Bridgewater's Greg Jensen sets his AI compute oversight threshold at 5% of global capacity

A SIFI designation is a claim on a balance sheet. Capital requirements make a firm hold loss-absorbing equity against exposures a supervisor has measured, and stress tests run those exposures through a scenario the supervisor wrote ^[3]. Antitrust asks who owns what. Jensen's threshold, reported by Crypto Briefing after an interview with The Information, borrows the first set of tools for a market whose unit of exposure has not been settled ^[1][8].

Take the 5% line literally and it describes a world with at least twenty holders of compute ^[1]. Jensen pointed to projections that OpenAI and Anthropic together reach 35% to 50% of global capacity within roughly two years ^[2]. That is an average of 17.5% to 25% each, three and a half to five times the threshold ^[2]. If the pair land at the top of that range, the other half of world capacity would have to be split among ten or more owners for nobody else to be designated ^[3].

Those two labs are the companies the account puts most directly in scope, and both are privately held, backed by some of the largest technology and venture investors ^[5]. It does not identify anyone else who would cross 5%. So the designation net as described falls on model developers, and a supervisor's first job would be measuring capacity that does not appear on a public balance sheet.

Definition decides the designee list. Crypto Briefing sets out four candidates: raw chip capacity, data centre square footage, model training runs, and inference throughput, each producing a different map of who controls what ^[7]. If the line lands on chip capacity, the designees are whoever owns the hardware. If it lands on training runs or inference throughput, the tenants are caught. Bank SIFI standards run through the Financial Stability Board and national implementation; compute has no equivalent body ^[9].

Bridgewater set up AIA Labs in 2023, and the unit now has more than 80 staff and about $4.5 billion under management in AI-driven strategies ^[10]. The firm has said the unit operates with deliberate human oversight built into its processes ^[11].

In my view the analogy breaks at the transmission channel. A bank is systemic because its liabilities are other people's money, redeemable on demand, which is why a capital charge is the right instrument. A compute holder that stops serving strands customers instead of triggering redemptions, and the part of the SIFI toolkit that answers continuity is resolution planning, not equity ^[3]. The stronger half of Jensen's case is the stress test, which is how a supervisor would find out who cannot switch providers and how quickly. I would drop the continuity reading if the capacity turns out to be financed with debt secured on the hardware and held across the same backers, because then a failure does travel through balance sheets.

Jensen's chosen comparison is February 2020, the quiet weeks before COVID-19 became a global emergency, and his argument as reported is that the window for building structures in advance is narrowing and that regulation arriving after an incident costs more and works less well ^[4][12]. Crypto Briefing also notes that concentration limits would complicate both labs' growth at a moment when each is making substantial infrastructure commitments ^[6]. The cheapest test of the proposal is the counting rule: the first official document that puts a share number on compute will name an industry, and it may not be the one Jensen named.