Five million barrels of oil, or: How I Learned to Stop Worrying and Love the Demonstration

Right now, we should be living through a global energy crisis unlike any other in living memory. The closure of the Strait of Hormuz by Iran should, according to every serious forecast, have caused an oil shock resulting in abject calamity. The economic contraction had the potential to define a generation. The sources that I read and rely on, in their reflection, say that the arithmetic was right, and yet the calamity simply never eventuated. We’re essentially fine. The reason, I am led to believe, was China.

I began reading about this story because the headline was an irresistible one (“China Quietly Saved The World Last Month”), and then I kept reading because the story felt to me to be substantially deeper than one just about how China responded to the closure of the Strait. And then I found myself reading more, and not about oil. The mystery, for want of a better two-word summary, was about Chinese altruism, and part of it has been chewing on the back of my mind for a year now.

Why do Chinese labs keep giving away their best AI models for free?

Stay with me. Oil and AI. Let me tell you the oil story first, because I’ve come to think that these two vines wrap around the same branch.

The global oil market runs with no slack in it. Roughly a hundred million barrels are produced every day, and roughly a hundred million are consumed, and the two numbers must match. So when the Strait closed and about a fifth of the world’s supply was disrupted, the deficit was a hole the size of several major economies’ entire consumption. Pipelines that bypass Hormuz could manage some of it. A coordinated release of strategic reserves (dozens of countries at once) clawed back a little more, but those reserves were quickly being depleted. The calculations couldn’t predict anything other than catastrophe. What no model contained (or could have contained) was what China did: China’s oil imports, the largest in the world and rising for four decades, fell by about five million barrels a day (almost halving) overnight and without announcement or explanation.

China could do this because it had spent years accumulating what is probably the largest oil stockpile on earth, much of it bought at a discount from sanctioned producers (Russia and Iran), paid for in yuan through the Bank of Kunlun (an institution that exists substantially for this purpose), entirely outside the dollar system that lets Washington see and veto the world’s oil trade.

That is a tidy version of the story, but the reality is that the mystery remains only partly solved across a few dimensions. China’s oil reserve (nearly 1.4 billion barrels, according to US EIA reconstructions based on pre-war data) is monitored via satellite by analysts, and they did not observe visible drainage of China’s oil tank farms the way the simple version of the oil story requires. Further, China’s motive for cutting its oil imports is entirely unestablished. The best account I found is ChinaTalk’s, whose theories I will borrow here. They run from the flattering (China wanted to save the world), to the transactional (China made a secret deal with Washington), to the ominous (a rehearsal for a Taiwan invasion, in which America closes the Strait of Malacca and China carries on without incident), to the boring one (that Beijing pulled a domestic lever for domestic reasons, and the rescue of everyone else was mostly a side-effect). China has not offered a comment. Perhaps the theories are all partly true. Perhaps Xi Jinping himself could not precisely rank them.

The part I keep returning to is that the motive is a mystery, but the demonstration is absolutely not. Whatever Beijing intended, every government and every trading floor on earth now ingests the same new fact: China can switch five per cent of global oil demand on and off at will. Or alternatively: China now has a hand on the price of the world’s most important commodity. This conclusion is huge. The demonstration outlives any motive.

Before going further I should say what I am and am not. I am not a China analyst, and the people who are cannot agree about any of what I have just described, so take my geopolitical interpretation with the grain of salt it deserves. This essay is not a prediction about who “wins” anything, or what “winning” really even means. I am a commercial technology lawyer, and the relevant thing about my job (in this context) is narrow: you spend a lot of time thinking about leverage. You watch what happens to a price when one side controls the marginal unit of supply, and you watch what happens to a negotiation on the day a good-enough free or cheaper alternative appears. The reason I could not put the oil story down is that it pulled the same thread as the other question that has been unravelling in the back of my mind all year.

For several years now, China’s leading AI labs have been doing something the standard arms-race story cannot comfortably explain. DeepSeek, Alibaba’s Qwen, Moonshot’s Kimi, Zhipu’s GLM — all of these are models within touching distance of the American frontier, and their makers just give them away. Not the product, but the weights themselves. The billions of numbers that constitute everything the model learned. Once those are published, anyone with the hardware can run the thing, copy it, modify it, build on it, forever. If you believe AI is a weapons race, this is madness. Nobody open-sources their warheads. It stops looking mad when you stop thinking about weapons and start thinking about oil.

Firstly, look at what kind of instrument each power has built. America’s move against China has been the Strait of Hormuz move: export controls on advanced chips, which is a chokepoint attack on the input to intelligence. China’s answering move is the opposite: not a blockade, but a flood. Every near-frontier model released for free recalibrates the world’s price for that level of capability, for everyone, permanently. The oil weapon is a valve: Xi can turn the imports back on tomorrow, which is exactly what makes it useful as leverage. Open weights are a ratchet. A released model can never be un-released, so the instrument fires only forward. It’s useless for coercion at a G20 summit, but it’s built for only one job, which is making sure that nobody, anywhere, ever again charges a durable rent for last quarter’s intelligence.

Secondly, look at where each instrument lives in relation to the law. A closed AI model, accessed through an API, is a dollar-cleared transaction in every sense that matters: it runs on someone else’s infrastructure, under someone else’s terms of service, through an account that can be closed, in a jurisdiction that can compel it. There is a kill switch, and we have seen it get used. In June, the US government ordered Anthropic to bar every foreign national from its two most capable models, and because the company had no way of checking who was who, for eighteen days the most advanced artificial intelligence in the world was simply switched off, for everybody, Americans included, by administrative order. It came back on, but the lesson left its mark. The episode showed the whole world that America can revoke access, which carries a force that pulls nations and individuals alike toward open models whose weights are not revocable. The same day, the Chinese lab Z.ai launched GLM-5.2, publishing the weights under an open source licence a few days later. Open weights are the yuan settlement of intelligence (the commodity moved beyond the control of the reigning power). What China is exporting is not so much intelligence as unsanctionability.

And I’ll quickly address one potential distraction: oil dependence is absolute, and AI dependence is, for now, a convenience. Flights do not fall out of the sky if the models go away. This is true, but it misses what the players are actually betting on, because the right comparison is not oil in 2026, but oil in 1906. In that era, oil was a product for lamps and early motoring, decades before the world had built itself around it so completely that a single closed strait could threaten to define a generation. Every dollar of the current buildout is a wager that the same accretion is underway for intelligence. Open weights are a bet placed before the dependence hardens, and an attempt to shape what the world will depend on while the habits are still forming. And there is a strange generosity in the mechanism, because a downloaded model makes you dependent on the technology without making you dependent on the supplier; it cannot be taken back. However, weights go stale, and so what the world will actually rely on is the stream of next releases. The stream remains discretionary; unlike a published file, it can be stopped.

None of this is metaphor, and none of it is hidden: the price of any fixed level of AI capability has been collapsing at somewhere between tenfold and several-hundred-fold a year depending on the task. Performance that cost $37.50 per million tokens in 2023 costs a few cents now, and the free Chinese alternatives enforce that trend. On OpenRouter (one of the main marketplaces through which developers route their AI traffic), Chinese models went from a rounding error to roughly half of all token consumption in eighteen months, and Qwen has overtaken Meta’s Llama as the most downloaded open-model family on Hugging Face (crossing a billion cumulative downloads this July).

We should also remember that twice now open weights releases have moved whole markets. In January 2025, DeepSeek published a near-frontier model as open weights (reportedly trained at a fraction of what the frontier was assumed to cost), and Nvidia lost $589 billion of market value in a single trading session (the largest one-day loss by any company in market history). And then in July this year Moonshot released Kimi K3 (the largest open model ever published, benchmarked behind only the top two closed American systems) and global chip stocks shed trillions in the days that followed (although in fairness the attribution might be a stretch; Kimi K3 was the trigger more than the cause). Two weeks later OpenAI cut the price of its lightweight GPT-5.6 model by eighty per cent and its mid-tier by twenty, and the headline that followed needed only two words: “OpenAI blinks.”

If you stand back and look at the AI landscape as a whole, an obvious objection appears, and it is a good one. If Chinese labs are commoditising intelligence, why are the American rent-collectors the fastest-growing companies in commercial history? Anthropic went from a one-billion-dollar run-rate to forty-seven billion in seventeen months. Enterprise surveys show the share of business workloads running on open models is actually falling even as spending explodes. Ben Thompson (Stratechery, worth your attention) argues it very well: tokens are not the commodity, intelligence is; the frontier labs may well produce frontier-grade intelligence at the lowest true cost; and enterprise demand is for the harness, the trust, and the accountability that OpenAI and Anthropic specifically provide, not for whatever is cheapest this week.

All of that is true, and I think it is the key to the whole picture, because the legal profession ran this experiment on itself.

Legal knowledge was once the rent. Knowing what an indemnity actually did was a thing clients paid for by the hour. Then came templates, precedent databases, document generators, and now the models that draft better than most juniors. But the profession did not die, and its revenues did not fall. The rent migrated, away from knowing things and toward judgement, trust, and the willingness to put your name on the advice and own the liability. Clients stopped paying us to know what the clause says, and they kept paying for someone to be accountable for it. I think that is what is happening to intelligence at large. The floor (the price of any fixed capability) is collapsing, and the ceiling (the frontier, the trust, the accountability) is where the rent went, and both things are true at once. The July price cuts were simply the first time the rising floor visibly tugged at the ceiling.

You can see the labs pricing this in themselves if you watch what they build. OpenAI’s first consumer device (reported to be a screenless home speaker with a camera and moving parts, an ambient companion that lives in the home) is what a software-rent company might produce when it suspects the rent is mortal. A model can be copied, but the object in your kitchen cannot, and neither can the context it accumulates about your household. Whether the device succeeds barely matters. What matters is the direction of travel: an American lab drifting downstream toward hardware and complements, onto the terrain where China already dominates manufacturing.

So why does China do it? None of the answers below are mine. To get at them, let’s follow the oil story and catalogue the theories, because no confession is coming.

Theory one: pace control. America’s frontier is financed by private capital chasing expected rents. Destroy the expectation and you slow the machine. This is the play Saudi Arabia ran against the financing of American shale in 2014, though it is worth remembering that shale cut its costs and survived.

Theory two: the value was never going to live in the model layer anyway. Nobody ended up owning rents on electricity. Instead, the century went to whoever industrialised with it fastest, and a US congressional commission has itself concluded that China’s open-model strategy and its manufacturing dominance are mutually reinforcing: cheap intelligence as a free input to the factory floor of the world.

Theory three: the floor controls the value of the lead. A frontier lead is worth exactly its distance from the best free (or cheap enough) alternative. That distance is what you can charge for, and what you can coerce with, and whoever owns the floor can compress the value of a rival’s lead, release by release, without ever catching up. The prize is not victory, but the denial of a decisive victory.

Theory four is domestic: open weights ensure no Chinese lab accumulates the kind of private power OpenAI holds, in a system that has already shown what it does to over-mighty platforms.

And theory five: they do it because they must. American export controls deny Chinese labs the chips to serve a global closed API at scale, and no Western enterprise would trust one anyway. A lab that is behind gives its work away for mindshare and talent, as Meta did with Llama. Trailing players open; leaders close. Within China the same split runs on business lines: the pure-play labs open their best, while ByteDance keeps its frontier models closed and even Baidu opens only its trailing generation.

I won’t pick among them, and I’d politely doubt anyone who picks with confidence. But one historical episode moderates my optimism about all five, because it is where this play has been run to completion. China commoditised solar panels. Prices fell by something like ninety per cent, Western manufacturers died at industrial scale, and the world got cheap clean energy. And then the world discovered that it had traded a price for a dependence, because once the alternatives were dead, the supply chain itself was the leverage. The sequence belongs alongside every theory above: patient commercial flooding first, dominance second, and the discovery, later, of what the dominance could be used for.

Whether the AI version ends differently is a live question, and it comes with a test anyone can watch. If this is doctrine, the best Chinese models keep coming open even as they reach the frontier. The week this essay publishes, Alibaba has promised the open weights of the most capable model it has ever built. As I write, I have no reason to doubt that Alibaba will follow through, but I am interested to see whether these promises continue to be made and honoured, because the pattern of the hosted product shipping first, followed by the weights, has repeated throughout this year for near-frontier Chinese models. Although the test currently still holds, recent revenue-triggered licence terms and paid-subscription-only periods have been more common for Chinese labs. It is possible that “open” is acquiring fine print.

And there is one further oddity: America cannot run the open weights play in reverse. An American lab that gave away its frontier would be destroying its own financing rather than its rival’s. The strategy is only available to a side that never needed the rent.

Which brings me to the frame itself. There is a reading of all this (and it’s a lens, not a finding) in which the AI war is being misread the way the West misread TikTok. The TikTok panic was about how the app was accumulating the habits, locations and appetites of 170 million Americans, and the fear (the very ground on which a unanimous Supreme Court upheld the divest-or-ban law) was that this hoard was a state leverage play in waiting. The ban itself never took hold: a divestiture deal closed this January, and the app plays on. And what the data has observably been used for, so far, is advertising. The underlying point was made years ago (An Xiao Mina and Xiaowei Wang): the Chinese internet was never a town square that might be weaponised; it was built as a great shopping mall. The West projected the town square onto it because that is what the West has built. Perhaps we are making the same projection error at a larger scale: one side playing a race with an endpoint, the other simply trading.

I hold this lens carefully. Firstly, because the mundane reading carries no warranty about the future: data collected to sell things remains data, and a hoard assembled for advertising can still be requisitioned for leverage. The motives are not exclusive, and the shopping mall does not explain away the chip stockpiles or the Taiwan planning. And secondly, because “the arms race is a misreading” happens to be Beijing’s own preferred editorial line, which I don’t want to present as verified truth.

From where I sit, these distinctions have never been philosophical anyway. Australia sells iron ore to China and shelters under a security treaty with the US. Being caught between the merchant and the crusader is not a thought experiment in Australia; it is the national condition. Two decades of it leave you with a deflating suspicion rather than a conclusion: that a merchant is simply what a crusader looks like while the alternatives are still alive.

There is one more place the floor is rising. The same arithmetic that prices a frontier lab’s lead prices your own. Whatever premium you charge for what you know is worth its distance from the best cheap-enough alternative, and that distance is being compressed task by task, and in more places every quarter. I don’t say that to alarm you. The legal profession’s experiment is trending towards migration, not extinction, and the value has gone to the part that cannot be downloaded: judgement (about the quality of work and what is worth doing), the trust of the people you serve, and accountability for the outcome. If the floor under what you know is rising, the move is up.

Nobody knows what China intends with any of this, possibly including China. But the oil story taught me that the demonstration matters more than the intent, because a demonstrated power does not need to be exercised again in order to work. It only needs to be priced in, and it already has been: by every government, every trading floor, and every planning model that now carries the new fact. The same is true of the AI story, with one key amendment. The oil weapon is a valve. Imports switched off can be switched back on, and a price pushed down can be let back up. Open weights have no valve. China could stop the stream of releases tomorrow. But a stopped stream only freezes the floor where it lies. Every model already released holds its floor forever, under everyone: under OpenAI’s margins, under Beijing’s own labs, and under your salary and mine. Whatever the motive turns out to have been, the demonstration has been made. And this one, unlike the one in the Strait, only moves in one direction.

Addendum (14 August 2026): The Qwen3.8-Max weights went live shortly after this essay was locked for release. What was published is a reduced variant of the hosted model (text only, no vision, mandatory thinking mode, and a shorter context window), released under a custom Alibaba licence.