How Is China's AI So Cheap? Inside the Price War

Is Chinese AI cheap because the state is paying for it? No. Subsidies exist, and they explain almost none of the gap. DeepSeek's April V4 release made the counterexample unusually concrete: V4 Flash's official API price is $0.22 per million cache-miss input tokens and $0.66 per million output tokens off-peak, doubling to $0.44 and $1.32 in peak hours, while the downloadable weights carry an MIT licence. You don't get there with cheap money alone. You have to build well, activate a small share of a large model and serve it on a cheaper stack. The cheapness is real, it's mostly structural, and it comes from five things stacking. Only one of the five is temporary.

Key Takeaways- Chinese AI is cheap for five stacking reasons, not just subsidies: open weights, genuine efficiency, a price war, cheap energy, and deep talent.- Open weights remove API margin entirely. You self-host, so there's no per-token rent to a vendor.- Cheap does not mean small or old. DeepSeek V4 Flash is a 284-billion-parameter mixture-of-experts model with 13 billion active parameters and a 1-million-token context window.- Not every Chinese frontier model is cheap in the same way. Kimi K3's official cache-miss API price is $3 input and $15 output per million tokens, so architecture, deployment and commercial positioning all matter.- Cheap energy is the most structural driver: China's monthly solar generation hit a record 96 TWh in April 2025, comparable to a full month of US nuclear output.- For buyers, "good-enough at a fraction of the cost" rewrites the build-versus-buy maths for most workloads.

The driver everyone fixates on is the price war, and of the five it's the one I find least interesting. I write this from Hangzhou, where the labs behind DeepSeek and Qwen both sit, and the talk here is rarely about who undercut whom this month. It's about what the next architecture buys you. If you want the wider argument, this sits under my pillar on whether China is winning the AI race. The narrower question is the practical one: why does it cost so little, and how much of that holds once the noise dies down?

Driver 1: open weights kill the API margin

The biggest reason is also the dullest. China's leading models are open-weight, so there's no API margin to pay in the first place. By 21 January 2026, Qwen, Alibaba's model line, had passed 1 billion cumulative downloads and more than 200,000 derivative models on Hugging Face, by Alibaba's own count. A few months earlier, MIT Technology Review had it at over 600 million downloads and more than 40% of all new Hugging Face derivatives. You download the weights and run them yourself.

It's worth being precise about what a closed API charges you for. Compute, obviously. But also the provider's margin, the cost of recovering their research spend, and a risk premium sitting quietly on top of both. Open weights collapse all of that into one line on your bill: the raw cost of the hardware you run the model on.

You don't really feel this until you watch your own invoice move. When I shift a workload off a hosted frontier API onto an open Chinese model on my own GPUs, I'm not negotiating a better rate, I'm getting rid of the rate. The cost per token doesn't just shrink, it changes category: you're paying for electricity now instead of rent.

Open weights also compound downward. Chinese open-weight models overtook US ones in their share of global Hugging Face downloads during 2025, and Hugging Face's own Spring 2026 report has them at 41% of all platform downloads. Every derivative and quantisation, every community tutorial written at two in the morning by someone who just wanted the thing to run, makes the next deployment cheaper. A closed API gives you none of that. You just keep paying the rent.

The leading Chinese models ship as open weights, removing API margin from the cost entirely. By January 2026 Qwen alone had surpassed 1 billion downloads and over 200,000 derivatives, and Chinese open-weight models have overtaken US models in their share of global Hugging Face downloads, reaching 41% of all platform downloads by spring 2026. Self-hosting replaces a per-token vendor rate with raw hardware cost.

If you want the head-to-head, I compare the main options in Qwen vs DeepSeek vs Llama.

Driver 2: real efficiency, not just subsidies

This is where the subsidy story falls apart. A large share of the cost advantage is genuine engineering. V4 Flash has 284 billion total parameters but activates only 13 billion for each token. V4 Pro is much larger at 1.6 trillion total and 49 billion active. The March 2026 snapshot in Stanford HAI's AI Index already showed the leading US and Chinese models close on one public leaderboard, but it predates V4, Qwen3.8, Kimi K3, GLM-5.2 and MiniMax M3. It's trend evidence, not a current price-adjusted ranking.

The clearest lever is architecture. Mixture-of-experts designs route each token through only a fraction of the model's parameters, so you get the capability of a huge model at the inference cost of a much smaller one. Several leading Chinese models lean hard on this. It's a cut to cost-per-token baked into the model rather than bolted on afterwards.

Then there's the DeepSeek story, which is where the phrase "the DeepSeek moment" comes from. I'd treat the headline training-cost figures as reported rather than audited. The exact numbers are contested, and anyone quoting them to three significant figures is guessing. The direction isn't in doubt, though: competitive models built on a fraction of the compute budget everyone assumed was the price of entry.

The efficiency angle matters more than the subsidy debate for one simple reason: efficiency travels and subsidy doesn't. A subsidy props up a price in one market for as long as the money flows. An architectural win, fewer active parameters per token or a better training recipe, lowers the true cost floor everywhere, for everyone who copies it. That's why the open-weight release of these efficient models is the part that counts. It exports the cost advantage, not just the product.

The adoption data says it's working. Chinese models overtook US ones inside OpenRouter's top 10 in late February 2026, taking roughly 61% of top-10 token consumption, and reached a full-platform majority around April 2026 on OpenRouter's rankings. Tokens are real work rather than press releases, and developers moved theirs.

A large part of the cost advantage is genuine efficiency, not subsidy. DeepSeek V4 Flash activates 13 billion of its 284 billion parameters per token, and its official cache-miss API price is $0.22 per million input tokens and $0.66 per million output tokens off-peak, doubling in peak hours, with the architecture on the model card. Mixture-of-experts design lowers the compute used for each token even when the full model is large.

Driver 3: a domestic price war

Some of the cheapness isn't built into the models at all. It's hammered in by competition. The domestic market is a brawl: Alibaba, DeepSeek and several others, all well funded, all chasing the same developers. That fight drags API and hosting costs down across the board. Plenty of the cheapness is simply rivals undercutting rivals, which is the part the subsidy story never accounts for.

The mechanism is plain enough. When everyone offers a comparable open-weight model, the only levers left are price and convenience. Nobody can charge a premium for capability the lab next door matches and then gives away for nothing. So the floor keeps dropping, and the saving reaches anyone hosting these models, inside China or out.

The open-weight releases are partly a weapon in that fight. Putting strong weights out for free wins mindshare, seeds an ecosystem, and denies a rival the chance to monetise. Qwen's 200,000-plus derivatives are partly the fruit of exactly that strategy, and nobody involved in it would call it charity.

This is the driver I'd trust least, and buyers should plan for it to fade. Price wars don't last. Some labs consolidate, some fail, and the survivors will eventually try to claw margin back. What holds the floor down is that they can't un-release open weights. The models already in the wild stay cheap to run whatever the market does next.

China's domestic AI market is intensely competitive, and that price war pushes hosting and API costs down for everyone, not just buyers inside China. Free open-weight releases are partly a competitive weapon, helping build ecosystems like Qwen's 200,000-plus derivatives. The price war is temporary; the open weights it produces are permanent.

I compare the survivors worth building on in Qwen vs DeepSeek vs Llama.

Driver 4: cheap energy and manufacturing scale

Underneath the software sits physics, and this is the most structural driver of the lot. A served token is, in the end, a small quantity of electricity. China's monthly solar generation hit a record 96 TWh in April 2025, comparable to the roughly 65 TWh a month the US gets from nuclear, about 775 TWh a year. Cheaper, more abundant power means a lower floor under both training and inference.

This is the input no model release can match. If your power is cheaper and your grid is growing faster, every layer of the stack gets cheaper at once: the training runs, the inference serving, the buildings the racks sit in. It's a quiet, compounding advantage rather than a headline one, which is probably why it gets discussed least and matters most.

Manufacturing scale stacks on top. The industrial base that makes hardware cheap also makes data-centre build-out cheap. China isn't only running models on cheap power, it's building the racks, the cooling and the facilities at domestic-supply-chain cost. That's a capital-expenditure edge, and it flows straight through to the price of a token.

Of the five, this is the one I'd still expect to be true in five years. You can copy an architecture in a quarter and end a price war in a year. Doubling national generation takes a lot longer, and you can't conjure a domestic manufacturing base at all. If I were building a long-term plan around any single driver here, it would be this one, because it's the hardest for anyone to reverse.

The most structural cost driver is energy. China's monthly solar output hit a record 96 TWh in April 2025, comparable to a full month of US nuclear generation (about 65 TWh). Cheaper power lowers the cost floor for training and inference together, and manufacturing scale makes data-centre build-out cheaper too. This advantage is the hardest to reverse.

I go deeper on the grid story in China's AI energy advantage.

Driver 5: a deep talent pool

The last driver is people, and the supply numbers are striking: China educates 38% of the world's top AI researchers, measured by NeurIPS 2024 authors, up from 27% in 2017. A deep, affordable pool of elite engineers is precisely what you need to chase the efficiency gains that make models cheap to run.

Efficiency work is labour-intensive in a specific way. Squeezing cost out of a model, whether through better architectures or smarter training or aggressive optimisation, takes a lot of skilled researchers iterating fast. A country with a deep domestic supply can put more capable people on the problem, at lower cost, than one importing scarce specialists. The talent pool feeds Driver 2 directly.

When I read the technical reports behind these models, the thing that stands out isn't a single breakthrough. It's the sheer volume of careful, incremental engineering. Dozens of small optimisations that only ever get done when you have enough hands to do them. That's a talent-depth story, and it's a large part of why the efficiency keeps improving release on release.

One caveat I won't skip: a big share of China-educated elite researchers have historically worked abroad, so this isn't a clean home-team advantage. But the trend is upward, 27% to 38% in seven years, and a growing roster of world-class labs back home gives that talent somewhere to land.

China's deep talent pool feeds the efficiency engine. It educates 38% of the world's top AI researchers, up from 27% in 2017. Cost-cutting model work is labour-intensive, so a large, capable engineering pool is precisely what lets Chinese labs keep lowering the real cost per token, release after release.

Illustrative, directional assessment, not precise measured values. Positions reflect the reported pattern: Chinese open-weight models cluster in the high-capability, low-cost quadrant, with the top closed US frontier still highest on capability but at higher cost. Capability gap context: Stanford HAI AI Index 2026.

What this means for your buying decision

Here's where it stops being trivia and starts changing decisions. When a model with a 1-million-token context window is offered at DeepSeek V4 Flash prices, the build-versus-buy maths flips for many routine workloads. You're no longer paying a large premium by default. You're deciding whether a more expensive model produces enough better work on your own evaluation to justify it.

Map it to the actual work. Summarisation, extraction, classification, drafting, routing: none of these need the single best model on Earth. They need one that's dependable, cheap enough to stop thinking about, and running somewhere you control. Open-weight Chinese models hit that target, and because you self-host them, the recurring cost is hardware rather than a per-token meter that scales up with your own success.

In my own stack, the cheapness has quietly changed which projects are even worth attempting. Workloads I'd have shelved as too expensive to run at scale on a frontier API are now trivially affordable on an open model I host myself. Cheap doesn't only save money on the work you're already doing. It makes new work viable, and that second-order effect is the one I'd plan around.

None of which is a reason to skip the diligence. Provenance, data governance, censorship behaviour, licensing: all of it needs checking before anything touches production. Cheap is a reason to evaluate something properly, not to wave it through. What you want is a structured decision. Which workload, what quality it tolerates, and what the risk assessment says about the specific model you're pointing at.

When capable open models cost a fraction of many frontier APIs, build-versus-buy flips for routine workloads. DeepSeek V4 Flash combines a 1-million-token context window with official prices of $0.22 per million cache-miss input tokens and $0.66 per million output tokens off-peak, doubling in peak hours. Self-hosting replaces the vendor meter with hardware and operational cost, after a proper risk assessment.

For the full framework, see Is DeepSeek safe for enterprise?, then compare the main options in Qwen vs DeepSeek vs Llama.

FAQ

How is China so cheap at AI, in one sentence? Open weights remove the API margin, efficient architectures cut the cost per token, a price war drives prices down, cheap energy lowers the floor, and a deep talent pool keeps it all improving. Subsidies are a minor factor, not the main one.

Why is DeepSeek so cheap? Two reasons stack. It ships downloadable MIT-licensed weights, so there is no mandatory API margin if you self-host, and its mixture-of-experts models activate only a fraction of their total parameters per token. V4 Flash activates 13 billion of 284 billion parameters; V4 Pro activates 49 billion of 1.6 trillion, per the official model cards. Treat historical training-cost headlines as reported rather than audited, but the current inference architecture and published API prices are concrete.

Are Chinese AI models cheaper than OpenAI? Some are, often substantially, but the family name is not enough. DeepSeek V4 Flash is extremely cheap on its official API, while Kimi K3 is priced much higher. Self-hosting removes a per-token vendor price but replaces it with hardware, energy and operations. Compare total cost on the exact workload rather than quoting a universal multiple.

Is cheap Chinese AI lower quality? Price alone cannot tell you. The leading Chinese families now include huge long-context and multimodal models, but their strengths differ and provider benchmark tables are not neutral comparisons. DeepSeek V4 Flash is exceptionally cheap; Kimi K3 is materially dearer at $3 per million cache-miss input tokens and $15 per million output tokens. Test the exact model on your workload.

Will Chinese AI stay this cheap? Mostly, yes, for structural reasons. The price war is temporary and may ease, but open weights, efficient architectures and cheap energy aren't easily reversed. Once a strong model is released openly, it stays cheap to run forever, whatever the market does next. I'd plan on cheap continuing.

The bottom line

China's AI is cheap for five reasons that stack, and "it's all subsidies" explains almost none of them. Open weights delete the API margin. Genuine efficiency, meaning mixture-of-experts and the DeepSeek story, lowers the true cost per token. A domestic price war drags prices down further. Cheap energy and manufacturing scale set a low floor. A deep talent pool keeps the efficiency improving. The price war is the part that fades. The rest is structural.

The current releases also kill the lazy version of this argument. "Chinese AI is cheap" is not one price. DeepSeek V4 Flash is priced to be infrastructure. Kimi K3 is priced like a premium agentic model. Qwen3.8's open weights landed in August under a custom licence, while the hosted Qwen3.8-Max is tied to a token plan. Compare the exact checkpoint, licence, serving route and measured task quality, not the flag on the lab.

For your business, the takeaway is plain. Some of the cheapest capable models are Chinese and downloadable, while other Chinese frontier models are premium services or enormous self-hosting projects. So I wouldn't default to them blindly, but the exact checkpoint deserves a proper look, because the build-versus-buy maths really has changed.

If you want the strategic context, start with whether China is winning the AI race. When you're ready to act, Is DeepSeek safe for enterprise? is where this hands off into the real decision.