AI Capex Hits Utility Scale: Power Scarcity Forces Vertical Integration
The mobilization of $500 billion in private capital via Goldman Sachs' consortium with Apollo, BlackRock, and KKR signals that AI infrastructure financing has exited venture speculation and entered utility-scale securitization. This capital influx, paired with Nvidia's $3 billion equity stake in power developer Lancium, confirms that chipmakers can no longer fund demand alone; they must anchor energy supply to unlock compute value. Simultaneously, Riot Platforms' $9.1 billion binding lease with Anthropic for 191 MW validates that specialized neocloud financing models have achieved parity with hyperscaler balance sheets, proving that gigawatt-scale capex can be secured through long-term off-take agreements rather than pure equity dilution.
Power scarcity has superseded chip allocation as the primary bottleneck, driving aggressive vertical integration among infrastructure winners. Amazon's construction of a 7.65-gigawatt natural gas plant authorizing 33 million tons of annual CO₂ emissions demonstrates that megawatt economics now force data center operators to own on-site generation to secure grid access. This dynamic rewards consolidated power-constrained suppliers like Core Scientific, which is advancing its Muskogee facility to 1.5 GW following a $444 million acquisition of Polaris DS. Meanwhile, JLL's report of 25 GW of record US absorption against a 1% vacancy rate underscores a structural supply-demand imbalance that will sustain premium lease-rate economics through 2027, further entrenching incumbents with pre-secured power permits.
Within the hardware stack, the winners are those controlling bottlenecks beyond the GPU core itself. Broadcom's $56 billion full-year AI revenue guidance for 2026, up 180% year-over-year, establishes cluster networking as a structural $50 billion-plus market, proving interconnect fabric is no longer commoditized. In memory, SK Hynix's approval of $38.1 billion in capital expenditure for two new fabs highlights HBM scarcity as a multi-year scaling constraint requiring massive fab-level commitment. Conversely, CoreWeave's $2.58 billion Q2 revenue and $104 billion backlog validate that legacy A100 GPUs remain highly profitable nine years post-launch, de-risking neocloud unit economics through high utilization on existing fleets while Cerebras captures structural fragmentation in inference by powering OpenAI's Ultrafast mode with 14x speed gains.
Despite the capital inflows, capital discipline is tightening, exposing leverage risks for operators relying on subsidized guarantees. Nvidia's reduction of its financial guarantee commitment for OpenAI's upcoming complex marks a pivot where chipmakers are shifting financing risk back to cloud operators, signaling that near-term margin cycles cannot indefinitely subsidize hyperscaler buildouts. This pressure is visible in neocloud equity performance, where CoreWeave, Nebius, and others face valuation headwinds despite strong fundamentals, as Wall Street begins pricing in the cost of capital within these new securitization structures. Even top-tier operators face margin erosion from energy costs, as evidenced by SpaceX's soaring revenue overshadowed by billions in AI spending burn, reinforcing that infrastructure density is becoming a liability without efficient power procurement. Investors should monitor whether the $500 billion financing vehicles successfully lower borrowing costs enough to offset rising energy premiums, or if the shift toward asset-backed debt triggers margin compression across the hosting layer as interest rate exposure becomes embedded in per-GPU pricing.