Institutions Call to "Sell Mag 7, Buy China AI" as Cost Gap Reaches 100x
N.R. Finch
Quant firm Quantum Strategy this week told clients to exit U.S. Mag7 stocks and rotate into Chinese AI names, driven by an output-price gap exceeding 100× between U.S. and Chinese large models — not a short-term blip, but a structural cost inversion already redirecting capital flows.
How strong is the "sell U.S., buy China" signal?
Quantum Strategy on Monday explicitly recommended exiting Mag7 and adding Chinese AI stocks. Chinese tech rose the same day.
This is not an isolated call: Mag7 dropped roughly 10% in June alone, shedding about $2.3 trillion in market cap. Markets began punishing the "spend-to-build-compute" thesis.
Bloomberg earlier reported that hedge funds in May were cutting Mag7 positions while buying Chinese ADRs — American depositary receipts, shares of Chinese firms listed in the U.S. Top picks included Alibaba, PDD, and Baidu.
How wide is the pricing gap between U.S. and Chinese models?
OpenAI's GPT-5.5 costs roughly $5 per million input tokens and $30 per million output tokens. DeepSeek's V4-Pro charges about $0.44 and $0.87; its lightest Flash tier drops to $0.14 and $0.28.
This means → the cheapest output tier is over 100× cheaper than GPT-5.5. That is not a discount — it is a different cost universe.
Zhipu's GLM-5.2 prices at roughly $1.40 and $4.40 — higher than DeepSeek, still far below U.S. mainstream models.
With prices that much lower, how big is the performance gap?
On SWE-bench coding tests, several Chinese models scored within one percentage point of Western frontier models.
In plain terms = performance is nearly level, yet pricing differs by one to two orders of magnitude — the performance gap simply cannot justify the premium.
In a test spanning 18 real-world coding tasks, GLM-5.2 completed them for roughly $2.74; GPT-5.5 cost about $16.10 — a gap of roughly 6×.
Why is open-source licensing another competitive moat?
Both DeepSeek and GLM-5.2 ship under MIT licenses — a permissive open-source agreement — with downloadable weights. Firms can deploy on their own hardware.
This means → enterprises can bypass per-token cloud billing entirely, pushing inference costs down to hardware depreciation.
A boundary worth noting: content-moderation filters embedded in the model weights (refusals on Taiwan, Tiananmen, Xinjiang topics) cannot be removed by self-hosting; fine-tuning is required.
What does Google's flagship delay reveal?
Google's next flagship, Gemini 3.5 Pro, was previewed at the May I/O conference with a 2-million-token context window and a "deep think" reasoning mode — but remains limited to an enterprise preview with no release date, no pricing, and no public benchmarks.
The original June launch date has slipped. The latest rumored target is July 17, but industry trackers say that date is also uncertain. Google declined to comment.
This reflects a deeper issue: the delay traces back to the extreme compute load that "deep think" mode generates on multi-step tasks — the very arena Chinese open-source models are cracking at a fraction of the cost.
Can this rotation sustain?
The logic chain driving capital from Mag7 into Chinese AI stocks is clear: converging performance + a 100× cost gap + open-source models that bypass paid APIs.
This means → this is not a sentiment-driven sector swap. The market is repricing who actually wins the AI application-layer cost race.
The core test ahead: whether U.S. AI leaders' cost structures are beginning to invert — if the performance edge bought by massive spending gets erased by low-cost open-source rivals, this rotation will not stay confined to hedge-fund portfolios.
Content is for reference only, not financial advice.