Nvidia Conquers the Stack While AI Returns Fall Short

This week's Computex announcements collectively represent something more than a product cycle: Nvidia is executing vertical and horizontal platform expansion simultaneously. The Vera CPU wins at Anthropic, OpenAI, and SpaceX close the CPU socket alongside the GPU, eliminating the remaining architectural opening for Intel and AMD in AI-native data centres. The RTX Spark PC platform with 128GB unified memory extends the same silicon authority to the edge device layer, directly challenging Qualcomm and Apple Silicon on their own ground. The Unitree robotics partnership seeds Nvidia's software stack into physical AI at the research level. Each move compounds the others: a developer ecosystem built on Nvidia's data centre architecture now extends naturally to Nvidia-powered laptops and Nvidia-seeded robotics platforms.

The supply chain implications are under-appreciated. This degree of platform consolidation concentrates semiconductor dependency on TSMC's advanced node capacity and on HBM memory from a supplier base that is already structurally tight and now partially equity-aligned with a single frontier lab. The AI hardware stack is verticalising faster and more completely than anything seen during the cloud buildout era — and the distributed, multi-vendor resilience model that made cloud infrastructure robust is being traded for integration efficiency at a pace that leaves regulators and competitors with little time to respond.

SoftBank's commitment to France is the clearest public confirmation yet of a dynamic infrastructure analysts have been tracking for 18 months: site selection for hyperscale AI capacity is now following electrons rather than economics. US markets including Northern Virginia, Phoenix, and Dallas face grid interconnection queues measured in years, making greenfield capacity effectively unavailable within planning horizons. France's 70-plus percent nuclear generation share offers dispatchable baseload at scale with lower carbon intensity — the combination that no US market can currently replicate. The strategic implication extends beyond this single commitment: European nations with nuclear infrastructure are acquiring a durable competitive advantage in AI compute hosting that will compound as US constraints worsen, and governments that understand this are accelerating reactor programmes partly on AI hosting economics.

The financial context demands scrutiny. SoftBank carries over $130 billion in debt, and the gap between announced AI infrastructure commitments and capital actually deployed has been substantial across the industry. But even as an announced intention rather than a confirmed build, the France commitment validates the thesis for other capital allocators and will accelerate competition among European nations to offer power access as an AI investment incentive. The SoftBank AI capital structure — which also includes the OpenAI bridge loan and a portfolio of AI equity positions — is itself an open question for investors assessing whether the infrastructure ambition is funded by durable capital or leveraged positioning.

The pattern across today's reporting is consistent and reinforcing. Infrastructure-layer AI investments are generating measurable, confirmed returns: SK Hynix has rallied approximately 1,000% over the prior year, South Korea's export growth hit a four-decade high on AI memory demand, SoftBank overtook Toyota on Arm exposure, and TSMC's local shares are narrowing their discount to the ADR. At the same time, Bain's global survey of large companies finds that AI cost savings are broadly missing projections at the enterprise level — investment is committed but proportionate returns have not arrived. This divergence is not new, but the Bain data suggests it is widening rather than closing as agentic AI deployment, the mechanism most often cited as the path to scale returns, remains predominantly in early stages.

The export control dimension adds a further structural complexity. The confirmed move to halt Nvidia chip shipments to Chinese firms operating outside China narrows the addressable market for the highest-margin data centre products precisely as new segment revenues — PC silicon, robotics — are still being established. For capital allocators, this creates a tactical case for continued infrastructure overweight while building a medium-term contrarian case for AI application vendors that can demonstrably close the ROI gap. China's parallel acceleration of domestic AI capital formation through STAR Market listings and domestic venture funding suggests the global AI investment ecosystem is bifurcating into distinct pools with separate valuation benchmarks — a structural shift that compounds the challenge of pricing AI opportunity on a globally-integrated basis.