Semiconductors

AI CapEx supercycle is rewriting semiconductor capital formation

Intel's first equity raise since its 1979 IPO — a signal cited as evidence of the AI infrastructure supercycle's unprecedented scale — coincides with $500B in private debt from Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR flowing into AI compute, and a $55B TerraFab semiconductor consortium raise. Bond market cover for AI infrastructure debt has already fallen from 5x to below 2x, with 86% of bonds trading at higher yields than issuance, suggesting the market is beginning to price in execution risk. Amazon leads hyperscaler AI capex at $220B in 2026 guidance alone, while GPU securitizations are spreading exposure to insurance companies, pension funds, and sovereign wealth funds. This is no longer venture-scale capital — it is sovereign-and-institutional-scale capital restructuring the semiconductor supply chain.

Anthropic is building in-house chip design capabilities and buying TPUs outright for its own data centers, converting compute from marginal cost to capital — a structural shift that mirrors Amazon's Trainium custom silicon strategy. Google's Tensor G6 chip powers the Pixel 11, and Google Cloud revenue has accelerated from 32% to 82% YoY with margins expanding from 21% to 36%, validating vertical silicon integration as a margin lever. Broadcom continues as a custom chip co-developer, co-signing a 5-gigawatt TPU deal with Google for Anthropic.

Why it matters · As hyperscalers vertically integrate silicon, merchant chip vendors face a shrinking addressable market at the top of the demand curve, compressing long-term pricing power.

A $2B growth round valuing Etched at $10.5B (backed by Blackstone, Jane Street, Coatue, and Nvidia) and a $1.1B Series B co-led by Nvidia and AMD Ventures underscore that inference-specific hardware is now commanding late-stage, large-cap valuations. Groq and Cerebras are publicly identified as the ongoing combatants in the inference chip fight, while XCENA raised $135M Series B at a $570M valuation targeting memory-adjacent AI inference. Tenstorrent, with over $1.8B raised, is now in takeover talks with both Intel and Qualcomm, signaling that M&A is becoming the exit path for inference chip startups.

Why it matters · The inference chip market is consolidating rapidly through both large funding rounds and M&A, meaning early-stage inference chip investment windows are closing fast.

Nvidia appears in five of the largest recent funding rounds — a $2B growth round, a $1.1B Series B, a $1.1B unknown-stage round (with AMD Ventures, General Catalyst, YC, and Temasek), and as the top investor by deal count at 43 deals in the period. Nvidia's GR00T N1 is setting the baseline for generalist robot foundation models, and the open-sourcing of Cosmos (training frameworks, synthetic data, model weights) extends Nvidia's platform lock-in beyond hardware into the physical-AI software stack. Jensen Huang's publicly documented '30 days to bankruptcy' operating mindset reflects the urgency behind this multi-front expansion.

Why it matters · Nvidia is transitioning from chip vendor to platform orchestrator across inference, robotics, and synthetic data — companies building on Nvidia's stack are deepening dependency, not reducing it.

Euclyd, backed by former ASML CEO Peter Wennink and Intel microprocessor pioneer Federico Faggin, claims up to 100x greater energy efficiency than Nvidia's chips by minimizing data movement between memory and compute — seeking €100M to scale. Schematik raised $4.6M pre-seed led by Lightspeed to generate complete hardware projects from natural language, collapsing the firmware-to-wiring design cycle. An unannounced CPU startup (signal [36]) is developing a new architecture specifically for LLM-generated code, reflecting a broader shift toward hardware designed around AI workload characteristics rather than legacy ISAs.

Why it matters · Co-design startups that can demonstrate verified efficiency gains at silicon level are becoming attractive both as standalone investments and as acquisition targets for larger semiconductor players.