Chinese AI labs train frontier models for millions and sell access for pennies

DeepSeek trained V3, the predecessor to its trillion-parameter V4, in under two months for about $5.5 million. It rented 2,048 H800 graphics processing units (GPUs) at roughly $2 an hour.

US labs spend well over $100 million on a frontier training run. The price gap between the two countries' flagship APIs (application programming interfaces) is steeper still.

Kyle Chan, a fellow at the Brookings Institution's John L. Thornton China Center, put it plainly. "Chinese AI models are particularly attractive to American companies now as AI costs skyrocket," he told CNBC.

"Where previously US companies were prioritizing AI adoption regardless of model, now they're getting more cost-conscious."

That cost-consciousness shows in the numbers. US companies have kept more than 30% of their OpenRouter token volume on Chinese models every week since February 8, with a peak near 46%. The average for the previous twelve months was 11%.

I spent time tracing where the savings come from. For the most part, the cheap price is four separate advantages stacked on top of each other.

Sparse models that only wake part of the brain

Architecture is the biggest saver. DeepSeek V4 is a Mixture-of-Experts (MoE) model: a design where only a slice of the model activates for each answer.

It holds about 1 trillion parameters, but only roughly 32 billion wake up per token. You get the intelligence of a massive model and pay the compute of a small one.

Chinese labs also compress attention. DeepSeek's Multi-head Latent Attention (MLA) shrinks the memory the model must hold during long conversations, cutting the cost of a million-token context from a quadratic climb to something close to linear.

Low-precision math is native to the pipeline. Models like Alibaba's Qwen 3.5 train and serve in FP8 (8-bit floating-point) format, which roughly halves the memory footprint. Less memory means cheaper serving and faster answers.

I get the sense the least-reported advantage is the thinking budget. DeepSeek and its peers ship thinking and non-thinking modes at the API level, so a simple question does not burn a long reasoning chain. Enterprises pay for the answer, and the model decides how much work it needs.

Research shortcuts and the hardware squeeze

Distillation cuts the research bill. There is strong evidence Chinese labs train on the outputs of top US models, skipping the expensive trial-and-error of building high-quality reasoning data from scratch. US officials have accused Chinese labs of exactly this.

Shared artifacts matter too. US labs run in silos, each paying for its own failed experiments. Chinese researchers build on each other's open discoveries, which stops a large share of duplicated data-center spending.

Export controls forced the rest. Denied NVIDIA's newest chips, Chinese engineers hyper-optimized code for the older H800 and domestic silicon. The models were co-designed with the hardware they run on, which is how a $5.5 million training run could match models that cost a hundred times more.

State money is the floor under all of it. AI infrastructure in China is backed by state-aligned capital and conglomerates that treat losses as the price of future cloud dominance. Pricing APIs at a twentieth of US levels is a deliberate move to make the model layer a commodity and route global high-volume work to Chinese systems.

The sticker price is not the whole bill

On raw pricing, the Chinese camp owns the floor. GPT-5.6 Sol charges $5.00 per million input tokens and $30.00 per million output tokens.

Claude Opus 4.8 charges $5.00 and $25.00. DeepSeek V4 Flash charges $0.14 and $0.28. The output gap is about a hundredfold.

Almost every leading Chinese model ships open weights, from DeepSeek V4 and Alibaba's Qwen 3.6 Max to Zhipu's GLM-5.2 and Moonshot's Kimi K3. Developers can download and run them on their own infrastructure. US flagships stay closed.

US models still hold the tooling and reliability lead. OpenAI and Anthropic offer mature developer tooling, enterprise governance, and committed capacity, where cheap Chinese APIs have throttled traffic during peaks. Multimodal work still favors US systems like Gemini 3.1 Pro.

OpenAI has also answered the price floor. It cut GPT-5.6 Sol's API cost and Anthropic introduced cheaper tiers within weeks of V4 Flash's launch, a direct competitor response to a Chinese model that matched frontier benchmarks at a tenth of the price.

Harpreet Arora, head of agentic infrastructure at Vercel, said teams are choosing on price. "Price is doing the work here," he told CNBC. "When a task doesn't need the best model, teams are beginning to route it to the cheapest one that's good enough."

That routing logic is the correct lens. Chinese models are not automatically the cheapest per finished job. Kimi K3 costs $3.00 input and $15.00 output, more than xAI's Grok 4.5 at $2.00 and $6.00.

On tasks where a US model finishes in fewer tokens, the total cost can flip.

Cache pricing narrows the gap further. DeepSeek drops cached input to $0.0028 per million. Anthropic discounts cached reads by 90% and OpenAI offers a 10x reduction, so teams running repetitive prompts see the US bill fall toward the Chinese number.

I think most teams will end up running both. Route high-volume, repetitive work to the cheap Chinese stack, and keep a US model for the reasoning tasks where it finishes first. The companies that win the next year will be the ones that measure cost per completed task, because sticker price and total cost have split apart.

Source