IMF warns an AI share crash threatens global financial stability
The fund's half-yearly stability report names an AI asset-price crash as one of three near-term threats to the global financial system. Also today: Amazon crosses a million warehouse robots, and the vLLM community goes open-weights on Jev's turf.
The IMF's October Global Financial Stability Report says an AI share-price correction is a plausible near-term catalyst for a sharp tightening of financial conditions — "with global spillovers." Released Tuesday during the fund's Annual Meetings in Bangkok, the report names three salient risks: an AI asset-price correction, rising government borrowing costs, and a large cyber-attack. On AI, the fund points at the leverage stacked around the sector — roughly $300 billion of equity and $650 billion of debt flowing between hyperscalers and weaker AI firms in exchange for $1.4 trillion of future revenue commitments, across 14 circular-financing deals since 2023 — plus hedge-fund positions and retail investors borrowing to bet on AI stocks, all of which could force selling into a downturn. The IMF also cited "growing evidence" that AI's payoff may not require spending on the current hyperscaler scale, and estimated that a stress scenario would push banks holding 22% of global assets below their regulatory capital limits. Read plainly: the fund isn't predicting a crash, it's saying the plumbing now transmits one — the same worry we made the market-side case for in AI stocks will correct even if the boom is real.
Amazon now runs more than 1 million robots in its warehouses, working alongside 1.5 million humans. Ars Technica's deep dive into the company's robotics program is the best read on where physical AI actually stands: robotic fulfillment centers cut severe injuries by 40% but raised minor and strain injuries by 77%, Amazon keeps about 30% more maintenance and engineering staff at robotized sites, and its fleet of robotic arms — still only in the single-digit thousands — is set to more than double this year, per the company's July earnings call. The staffing math is the open question: internal projections Ars cites point to 160,000 fewer US hires by 2027. The largest robot deployment on earth is no longer a pilot program; it's shaping the hiring plan.
The vLLM community has open-sourced Decision 3.0, a family of six "decision" models that classify, score and route instead of writing prose — the first to read images and video with open weights. The Apache 2.0 checkpoints run from 0.6 billion to 27 billion parameters, the flagship built on Qwen3.8, and the project's card claims the top of the Jev Decision Index in both text and vision — with the caveat that those d3 scores are labeled internal evaluation and no independent reproduction exists yet. This is the first serious open-weights answer to a category Jev defined as a hosted API, and we saw the wave building in Open Source Radar — September 21: decision models go local. Reported medians of 55 ms for a text decision on a single GPU suggest these become cheap routing layers inside multi-model stacks rather than chatbot contenders.
What to watch: whether other central banks echo the IMF's framing before hyperscaler earnings — and whether Jev's index gets independently rerun now that open weights claim to beat it.
Is an AI asset-price correction a question of when, not if? Tell us in the comments.
