When Nvidia’s CEO blessed OpenClaw from the GTC stage, two Chinese AI stocks added billions in hours. But the real story isn’t the rally — it’s the trillion-token arms race hidden inside it.
Jensen Huang has a gift for picking winners out loud. At GTC 2026 on Tuesday, he stood in front of thousands of developers and called OpenClaw “definitely the next ChatGPT” — comparing it in the same breath to the launches of Windows and HTML. By Wednesday morning, Chinese AI stocks were up by double digits and still climbing.
MiniMax surged 29% to a record high. Zhipu (Knowledge Atlas) jumped 23%. Cloud provider UCloud added 13% in Shanghai. The so-called “AI tigers” — China’s crop of frontier model startups — just got a very public endorsement from the man who sells the chips powering the entire industry.
What is OpenClaw, and why does it matter?
OpenClaw is an open-source autonomous AI agent platform built to run locally on your device. Think of it less like a chatbot and more like a digital coworker that reads your files, handles your emails, executes code, and manages tasks — without you supervising each step. Developer Peter Steinberger (formerly of PSPDFKit) launched it in late 2025, and it exploded to over 150,000 GitHub stars within weeks. In China, fans have dubbed the movement “Crustacean Summer,” a nod to OpenClaw’s lobster mascot, Molty.
NVIDIA didn’t just praise OpenClaw from the stage. The company launched NemoClaw, an enterprise-secured fork of the OpenClaw platform designed for corporate deployments with data governance controls. That announcement landed alongside Vera Rubin — Nvidia’s next-generation chip architecture, promised to cut the cost of agentic inference tokens by 10x by late 2026. The stage was set, deliberately.
The hardware problem nobody mentions
Here’s the fine print on the “run AI locally” dream: it’s currently expensive in hardware terms. Running Zhipu’s GLM-5 at full 744 billion parameters demands close to 1.5 terabytes of VRAM — a configuration that exists only in server rooms and research labs. For almost every regular user, “local AI” still means a cloud subscription, usually billed by the token. That gap between the marketing and the reality matters. But it also quietly explains why MiniMax and Zhipu stocks are soaring.
- 1×tokens used by a standard chatbot query
- 5–10×tokens consumed per agentic loop
- 10×cheaper with Vera Rubin (est. late 2026)
AI agents like OpenClaw don’t just generate one response — they “think” in loops, calling tools, checking results, correcting errors, and retrying steps. Each loop burns tokens. A task that costs pennies as a chatbot interaction can cost dimes or dollars as an agentic workflow. This is the inference inflection point: as the world shifts from chatbots to agents, the companies selling tokens become fuel suppliers to a fast-growing engine.
Regulatory frictionThe Chinese government has quietly restricted state-owned enterprises (SOEs) from deploying OpenClaw, citing data exfiltration risks from an open-source tool with foreign origins. This limits the addressable market in China’s large state sector — a nuance the stock rally hasn’t fully priced in.
The AI tigers, by the numbers
So are the AI tigers genuinely competitive, or just riding Huang’s coattails? The benchmark data suggests the former, at least in coding and engineering tasks. Zhipu’s GLM-5 currently ranks first among open-weight models on SWE-bench Verified, the industry’s most respected agentic coding benchmark. It’s not catching Claude Opus 4.5 yet, but it’s closer than most Western analysts expected.
| Model | SWE-bench verified | Active params | Key strength | Status |
|---|---|---|---|---|
| Claude Opus 4.5 Anthropic |
80.9% Gold standard | Proprietary | Coding & reasoning | Proprietary |
| GLM-5 Zhipu |
77.8% #1 open-weights | 40B active / 744B total | System engineering | Open weights |
| M2.5 MiniMax |
75.2% | Undisclosed | Office/productivity | Closed |
MiniMax’s focus on deep integration with productivity tools — Office suites, calendars, corporate workflows — makes it a practical play for enterprise OpenClaw adoption. Zhipu’s open-weights approach, meanwhile, makes GLM-5 the model most likely to power third-party OpenClaw forks and regional deployments. They’re running different races and, for now, both are winning.
What the rally actually means
The fundamentals haven’t caught up to the excitement yet. MiniMax brought in just $79 million in revenue across all of 2025 — a thin foundation under a market cap that now pushes into the hundreds of billions of Hong Kong dollars. Investors are pricing in a future where MiniMax and Zhipu are infrastructure — token utilities feeding millions of OpenClaw agents running around the clock. That future could arrive. Vera Rubin’s cost reductions, if they land as promised, could unlock agentic AI for tens of millions of users who find it too expensive today.
Moody’s noted this week that China’s rapid AI adoption reinforces its standing as one of the world’s leading AI markets. The OpenClaw ecosystem — Tencent, Alibaba, and Baidu have all rushed to offer it to enterprise customers — is growing faster than most Western observers tracked. And Huang’s GTC comments weren’t offhand praise. NVIDIA built NemoClaw, timed the Vera Rubin announcement alongside OpenClaw’s rise, and explicitly framed agents as its next decade-defining market.
The lobster revolution has Nvidia’s blessing, China’s infrastructure, and a genuine benchmark story to tell. Whether the revenue follows is the only question that matters now — and the market has decided, at least for today, that the answer is yes.
Related: 1.5 Million AI Agents, 17,000 Humans: The Security Nightmare Inside Moltbook
