How much does your chip cost to manufacture?

Strip a rented GPU-hour down to its cost basis and the hardware barely matters. At 70% fleet utilization, four-year amortization of a fully burdened H100 server comes to $1.12/hr, and energy plus facility adds another $0.33/hr — against a $3.99 on-demand price at Lambda. The residual $2.54/hr — 64% of the price — is demand premium: scarcity rent, not cost recovery (SA estimate). That premium splits the market into two pricing regimes. Neoclouds price per unit of compute: Lambda's B200 rents at 1.38x its H100 price for 2.27x the FP8 (sparse) throughput — about 39% cheaper per effective PFLOP. Hyperscalers price scarcity: AWS capacity blocks put B200 at 2.38x H100 — per-PFLOP parity, meaning the newest silicon carries zero performance discount. How fast reserved-tier discounts erode that premium is the most useful forward indicator of AI compute supply catching up with demand.

The four US hyperscalers — Microsoft, Alphabet, Amazon, and Meta — purchased $433.9B of property and equipment in the four quarters through March 2026 (SA estimate: sum of reported cash-flow figures), against roughly $149B of reported depreciation over the same span. Quarterly combined capex reached $129.8B in 1Q26, up 80% year over year, and company guidance for calendar 2026 sums to roughly $700B of intended spend. The capex-to-depreciation gap is structural: depreciation recognizes spend over 5-6 year server schedules and 25-40 year building schedules, so today's income statements carry only a fraction of today's build-out. That wave arrives regardless of what AI revenue does. It also explains the borrowing: Meta priced $30B in October 2025 (the largest investment-grade deal of the year) plus a ~$27B off-balance-sheet SPV, Alphabet raised ~$25B in November 2025 and ~$31B more in February 2026, and Amazon added $24.9B this week after $15B in November — all while generating record operating cash flow. Amazon's trailing-4Q capex now exceeds its trailing-4Q operating cash flow (102%, SA-computed). For the supplier chain, buyer-side capex is the demand signal; the depreciation drag is the 2027-2028 risk to its durability.

The AI infrastructure battleground has shifted from training throughput to inference unit economics. OpenAI's Jalapeño — a custom inference accelerator built with Broadcom — is not primarily a competitive strike against NVIDIA; it is a structural bet that owning inference silicon is the only way to make gigawatt-scale LLM deployment economically sustainable. Enterprise buyers who treat GPU procurement as their only inference lever are already behind the curve.

Terafab is not a Tesla fab — it is an Intel foundry engagement dressed in Musk-brand capital formation, with Tesla and SpaceX providing the demand anchor that Intel's foundry turnaround desperately needs. The 14A licensing arrangement tests a structurally new US fab model: committed captive demand substituting for the open-market customer base that eluded Intel Foundry Services. Whether the model survives execution depends on talent retention, CHIPS Act disbursement timing, and whether Gary Jiang's operational playbook from Tesla can translate into semiconductor manufacturing discipline.