One datasheet per company.
Latest round, investors, valuation, leadership, customers and sources. Filter by status, sort by size, and star the ones that matter; stars stay on this device.
Who signs for the compute.
The labs, clouds, trading firms and states whose orders decide which chipmakers survive. Most now run three or more silicon suppliers at once.
Point at a supplier to light it up across every buyer.
Acquisitions ledger
| Buyer | Target | What they got | Value | When |
|---|
The models that change what chips need.
Companies replacing attention with state-space, linear-attention or diffusion designs. Each datasheet says what it means for silicon.
Less memory per token
Hybrids keep a small fixed state instead of a cache that grows with context. IBM reports over 70% less RAM for Granite 4.
Compute, not bandwidth
Diffusion models decode many tokens at once, moving the bottleneck from memory to raw compute.
Single-architecture risk
Chips tuned for one model shape carry that bet. Etched has already moved its systems to run any frontier model.