Amazon weighs an $8B SPV to move Grace Blackwells off its books
A chip-financing story leads the morning: Amazon is reportedly exploring the same off-balance-sheet trick the GPU-leasing world has been leaning on all year. Elsewhere, arXiv pulls the emergency brake on its flooded submission queue, and ServiceNow publishes how it manufactures training data for enterprise agents.
Amazon has held talks with investors about spinning roughly $8 billion of Nvidia Grace Blackwell chips into a special-purpose vehicle and leasing them back for its US data centers. The Financial Times reported the discussions, which people familiar with the matter said have taken place in recent weeks; Bloomberg separately reported that the company is seeking to move the chips off its books. The vehicle would raise money from outside investors through debt and offer an equity stake of up to 10%, while the hardware — thousands of Grace Blackwell units deployed across more than a dozen US data centers in five states, including Nevada and Virginia, per the report — stays exactly where it is. Amazon and Nvidia have not commented. The pattern should look familiar: Broadcom's $42 billion loan to Anthropic rests on the same logic — Broadcom lends Anthropic $42B to lease the chips it designs — and it tells you how heavy AI capital spending has become when the world's largest cloud operator wants depreciating GPUs off the balance sheet while keeping every one of them running.
arXiv has capped every submitter at two new preprints per calendar month, effective October 1, blaming AI tools for a record flood of submissions. The announcement, from arXiv's Kat Boboris, puts September's intake at 40,363 submissions — double the same month two years ago and four times its 2016 level — which generated almost 9,000 support tickets for staff and volunteer moderators, with the cs.AI category alone growing more than six-fold over two years. The cap, paired with a three-active-submission limit, is explicitly a stopgap while arXiv upgrades its moderation tooling, and it lands weeks after the platform locked in its funding as an independent nonprofit — we covered that in arXiv just funded its independence — as AI floods its queue. Our read: this is the first hard constraint AI has placed on the machinery of science itself — when generation is nearly free, the queue becomes the bottleneck, and now every author waits behind it.
ServiceNow's CoreAI lab published AutoSynthData, a pipeline that turns an agent model's own failures into synthetic training tasks. It runs the target model in an environment, has a stronger teacher solve what the target cannot, distills that gap into capability specifications, then generates and verifies new tasks at scale; on ServiceNow's EnterpriseOps Gym benchmark, the team reports a fine-tuned Gemma checkpoint gaining 7.2 percentage points in pass rate — a 35% relative improvement — from 2,000 samples generated in about 18 hours, with a second domain improving from 18.77% to 27.18%. Those figures are self-reported and not yet replicated, but the underlying benchmark is independently run by Artificial Analysis, and if the method holds, it argues the scarce input for training enterprise agents is no longer raw data but well-verified tasks sitting right at the model's ability edge.
What to watch: whether Amazon confirms the structure or any investors bite — the first cloud SPV at this scale would turn GPU off-loading into a template rather than a one-off.
Is a two-preprint-a-month cap a reasonable line to hold against AI slop, or is arXiv fixing its moderation budget problem on its authors' backs? Tell us in the comments.
