AI Capex Inflation Bites As Frontier Valuations Defy Gravity

The Q1 2026 earnings cycle produced an unusually clean verdict. Alphabet, Amazon, and Microsoft were all rewarded by markets after reporting AI infrastructure investment constrained by capacity rather than demand — meaning their returns would have been higher with more hardware online, validating continued aggressive spend. Meta was penalised despite record revenue growth precisely because its $125–145 billion capex revision, accompanied by an explicit acknowledgement of higher component pricing, could not be anchored to an equivalent enterprise monetisation story. The 20 million paid Microsoft Copilot users and Google Cloud's $20 billion quarterly milestone gave investors concrete throughput metrics; Meta offered a build-out rationale without a comparable demand signal.

The supply-side implication is at least as significant as the investor reaction. Samsung's 48-fold chip profit surge and Murata's data center-driven beat confirm that the component stack from passive components through HBM is in a genuine supply-demand imbalance that is extracting rents from every hyperscaler simultaneously. Meta's disclosure names this dynamic explicitly. For infrastructure planners, two parallel constraints are now confirmed: GPU and HBM scarcity at the accelerator layer, and — as evidenced by Meta's multi-billion-dollar Graviton5 deal with AWS — high-core-count CPU availability for agentic inference workloads. The capex cycle is not decelerating; it is accelerating into a tighter and more expensive component environment.

Anthropic's reported $850–900 billion funding discussions and SoftBank's simultaneous $40 billion leveraged bridge loan on its OpenAI stake represent two ends of the same repricing event: frontier AI model developers are being valued on the assumption that the market will consolidate around two or three players, and that assumption is now being financed with real leverage. Hut 8's $3.25 billion bond issuance to fund the River Bend campus — with Anthropic and Google as anchor tenants — completes a three-part picture in which capital risk is migrating from hyperscaler balance sheets to infrastructure operators, debt markets, and levered holding vehicles. Each layer is individually defensible; the aggregate concentration on AI valuations holding is not.

OpenAI's infrastructure pivot reinforces the pattern. By preferring leased compute to owned assets, OpenAI has transferred stranded-asset risk to third-party operators and their bond investors while retaining operational flexibility. The Stargate headline capacity numbers remain real in the sense that OpenAI can access the compute — but the financial resilience of that capacity now depends on the solvency of a fragmented set of counterparties rather than OpenAI's own balance sheet. If demand growth disappoints or interest rate conditions tighten, the infrastructure operator class — newer, less diversified, and more leveraged than the hyperscalers — represents the most exposed node in the AI capital stack.

China's domestic AI chip ecosystem is crossing a commercially meaningful threshold. Cambricon's doubling revenue, Huawei Ascend's growing role as DeepSeek's preferred compute platform, and SenseTime's deliberate release of an open-source image model tuned for domestic silicon collectively describe a reinforcing loop: policy-backed procurement creates demand, commercial revenue validates the ecosystem, and model-hardware co-optimisation closes the performance gap at the inference layer. Lisuan Tech's WHQL certification — making it only the fourth GPU maker globally to hold the designation — adds a software ecosystem milestone that the training compute gap cannot offset. Export controls remain effective at slowing frontier training capacity additions; they are proving insufficient to prevent a competitive inference and deployment ecosystem from emerging.

The sovereign infrastructure picture is equally differentiated among US allies. Japan's NTT-Rapidus partnership represents a credible manufacturing anchor — Rapidus is targeting 2nm production with a live infrastructure customer validating the roadmap, and TSMC's Kumamoto fab provides a near-term backstop. The UK's announced AI Hardware Plan has no equivalent manufacturing anchor: it enters the race from a demand-side position without a domestic fab, packaging ecosystem, or chip champion. For multinational enterprises making long-range infrastructure sourcing and compliance decisions, the bifurcation is now structural, not theoretical — procurement teams need to model both a US-aligned and a Chinese-domestic scenario for AI infrastructure dependency.