TL;DR: Picking an AI model is picking a geopolitical stance. American bloc leads on capability but costs 5-10x more and keeps weights closed. Chinese bloc is 8 months behind but 16x cheaper with fully open weights. European bloc sells sovereignty — data processed under EU law on EU infrastructure — at a capital disadvantage. Six questions (capability, cost, architecture, jurisdiction, supply chain, ownership) determine which fits your workload.
The Airbnb Letter
In October 2025, Airbnb’s CEO said publicly that they rely heavily on Alibaba’s Qwen model in production — it’s good, fast, and cheap. They use OpenAI’s latest models but “typically do not use them that much.” Six months later, the US House Homeland Security Committee sent Airbnb a formal letter asking for a national security justification.
On the surface, geopolitics. Underneath, cost control. Qwen costs roughly 1.80 per million output. Comparable US models run 30 output — a 16:1 ratio on output. At Airbnb scale, that’s not a margin detail, that’s a budget decision.
The point is not that Airbnb made the right or wrong call. The point is that every time you pick a model, you’re committing to a bloc with different capabilities, costs, legal reach, and supply chains. Most companies have already done this without framing it that way.
American AI — Capability, Closed, Expensive
The American bloc leads on raw capability. In April 2026, NIST’s Center for AI Standards and Innovation evaluated DeepSeek V4 Pro, the strongest open-weight model from China, and concluded it lags the frontier by about 8 months. That’s the closest objective measurement available — not negligible, but not insurmountable.
The defining characteristic is closed weights. OpenAI, Anthropic, and Google keep their best models behind an API. You rent capability; you don’t own it. Meta, previously the anchor of Western open-weight efforts, released Muse Spark in April 2026 with weights closed for the first time. The direction is unmistakable.
The most striking example: Anthropic completed a frontier model called Claude Mythos and chose not to release it publicly. It was restricted to a small group of cybersecurity partners under Project Glasswing because it demonstrated meaningful progress on autonomous multi-step cyber attacks — completing 3 out of 10 full 32-step attacks in UK AI Security Institute testing. The model exists, but only select partners can access it. That’s closed-weight philosophy in practice: the lab holds the dial.
Cost is the third factor. American frontier models run 5-10x more per output token than comparable Chinese alternatives. For small workloads, invisible. For enterprise scale, decisive.
Finally, the Cloud Act. A 2018 US law that allows authorities to compel US-headquartered companies to hand over data regardless of physical storage location. Applies to every major US cloud and AI provider. For EU-regulated industries — healthcare, banking, government — this is a procurement constraint, not a theoretical concern.
| Trait | American Bloc |
|---|---|
| Capability | Frontier leader |
| Architecture | Closed weights, API-only |
| Cost | Highest (5-10x Chinese) |
| Jurisdiction | US Cloud Act applies |
| Key players | OpenAI, Anthropic, Google, Meta |
Chinese AI — Cheapest, Open, Geopolitical
Chinese models are roughly 8 months behind on capability but one to two orders of magnitude cheaper. Weights are open — download, self-host, never speak to the lab again. The architectural choice is inverted: the lab loses control on release day.
The adoption is already happening. Cursor’s Composer 2 launched in March 2026 without mentioning its underlying model. Users discovered it runs on Kimi K2.5 from Moonshot AI. Cursor is used by 64% of Fortune 500 companies — meaning two-thirds of America’s largest companies use a coding tool built on a Chinese foundation model. Perplexity integrated DeepSeek as a core option. The list grows.
How did Chinese labs get here? Hardware constraints. They can’t legally buy the most advanced US chips, so they built more efficient architectures, trained smaller models harder, and squeezed more performance per watt. The result is some of the most efficient AI engineering in the industry.
Three risks to factor in:
- Hardware dependency — DeepSeek V4 was still largely trained on Nvidia GPUs. Chinese labs are building domestic compute with Huawei, but they’re not there yet. Tighter US export controls could slow progress.
- Data jurisdiction — API traffic routes to China, subject to PRC data laws. Most enterprises mitigate by self-hosting open weights, which adds operational overhead but eliminates data exposure.
- Geopolitical scrutiny — The Airbnb letter shows even consumer companies attract attention. Defense, federal agencies, and healthcare face sharper scrutiny. Expect more as the political environment hardens.
| Trait | Chinese Bloc |
|---|---|
| Capability | ~8 months behind frontier |
| Architecture | Fully open weights |
| Cost | Lowest (1.80 per million tokens) |
| Jurisdiction | PRC data laws (API); self-hosted = your choice |
| Key players | Alibaba (Qwen), DeepSeek, Moonshot (Kimi), Zhipu |
European AI — Sovereignty as the Product
The European ecosystem is smaller by every metric except one: regulatory alignment. Mistral is the center of gravity — strategic partnerships with the French Ministry of Armies, BNP Paribas, ASML, and SAP. Revenue went from ~400M by February 2026. 20x growth in one year.
Mistral sells what neither US nor Chinese labs can: sovereignty. Data processed under European law, stored on European infrastructure. They’re building 13,800 Nvidia GB200 GPUs into a new data center near Paris, with another €1.2B facility in Sweden, and a partnership with SAP to embed sovereign AI into European government services.
But the capital gap is enormous. US hyperscalers spent ~14B versus OpenAI at ~380B.
And American capital is quietly absorbing European labs. By latest funding rounds, 73% of Europe’s AI companies have American-led investors. When Finland’s Silo AI was acquired by AMD in 2024, the lab didn’t disappear — it stopped being European AI and became American AI with a European office.
| Trait | European Bloc |
|---|---|
| Capability | Behind frontier |
| Architecture | Mixed (open and closed) |
| Cost | Mid-range |
| Jurisdiction | EU law, European infrastructure |
| Key players | Mistral, Aleph Alpha, Black Forest Labs |
The Six-Question Framework
The video proposes six questions for model selection. Here they are structured as a decision matrix:
1. Capability
How much does the absolute best model matter for your workload? Frontier research and advanced agentic workflows need the top tier — the 8-month gap matters. Customer service inference, content generation, and classification tasks don’t.
2. Cost
How sensitive is your workload to per-token pricing? The question isn’t “which is cheaper” — it’s “how much of my AI spend scales linearly with usage?” At small volumes, price differences are noise. At Airbnb scale, they’re the budget.
3. Architecture
Open weights or closed? Closed means the lab holds the dial — they can change pricing, deprecate versions, restrict use cases, or take models offline. Open means you own the weights but take on hosting. Self-hosting exists in all three blocs but restricts you to specific providers.
4. Jurisdiction
Where can your data legally live, and who can compel access? American AI = US legal reach. Chinese AI API = PRC reach. European AI = the only bloc built for jurisdictional choice. If you handle regulated data, this question is mandatory.
5. Supply Chain
Every block depends on TSMC, which manufactures ~90% of the world’s most advanced chips. How sensitive is your business to a Taiwan disruption? Self-hosting provides more resilience than API access because you own the compute already running, but you’re still locked to specific model providers.
6. Ownership
Will the lab you’re building on still be the same lab in 3-5 years? Google DeepMind and Silo AI both shifted blocks through acquisition. The labs didn’t disappear, but their alignment changed. Price this risk into contract duration.
The Hidden Lock-In
The video’s core argument: once you embed a bloc choice into your infrastructure, workflows, and contracts, it becomes very hard to reverse. ChatGPT users are structurally committed to the American bloc. Companies self-hosting Qwen are committed to the Chinese bloc. EU regulators mandating sovereign AI are committing to the European bloc.
The lock-in isn’t technical — it’s organizational. Training pipelines, compliance reviews, vendor contracts, and team expertise all accumulate around one choice. Switching blocs mid-project is a rewrite, not a config change.
References
- Should You Pick American, Chinese, Or European AI? — Ali H. Salem, YouTube (May 25, 2026) — https://www.youtube.com/watch?v=yKydIeHoVbY
- CAISI Evaluation of DeepSeek V4 Pro — NIST (May 2026) — https://www.nist.gov/news-events/news/2026/05/caisi-evaluation-deepseek-v4-pro
- Project Glasswing — Anthropic — https://www.anthropic.com/glasswing
- Introducing Muse Spark — Meta AI Blog (April 8, 2026) — https://ai.meta.com/blog/introducing-muse-spark-msl/
- Mistral AI raises $830M for Paris data center — Data Center Dynamics (March 30, 2026) — https://www.datacenterdynamics.com/en/news/mistral-ai-raises-830m-in-debt-financing-for-data-center-in-paris-france/
This article was written by Hermes Agent (Qwen3.6-27B | llama.cpp), based on content from: https://www.youtube.com/watch?v=yKydIeHoVbY
