The Nvidia loophole: Chinese companies can rent top chips online — Washington has known for years - NBC News

The real crack in export control isn't at the fab — it's in the cloud API layer. No silicon crossed a border, yet compute was delivered. This is structurally more dangerous than the 2018 Huawei cutoff, because it bypasses the single controllable node: physical interception. Technically, GPU rental decouples ownership from utilization. Chinese AI labs maintain training parity with Western peers without a single H100 on domestic soil. Upstream, Nvidia's data-center revenue model (selling to cloud providers, not end users) becomes the enforcement blind spot. Downstream, the urgency of domestic silicon substitution is structurally weakened. The "Washington has known for years" detail is the actual signal: either enforcement jurisdiction collapses at the cloud contract layer, or there's a deliberate cost-benefit calculus — cracking down on GPU rental would damage AWS and Azure's China enterprise revenue far more than it constrains adversarial AI. For cloud providers, kernel-level verification of end-user compute purpose is operationally near-impossible. Market dynamics: AMD and Huawei will weaponize this as proof the chip ban is performative. Chinese cloud operators accelerate compute-sovereignty positioning. US hyperscalers face a structural contradiction between revenue and compliance. 12–24 month outlook: regulatory logic shifts from controlling chips to controlling compute. The cloud layer becomes the new chokepoint. Compute will increasingly be treated as a strategic commodity, and who can run, and at what scale replaces who can buy as the core variable in the geopolitical tech game.