ACE Emerging-Tech Insight – 24 June 2025 – ≈ 6-Minute Read
NVIDIA used VivaTech to make a conceptual pivot: chips are no longer the product, “AI infrastructure is the product.” Jensen Huang rolled every flagship SDK—CUDA-X, Omniverse, NeMo, Isaac, DriveOS—into a single modular platform he called NV-Stack. Hardware becomes the substrate; the profit engine moves to recurring software licences and cloud consumption fees.
The logic is straightforward. AI workloads scale faster than Moore’s Law, but margins on GPUs compress as hyperscalers design custom accelerators. By collapsing its toolchain into an “AI operating system,” NVIDIA seeks to lock developers at the API layer, harvest network effects and turn each new model, robot or digital twin into annuity revenue. Silicon sales fund the flywheel; proprietary middleware captures the upside. Europe is the first proving ground. A dozen “AI factories”—liquid-cooled data-centre pods packing Grace-Blackwell superchips—will anchor sovereign-cloud projects for France, Germany and the Nordics. Regional firms gain instant scale for LLM training, but dependence shifts from U.S. hyperscalers to a single U.S. vendor embedded in every software layer. For Brussels this raises two questions: does NV-Stack constitute gatekeeper power under the Digital Markets Act, and how will regulators square industrial AI growth with a grid already strained by electrification targets? Strategically the platform move hardens supply-chain asymmetries. Washington can exert export-control leverage not just through advanced chips but through a software stack woven into defence, health and mobility sectors worldwide. Beijing, already stockpiling H100s, now faces the prospect that future algorithm innovation will sit behind U.S. licence keys. Expect China to accelerate open-source MLIR and RISC-V accelerators, and to court AMD and domestic fabs for leverage. The competitive counter-strategy in Silicon Valley is clear: disaggregate CUDA lock-in. Google and Intel are pushing MLIR as a universal compiler front-end; Meta sponsors PyTorch 3.0 to abstract kernel calls; AWS doubles down on Trn-series silicon plus an “exabyte-class” parameter-store that bypasses NVIDIA’s memory hierarchy. None of these moves erodes the installed base overnight, but each chips away at the inevitability narrative Jensen Huang is crafting.Watchpoints for the next 12 months
- NV-Stack public beta (Q4 2025) – Real test of developer uptake and ease-ofporting from standalone SDKs.
- First industrial AI factories online (early 2026) – BMW and Airbus pilots will show whether the economics beat bespoke hyperscale contracts.
- EU gatekeeper designation – A formal investigation would force interface unbundling and data-portability guarantees, cutting into platform stickiness.
- Energy-intensity backlash – If regional grids approach capacity, look for mandatory efficiency targets or carbon-based GPU quotas that blunt deployment speed.