April 25, 2026

The Custom Silicon Rush

Google just dropped $40 billion on Anthropic. Amazon and Meta inked a multibillion-dollar Graviton deal. And Intel — Intel — is up 124% YTD with a 23.6% single-day rally that hasn't happened since 1987. The chip narrative has flipped hard, and if you're not paying attention, you're missing the big picture.

Here's what's actually happening: the hyperscalers are done waiting for Nvidia's scraps. They want their own silicon, built for their own workloads, at scale. The $40B Google commitment to Anthropic is the headline, but the real story is in the infrastructure underneath.

TPU 8th Gen: Google's homegrown bet

At Google Cloud Next 2026, Google unveiled its eighth-generation Tensor Processing Unit — the TPU 8t for training, the TPU 8i for inference. Both ship later this year, and they're not playing nice with the ecosystem. Google says 75% of new code inside the company is now AI-generated and human-reviewed. That's up from 50% last fall.

The TPU has always been the quiet competitor to Nvidia's A100 and H100. But with the new generation, Google is making a play for external customers who got burned by the GPU shortage. AI startups have been struggling to access Nvidia GPUs as Microsoft and other cloud providers divert supply to internal teams and large customers. The pipeline is clogged. Custom silicon is the workaround.

Amazon's Graviton goes big

Meta and Amazon just signed a multibillion-dollar, multiyear deal for hundreds of thousands of Amazon's Graviton chips. This isn't a pilot — it's production inference at scale. Meta is renting Amazon's silicon to run its AI models, and that's a massive validation for a chip line that was always playing third fiddle to Google's TPUs and Nvidia's GPUs.

The knock on custom silicon has always been: it's optimized for one thing, and AI moves too fast. But the market is maturing. The models are stabilizing. And when you're running inference on hundreds of millions of users, you don't need the latest breakthrough chip — you need efficiency at scale. That's where Graviton wins.

Intel's comeback tour

Let's be honest: nobody expected Intel to matter in this cycle. But the US government's stake has quadrupled to ~$36B since August 2025. Intel's upbeat outlook suggests CEO Lip-Bu Tan is making progress on a turnaround. The stock is up 124% YTD. That's not noise — that's a structural bet that Intel's domestic manufacturing matters for AI infrastructure.

The US government has a stake in Intel's success. That's not normal. But when your supply chain resilience depends onTSMC and Samsung, and either of those gets caught in a Taiwan Strait incident, you need a Plan B. Intel is that Plan B.

The implications

  1. Nvidia stays dominant, but the moat is narrowing. The big players have the capital to build their own chips. They'll still buy Nvidia for training, but inference is migrating to custom silicon.
  2. Silicon differentiation is dead for startups. If you're an AI startup and you're building custom chips, you're too slow. The hyperscalers have already done it.
  3. The GPU shortage solves itself. As inference shifts to TPU 8i, Graviton, and custom silicon, the bottleneck at Nvidia eases. That helps startups actually ship.

The chip wars aren't over. They're just entering a new phase. For the first time, the customers are building their own weapons.

Based on TEXXR signals from April 25, 2026. View original data →