of 140 US data center projects in 2026, 30% to 50% face delay or cancellation. Only about 5 GW actually broke ground.

The four hyperscalers will spend $720–745 billion. Money is not the problem. Chips are not the problem. The land is approved. The problem is when the power arrives.

China leaks at iron

The same eight-GPU B300 server lists at about $572,000 in the US. In Chinese channels it quotes at 14.8 to 15.3 million yuan.

At market exchange rates, nearly four times as much.

Which means the savings from construction — that $7.1 million per MW, that twelve-month delivery — get eaten entirely by the chip premium.

This is the bitter core of "24 GW against 56 GW": same wall, same substation, same cooling towers. Different contents.

Two leaks: the castle leaks at power; the ridge leaks at iron

Not for lack of will. For lack of reach. America cannot strike China's power grid and manufacturing base. China cannot strike America's chip and software ecosystem.

So the real form of this war is a race to patch your own weakness. Whoever fails first, loses. Not whoever lands the first blow.

The jianghu has never seen a fight like this. Not a duel. An endurance match.

IV. The Art Leaks Out

Every previous industrial revolution bundled the craft with the machine.

Buy a steam engine, and the drawings came with the iron. Buy a chemical line, and the formula came with the equipment. If you wanted the skill, you bought the hardware.

For the first time, that is no longer true.

The art can be given away. The vessel cannot.

Open weights: copy them to a USB stick, commercial use included. Per Hugging Face's spring 2026 report, Chinese open models account for 41% of platform downloads.

The vessel, however, cannot be given away. Eight-GPU systems carry a 4x premium. The training hall is priced by the gigawatt.

So a configuration appears that the jianghu has never seen: the one giving away the art earns nothing; the one renting the hall earns everything.

This is not stupidity. It is the first time knowledge and the vessel that carries it have been split into two separate markets.

And their pricing logics are opposite by nature: knowledge must be free to spread; the vessel must be scarce to rent.

One is destined to fall. The other is destined to rise. Same board, two opposing forces.

The art is given away; the vessel collects the rent

Two snakes, each biting the other's tail

America pushes the price of the vessel into the sky: over $700 billion a year from four companies, betting in gigawatts.

That push forces China onto the low-cost path — not out of conviction, but out of necessity.

China drives the price of the art to the floor: 5.5x the weekly volume, API prices one to two orders of magnitude lower.

And the floor price eats American margins. On OpenRouter, the US share fell from 72% to 33% in a year.

Which produces the sharpest layer of all:

• Every gigawatt America builds proves China's point: this business will eventually be crushed by cost.
• Every price cut China makes proves America's point: compute is still hard currency.

This is why there is no result. Not because the sides are balanced, but because each side's success underwrites the other side's argument.

The more you win, the more right they look.

V. Three Clocks

The real opponent is not across the table.

This contest runs on three clocks, and they do not agree.

• The money clock: three years. Wall Street is already asking when this pays back. Morgan Stanley estimates global AI infrastructure will need about $1.5 trillion in external financing by 2028 — the game has moved from "who can raise" to "who can generate cash."
• The planning clock: five years at a time. The 15th Five-Year Plan is the fifteenth of its line, covering 2026–2030; 2035 is a long-range vision, not this term's deadline. It does not read quarters — it reads succession, one plan after another, for decades.
• The technology clock: a few months. In 2026 a frontier generation lasts three to six months, with minor versions on a weekly cadence — in April 2026, three frontier models landed within eight days.

Three clocks out of sync — but one thing must be said first: the planning clock was never built to chase the technology clock.

The five-year plan is a long-horizon ledger. It governs power, land, and production lines — things that take a decade to pay back and must not be built on a model-release tempo. Model iteration was never the plan's job; it is done by companies fighting in the market. DeepSeek ships a new generation every six months and patches monthly — its craft clock runs no slower than Silicon Valley's.

So these two clocks do not throttle each other. America can die of the money clock's impatience: $700 billion committed, a third to half of projects slipping, and capital will brake on its own. On China's side, the plan is not unable to turn — five-year plans carry mid-term reviews and rolling adjustments, and after DeepSeek's breakout, the pivot across the country was visible to the naked eye. The real danger is the stampede: once a direction is set, every province and every fund crowds into the same lane at once — you build an AI compute park, so do I. The planning clock's danger was never being too slow to turn. It is that when it turns, everyone turns together.

VI. Three Thousand Disciples

Everything above is dead matter. This layer is alive — and it is the only one that walks away on its own.

Jensen Huang once said something widely repeated: "Half the world's AI researchers are Chinese."

Half of that is right. The right half deserves detail. The wrong half, followed through, leads somewhere bad.

First, the numbers: over-represented, not total

By AMiner's count, ethnic Chinese scholars make up 598 of the world's top AI researchers — 29.9% — while China holds under a fifth of world population. Clearly over-represented. Not universal.

The Paulson Institute's 2022 figure is sharper: in the top 20% of US AI institutions, ethnic Chinese researchers were 38%, overtaking the 37% who were US-born.

At team level: five of xAI's twelve founders; seven of the first eleven in Meta's Superintelligence Lab; nine of ChatGPT's original eighty-seven.

Harder still, at the hardware layer: Nvidia's Jensen Huang, AMD's Lisa Su, Broadcom's Hock Tan, Intel's Lip-Bu Tan — all four semiconductor giants run by ethnic Chinese executives.

Four filters, and none of them is called "mindset"

People like to explain this with culture. I do not buy it. Four filters do the work.

One: the gaokao and the olympiads. A pure-score mathematical sieve that strips out wealth, connections, and extracurricular polish, leaving raw abstraction. And AI's core requirement for its first decade was exactly mathematical fluency plus engineering stamina. The filter happened to match the era.

Two: the scale of engineering education. About 5 million STEM graduates a year in China; about 500,000 in the US. A factor of ten is not mysticism.

Three: density of real deployments. A billion-person single-language market plus a complete industrial chain. Recommendation, video understanding, commerce search — these are trained in live systems, not on leaderboards.

Four: double selection at the border. A Chinese researcher standing in a core Silicon Valley team has already been filtered twice — once by the gaokao, once by visas and admissions. The "Chinese researchers are strong" you observe is a sample that was heavily pre-selected.

That is selection bias, not group trait. The same logic explains three other over-representations: Indian-Americans in Silicon Valley management, Israelis in cybersecurity, Russians in algorithmic competition. You can invent a "national character" story for each — and an unfalsifiable explanation has no operational value.

Three classes of disciple

The real problem with "both sides are Chinese" is not that it is wrong. It is that it collapses three utterly different groups into one.

• Retained guests: raised in China, recruited to America. Highest skill, most fragile status. One visa memo and they are gone.
• Household members: ethnic Chinese born and raised in America. They run all four chip giants — but they are Americans, not "that Chinese cohort." When Jensen Huang lobbies Washington on export rules, that is his country, not his ancestral home. Ethnicity is not a position; passports and interests are.
• Mountain folk: those who never left. Of DeepSeek's 356 researchers, 145 — 53.5% — have never in their careers been affiliated with any institution outside China. Of the 31 core researchers, 10 have never left the country.

For the first time in history, the core of a Chinese frontier team did not have to go abroad.

That matters far more than how many Chinese researchers America has.

What is breaking is the inflow, not the stock

One widely distorted fact needs correcting.

"A mass return of top scholars" is an exaggeration. Carnegie's tracking: of NeurIPS authors who were in the US in 2019 with Chinese undergraduate degrees, 87% were still in the US in 2025.

But on the other side: the flow of AI talent to the US has fallen 89% since 2017. Chinese enrollment in US math and CS programs was 63,000 in 2024–25, down 20% in five years. And the share of China-educated top researchers staying home rose from 57% to 69%.

Zoom out: The Economist's analysis shows China's share of the world's top-tier AI researchers rising from 29% in 2019 to 51% in 2025, while America's fell from 20% to 12%.

The conclusion is precise: the stock is still there, so "America hollowed out" is false; the inflow has stopped, so "America is bleeding" is true.

And bleeding is slow. You only see it a decade later.

A reversal: the variable is not the country, it is the courtyard

Here is a fact that overturns everything above.

ResNet was built by Kaiming He in 2015 at Microsoft Research Asia — in Beijing. All four authors were Chinese. All were in Beijing.

LoRA came out of Microsoft, by Edward Hu. CycleGAN, by Jun-Yan Zhu.

So "Chinese researchers can catch up but cannot originate" is simply false. They have originated era-defining work — some of it inside China.

The variable is not brains, and it is not territory. It is the courtyard.

The same person, placed in an institution that tolerates inefficient exploration, produces work that defines the problem. Placed in an institution with quarterly targets and mandatory output, that person produces work that solves the problem better.

MSRA sits in Beijing. It is not a "Chinese-style courtyard."

Good news and bad news. Good news: the people are not the problem. Bad news: the bottleneck is the courtyard, and courtyards are much harder to change than people.

VII. Stacking and Attracting

At this point we can name the contest at a higher level.

The five passes are matter. The four beams are people. And the two halves obey completely different physics.

Stackable: earth-fire, dark iron, the training hall, the internal art, technique — money, power, factories, chips, manpower. You can pile these up. China is extremely good at this, and getting better.

Only attractable: disciples, capital, provisions, rules — environment, opportunity, certainty, dignity, reputation. These cannot be stacked. America is extremely good at this, and has been for eighty years.

So the real main line emerges: not who takes whose mountain, but stacking versus attracting.

And the biggest variable is this — America is currently ruining its own capacity to attract.

Visa screening, Proclamation 10043, suspicion of ethnic Chinese scientists, shrinking university budgets, tighter immigration. The 89% collapse in AI talent flow to the US was not poached by China. America closed its own door.

You can build a gigawatt data center in a year. You cannot rebuild "this is where the world's smartest people most want to be" in a year.

Reputation takes twenty years to accumulate and three to destroy.

China's side has a deeper ailment in its stacking, and it is not about speed. On the road from ledger to the teams that can actually fight, money passes through many hands — those who write good grant proposals are not always those who can wage war, and those who can wage war are not always good at courtly navigation. Who holds the baton is another question: paper counts, patent counts, utilization rates — every metric gets optimized, and what gets optimized is often not the thing that needed optimizing. And accountability is structured around "no mistakes first": layers of sign-offs, gatekeepers at every step, and in the end, often no one answerable for the outcome.

As for talent — recruiting it is one thing; daring to hand a person a major project and let him go ten years without delivering is another. The ridge has no shortage of people. What it lacks is the rule that lets real craft get real funding, and real burdens fall on real shoulders. That rule cannot be bought with money. It is far harder to build than a compute park.

Fairness requires saying this too: the same system delivered high-speed rail, solar, and EVs — mobilizing concentrated strength is real, and so is the drift into form over substance. One hand, two kinds of results. Which one prevails depends on what the task demands: clear-targeted catch-up is its strength; open-ended originality is where it runs slow to warm.

So this is not a stalemate. It is both sides walking toward their own opposite — America ruining its attraction, China over-stacking: overcapacity, duplicated buildouts, idle compute, internal churn.

Not a war of blows. Mutual attrition, in different places.

VIII. The Narrow Ridge

This was never a two-party affair. Many sit at the table. They simply do not sit in the seat marked "can build the table."

Between the two mountains runs a narrow ridge. Four things sit on it: TSMC's forge, ASML's mirror-grinding table, Korea's stacked jade, Japan's quenching pool.

Every blade in the world passes through those four processes.

On the narrow ridge: two definers, one vital chain, many tenants

Tier one — full stack: the US and China

Both can supply all five passes, each with one leak. The real threshold is a product, not a sum. In four centuries, only these two clear it.

Tier two — holding the choke: Taiwan, China; the Netherlands; South Korea; Japan

The most comfortable tier, and the most dangerous. Comfortable because irreplaceable. Dangerous because narrow.

South Korea is the clearest-headed player in the entire game: don't gamble, run the casino. Samsung plus SK Hynix hold roughly 80% of global HBM; SK Hynix alone had 58% in early 2026. The HBM market is about $54.6 billion in 2026, up 58%.

Whoever wins the model race still has to buy Korean memory. No technology risk, and a toll on every traveler.

Japan is the awkward one. Rapidus aims at 2nm by 2027, with about $16.3 billion in government support. Meanwhile, of Japan's $550 billion investment pledge to the US, roughly $21 billion goes into two American fabs — money out the door, domestic capability not necessarily stronger.

Japan's real ace remains materials and equipment, plus robotics — the gateway through which AI enters the physical economy. But it stays a supplier, never a definer.

Tier three — buying a seat: the UAE, Saudi Arabia, Singapore, Norway

The UAE and Saudi Arabia hold over $3 trillion in sovereign wealth. The UAE is building a 5 GW campus, with Stargate UAE's first 1 GW bringing 200 MW online in 2026; Saudi Arabia's HUMAIN has committed roughly $100 billion across 11 data centers and 2.2 GW.

But all of it runs on US-approved silicon. The seat is rented. So is the security.

In 2026, Iranian strikes damaged AWS facilities in Bahrain and the UAE. Abu Dhabi was forced to redesign the entire 5 GW campus: dispersed siting, hardening, possibly underground.

An expensive lesson: the cloud is not in the cloud.

Tier four — trading rules for a seat: the European Union

In 2026 the EU holds about 5% of global AI compute; the US, 78%; China, 11%. Seventy percent of European cloud sits with AWS, Microsoft, and Google. Europe holds 3% of global AI patents (the US 70%, China 14%).

Most telling: European AI startups raised a record $23 billion in H1 2026 — but that is only 11% of the global total, and late-stage funding runs 9x below America's. Fifty-three percent of late-stage capital comes from foreign funds.

So the best companies get bought at Series B.

The €200 billion InvestAI program's seven "gigafactories" will not be operational until 2027–28. The four American hyperscalers will spend over $700 billion this year alone.

Europe regulated the interface and forgot to build the engine.

Tier five — scenarios only: India, Southeast Asia, Latin America, Africa

India is the variable worth watching. By June 2026, the IndiaAI Mission had aggregated 45,000 GPUs; national data center capacity reached 1,575 MW; Sarvam open-sourced 30B and 105B models.

Sounds respectable. Compare: Google alone holds the equivalent of roughly 5 million H100s; Meta plans 1.2 million high-end GPUs by the end of 2026.

India's entire national compute is smaller than one company's.

Its strategy is smart: trade the world's largest untapped market for compute, models, and supply-chain landing. But the ceiling on that path is power, land, and manufacturing depth — not cleverness.

And the most critical link on that chain — advanced process — sits directly on the world's most dangerous geopolitical fault line.

Which means chips were never merely a technical topic: the physical foundation of global AI is welded directly above that fault.

And here is the sharper point: China and the US stand on the two sides of that fault — not merely as rivals facing each other across it, but as the two bankers running it. One uses chip controls to turn the fault into leverage; the other lays industrial bets on the far side. Whether the fault tightens or loosens, holds or trembles, does not depend on the fault itself. It depends on how much the two bankers have staked, and how much patience they have left.

The ridge does not belong to anyone. Everyone stands on it. The more who stand, the thinner it gets. And the fault beneath the ridge is held by the two houses dealing from either end — the more impatient the bankers, the more dangerous the fault.

IX. All the Marbles

Now this contest can be named.

Not "a contest for supremacy." Supremacy is two men fighting over one chair.

Not "guarding a pass" either. Passes cannot be held — steamships bypassed canals, railways bypassed post stations, containers bypassed the docks. Everyone who collected a toll was bypassed.

Some in the jianghu call it "setting the price" — America sets the ceiling of compute: compute futures are queuing up in Chicago (CME had planned an October 2026 launch, with regulatory review extended to November); whoever fixes the benchmark price rules the commodity. China sets the floor of intelligence: open weights weld "cheap" into the standard, and when Chinese models trail by three to six months but cost one to two orders of magnitude less, customers vote with their feet. The height between ceiling and floor, they say, is the entire profit pool of this era.

It is an elegant framing. And it makes the contest look smaller than it is.

Price is the battlefield, not the stake. The real question is: whoever wins this war wins the whole world.

Look at the list of stakes and you see why this is winner-take-most: rules — benchmarks, agent protocols, license terms, written by the winner. Ecosystem — developers, data, compute procurement, all flowing toward the winner. Capital — the next round of money goes only to the side proven to be winning. Talent — the best generals always defect to the courtyard where things get done. These are not five prizes. They are five faces of one thing: the whole board.

And "setting the ceiling" or "setting the floor" is something you earn the right to do only after winning. Prices are always set by the winner. In the 1980s Japan held half the world's semiconductor market — one agreement took the pricing away. Kodak invented the digital camera itself — the pricing power still went elsewhere. The loser is not short of capacity or technology. The loser's price is simply anchored by the winner's price.

So the real picture is not a division of labor — one sets the ceiling, one sets the floor, profit in between. It is two bankers at one table, and the table holds only one stack of chips. If America wins, chip standards, cloud rules, and model protocols all fly its flag. If China wins, manufacturing standards, the open-source ecosystem, and the price benchmark fly its flag instead. Whoever wins sets the price. Whoever loses gets priced.

Then why has no one won yet? Because both can still hold their breath — on different clocks. The money clock forces America to show returns within three years. The planning clock lets China roll forward five years at a time. Time stands on different clocks: outlast the other's leak, and the winner's price becomes the world's price.

X. Lamplight

Back to that autumn in 1793.

The crates held a steam engine. The Emperor glanced at it and said the empire lacked nothing.

Two hundred and thirty-three years later, on the same coastline, the furnace burns brighter than anyone's.

But one thing must be said plainly: climbing the mountain is not winning.

In the late 1980s, Japan held roughly half the world's semiconductor market. It lost to a single agreement, and to missing the PC.

Britain led the entire first industrial revolution and lost the second.

Leaders are rarely defeated by the pursuer. They are defeated by the very way they won.

So "nobody can call it" is not politeness. It is mechanical: of the nine layers, two are deadlocked (dark iron, earth-fire), three are changing hands (training hall, internal art, technique), and the decisive one has not been named.

Anyone who hands you a "we'll know in five years" conclusion today — question their ruler first.

The real meaning of this round was never "can we win." It is whether the rules get written by someone else on your behalf.

The cost of the first three absences was not a few decades of lag. It was a century in which every bill for technological progress was written by others.

The five passes ask whose wall is higher. The four beams ask whose courtyard still has people willing to walk in.

In Closing

Compress the whole board, and it comes down to seven lines:

1. The five passes are serial, not parallel. You cannot count squares. The real score is not 5–0 or 3–2; it is two leaks — America's power, China's iron.
2. This is not one offensive war. It is two defensive ones. Neither side attacks the other's strength, because neither can reach. Whoever fails to patch first, loses.
3. The model is the fastest-depreciating card. A 2.7% gap earned through engineering and open weights can be earned by anyone. What cannot be copied is the industrial base, the power engineering, and "cheap" as a standard.
4. "Both sides are Chinese" must be split into three. Retained guests are liquid; mountain folk are fixed. The long game belongs to fixed assets.
5. The bottleneck is the courtyard, not the people. ResNet came out of Beijing, so the people are fine. But MSRA is not a Chinese-style courtyard, so the organization is the real variable.
6. The five passes can be stacked; the four beams can only be attracted. China stacks, America attracts — and America is currently ruining its own attraction. That is the biggest variable of all.
7. Price is the spoils. The world is the stake. Setting the ceiling or the floor is something you earn the right to do only after winning. Whoever wins this war wins the whole board — rules, ecosystem, capital, talent all flow to the winner. The loser's price gets anchored by the winner's.