UBS: China's AI Model API Prices Average Less Than 20% of U.S. Counterparts, With Comparable Gross Margins
Claire Weston
UBS finds Chinese AI models charge under 20% of US API prices with training costs below 10%, yet gross margins match US peers — the discount runs on structural cost advantage, not subsidies.
Where does the cost gap come from?
Three layers all point one way: smaller parameters, leaner algorithms, cheaper power.
On parameter scale, DeepSeek V4 has 1.6 trillion total parameters and Kimi K2.6 about 1 trillion; Claude Opus 4.6 and GPT-5.5 are estimated at roughly 10 trillion and 5 trillion — Chinese models run with far fewer parameters.
On compute precision, DeepSeek-V3 pioneered FP8 mixed-precision training; V4 pushed further to FP4. This means → the same GPU crunches more work per cycle, cutting unit cost directly.
On infrastructure, Chinese data-center power costs about 4.4 ¢/kWh versus roughly 7.9 ¢/kWh in the US — about 44% lower. GPU rental rates are likewise about 40% cheaper.
How is inference cost pushed so low?
Two key levers: lower MoE activation ratios and KV-cache compression.
US early MoE models — mixture-of-experts, splitting a large model into specialist modules and activating only a few per query — typically activated 15–30% of total parameters. Leading Chinese MoE models activate just 3–10%. In plain terms = for a model the same overall size, the Chinese version wakes up far fewer experts each time, slashing compute.
DeepSeek V4 cut its active-parameter ratio from V3.2's roughly 5% to about 3%, yet its intelligence index rose from 42 to 52 — cheaper and smarter.
KV cache — the model's "memory strip" for prior conversation turns — accounts for about 70% of inference cost in multi-turn agent tasks. DeepSeek V4 compressed its cache to roughly 10% of V3.2's size, shaving another ~10% off total cost versus V3.
What does margin parity really tell us?
MiniMax's end-to-end infrastructure team achieves over 75% GPU utilization, against an industry average of 40–50%.
Its M2.7 model carries a gross margin above 40%, essentially matching Anthropic's roughly 40% API gross margin in 2025.
This means → the low price is not a loss-leader play. The cost structure itself sits a tier below, and profit margins are intact.
How large is the remaining performance gap?
In 2023 China's frontier models scored about 60% of top US models on composite intelligence; by 2025 that ratio reached roughly 90%.
By domain: text models are near the 90% mark; four of the global top-five video-generation models are Chinese; in AI coding, leading Chinese models match the prior US generation but still trail the latest frontier.
This reflects a deliberate strategy — "good enough plus radically cheaper." In most commercial use cases, 90% capability at 20% of the price makes the procurement decision straightforward.
Why does open source accelerate the catch-up?
Zhipu (智谱) and MiniMax spent a combined ~$800 million on R&D in 2025 — roughly one-tenth of Anthropic's outlay.
Yet Chinese teams heavily reuse each other's validated techniques: Kimi K2 and GLM-5 adopted DeepSeek's MLA latent-attention design; DeepSeek V4 incorporated Moonshot AI's Muon optimizer; Qwen3 and GLM-4.5 use DeepSeek's GRPO reinforcement-learning framework.
In plain terms = US closed-source labs each reinvent the wheel; China's open-source ecosystem runs on "experiment collectively, benefit individually" — marginal R&D cost is systematically spread.
What does UBS expect — and what is the big unknown?
UBS estimates the global long-term AI market could exceed $10 trillion and sets a base case — analogous to EVs and smartphones — where Chinese models capture 30–50% of global share.
Bull case (solar-panel analogy): if capabilities converge and competition turns purely on cost, share could top 80%. Bear case (cloud / OS analogy): if US frontier models keep a clear capability lead and embed deeply into enterprise workflows, China's share may stay in the low single digits to under 10%.
Goldman Sachs One-Delta desk head Rich Privorotsky frames the core tension as "the trillion-dollar question": does cheaper intelligence create more demand, or destroy more pricing power?
He cites an OpenRouter experiment: a blend of Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro beat GPT-5.5 and Opus 4.8 running solo across benchmarks — at roughly half the cost — closing the gap to within 1% of Fable 5. This reflects a verdict whose direction will determine whether the current multi-trillion-dollar AI sector valuation holds up.
Content is for reference only, not financial advice.