Yvette Teo

Open weights are now a security decision, and the hub has a new owner

24 August to 30 August 2026

Two themes ran through the week, and they are more connected than they look. A Chinese lab ran a stealth test of a frontier-grade model, open-sourced it, then held back its cyber-capable cousin. And the money behind AI and robotics turned from cheques to debt. Both are worth an operator’s attention.

Quick terms, so the rest is clear:

  • Multimodal means the model can understand and generate more than plain text, including images, audio and video.
  • A neocloud is an AI-focused cloud company, usually independent of the big hyperscalers, that rents out computing power for training and running models; Lambda is one.

Solved: Z.ai’s GLM-5.3-Flash was the anonymous model

Last week’s anonymous Ox Alpha turned out to be Z.ai’s GLM-5.3-Flash: released under MIT, run on Chinese chips, 1M-token context, at $0.075 in / $0.25 out per million tokens ($0.15/$0.50 at list).

The playbook matters more than the model. Z.ai launched it anonymised, watched it top OpenRouter for a week, then revealed themselves and shipped the weights: a label-blind product test converted into an open release. The anonymous launch lane is now a recognised way to ship.

The same week, the same lab gated its more capable GLM-5.3 while hardening finished. Its review logged 2,436 vulnerabilities and ExploitBench jumped from 24.4% to 54.4%. Open weights are now a security decision as much as a licensing one: the gate can move, the licence can change, the branch can die.

The platform that hosts them changed owners this week: Nvidia agreed to buy Hugging Face for about $13 billion, putting the hub where open weights get published in the hands of the company that sells the silicon those models run on. If it closes, the answer to who sets the open-weights rules starts with Nvidia.

Money rotated to physical AI, and turned to debt

The capital this week was structural, not spare change. Xpeng Robotics, the robotics arm of EV maker Xpeng, raised over $900 million at above a $6.3 billion valuation, a record for Chinese embodied AI, with Tencent and Alibaba both writing checks.

And the financing shifted from equity to debt. Lambda, the NVIDIA-backed neocloud, secured a $1 billion debt facility arranged by JPMorgan to buy chips, part of a year in which more than $400 billion of AI buildout debt is on the books. That same week Lambda was reportedly in talks for a $3 billion pre-IPO round, after raising $1.5 billion at a $5.43 billion valuation last November: debt to buy the hardware now, equity to fund the company toward an IPO later.

When infrastructure financing moves from equity cheques to interest-bearing loans, the question stops being who writes the cheque and becomes who services the debt if compute demand wobbles.

Bitcoin crossed $80,000, and the dollar fought back

Bitcoin crossed $80,000 for the first time since mid-May on the strongest spot-ETF inflow streak in months, and it had a weak dollar to thank. DXY spent the week testing support near 98.7, the debasement trade from last week still running.

Then Friday tested that read. Fed Chair Kevin Warsh’s hawkish Jackson Hole remarks put a September rate hike back on the table. Markets repriced the odds from about 36% to 56%, the dollar jumped to roughly 99.5, and bitcoin faded back under $80,000 to close the week.

For an operator the lesson cuts both ways. The rotation into physical AI and the shift to debt were structural and still stand, but the speculative leg of the debasement trade just learned that the dollar’s floor is not guaranteed. When the confirmation wobbles the same week it happens, treat it as a risk to price in, not a thesis to retire.

One question I am carrying into next week: the same week a frontier model could be built and run entirely on Chinese chips, the West’s AI buildout leaned harder on loans than ever. Does the ability to train frontier models without Western silicon change who sets the open-weights rules, and does the debt shift change who carries the risk?

Sources