Commoditize the Model Layer

The common framing of Chinese AI competition is that China is trying to build models as good as OpenAI’s and Anthropic’s. That framing is too small, and it leads to the wrong conclusions about who is at risk and why.

The sharper reading is this. China is using open weights and aggressive pricing to commoditize the model layer itself: to weaken the economics of American proprietary models, establish Chinese models as the global default, and shift the returns from AI toward the parts of the economy where China already leads, which are manufacturing, robotics, energy systems and physical deployment.

That is recognisably the playbook from solar panels, batteries, electric vehicles, telecommunications equipment and industrial materials. But AI has a property those industries lacked, and it makes the strategy considerably more efficient.

This piece works through the mechanism, checks the numbers, and then examines the several places where the thesis needs qualifying, including one development from July 2026 that complicates it more than anything else on the table.

The industrial playbook

China’s recurring industrial strategy does not try to maximise the profitability of every Chinese producer. It has repeatedly tolerated too many domestic competitors, thin or negative margins, state-directed financing, excess capacity, brutal price competition, and consolidation and failure among Chinese firms.

The objective sits above the individual firm: drive down global prices, gain scale, take supply-chain position, weaken foreign incumbents, and make other countries dependent on Chinese production.

The scale of that willingness is documented. Bloomberg’s analysis put support for China’s electric vehicle industry at at least 231 billion dollars between 2009 and 2023. CATL, holding roughly 43 percent of the Chinese battery market and 37 percent globally in 2023, saw its government subsidies rise from about 77 million dollars in 2018 to about 809 million in 2023. A 2023 tax incentive package for electric and environmentally friendly vehicles was worth 520 billion renminbi over four years.

Those industries ended with Chinese firms holding commanding global positions and foreign incumbents unable to match the cost base. Several Chinese participants in each sector did not survive the process. That was an acceptable outcome, because firm-level profitability was never the target.

What makes AI different, and more efficient

With solar panels or electric vehicles, every additional unit of global market share has to be manufactured, financed, warehoused, shipped and pushed past tariffs. Distribution is expensive, and it stays expensive forever.

With an open-weight model, the sequence is different. Train once, at whatever cost. Release the weights. Then:

China does not have to subsidise a single downstream unit. The rest of the world distributes, improves and monetises the strategic asset voluntarily, because doing so is individually rational for every participant. Marginal distribution cost is approximately zero, and the distributors are foreign.

That is a materially better deal than subsidised solar panels, and it explains why this looks different from ordinary foreign competition.

The two loops, with numbers

The U.S.-China Economic and Security Review Commission published a report in March 2026 titled “Two Loops: How China’s Open AI Strategy Reinforces Its Industrial Dominance”. It describes exactly this mechanism: permissive licensing and pricing that undercuts American rivals, producing adoption; adoption producing derivatives, expertise and deployment data; that feeding back into the ecosystem and driving further adoption.

The measured adoption is not marginal.

Alibaba’s Qwen family constitutes the largest model ecosystem on Hugging Face, with more than 100,000 derivatives. By March 2026 it had passed one billion cumulative downloads, faster than any model family in history, overtaking Meta’s Llama as the platform’s most-downloaded open model. A partner at Andreessen Horowitz estimated that roughly 80 percent of American AI startups use Chinese base models to build their applications.

That last figure deserves a pause. If it is even approximately right, the American startup ecosystem is already substantially built on Chinese foundations, and the commoditization is not a forecast. It has happened at the base-model layer and is now working upward.

The report’s other point is the one most often missed: this ecosystem lets China innovate close to the frontier despite real compute constraints. Openness is partly a workaround for export controls. Distribution compensates for silicon.

How this attacks the proprietary model business

It removes the price umbrella

A proprietary lab can charge a large premium while the alternatives are clearly inferior. Twenty percent better on a demanding task can justify a substantial multiple on price, because there is nothing else that does the job.

That holds until an open model is 90 to 98 percent as capable for ordinary work, customisable, locally deployable, available from several hosting providers, and much cheaper. At that point the premium survives only where the marginal capability actually matters.

Customers do the obvious thing: route the hardest work to the frontier model and everything else to Qwen, Kimi, DeepSeek, GLM or a derivative. The frontier lab keeps the most valuable workloads and loses the enormous volume of ordinary inference underneath them.

This does not eliminate revenue. It compresses revenue per token, gross margin and, most importantly, bargaining power.

It attacks terminal value rather than current revenue

This is the part that makes the “but revenues are growing” rebuttal beside the point.

A company can grow revenue very quickly while its long-run valuation assumptions deteriorate, because those assumptions are about durable pricing power rather than this quarter. The question that matters is not whether revenue is growing. It is whether raw model intelligence will still command scarcity pricing once open models are good enough for most commercial work.

Chinese models do not need to capture most current revenue to damage that. They only need to persuade the market that model intelligence will end up resembling cloud compute, bandwidth or commodity infrastructure rather than a scarce proprietary product.

The capital intensity makes the timing dangerous

Here the verified numbers matter, in both directions.

Claims that Anthropic is financially collapsing are simply wrong. It is on track for its first profitable quarter, with second-quarter 2026 revenue of about 10.9 billion dollars, more than double the 4.8 billion of the first quarter, and expected operating profit of 559 million dollars for the quarter ending in June. Demand for premium capability is real and is being paid for.

But the cost structure behind that is extraordinary. SpaceX’s IPO filing disclosed that Anthropic pays 1.25 billion dollars per month until May 2029 for compute access across the Colossus and Colossus II clusters, covering more than 220,000 Nvidia GPUs and over 300 megawatts of capacity. That is roughly 15 billion dollars a year of contracted spend. Some analysts argue the reported operating profit is flattered by a ramp-up discount on that very deal, which is worth holding in mind before treating the profit as a structural turn.

OpenAI is more exposed still. It raised planned compute spending to 750 billion dollars through 2030, up from a 600 billion target set in February. The commitments are binding contracts across Microsoft, Oracle, Amazon and CoreWeave, including a reported 300 billion dollar Oracle agreement starting in 2027, plus 20 billion for a Georgia data centre. Against 2025 revenue of 13.07 billion dollars and a reported net loss of 38.5 billion, the commitment-to-revenue ratio is roughly 57 to 1. OpenAI’s own chief financial officer has reportedly raised concerns internally about the company’s ability to honour those contracts if revenue growth does not keep pace.

Now put the two halves together, because the combination is the actual risk:

Chinese open models push the price of intelligence downward at precisely the moment American frontier labs are locking themselves into fixed capacity obligations measured in hundreds of billions.

Falling unit prices are survivable with a variable cost base. They are considerably less survivable when the cost base is contracted years ahead. The strategy does not need to beat the frontier labs on capability. It needs to move the price curve while their obligations are inflexible.

Every frontier breakthrough becomes a teacher

There is an asymmetry in how capability propagates.

An American lab spends billions reaching a new capability frontier. The model is then exposed through products and APIs. Competitors study its behaviour. Chinese labs release a cheaper or open approximation through legitimate research, benchmarking, synthetic data generation, or distillation. The economic life of the American premium shortens. The American lab must spend billions again to reach the next frontier.

This should not be overstated. Distillation is not alchemy, and Chinese labs still need researchers, compute, pretraining and serious engineering. But it plausibly shortens the expensive final step between a strong model and a near-frontier one, and that step is where much of the money goes.

The consequence is not that American labs fall behind. It is that they stay ahead while getting less and less time to monetise each lead.

It also suppresses American open-weight competitors

This effect gets less attention and may matter more.

An American open-weight startup must raise hundreds of millions, pay market rates for compute, and eventually show a path to profit. A Chinese lab backed directly or indirectly by a large corporation, local governments, state procurement or strategic capital can release weights and accept poor near-term returns for years.

A developer choosing a foundation asks a simple question: why adopt the merely respectable American open model when a stronger Chinese one is already free? The a16z estimate above suggests most have already answered it.

The perverse outcome is a market where American proprietary models stay expensive, American open-model companies struggle to finance themselves, and Chinese open weights become the default substrate of the global open ecosystem. That is strategically worse than losing a fair fight, because it removes the domestic alternative rather than merely beating it.

Commoditize your complement

The strategy only makes sense once you see where the returns are meant to land.

In July 2026 Xi Jinping used the World AI Conference in Shanghai to tie open-source AI explicitly to national strategy. Reuters characterised the speech as casting Beijing as champion of a new global AI order, urging countries to seize the “historic opportunity” of open-source AI, pledging to help developing nations build capability, and warning against “new historical injustices” arising from unequal access. The day before, 29 countries signed on to establish the World Artificial Intelligence Cooperation Organization, headquartered in Shanghai.

The connection to the physical economy is the point. China does not need large margins on API tokens if it captures the value through robotics, autonomous factories, industrial inspection, logistics, electric vehicles, drones, grid management, materials processing, pharmaceutical and chemical manufacturing, domestic semiconductor development, and military and surveillance systems.

That is the classic complement strategy, stated plainly:

The two positions are structurally opposed, and no amount of good behaviour by either side resolves it. A model that is unprofitable as a standalone product can be extremely valuable as an input to an industrial base, and the USCC report makes the reinforcing version of this argument: low-cost deployment across factories, logistics and robotics generates specialised real-world data that feeds further model improvement.

The standards layer

There is a second prize that does not show up as revenue at all.

Once institutions build on Qwen, Kimi or DeepSeek derivatives, Chinese architectures become familiar, Chinese tooling becomes embedded, Chinese standards gain legitimacy, local engineers accumulate expertise in Chinese ecosystems, China earns diplomatic credit for supplying affordable technology, and American sanctions lose force because alternatives already exist.

WAICO, headquartered in Shanghai with 29 founding signatories, is an institutional expression of exactly that. The loss here is geopolitical and can occur even when no money reaches a Chinese model company at all.

Where the thesis needs qualifying

A thesis this strong attracts overstatement. Four qualifications, the last of which is the most important and the most recent.

Frontier labs are not solar manufacturers

A solar panel is an interchangeable physical unit. A frontier lab is not, and it has real defensive surfaces: superior capability at the top end, consumer distribution, enterprise contracts, agents and complete workflow systems, memory and connectors, reliability and support, specialised models, proprietary data loops, moving up into applications, and keeping some capabilities internal.

Anthropic’s revenue trajectory and approach to operating profitability show that demand for premium capability is not theoretical. The likely outcome is not that OpenAI and Anthropic disappear. It is that they become vertically integrated AI service companies rather than firms collecting extraordinary margins indefinitely on access to raw intelligence. Which, incidentally, is exactly the vertical move that gives their own enterprise customers a reason to keep them at arm’s length.

Chinese labs can bleed too

China’s industrial strategy has produced global champions, and it has also produced wasteful investment, domestic price wars, weak profitability, bad debt, duplicative capacity, bailouts and trade retaliation.

An ecosystem of dozens of labs training expensive models and releasing them at close to zero price is not automatically sustainable. The strategy may transfer enormous value to manufacturers and users while destroying it for Chinese model developers. The state may regard that as an acceptable trade. Private investors in those labs may not, and the difference between those two views is where the strategy could fracture.

Open weights help America too

American companies can take Chinese weights, improve them, strip out unwanted components and run them in American data centres. China cannot fully control downstream value once weights are released. Open weights also lower costs for American startups and enterprises, which is most of the American AI economy by count.

This has a direct policy implication. A blanket restriction on open models would protect two American companies at the expense of thousands of American builders. Suppressing open weights is defending the incumbents, not the country.

And now the complication: China may be closing the window

This is the development that most complicates the thesis, and it is only weeks old.

Reuters reported on 7 July 2026 that Chinese authorities had spent the previous month meeting Alibaba, ByteDance and Z.ai about limiting overseas access to China’s most advanced AI models, including models not yet released, in talks led by the Ministry of Commerce. Two questions sit at the centre: whether training data may be transferred abroad, and whether foreign users should continue to freely download the weights of Chinese models. Access as a service, through APIs and cloud, would remain. Qwen, Doubao and GLM-5.2 are among those named.

Nothing has been decided. Scope, legal mechanism and timing are all unsettled, and any restriction might apply only to future models. But the direction of the conversation matters in three ways.

It confirms the strategic reading while undermining the permanence one. You do not convene the Ministry of Commerce about an artefact you consider commercially incidental. The willingness to contemplate controls is evidence that open weights are understood in Beijing as strategic leverage. That raises confidence that the openness was instrumental, and lowers confidence that it is permanent.

It converts the open weights from a free good into a dependency. Every organisation that built on Qwen or GLM did so on the assumption that the next version would arrive on the same terms. If weights stop flowing and only API access remains, those organisations are left with a frozen base model, an ecosystem oriented around it, and a service relationship with a provider in another jurisdiction.

And it makes the symmetry unavoidable. The argument for keeping your memory, orchestration and feedback loop out of an American frontier lab’s hands is that the provider can change access rules, retention, pricing and availability unilaterally, and might become a competitor. Every word of that applies to a Chinese open-weight provider that is now discussing export controls with its government, with the addition of a second sovereign able to change the terms.

The lesson is not that Chinese models are more dangerous than American ones. It is that provider dependency is provider-agnostic. An architecture that treats any single model supplier as replaceable survives both an American lab moving up into your vertical and a Chinese ministry restricting downloads. An architecture that does not, fails on whichever arrives first.

Probability assessment

These are judgements rather than findings, and worth stating as such. The right-hand column notes how the July export-control reporting bears on each.

Proposition Estimate Effect of the July reporting
China treats open AI as a national strategic instrument 90% Raises it. Contemplating export controls is itself evidence of strategic framing.
Weakening US proprietary pricing power is an objective 80% Neutral.
Chinese open models materially compress frontier margins within five years 70% Slight reduction. Restrictions would slow diffusion at the margin.
Raw foundation-model access becomes substantially commoditized 75% Neutral to slightly lower. Commoditization is already well advanced.
OpenAI or Anthropic must materially change business model 70% Neutral. Vertical integration is already under way.
Chinese competition alone causes either to fail financially 20% Neutral. Capital commitments are the larger risk than competition.
China captures more value through physical deployment than model API profit 75% Raises it. Restricting weights while keeping APIs open concentrates value at home.

The single largest uncertainty is not any of these. It is whether the fixed compute obligations described earlier come due in a period of falling prices for the thing they produce.

What this means if you are building on top

The strategic contest is above most readers’ heads in the sense that they cannot influence it. What they can decide is how exposed they are to whichever way it resolves.

The requirements are the same ones that fall out of the vertical-encroachment argument about American labs, which is a useful sign that they are the right requirements rather than a reaction to one news cycle:

Bottom line

The thesis is substantially right, and phrasing it as China trying to bankrupt profitable American companies makes it weaker than it is.

The deeper strategy is that China is willing to sacrifice profitability at the model layer in order to make intelligence globally abundant, reduce the value of American proprietary AI, establish Chinese technology as the open global default, and capture the resulting gains in industry, robotics, standards and geopolitical influence. With open weights, the productive asset can be released into the world and foreign developers, clouds and enterprises will propagate it at their own expense. That is a considerably more sophisticated instrument than subsidised solar panels.

OpenAI and Anthropic are unlikely to disappear. Their worst case is no longer that China builds a better chatbot. It is that frontier-quality intelligence becomes abundant enough that neither can earn returns commensurate with the hundreds of billions they have already contracted to spend staying ahead.

And the July export-control reporting adds the final turn. The open-weight flood may not be permanent, because the same strategic logic that opened the tap can close it. Which means the correct response for anyone building on any of this is not to pick the winning provider. It is to build so that the question does not need answering.