Amazon hosts the Chinese AI model its own CEO is waiting to replace
Andy Jassy put Z.ai's GLM 5.3 on Bedrock's six-to-eight-model shelf, called the platform gigantic, then said the West is hungry for a Western replacement. Z.AI's Singapore shares fell 8.62% the next day.
Vincent Jiang · 4 min read
Andy Jassy used Amazon's−1.20% — Amazon, down 1.20 percent today third-quarter earnings call on October 29 to confirm that Bedrock, the company's model marketplace, now carries Z.ai's GLM 5.3 beside Moonshot's Kimi K3 and closed models from Anthropic$350B — Anthropic, private, latest valuation $350B and OpenAI$500B — OpenAI, private, latest valuation $500B 1. GLM 5.3 is the open-weight model of Z.ai, the Beijing lab known as Zhipu that Tsinghua professor Tang Jie spun out of his university lab in 2019 2. Jassy called Bedrock's potential "gigantic," a service he wants to grow into something like EC2, AWS's core compute product 1.
The shelf that decides the money
The economics sit on a narrow shelf. Jassy expects six to eight models to handle more than 95% of Bedrock workloads within two to three years 1. GLM 5.3 holds one of those seats 1.
The same CEO said the Western world is "very hungry" for additional Western open models, and that he believes they will arrive over time 1. On Jassy's own words, the seat is political as much as contractual: the host is openly waiting for a model it likes better.
A month of security studies
The seat also carries a record. Anthropic's Frontier Red Team reported that GLM 5.3's safeguards can be bypassed in 64% to 100% of simulated tests, and that the model built working exploits in 50 of 410 benchmark attempts, six short of the 56 Anthropic's own guarded model managed 3. NIST's Center for AI Standards and Innovation called GLM-5.3 the most cyber-capable open-weight model released to date, about four months behind the US frontier 3.
Days before AWS's call, researchers reported the model exploited six versions of Firefox working from security fixes that had not yet shipped, and Mozilla patched 4. More than 1.3 million copies of the weights sit on Hugging Face and cannot be recalled 4. AWS's wrapper protects customer inputs and outputs; it does not patch the weights 1.
The headline lands on the name anyway
Two qualifications before the numbers scare anyone. The breach figures are Anthropic's, a competitor's, measured against its own model, and the Firefox work ended in a patch, not a breach 4. And the name travels with or without a fault: The Inference reported on October 27 that Cloudflare−9.41% — Cloudflare, down 9.41 percent today disclosed 155 GLM-5.3 responses leaking fragments of other customers' prompts, pinned the failure on Cloudflare's own cache configuration and the SGLang engine, not the model, and watched the headline land on Z.AI anyway 5. That wrapper logic now sits inside AWS.
The tape is already voting
Z.AI's Singapore-listed shares closed at S$0.795 on October 30, down 8.62% in the session after the call, on a listing with no traded options 6. That day is small against the bigger slide: the stock trades near $32 billion after an 81% fall from its June high 5. Chosun reported Chinese AI stocks deflating as semiconductor shares plummet on November 1 7. The STAR 50 index sits 40% below its July high, about $480 billion erased, and Morgan Stanley−2.74% — Morgan Stanley, down 2.74 percent today expects another 10% to 15% of downside in offshore Chinese stocks 8.
China's AI names sit 40% to 50% below their recent highs; Z.AI just had its worst day
- Estimate
Data
| Value | |
|---|---|
| STAR 50 index | -40% |
| KSTR ETF, US-listed | -42.63% |
| Unitree, robots (estimate) | -50% |
| Moore Threads, GPUs (estimate) | -50% |
| Z.AI, one session | -8.62% |
The case for the model
The other side has names on it. Nvidia−3.89% — Nvidia, down 3.89 percent today chief Jensen Huang called Chinese models "excellent" and argued their adoption raises demand for his chips 1. Eric Schmidt put China six months behind the US but warned free open-weight models could win developing markets 1.
Tang Jie told the World Laureates Forum that recursive self-improvement, AI improving AI, is Z.ai's "main research problem" on the road to AGI, and that GLM-5.3 helped optimize its own inference infrastructure for Chinese chips at a scale not done before 2. One model, serving Western shelves and the domestic silicon bloc.
Two tickets, two assumptions
Two things settle the seat: whether a Western open model reaches Bedrock's shelf inside Jassy's two-to-three-year window 1, and whether GLM 5.3 survives the next round of security studies 34. The next scheduled readings are the January free float of about 178 million Z.AI shares and the next red-team report 5.
A buyer of HZAD.SI holds the seat and the risk of losing it, and assumes cheap Chinese open weights keep winning Western shelves. A buyer of AMZN holds the toll booth and collects whichever model sits there, including the Western one its own CEO is waiting for 1.
Get Model by email, free
The AI labs, their models, and the business behind them. In our daily email, every weekday.
