Intel and Qualcomm aim at inference, where two thirds of compute now lives
Intel plans to launch its Crescent Island AI datacentre GPU by the end of 2026. Qualcomm's AI200 ships this year with AI250 following in 2027, and its Dragonfly C1000 datacentre CPU — with Meta named as a customer — is expected commercially in 2028. Inference is projected at roughly two thirds of all compute in 2026.
Training is a market with one dominant supplier and enormous switching costs. Inference is a different problem — latency-sensitive, cost-sensitive, and far more tolerant of alternative silicon because the software surface is narrower. That is where the challengers are aiming, and it is the correct place to aim.
The size explains the interest. Inference is projected at roughly two thirds of all compute in 2026, with the market for inference-optimised chips growing past $50bn. The four largest cloud providers spent about $410bn on AI infrastructure in 2025 and have guided over $700bn for 2026.
The entries are staged rather than simultaneous. Qualcomm's AI200 this year, AI250 in 2027, both positioned for inference; the Dragonfly C1000 CPU commercially in 2028 with Meta named under a multi-generation collaboration. Intel's Crescent Island by year end. Naming a hyperscaler customer years ahead of availability is how you make a roadmap credible enough to design against.
What none of it resolves is software. Alternative accelerators have failed repeatedly not on FLOPs but on the toolchain, and inference-only positioning helps precisely because the surface is smaller. Whether it is small enough is the question the next two years answer — and the labs now measuring themselves on cost per served token have more reason than ever to find out.
Barchart — Intel Sets Sights on Nvidia and AMD With Upcoming AI Data Center Chip Launch by Year End → · Deloitte — Compute power and AI →