If Apple Can Use Qwen, What's Your Excuse?

East/West Alpha · Edition 7 · Wednesday, September 9, 2026 · Day 1038
Automate Work. Amplify Life. East/West AI, for the people betting real capital and real careers on it.

On July 15, 2026, China's Cyberspace Administration registered Apple Intelligence for use inside China alongside Samsung Galaxy AI. The model behind Apple's service is not Apple's own.

It is Qwen 阿里通义千问, built by Alibaba 阿里巴巴. Apple co-trained it for its devices. Apple's own model handles on-device tasks. Qwen handles complex queries in the cloud through a domestic partner. The Qwen integration is China-specific. Apple is not using Qwen for Apple Intelligence outside China.

This partnership operates at the operating-system level. The China rollout is expected with Apple's fall OS cycle, widely anticipated as iOS 27, though Apple has not named a commercial on-date. The target market spans roughly 200 million iPhones in China.

Answer first. Apple may be providing the strongest credibility signal yet about models from Chinese labs. Too capable, too cheap, and too strategically important for Western companies to ignore. That forces a sharper question. If Apple can partner with Qwen, what prevents other companies from evaluating these models?

The surface reading is regulatory compliance. Beijing required a local partner, so Apple found one. That interpretation misses the deeper story. The company with one of technology’s most valuable privacy brands evaluated models through a three-year, multi-lab process. Then it embedded the winner inside its flagship product.

If you read only this far: the top models from Chinese labs trail the Western frontier by five to ten benchmark points. They cost one to two orders of magnitude less. Companies from Airbnb to Singapore's government already run production workloads on them. Apple's partnership represents permission, not precedent. The remaining barriers are legal boundaries and political risk, not technical capability.

The performance gap has compressed. Alibaba reports its Qwen 3.5 flagship at 91.3% on AIME mathematics and 88.4% on GPQA Diamond scientific reasoning, with wins over GPT-5.2 on instruction-following tasks. Those are vendor-reported scores. DeepSeek 深度求索’s V4 Pro achieves 80.6% on SWE-bench coding challenges. The frontier gap measures five to ten points, not the twenty-to-thirty-point spreads of eighteen months ago.

Cost advantages run deeper. DeepSeek remains materially cheaper than comparable Western models, although its pricing increased after an August 16 adjustment. The exact gap now varies by model tier and workload. Across OpenRouter's enterprise routing, models from Chinese labs have priced 60% to 90% below U.S. alternatives. Training economics follow the same pattern. DeepSeek reportedly trained its V3 model for roughly $6 million. One caveat: DeepSeek has signaled price increases, and the widest gaps live at the efficient tier, where most workloads live.

Scale shows adoption velocity. Hugging Face shows roughly 2.05 to 2.06 billion Qwen downloads and 151,448 derivative models. Separately, Alibaba has reported roughly 3 billion downloads across channels and more than 300,000 derivatives/models. Two different measures from two different sources. During peak periods, Chinese-origin models handled 46% of OpenRouter enterprise tokens. U.S. models drew 35.7%, per OpenRouter data tracked by Digital in Asia.

Western providers are already responding. OpenAI has cut prices twice since mid-August 2026, per reporting tracked by SwadeshiSync. Its flagship dropped more than 20%.

This is not a rising ecosystem. It is a risen one.

The selection matters more than the deal. Apple ran an elimination process, not a vendor review.

According to 36kr's reporting, Apple evaluated multiple Chinese labs over roughly three years. Baidu 百度 was the early favorite and lost. Apple had subsidized Baidu's model training, but Baidu's search business conflicts with Siri's role. ByteDance 字节跳动 could not overcome its App Store policy disputes with Apple. Tencent 腾讯 never reached the final round. DeepSeek, the benchmark leader, saw its service buckle under Spring Festival traffic. That raised a harder question: could it serve a hundred million devices at once?

Alibaba won on sustained performance and infrastructure reliability. No competing operating system. No rival super-app. Cloud scale proven at Double 11 traffic peaks. A compliance record that survived a thirteen-month filing. Joe Tsai put it simply in Dubai in February 2025: Apple is extremely picky, and it finally chose us.

The selection logic, in one line from the Chinese coverage: benchmark scores are the facade. Service capacity, interest boundaries, and compliance qualifications are the core. Apple ran one of the most demanding procurement processes in the industry. A model built in China cleared it.

Note the boundary of the deal. Baidu keeps separate roles in Apple's China stack for search and voice services. The Qwen partnership covers Apple Intelligence specifically, not the broader platform.

And Apple is not the outlier. Airbnb runs customer service on Qwen. CEO Brian Chesky called it very good, fast and cheap. Cursor built its Composer 2 feature on Kimi 月之暗面, from Moonshot AI. Social Capital deploys Kimi K2 for investment research. Mira Murati's Thinking Machines Lab used Kimi K2.5-generated synthetic data in Inkling's post-training, while Inkling's architecture closely follows DeepSeek-V3. Pinterest is experimenting with open-weight alternatives for recommendation systems. Singapore, a U.S. ally, selected Qwen 3 as the foundation of its national language model. Sovereign AI strategy is now betting on Chinese model architectures. The difference with Apple is scale and visibility, not pioneering adoption.

If Apple can work with a model built in China, what argument is left for smaller Western companies that refuse to evaluate one? Three objections persist: legal compliance, political risk, and frontier capability.

The legal exposure is real but bounded. Export-control counsel at Sheppard Mullin maps four risk categories. Controlled technical data flowing into models. Entity List restrictions, with GLM 智谱's parent listed since January 2025. Procurement bans, including the FY2026 defense authorization barring DeepSeek from Pentagon systems. And revocability, with Beijing reportedly weighing limits on overseas access. Federal contractors and regulated industries face explicit restrictions. Commercial companies outside federal procurement keep broad flexibility.

The political exposure is real but manageable. House committees opened inquiries into Airbnb and Cursor this spring over their use of models from China. A CFO's cost optimization can become a congressional headline six months later.

The frontier gap is real but narrow. Anthropic's Claude Opus 4.7 tops SWE-bench at 87.6%, ahead of DeepSeek's 80.6%. For the small share of tasks where failure is catastrophic, the premium buys insurance.

Every argument has a boundary. Open weights self-hosted on your own infrastructure keep data local. That was Chesky's exact defense. Most enterprise workloads operate well below the frontier, where cost advantages outweigh capability gaps. The Apple signal does not say the objections are wrong. It says they are no longer a reason to skip the evaluation. The world's most brand-cautious company priced the risks, engineered around them, and decided the economics won anyway. Smaller companies cannot afford to be less rigorous than Apple about that cost-performance math.

For Western AI, the repricing has started. Frontier premiums compress toward the tasks that genuinely need them. Volume applications shift toward lower-cost alternatives. Scrutinize any AI revenue stream that is volume business priced at frontier margins.

For Eastern AI, capital markets have voted at least once. DeepSeek reportedly approaches a $74 billion valuation in pre-IPO funding, though final terms remain unconfirmed. Moonshot AI has reported a valuation near $18 billion. MiniMax 米奈 has prepared a Hong Kong listing at a reported $15 billion. Comparable benchmark output, a fraction of Western valuations. Three signals define the opportunity. Valuation compression. Ecosystem lock-in from downloads and derivatives. Structurally captive demand, since China's rules reserve a protected market for compliant domestic players. The Apple deal is exhibit A.

The risks are equally concrete. DeepSeek's commercial backers reportedly accepted no voting rights and long lock-ups. Sanctions exposure runs in both directions. The cost advantage is already being repriced by DeepSeek itself. And the last-to-evaluate question gets more expensive each quarter. Early adopters bank cost advantages while late movers face compressed margins and higher switching costs.

The same request now returns a different answer in Beijing than in Boston. Not because Apple's values changed. Because the model behind the interface and the rules around it are different. The iPhone in Shanghai is set to run Qwen. Outside China, Apple uses a different AI stack. The hardware is uniform. The intelligence is not.

For Apple, the deal supports a regional recovery. Apple's Greater China revenue grew about 22% year over year in the quarter ended June 27, 2026, per the company's reported results. Competitive pressure continues. Huawei captured 23% of China's smartphone market in Q2 2026 versus Apple's 19%, per Omdia. Approval removes a binary overhang, but the template risk remains. If the EU, India, or Brazil writes similar rules, the China pattern becomes the default. Local partner, local model, local compliance layer. Apple paid for the key with a piece of the house.

Commandment Two: Own Your Space.

Qwen is set to occupy the default AI slot on Apple's premium iPhone in one of the world's largest smartphone markets. It did not rent that position. It earned it through an elimination process run by one of the most demanding buyers in technology. The commandment cuts both ways for the reader. Western labs renting attention at frontier prices own less of their space than their benchmarks suggest.

Know a founder, CFO, or investor still refusing to evaluate Qwen, DeepSeek, or other models from Chinese labs on technical merit? Forward them this edition. The evaluation costs nothing. Refusing it might.

San Eng

WSJ and USA Today bestselling author

$500M+ deployed across East and West

Disclaimer: East/West Alpha is for informational and educational purposes only and does not constitute financial, investment, legal, or tax advice. Nothing in this newsletter is a recommendation to buy or sell any security. The views expressed are our own, may be incomplete, and can change without notice. Do your own research and consult a licensed professional before making investment decisions. WorkOptional.ai, its writers, and affiliates may hold positions in companies or assets mentioned. Past performance is not indicative of future results.