Intel previews Crescent Island GPU and ships Xeon 6+ at Computex 2026 — 480 GB LPDDR5X inference card and 288-core 18A Clearwater Forest CPU bet the inference floor on memory density
Intel used the June 1 Computex 2026 keynote to unveil two datacenter products. Xeon 6+ (Clearwater Forest) is the industry's first 18A-node CPU at up to 288 E-cores, claiming 2.5x performance and 45% better perf/watt versus the prior generation. Crescent Island is a preview of a 350W air-cooled PCIe inference GPU on Xe3P with up to 480 GB of LPDDR5X — a memory-density bet that explicitly targets agentic inference workloads where context window is the binding constraint.
The interesting piece is what Intel did not announce. There is no HBM Crescent Island variant, no NVLink-equivalent fabric, no claim of frontier-training viability. Intel is conceding the high-end of the market that Nvidia just shipped Vera Rubin into and competing on a different axis entirely — memory capacity per dollar per watt at the inference edge. 480 GB of LPDDR5X on a single PCIe card holds a 70B-class model with several hundred thousand tokens of context resident, at a thermal budget (350W, air cooling) that fits into existing enterprise datacenter footprints without a liquid loop retrofit. The product specification reads like a deliberate inverse of the NVL72.
Xeon 6+ is the more strategically loaded announcement. Clearwater Forest is the first datacenter processor built on Intel's 18A process — the node that the company's entire foundry recovery thesis depends on. Up to 288 E-cores per socket and the 45% perf/watt claim are competitive against AMD's Turin-X and Nvidia's just-announced Vera CPU, but the marketing axis is execution risk: shipping any 18A product into volume on the announced schedule is itself the proof point Intel needs. CEO Lip-Bu Tan walked the keynote himself, and the stock reaction (INTC closed down on the day despite the announcements) suggests the market is pricing the execution-risk discount rather than the spec-sheet wins.
The combined Xeon 6+ + Crescent Island + E835 Ethernet stack is positioned as a complete inference-tier alternative to the Nvidia full-stack rack pattern. Intel is not claiming it competes with Vera Rubin on a trillion-parameter frontier-training workload. It is claiming it competes on the 70B-class inference workload that thousands of enterprises actually deploy — and on those buyers' procurement terms (PCIe form factor, air cooling, existing rack footprint, no liquid plumbing). Whether that pitch lands depends on Crescent Island shipping on the previewed timeline, which Intel did not commit to publicly.
Tom's Hardware — Watch Intel's Computex 2026 keynote — CEO Lip-Bu Tan in Taipei at 10:30pm PT June 1 → · Hardware Upgrade — Nuovi Xeon 6+, Ethernet E835 e GPU Crescent Island — Intel aggiorna la piattaforma per l'AI nei datacenter → · Cloud News — Intel aims to regain ground in AI with Xeon 6+, 200G Ethernet, and Crescent Island →