Nvidia's Capex Leverage Peaked; Power Grid Is Now the Bottleneck

Nvidia's capex-cycle leverage peaked when AI infrastructure demand hit power grids, not GPU supply—validating AMD alternatives and forcing hyperscalers to own infrastructure.

Nvidia responded to supply-chain pluralism this week with a $250 billion financing guarantee for OpenAI's Ohio buildout and $50 billion in signed Texas leases. These are not dominance signals; they are ceiling signals. Nvidia has held capex leverage for three years by controlling supply in an insatiable market. But supply-chain pluralism is now real. Anthropic committed 2 gigawatts of AMD Instinct capacity (with $5 billion AMD co-investment), OpenAI's Stargate project built around AMD Helios for 6 gigawatts, and Meta deployed custom AMD MI450 GPUs in Temple, Texas as lead customer for another 6-gigawatt rollout. AMD's Helios racks deliver 15% higher compute density than Nvidia Vera Rubin, and cost structures are comparable. That is 13+ gigawatts of named, production-validated AMD capacity—structural market share loss that financing guarantees cannot repair.

The real constraint is no longer GPUs but power. A backlog of $130 billion in AI data centers is stalled by grid consent and interconnection bottlenecks, not capital or equipment shortages. Alphabet's $205 billion capex commitment, OpenAI's $750 billion Stargate forecast, and Meta's $14 billion El Paso venture are all competing for the same scarce resource: geography with available power and regulatory approval. Nuclear power is bankable again. Institutional capital—MGX, AIP, BlackRock's GIP—is consolidating the data center asset class via the Aligned Data Centers $40 billion acquisition on assumption of 5-10 year holds. These investors are betting on grid-constrained real estate, not accelerator brands.

Hyperscalers have stopped renting compute and started building owned infrastructure. Meta brought 1.2 gigawatts online in Temple; OpenAI locked Georgia Power into long-term PPAs through 2032; OpenAI Stargate includes owned Helios racks, not third-party GPU cloud access. This ownership shift eliminates cloud provider optionality and transfers grid risk to hyperscalers. Applied Digital's $36 billion backlog signals GPU cloud providers remain viable, but only if they can secure power and permitting—constraints they cannot control. The capital entering data centers is locking into 5-10 year utility-like yields on grid-constrained assets, not venture returns.

Winners: AMD (validated structural competitor), land and power infrastructure (scarcity premiums), and institutional capital (consolidating grid assets). Losers: Nvidia (leverage declining as alternatives scale), cloud GPU providers (competing for power, not accelerators), and hyperscaler capex velocity (power bottlenecks limit deployment). OpenAI's CFO already signaled the peak: $750 billion capex forecast paired with "uncertainty on financing and repayment viability." That is a capex cycle ceiling. Nvidia will not collapse—it has 80%+ installed base and switching costs are real—but its ability to command capex cycle premiums has ended.

Watch three metrics into Q4 2026: power interconnection queue depth (are $130 billion of stalled projects moving?), institutional data center M&A flow (will Aligned precedent repeat?), and hyperscaler capex guidance revisions (will CFO uncertainty spread?). Custom silicon economics (Amazon's $20 billion play) and AMD production validation are structural; but if power constraints or LLM monetization stalls, the $1 trillion capex narrative compresses sharply. The capex cycle peaked not on technology but on grid scarcity.