AI chip startups shift toward inference and interconnect bottlenecks

The AI chip startup market is fragmenting beyond a single Nvidia-challenger race. Cerebras leads on business proof after reported 2025 revenue around $510 million, up about 76% year over year, along with IPO scrutiny and cloud demand. d-Matrix is the strongest production-stage signal after raising $275 million at a $2 billion valuation in November 2025 and moving Corsair into full production in June 2026.

Investor appetite is concentrating on focused bottlenecks rather than generic accelerators. MatX raised $500 million in February 2026 for large language model chips, while Etched reportedly raised around $500 million at a $5 billion valuation in January 2026 for transformer-specific hardware. Positron and Fractile are targeting inference economics, memory pressure and latency, with Positron rising quickly after a $230 million Series B and Fractile drawing $220 million in May 2026.

Memory movement and interconnect are becoming core infrastructure categories. Ayar Labs, Lightmatter and Celestial AI are positioned around optical I/O, photonic interconnect and memory fabric, with Ayar Labs raising $500 million in March 2026 to expand high-volume production and test capacity. Edge and sovereign AI remain active but more fragmented, led by Axelera AI, SiMa.ai, Rebellions, FuriosaAI and EnCharge AI across local inference, physical AI and national infrastructure.