Skip to content
Capital & Compute
· ai· china· economics· open-source· pricing

Why Are Chinese AI Models So Cheap? The Real Economics

Chinese AI models list output tokens up to 57x below US flagships. The verified economics of efficient training, cheap power and open weights as strategy.

By Capital & Compute

Chinese AI models are cheap for three compounding reasons. They are cheaper to build: sparse mixture-of-experts architectures, FP8 training and curated data cut compute per unit of capability. They are cheaper to run: efficient inference plus deliberately cheap power in western data-center hubs. And they are cheap on purpose: an open-weight strategy that prices for adoption, not margin. On verified list rates, the gap to US flagships runs from 4x to 57x.

That last sentence is doing a lot of work, so here is the evidence behind it, and the strategy behind the evidence.

57x
Claude Fable 5 vs DeepSeek V4
output list price per Mtok
$5.576M
DeepSeek V3 training run
reported GPU cost, with caveats
370 GW
solar China added in 2025
about two thirds of global additions

How much cheaper are they, really?

Headline claims get sloppy here. A widely shared UBS analysis, reported by Business Standard in July 2026, put Chinese models at up to 50 times cheaper per token. Skeptics countered that like-for-like tiers are closer than that. Both are right, and the confusion dissolves the moment you compare specific models instead of national averages.

These are current list rates per million tokens, verified between June 20 and July 1, 2026 against each provider’s own price page (GLM-5.2 via live API resellers, pending an official Z.ai rate card). It is the same registry that powers our AI model price tracker:

Model Input $/Mtok Output $/Mtok vs DeepSeek output
DeepSeek V4 $0.435 $0.87 1x
Kimi K2.7 Code (Moonshot) $0.95 $4.00 4.6x
GLM-5.2 (Zhipu) $1.40 $4.40 5.1x
Qwen3.7 Max (Alibaba) $2.50 $7.50 8.6x
Gemini 3.1 Pro (Google) $2.00 $12.00 13.8x
Claude Sonnet 5 (Anthropic) $3.00 $15.00 17.2x
GPT-5.5 (OpenAI) $5.00 $30.00 34.5x
Claude Fable 5 (Anthropic) $10.00 $50.00 57.5x

So the honest answer has two halves. Compare the cheapest Chinese frontier model to the most expensive US flagship and the 50x headline is real: Claude Fable 5 lists output at 57 times DeepSeek V4. Compare peers instead, say Claude Sonnet 5 against GLM-5.2 or Kimi K2.7 on serious coding work, and the gap is 3x to 4x. Still enormous. A 4x input cost difference decides architectures, vendor contracts and whole product lines. It just is not 50x.

Output token prices: Chinese models vs US flagshipsA log-scale dot plot of output price per million tokens. DeepSeek V4 at $0.87 is the baseline. Kimi K2.7 Code $4.00, GLM-5.2 $4.40, Qwen3.7 Max $7.50, Gemini 3.1 Pro $12, Claude Sonnet 5 $15, GPT-5.5 $30, Claude Fable 5 $50.$1.00$2.00$5.00$10.00$20.00$50.00output price per million tokens, USD (log scale)DeepSeek V4DeepSeek V4$0.87 DeepSeek V4Kimi K2.7 Code$4.00 ×5GLM-5.2$4.40 ×5Qwen3.7 Max$7.50 ×9Gemini 3.1 Pro$12.00 ×14Claude Sonnet 5$15.00 ×17GPT-5.5$30.00 ×34Claude Fable 5$50.00 ×57
Output token prices: Chinese models vs US flagships
ToolCost per taskMultiple of baseline
DeepSeek V4$0.871.0x
Kimi K2.7 Code$4.004.6x
GLM-5.2$4.405.1x
Qwen3.7 Max$7.508.6x
Gemini 3.1 Pro$12.0013.8x
Claude Sonnet 5$15.0017.2x
GPT-5.5$30.0034.5x
Claude Fable 5$50.0057.5x
Output list price per million tokens on a log scale, July 2026. Every row is annotated with its multiple of DeepSeek V4. The four Chinese models cluster below $8 while US flagships run $12 to $50.Source: Provider price pages, verified June 20 to July 1, 2026; full registry on the model price tracker

And the sticker rates undersell it. DeepSeek bills cache hits at $0.0036 per million tokens, a roughly 99 percent discount that makes repeated context nearly free; Anthropic’s cache reads on Fable 5 cost about 278 times that. Alibaba is running a 50 percent promotion on Qwen3.7 Max until July 23, 2026. For agent workloads that resend large prefixes on every turn, the effective gap is wider than the table shows.

Reason 1: They are cheaper to build

The clearest public evidence is the DeepSeek-V3 Technical Report, a December 2024 arXiv preprint from the DeepSeek team. The model has 671 billion total parameters but activates only 37 billion per token. That sparse mixture-of-experts design means each token pays for about 5 percent of the network. The team trained in FP8 mixed precision, which halves memory against the BF16 that most Western labs still used at the time, on 14.8 trillion curated tokens.

The report puts the full training run at 2.788 million H800 GPU-hours, or $5.576 million at an assumed $2 per GPU-hour rental. Quote that number carefully. The paper itself says it covers the final training run only, excluding all prior research and ablation experiments, and it is the lab’s own accounting in a preprint, not an audited figure. The real all-in number is some multiple of it. But even at ten times the sticker, it is nowhere near the hundreds of millions US frontier runs are reported to cost.

Necessity did the teaching here. US export controls cut Chinese labs off from top-tier accelerators, so the engineering culture optimized for performance per chip-hour instead of scale at any price. The result, sparse models that are cheap to serve, is now a structural cost advantage rather than a workaround. The same constraint is pushing training onto domestic silicon, a shift we covered in China’s AI chip companies in 2026.

Reason 2: They are cheaper to run

Here is where most explainers get lazy, so let this one be precise.

The targeted part is a policy called Eastern Data, Western Computing: route AI workloads to hub zones in Guizhou, Ningxia, Inner Mongolia and Gansu, where hydro, wind and solar are abundant and demand is thin. A February 2026 comment from the Oxford Institute for Energy Studies, The China data centre advantage, documents western distributed solar settling as low as 0.19 yuan per kilowatt-hour (about 2.7 US cents) against up to 0.43 yuan in eastern regions. Inference is an electricity business. Serve it from the cheap zones and your marginal cost per token drops with it.

The trajectory matters more than the snapshot. Per the IEA’s Renewables 2025 report, China commissioned roughly 370 GW of solar and 117 GW of wind in 2025 alone, about two thirds of everything added on Earth. Ember’s Global Electricity Review 2026 recorded renewables passing coal in the global generation mix for the first time in a century, with China supplying most of the growth. Meanwhile US data-center buildouts are bidding up power in constrained grids. One side of this competition is adding generation faster than it adds demand. The other is not.

Read that as strategy, not luck. A country that dominates solar manufacturing, deploys generation at that pace and prices hub-zone power below anyone’s marginal cost is building the energy layer of AI the way OPEC once ran oil: control the cheapest supply and let everyone else set their cost structure against yours. That is analysis, not a sourced fact, but the capacity numbers above are what it rests on.

Stack the rest on top: engineer salaries a fraction of Bay Area packages, subsidized land and tax treatment in the hub zones, and state-backed compute clusters. None of these show up on a price page. All of them show up in the price.

Reason 3: They are cheap on purpose

Every major Chinese frontier lab ships open weights: DeepSeek under MIT-style licensing, Qwen, Kimi and GLM under permissive terms of their own. US labs treat weights as the crown jewels. Chinese labs give them away and charge less for hosted inference too. That only looks irrational until you see what it buys.

It buys distribution. Hugging Face’s own State of Open Source report from March 2026 counts more than 113,000 derivative models built on Qwen checkpoints and puts Chinese models at 41 percent of all platform downloads in 2025, the year China passed the US in monthly downloads. Every fine-tune and startup built on those weights is switching cost accruing to the ecosystem. We traced the adoption flip itself in how open-source LLMs overtook proprietary in 2026.

It also buys policy alignment. China’s Global AI Governance Action Plan, released at the World AI Conference in July 2025, explicitly promotes an open and inclusive AI ecosystem as a pillar of its international pitch. The State Council’s AI Plus opinions from August 2025 (available in translation from Georgetown’s CSET) make a flourishing open-source ecosystem a named national objective. Open weights are not a lab quirk. They are industrial policy.

And it buys a lever on competitors’ margins. The classic move is commoditizing your complement: if you cannot yet beat the leader’s best model, make model access itself cheap and abundant, and the leader’s pricing power evaporates. US closed labs fund research from inference margin. Every open Chinese release marks that margin down.

Why China specifically, and not anyone else

Each reason above is copyable in isolation. A French lab can open-source a model. Texas can build solar. What no other player holds is every layer of the stack at once, with each layer feeding the next.

Start at the bottom. China does not just deploy two thirds of the world’s new renewable generation; it builds the machinery of the buildout. The IEA’s Solar PV Global Supply Chains analysis puts China’s share above 80 percent at every manufacturing stage, from polysilicon to finished modules, and near 95 percent for wafers. Cheap power is usually something a country is lucky enough to sit on. China is the only country that mass-produces it. Which is why the OPEC comparison, if anything, undersells the position: OPEC controlled a resource in the ground, while China controls the factories that turn capital into new energy supply anywhere on Earth, including in the countries it sells panels to.

One layer up sits the silicon. Export controls locked Chinese labs out of NVIDIA’s best parts, and the response was a domestic accelerator push co-designed with the open models: DeepSeek ships FP8 formats positioned for domestic chips, and its V4 is reported by Fortune to integrate Huawei’s Ascend hardware, a shift we mapped in China’s AI chip companies in 2026. Notice how the layers connect. A model that activates 5 percent of its parameters per token is exactly the model you can serve on a mid-tier domestic chip. The efficiency habit and the hardware constraint are the same story told twice.

Then the model layer, where being behind became the asset. A lab defending frontier margins cannot rationally give its weights away. A lab without those margins loses nothing by doing it, and gains the world’s developers. Open weights are the one competitive weapon that only works from behind, and China is the only AI power positioned behind.

And on top, coordinated demand. The State Council’s AI Plus opinions set adoption targets of 70 percent penetration in key sectors by 2027 and 90 percent by 2030. The state subsidizes the supply side and mandates the demand side of the same market.

The US holds the stronger position at exactly one layer, the frontier models themselves. Nearly every other input is rented: power from grids where data-center demand is bidding up prices, leading-edge fabrication from TSMC, distribution through metered APIs. Rented inputs do not get cheaper as you scale. China’s inputs do, because they feed each other: cheap power lowers inference costs, cheap inference drives open-weight adoption, adoption pulls the domestic chip ecosystem along, and the energy buildout keeps the loop supplied. That flywheel, not any single subsidy, is the part competitors cannot copy.

The price war is working

Look at what US labs have done since the Chinese cluster formed at the bottom of the price ladder. OpenAI positioned GPT-5.6 Terra at GPT-5.5-level performance for half the price. Anthropic shipped Claude Sonnet 5 with introductory pricing of $2 input and $10 output through August 2026, a third off its own list. The mid-tier is compressing exactly the way you would expect when a competitor sells adequate capability at a fifth of your rate. We unpacked the mechanics in did OpenAI actually cut prices with GPT-5.6, and the pattern holds across the coding-agent value leaderboard: score per dollar keeps migrating toward the cheap end.

Adoption data says buyers noticed. Rest of World reported in June 2026 that DeepSeek hit 17 percent of token volume on Vercel’s AI platform in May while collecting about 1 percent of the revenue. The same UBS analysis cited earlier found roughly 60 percent of companies that track AI spend actively migrating workloads toward cheaper models. On serious agentic coding, our own modeled numbers in Qwen 3.7 Max vs Claude found the premium model still wins on capability while the challenger wins on cost per task, which is precisely the squeeze.

US flagships carry a political tax now too. Claude Fable 5 spent most of late June 2026 suspended for foreign users under a US export-control directive, and OpenAI previewed GPT-5.6 behind a government-gated allowlist, episodes we covered in the GPT-5.6 release gate. Weights you can download and run yourself cannot be switched off by someone else’s regulator. For a CTO outside the US, that is not an abstract point in favor of open models. It is uptime.

What this means if you are buying AI

Titans are fighting above you, and the spoils land on your invoice. Frontier capability from US labs, near-frontier capability from Chinese labs at a fraction of the price, and both sides cutting: buyers have never had it better. The rational play is the one enterprises in the UBS data already run, routing routine volume to cheap models and reserving flagship rates for work that actually needs the last few benchmark points.

The caveats are real but manageable. Sending data to China-hosted APIs is a compliance question your counsel answers, not your benchmark suite; most Western deployments of Chinese models run the open weights on US or EU infrastructure instead, via the hosts in our inference provider directory. Benchmark claims from any lab deserve suspicion. And subsidized prices can rise once share is won, a risk we track by verification date on the price tracker rather than assuming rates hold.

The bigger picture is national strategy colliding with market structure, a collision measured country by country in our AI nations data pillar. China is spending policy, power and open weights to commoditize the layer where US labs make their money. US labs are responding the only way arithmetic allows, with lower prices. You do not have to root for either side to bank the difference.

Frequently asked questions

Why is DeepSeek so cheap?
DeepSeek V4 activates a small fraction of its parameters per token (its V3 predecessor used 37B of 671B), trains in FP8, and bills cache hits at a 99 percent discount. On top of the efficiency, DeepSeek prices for adoption: $0.435 input and $0.87 output per million tokens as of July 2026, with open weights available to self-host.
Are Chinese AI models as good as ChatGPT or Claude?
At the mid-tier, close: GLM-5.2, Kimi K2.7 and Qwen3.7 Max sit near US mid-tier models on coding and reasoning benchmarks at a third to a fifth of the price. The top US flagships still lead on the hardest agentic work. The gap is a pricing decision: whether the last few points are worth 17 to 57 times the output rate.
Is it safe to use Chinese AI models?
Split the question. Calling a China-hosted API routes your data through Chinese infrastructure and its legal regime, which many companies exclude. Running the open weights yourself, or through a US or EU inference host, keeps data under your jurisdiction and is how most Western production deployments use these models.
Why does China open-source its AI models?
It is stated industrial policy, not charity. The 2025 Global AI Governance Action Plan and State Council AI Plus opinions both promote an open ecosystem. Open weights buy global distribution and developer lock-in, compensate for restricted chip access, and undercut the inference margins that fund US closed labs.
Will US AI prices come down?
They already are, selectively. OpenAI priced GPT-5.6 Terra at GPT-5.5 performance for half the cost, and Anthropic gave Claude Sonnet 5 introductory pricing a third below list. Expect continued compression in the mid-tier, where Chinese models compete directly, while true frontier rates stay premium.

Sources

Subscribe to Capital & Compute

Source-backed analysis of what AI compute really costs, sent when a new post goes live.

No spam. Unsubscribe anytime.

← Back to all posts