> 📥 The full analysis is available as a PDF (Danish, 8 pages). Download it at the top of the page — all numbers, prices and sources in one place.
I teach the American frontier models. At ITU and DIS Copenhagen, the curriculum is mainly OpenAI, Anthropic, Copilot and Google. Those are the ones students meet when they graduate. Those are the ones my clients buy. And those are the ones I work in every day.
Then comes the question I now get in almost every client conversation: "Are we allowed to use DeepSeek?" It's rarely asked out of curiosity. It's asked because someone in the organization has already done the math on price.
I spent the summer figuring out the right answer. Here it is: who the Chinese players are, how good their models actually are, what they cost, and what you should concretely say. Along the way my own answer moved somewhere I hadn't expected. That comes last.
1. Week 29 was no coincidence
On July 16, Moonshot AI released Kimi K3, the largest open model ever at 2.8 trillion parameters. On July 19, Alibaba answered with a preview of Qwen3.8-Max at the World AI Conference in Shanghai. On July 24, DeepSeek retired the old API names, forcing everyone running DeepSeek in production to migrate to V4.
And while the models landed, the politics moved. On July 21, the Financial Times and Reuters reported that Beijing is considering export controls on Chinese model weights. On July 24, 77 companies signed an open letter defending open weights — Nvidia, Google, Meta, Microsoft, Mistral and OpenAI among them. Anthropic did not sign. On July 29, Moonshot closed a 3.5 billion dollar round at a 35 billion valuation. On August 3, Alibaba launched Qwen3.8-Max in full, and the stock rose 9 percent.
The story of Chinese AI as cheap imitation died somewhere in 2025. This is a race where number two keeps running faster — and it is nearly free.
2. DeepSeek: the hedge fund that cost Nvidia 590 billion dollars
DeepSeek looks like no other AI company in the world. It was founded in 2023 in Hangzhou by Liang Wenfeng, the man behind the quant hedge fund High-Flyer, and financed by his own fund profits. Around 150-170 employees, no KPIs, young researchers from elite Chinese universities, and a culture closer to a research institute than a company.
The world discovered them on January 20, 2025, when the R1 model shipped and the DeepSeek app topped the US App Store. A week later Nvidia fell 17 percent in a single day, roughly 590 billion dollars in market value — at the time the largest single-day loss for one company in US market history. The point that scared the market: the final training run behind V3 cost about 5.6 million dollars in GPU time. That figure covers only the final run, not all the research and failed attempts — but even with that caveat it stands in stark contrast to American frontier budgets in the hundreds of millions.
In June 2026 DeepSeek took outside capital for the first time: 7.4 billion dollars at a valuation above 50 billion. Note the structure: most investors bought into a limited partnership controlled by Liang himself, without voting rights and with a five-year lock-up. The only investor with voting rights and no lock-up is China's state AI industry fund.
DeepSeek V4 comes in two versions: V4 Pro with 1.6 trillion parameters (49 billion active) and V4 Flash with 284 billion (13 billion active), both with 1 million tokens of context and both under a genuine MIT license. The price is the wild part: Flash costs 0.28 dollars per million output tokens. OpenAI's GPT-5.6 Sol charges 30 dollars for the same.
The weaknesses are just as concrete: first-party hosting in China under Chinese law, documented censorship on politically sensitive topics, weaker multimodality, and an R2 model that never shipped, reportedly because training on Huawei chips failed. And note: DeepSeek's own API trains on your data. The Western hosts serving the same model do not.
3. Moonshot: the model so popular they closed signups
Moonshot AI was founded in Beijing in 2023 by Yang Zhilin, a researcher with a PhD from Carnegie Mellon. Alibaba and Tencent are among the investors. The valuation went from 2.5 billion dollars in February 2024 to 20 billion in May 2026, and on July 29 they closed 3.5 billion dollars at a 35 billion valuation.
Kimi K3, released July 16, is their masterpiece: 2.8 trillion parameters, 104 billion active, 1 million tokens of context, native multimodality. The day it shipped it entered at number 1 in Arena.ai's Frontend Code Arena with 1,679 points, ahead of Claude Fable 5 (1,631) and GPT-5.6 Sol (1,618). A jump from 18th place to first. Three days later Moonshot had to close signups, in their own words: "Kimi K3 has received far more love than we expected, and our GPUs are feeling it."
On July 27 the weights arrived, and K3 became the fastest-growing release in Hugging Face history with nearly half a million downloads in the first month. From that moment anyone can host the model themselves. But read the license before you plan anything: K3 sits under a self-written "Kimi K3 License", not MIT or Apache. Open weights are not the same as free terms.
Two honest caveats. Moonshot themselves write that K3 still trails the best American closed models on overall capability. And the coding lead didn't hold: K3 sits at number 2 in the Frontend Code Arena today with 1,676 points, overtaken by Claude Opus 5 Max at 1,705. Three weeks in first place.
4. Alibaba: Qwen is the foundation under most open source AI
While DeepSeek and Moonshot take the headlines, Alibaba has built something more structural. The Qwen family passed a billion downloads on Hugging Face in January 2026, with around 200,000 derivative models. It's the world's most used open model family, and it overtook Llama along the way. When Apple Intelligence received regulatory approval for China on July 15, Qwen was the named partner in the stack.
Qwen3.8-Max was previewed on July 19 and launched in full on August 3: 2.4 trillion parameters, multimodal, 1 million tokens of context. Alibaba's own claim is that only Claude Fable 5 beats it. An independent measurement placed it in Arena.ai's Frontend Code Arena on August 2 at number 4 with 1,668 points, only 8 points behind Kimi K3. But there is still no published license and no full model card, so you can't plan self-hosting on it yet.
Alibaba's real strength is the portfolio: models in every size, a cloud business that makes money on inference, the Apple deal as consumer distribution, and an investment program of 380 billion yuan (about 53 billion dollars) over three years. The weakness is talent and trust: Qwen's technical lead Junyang Lin abruptly left the company on March 3, the day after a major model release.
5. Behind the big three is a whole league — and it's publicly listed now
Zhipu AI became the world's first publicly listed LLM company on January 8. Market cap peaked above 1,000 billion Hong Kong dollars (about 128 billion dollars) in late June after GLM-5.2. In June, GLM-5.2 was the best open model on Artificial Analysis' intelligence index and leads PostTrainBench for agentic coding. Both GLM-5 and GLM-5.2 are, by the company's own account, trained exclusively on Huawei Ascend chips, without a single Nvidia GPU. That's a publisher claim, not independent verification — but if it holds, it means American chip export controls have a back door.
MiniMax listed the day after Zhipu and doubled on debut. ByteDance runs the Seed models with aggressive pricing, Tencent's Hunyuan has gone open source under an Apache license, Baidu's Ernie is trained on their own Kunlun chips, and StepFun runs the efficiency track with small, fast open models. Two years ago people talked about "the six AI tigers". Today the picture has consolidated, and the four tech giants have gone from sponsors to competitors. The weapon they share is the same: open weights.
6. The lead is 39 Elo points — and it has grown
You have to be careful with the numbers here, because two entirely different measurements keep getting conflated. Stanford's AI Index 2026 measures the difference between the best American and the best Chinese model on LMArena's Elo scale. In March 2026, Claude Opus 4.6 led with 1,503 points over the best Chinese model, Dola-Seed-2.0 Preview, at 1,464. A difference of 39 points, or 2.7 percent.
> How big is the gap — measured properly?
> 39 points: US lead on the Elo scale (2.7%), March 2026 — larger than the year before (Stanford AI Index 2026).
> 3-6 months: a constant gap behind the American frontier, neither growing nor closing (OpenRouter, 18 months of data).
> No. 4: Kimi K3 on the overall intelligence index — behind Opus 5, Fable 5 and GPT-5.6 Sol (Artificial Analysis).
The narrative that China is overtaking doesn't hold on the Elo measurement. In February 2025 the gap was 5 points; a year later 39. The US keeps the lead — but it's rented, not owned. The big "17-31 percentage point" gaps still circulating come from 2023 benchmarks (MMLU and others) — a different scale, not comparable.
There is one place to be especially skeptical: coding. There is still no independent SWE-bench Verified score for Kimi K3. Every percentage you see circulating is Moonshot's own run or a pure guess. Moonshot's own number is 67.5 on DeepSWE against Claude Fable 5's 70.0 and GPT-5.6 Sol's 73.0. Close, but behind, and measured by the publisher itself.
7. The math your developers have already done
A typical automation workload of 50 million output tokens per month costs very differently depending on the model:
| Model | Price per 1M output tokens | 50M tokens/month |
|---|---|---|
| OpenAI GPT-5.6 Sol | $30 | ~$1,500 |
| Claude Fable 5 | $50 | ~$2,500 |
| Claude Opus 5 | $25 | ~$1,250 |
| Google Gemini 3.1 Pro | $12 | ~$600 |
| DeepSeek V4 Flash | $0.28 | ~$14 |
Table 1. Be fair: cheap American options exist (Gemini). But the distance from $600 to $14 is still a factor of 40.
The market has reacted. On OpenRouter, the largest independent model router, Chinese models overtook American ones in token consumption in early June 2026. Today over 60 percent of routed traffic is Chinese, all five most-used models are Chinese, and Meta's Llama is below 1 percent. According to a Bloomberg tally, the American share has fallen from around 70 percent in June 2025 to around 30 percent a year later.
> Two caveats before you calculate further. OpenRouter measures developer traffic, not revenue — on revenue the American labs still lead clearly. And the cheapest prices are first-party from China, where your data goes to Chinese servers. Western hosts (Fireworks, Together, Azure) typically charge double the Chinese first-party price for the same model, without training on your data. Double of almost nothing is still almost nothing.
8. Both superpowers are considering closing the border
Washington is at war with itself over the answer. On July 24, 77 companies signed the letter "Open Weights and American AI Leadership" — Nvidia, Google, Meta, Microsoft, IBM, Hugging Face, Mistral, Mozilla, the Linux Foundation and OpenAI. Anthropic did not sign and stood alone. On July 27 they responded with a post opening with the claim that they have never argued for a ban on open weights, proposing instead chip export controls, an effort against distillation, and mandatory safety testing of all sufficiently capable models.
And then the mirror image, which European companies should read twice. On June 12, the US Department of Commerce ordered Anthropic to require an export license for any foreign national's access to Claude Fable 5 and Mythos 5, inside and outside the US. Anthropic shut off both models for all customers to comply. The control was lifted on June 30, and Fable 5 was globally available again on July 1. Eighteen days without access.
> The lesson. An American closed model can disappear overnight, without warning, for reasons that have nothing to do with you. Ironically, the episode made the Chinese open models more attractive — because no government can switch off a downloaded weights file. Beijing is now considering export controls on its own model weights. The window where frontier weights flow freely across borders may not stay open forever.
9. What to say when the question hits the leadership meeting
After a summer with that question, my answer boils down to three rules.
1. Chinese first-party APIs with personal data: no. Sending EU personal data directly to deepseek.com or kimi.com has no documented GDPR transfer mechanism. Italy's Garante blocked DeepSeek on January 30, 2025, seven German state authorities opened cases in February 2025, and the Danish Parliament's Presidium banned DeepSeek on all parliamentary devices in the winter of 2025. That track is closed for any serious European company.
2. The same models via Western hosts or self-hosting: a real option — but read the license. If you run the weights at a European or American provider with a no-train guarantee, your data never leaves your control. But "open" covers very different legal realities: DeepSeek V4 is genuine MIT, Zhipu's GLM-5.2 is MIT, Kimi K3 has a self-written license, and Qwen3.8-Max has no published license at all yet.
3. Governance before pilot — and now there's a deadline behind it. EU AI Act enforcement for general purpose models took effect on August 2, 2026. The Commission can now demand documentation, evaluate models, require withdrawal and issue fines of up to 3 percent of global revenue or 15 million euros. If you fine-tune or redistribute an open model, you may become a "provider" yourself with the obligations that follow.
Three questions I'm asking my clients right now: Do you know which models your vendors and tools actually run you on today? Have you calculated what your largest AI workload would cost on an open model via a Western host? And does your AI policy have an answer on Chinese models — or will you find out the day a developer already has them in production?
10. And then what I landed on over the summer
I still teach the American models at ITU and DIS, because that's what the market buys, and that's what students meet when they graduate. But something shifted. When I sit with a European client and work through data, law and vendor risk, it's no longer an American name I write first on the board.
The eighteen days when Fable 5 was switched off showed what vendor dependency costs in practice. The Chinese first-party APIs are off limits when personal data is involved. And since August, the AI Act has put a price on compliance that leadership can see in a spreadsheet. Two options remain, and they point the same way: open weights at a European host, or the European models themselves. For an EU company it's getting close to a no-brainer. I wouldn't have said that six months ago — and it isn't because the European models have become the best. It's because the rest of the math has changed. That's the article I'm writing next.
Sources
This analysis draws on a broad source base; the most important are (as of August 2026):
- DeepSeek: api-docs.deepseek.com · huggingface.co/deepseek-ai (specs, MIT license) · forbes.com 17/6 2026 · arxiv.org/pdf/2412.19437 (V3 training) · forbes.com & reuters.com 27/1 2025 (Nvidia loss)
- Moonshot/Kimi: huggingface.co/moonshotai/Kimi-K3 · x.com/arena 16/7 2026 · arena.ai/leaderboard/code/webdev · bloomberg.com 7/5 & 29/7 2026 · technode.com 3/8 2026
- Alibaba/Qwen: x.com/alibaba_cloud 21/1 2026 · marktechpost.com 19/7 & 3/8 2026 · techcrunch.com 15/7 2026 (Apple in China) · technode.com 4/3 2026 (Junyang Lin)
- Other labs: yicaiglobal.com & scmp.com (Zhipu) · deeplearning.ai/the-batch 26/6 2026 · huggingface.co/blog/mlabonne/glm-5 (Ascend) · cnbc.com 9/1 2026 (MiniMax)
- Race and pricing: hai.stanford.edu (AI Index 2026 & 2025) · artificialanalysis.ai · openrouter.ai/blog/insights 30/6 2026 · dataconomy.com 29/7 2026 · platform.claude.com/docs · developers.openai.com · ai.google.dev
- Policy: axios.com 17/7-22/7 2026 · ppc.land (the 77-signatory letter) · anthropic.com/news/position-open-weights-models 27/7 · /fable-mythos-access 12/6 & /redeploying-fable-5 30/6 · cnbc.com 8/7 & qz.com 31/7 (Congress) · reuters.com & ft.com 21/7
- Customer examples: scmp.com 22/10 2025 (Airbnb) · bloomberg.com 20/5 2026 · cursor.com/blog/composer-2-5
- EU/compliance: artificialintelligenceact.eu (ch. V, 2/8 2026) · garanteprivacy.it 30/1 2025 · datenschutz.hessen.de 20/2 2025 · ft.dk (the Danish Parliament's DeepSeek ban)
> ⚠️ Disclaimer. This is general practical guidance, not legal or compliance advice, and I am not a lawyer or compliance specialist. Don't use it as a basis for decisions on your own. Always involve your own compliance or legal department and your DPO, and verify dates and status against primary sources (EUR-Lex, the European Commission, your data protection authority). All prices are in USD excl. VAT and may have changed. Status as of August 2026.
📬 Subscribe to the newsletter "AI, Built Human" on Substack — weekly insights on AI in practice.
Stefano Vincenti · GenAI strategist and architect · External lecturer, IT University of Copenhagen & DIS Copenhagen · Cofounder & CTO BotTellMe · Partner, TryZone · aitrainer.dk