US AI firms route 30% of traffic to Chinese models despite blacklist

1. A sharp 1-2 sentence LEDE (no heading) that leads with the most important concrete
   fact and makes the stake clear.

U.S. AI companies are sending roughly 30 % of their model queries to Chinese providers, a share that has risen steadily despite growing U.S. export controls on those firms [1].

2. An "At a glance" KEY-FACTS TABLE — a 2‑column Markdown table whose header row is
   exactly | At a glance | |, then the separator |---|---|, then one row per fact
   (e.g. | Price | $1,735 |). Capture the company/product, the headline figure (raise / users / benchmark / price), and the stage (launch, raise, earnings, release). as 3‑4 rows, each a hard
   fact with its number.

3. The body as 3‑5 tight paragraphs, BROKEN INTO 1‑2 sections under short DESCRIPTIVE
   ## subheads that name the actual content (e.g. "## What drove the move", "## The
   competitive picture") — never generic labels like "Why it matters". what shipped or changed, the key numbers, how it stacks against rivals or the prior generation, and the competitive or market implication.
   Anchor every key number in context (vs. prior / expected / record), keep fact
   separate from claim, and cite each distinct fact once with [n].

Rising reliance on Chinese AI models

OpenRouter, a startup that aggregates multiple AI back‑ends for U.S. firms, reported that Chinese‑origin models now account for about 30 % of the queries it routes [1]. The shift reflects the 60‑90 % cost gap these models hold over leading offerings from OpenAI and Anthropic, a margin that can translate into substantial savings for early‑stage startups operating on tight budgets [1]. Some nascent companies have even migrated all of their traffic to Chinese providers such as DeepSeek, betting on the lower price and greater control over model behavior [1].

Regulatory risk and market friction

Beijing is reportedly preparing export controls that could restrict overseas access to its AI models [1]. If such controls materialise, U.S. firms that have become dependent on Chinese models could face sudden supply disruptions, echoing earlier concerns about Chinese hardware components flagged by the Pentagon [2]. Analysts note that the “inertia” of building applications on a particular model can make a rapid switch costly, a point underscored by Reflection’s VP of product, who warns of firms becoming “locked into” the Chinese ecosystem [1]. At the same time, the Chinese lidar maker Hesai, despite being on a U.S. defense blacklist, continues to supply sensors to major autonomous‑vehicle players, illustrating how commercial demand can outpace regulatory restrictions [2].

4. If the sources give a spec / benchmark / pricing comparison vs. a competitor or prior version, add a small Markdown table; otherwise skip it — never force one.

5. A ## What to watch section with 2‑3 specific, concrete, NON‑advice bullet items:
   a ship or launch date, the next earnings, a rollout milestone, or a likely competitive response. (Frame as what to monitor, never as what to do.)

What to watch

• Any formal export‑control announcement from Beijing that limits overseas licensing of Chinese AI models.
• Updates from OpenRouter on the proportion of traffic routed to Chinese providers in the next quarter.
• Responses from U.S. AI firms, such as diversification of model providers or new partnerships with domestic alternatives.

6. Close with one or two sentences delivering the real significance or the open
   question — concrete, not a generic wrap‑up.

The growing reliance on cheaper Chinese AI models highlights a cost‑driven trade‑off that could reshape supply chains, but the looming threat of export controls leaves U.S. developers facing a strategic dilemma: balance immediate savings against the risk of future access restrictions.