The Agent Stack #033 — Friday Signal

The AI infrastructure market just got real validation. Cerebras raised £4.3B in their IPO this week, with shares jumping 108% on debut.

Hardware becomes the new moat

Cerebras makes wafer-scale processors designed specifically for AI training. Their IPO success proves investors believe specialised AI chips will dominate general-purpose GPUs. The stock pop shows demand for alternatives to NVIDIA’s stranglehold on AI compute.

This matters for agent builders. Training costs have been the biggest barrier to custom models. Cerebras promises 20x faster training at lower power consumption. If they deliver, we’ll see more companies building domain-specific agents instead of relying on OpenAI’s general models.

The timing is perfect. OpenAI and Apple are fighting over integration terms. Microsoft is cancelling Claude Code licenses. The big platforms are becoming less reliable partners. Cerebras gives you another path to independence.

But don’t get carried away. Cerebras still costs millions for entry-level systems. Their customers are research labs and major corporations. For most of us, this is about the future direction, not immediate access.

The real signal here is market confidence in AI-specific hardware. Expect more IPOs from companies like Groq, SambaNova, and GraphCore. Competition will eventually drive prices down and performance up.

Your move: Start planning for a world where custom training becomes accessible. What would you build if model training cost 90% less?

Quick hits

• Notion launches developer platform - Connect AI agents directly into workspaces. Early access starting next month. Finally, proper agent integration beyond chat interfaces.

• IBM releases Granite Embedding R2 - Apache 2.0 licensed, 32K context, best sub-100M parameter retrieval quality. Open source embedding models just got competitive with proprietary ones.

• 70% of Americans oppose AI data centres - Gallup survey shows massive public resistance. Expect regulation and permitting delays to slow AI infrastructure expansion.

One thing to try

Set up a simple embedding benchmark using IBM’s new Granite model. Compare it against OpenAI’s text-embedding-3-small on your specific domain data. Open source might already be good enough for your use case, saving you API costs and giving you full control.

The infrastructure wars are heating up. Choose your dependencies wisely.