Carney wants a Financial Stability Board for AI
Monday's governance beat produced something the pacing debate has been missing: an actual institutional design. Then, a workstation card aimed squarely at local inference, and a look inside the cost culture behind China's cheapest humanoids.
Canadian Prime Minister Mark Carney called for a global "technology stability board" to oversee AI, modelled explicitly on the Financial Stability Board that coordinates financial regulation. "There's a need for coordination," Carney said in a Bloomberg interview on Monday, adding that a body along the FSB's lines "would make sense." The timing is the story: the proposal lands while Washington has dismissed the slowdown asks from Anthropic's Dario Amodei and other lab leaders, and as Carney opens a Toronto summit pitching more than 160 national projects — including a C$14.5 billion data-center campus in Alberta — to over 100 global investors. Our read: this is the first G7 government to put a governance structure on the table rather than a position, and a country with no frontier lab has the strongest claim to neutral ground. The referee argument this site has tracked all week just got its first serious draft, from the least obvious author.
Nvidia has listed the RTX PRO 5500 Blackwell Workstation Edition: a professional card with 84 GB of GDDR7 and the same 21,760 CUDA cores as the GeForce RTX 5090. Per Nvidia's own spec page, it runs at 600 W, delivers up to 1,398 GB/s of memory bandwidth — less than the 5090's 1,792 GB/s — supports Multi-Instance GPU partitioning into two isolated 42 GB instances, and ships in air- or liquid-cooled form for rack-mounted workstation deployment. Status is "coming soon," with no published price; the RTX PRO 6000 flagship, by contrast, now lists at $16,000 after doubling — Nvidia doubles RTX PRO 6000 Blackwell price to $16,000. The read for anyone running models locally: Nvidia is selling desktop VRAM capacity as enterprise IT while consumer cards are sold out at triple MSRP, and MIG partitioning means one 84 GB board can serve two people — the economics of a shared box, not a hobby rig.
Unitree founder Wang Xingxing personally signs off on any expense reimbursement above 100 yuan (about $15) and decides details down to screw lengths and material colours, according to a Caijing Magazine feature translated by ChinaTalk and reported by Ars Technica on Monday. The profile argues that exact obsession produced the cheapest walking robots on the market — a G1 humanoid at $13,500 and an R1 at $4,900 — but at a cost: employees describe an "extremely high" early rate of returns for repair, a penalty-heavy incentive system that scored every senior executive 1 on a 0-to-1.5 scale, and the highest attrition of core staff in the company's history across 2025 and 2026. Unitree has called the reporting "misinformation" without specifics. The financial backdrop is already rough — shares sit about 50 percent below the post-listing peak near a $30 billion valuation, which we covered in Unitree loses half its value as robot hype meets its GPT-2-era reality, and Chinese regulators have reportedly signalled that future humanoid IPO approvals should go to firms with recurring revenue.
What to watch: whether a second government backs the stability-board idea before the next multilateral AI meeting, and whether Nvidia prices the 5500 as a 5090 replacement or a data-center-lite product.
If a global AI referee existed tomorrow, whose standards should it enforce — the labs', the states', or an independent panel's? Tell us in the comments.
