Big Tech vs. Nvidia: The New Arms Race in AI Chips

Short take

Nvidia is the headline act in AI hardware, but the backstage has been rearranging itself. Over the last two years cloud giants and a handful of chip upstarts have pushed bespoke accelerators aimed at cutting inference costs and locking in consistent performance for large language models. For investors this is less a simple disruption story and more an ecosystem fight — margins, software lock-in, and who gets to set deployment rules are all at stake.

Why now

• Big Tech already knows how to build custom silicon: Google with TPUs, Amazon with Inferentia and Trainium, Apple with the M-series. The point isn’t novelty — it’s that hyperscalers can turn scale into a real silicon advantage.
• The economics are harsh. Running LLMs at scale burns power and cash; purpose-built chips can materially improve per-query economics.
• Software is where most of the value lives. CUDA still dominates the stack, but efforts to improve portability and new compiler projects could chip away at that lead. In practice, though, software moves slower than hardware prototypes.

How things are shaping up

• Winners: Nvidia still holds a structural edge thanks to a deep software stack and dominant share in training GPUs. Expect healthy near-term data-center revenue.
• Challengers: Microsoft, Google, Amazon and Apple will keep pushing tailored silicon to trim operating costs and create product differentiation. They have the scale to make it meaningful.
• Wildcards: Graphcore, Cerebras and SambaNova are experimenting with different architectures. They might win particular workloads or become attractive targets for acquisition.

Signals investors should watch

• Quarterly commentary: data-center revenue trends and gross-margin guidance from Nvidia and AMD. Those lines tell you more than product announcements.
• Hyperscaler capex and billing signals: more references to in-house silicon on cloud invoices or capex plans would suggest faster adoption.
• TSMC capacity and advanced-node allocations — foundry constraints still drive the tempo of the whole market.
• Software portability: broader support for ONNX, MLIR and related compiler work that reduces CUDA lock-in.

A few concrete scenarios

• Nvidia best case: continued dominance in training GPUs, plus a growing ecosystem for inference that keeps margins high.
• Hyperscaler best case: custom inference silicon meaningfully cuts per-query costs, shifting some demand away from GPU vendors.
• Middle ground: a heterogeneous market — GPUs for heavy training, a mix of custom accelerators and newer GPUs for inference — where multiple vendors remain viable. This middle outcome feels quite plausible.

A historical parallel (and its flip side)

It’s reminiscent of the early server CPU wars, when vendors tried to move away from Intel. Some custom efforts fizzled, but cloud scale made in-house designs rational for hyperscalers. The flip side is the power of software ecosystems: just as x86 won through software, CUDA could cement GPU leadership.

Investment implications

• Keep Nvidia as a core way to play AI hardware, but size positions with attention to valuation and execution risk.
• Add selective exposure to hyperscalers and foundries that benefit no matter which silicon model wins — think cloud providers and TSMC.
• Consider indirect exposure to niche accelerator startups through an M&A theme, but expect volatility and long time horizons.

Watchlist

NVDA, MSFT, GOOG, AMZN, AMD, INTC, TSM

Pedro Marini