Nvidia backs $500B data center deal with GPU value guarantee
The AI buildout's biggest financing machine just got its formal terms: Nvidia is guaranteeing part of the resale value of aging GPUs to keep half a trillion dollars of institutional capital flowing into data centers. Also in the feed: Ant Group's smallest open model yet punches above its weight, and Rubrik's month with Anthropic's restricted Mythos model redrew how it fixes security flaws.
Nvidia has put a price floor under aging GPUs to keep its $500 billion data-center financing machine running — and that guarantee is what spooked the bond market. The company announced this week that Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR will commit up to $500 billion through new AI Compute Infrastructure Financing Platforms, with Nvidia promising to cover up to 25 percent of any shortfall when chips used as loan collateral lose value faster than expected.
Huang took to X and business TV to calm credit markets rattled by the terms, arguing the plan brings "independent, long-term institutional capital" into the buildout — a direct answer to the circular-financing charge. Bloomberg counts another $750 billion in similar deals Nvidia has been structuring this summer, and the company has already backed buyers including OpenAI, Anthropic and CoreWeave, the neocloud that originated using Nvidia chips as collateral.
The deeper play is a secondary market for used AI hardware: with a guaranteed floor price, aging GPUs become financeable, liquidatable assets, which keeps demand for Nvidia silicon alive as it ages and hands startups and researchers a cheaper tier of compute. The danger is what financiers call "wrong way" risk — Nvidia's obligations would grow exactly when AI demand weakens — which is why the Lucent-era comparison keeps surfacing. Huang's counter is that AI factories should behave like railroads, not depreciating PCs; whether Wall Street buys it is now the market's argument.
Ant Group's inclusionAI released Ling 3.0 Flash, which Artificial Analysis ranks as the smartest open model under 124 billion total parameters — and it ships under an MIT license. The model scores 38 points on the AA Intelligence Index, matching Qwen3.6 27B with far fewer active parameters while trailing only DeepSeek V4 Flash (52) in its class, and its hallucination rate on the AA Omniscience test dropped from 97 to 44 percent versus the previous version, largely by refusing questions it can't answer reliably. It is live on the inclusionAI API, DeepInfra and Hugging Face, and undercuts comparably capable models on per-token price even though it burns more tokens on complex tasks — another sign of how fast the open-model price-performance curve is moving.
Rubrik says a month of scanning its code with Anthropic's Mythos Preview surfaced so many potential security issues that it rebuilt its remediation pipeline instead of hiring more reviewers. The data-resilience company, which got access through Anthropic's restricted Project Glasswing program, found most of the work went into the "harness" — the software layer that manages the model's tool calls and encodes where trust boundaries sit — leading CTO Arvind Nithrakashyap to conclude that "prompts drift; harness architecture doesn't." Rubrik deliberately automated only a subset of vulnerability classes and kept engineers on everything else, a stance he sums up as "in security, trustworthy automation and maximum automation pull in different directions"; Anthropic's program has logged more than 10,000 high- or critical-severity findings from partners, with a 90.6 percent true positive rate.
What to watch: whether the value guarantee actually attracts the institutional capital Nvidia is courting, and how the used-GPU market it implies prices aging hardware.
Does Nvidia's guarantee make the AI buildout more stable, or just better collateralized? Tell us in the comments.
Sources: TechCrunch · NVIDIA Newsroom · NVIDIA Blog · The Decoder · Ling 3.0 Flash (Hugging Face) · SiliconANGLE
