Semiconductors

Hyperscalers racing to own silicon end-to-end

The clearest structural shift in semiconductors is hyperscalers eliminating third-party chip dependency through massive proprietary silicon bets. Apple's $30B chip deal with Broadcom, Anthropic's rumored multi-billion TPU procurement from Google, and Amazon's Annapurna team — credited as 'the most talented silicon team at any hyperscaler' — developing Trainium 3 all reflect the same logic: controlling the full stack from wafer to model. As signal [18] captures, 'people who love software need to build their own hardware.' This vertical integration drive is compressing the addressable market for merchant silicon vendors and elevating Broadcom, TSMC, and ASML as the irreplaceable manufacturing layer regardless of who designs the chip.

Purpose-built inference hardware is attracting the largest individual rounds in the theme: the $2.5B Series C backed by Nvidia, Sequoia, Lightspeed, JPMorgan, and B Capital at a $27.5B valuation is the clearest proof point. Tenstorrent — in acquisition talks with both Intel and Qualcomm after raising over $1.8B — and XCENA's $135M Series B at a $570M valuation reinforce that inference-specific architectures are commanding premium valuations. Euclyd's claim of up to 100x greater efficiency than Nvidia by minimizing data movement, and Maddox being led by former Google TPU architect Reiner Pope, show the talent and capital concentration in this sub-segment.

Why it matters · The inference chip race is still open enough for well-capitalized challengers, but the window is narrowing as Nvidia's platform ecosystem and hyperscaler captive silicon harden their positions.

Despite $23B+ in capital deployed across 52 deals in 28 days, TSMC and ASML remain the non-negotiable bottlenecks. TSMC's demand outpacing supply even amid tens of billions in capacity expansion, ASML's status as the sole EUV lithography provider, and Rapidus's 2nm production pilot all confirm the supply chain choke remains geographic and equipment-constrained. Building a custom AI chip reduces Nvidia dependence but not dependence on the handful of companies capable of manufacturing advanced semiconductors — a structural reality no amount of VC can quickly dissolve.

Why it matters · Any semiconductor investment thesis that ignores foundry access risk is incomplete; geopolitical disruption to TSMC or ASML export controls would cascade across every AI infrastructure bet in the portfolio.

With 28 deals in the top-investor rankings — more than Amazon (13) and Google (12) combined — Nvidia is systematically co-investing across the AI stack to make its platform the default. Co-leading the $2.5B Series C alongside Sequoia and Lightspeed, and appearing in both seed-stage ($200M Kleiner/a16z/Nvidia round) and Series C ($800M, General Catalyst/Vista) rounds, Nvidia is buying strategic optionality at every maturity level. Despite losing roughly $1 trillion in market cap in under two months per signal [16], analysts continue raising profit estimates — and at ~40x trailing earnings versus Cisco's 150-180x peak, the valuation argument for a bubble remains weak.

Why it matters · Nvidia's co-investment presence creates a self-reinforcing ecosystem moat: startups that take Nvidia capital are incentivized to build on CUDA, entrenching GPU dependence even as alternatives proliferate.

Euclyd's proprietary CRAFTWERK architecture, Etched's approach of baking neural network weights directly into hardware, and Vinci's physics-accurate AI foundation model consolidating simulation and manufacturability validation onto a single platform all represent the same thesis: meaningful efficiency gains require co-designing algorithms and silicon simultaneously. NVIDIA's Isaac Gym running 62,000 parallel environments on RTX 5090s and a locomotion policy retrained in two hours on a single RTX 5090 with 4,096 parallel environments demonstrate how tightly coupled hardware and software have become even at the application layer.

Why it matters · Pure-play chip companies without a co-developed software layer will struggle to demonstrate the order-of-magnitude efficiency gains needed to displace Nvidia's deeply integrated CUDA ecosystem.