Intel used Computex 2026 in Taipei to share fresh details on Crescent Island, an inference-class AI accelerator built on its Xe3P architecture. The chip carries up to 480 GB of LPDDR5X memory, fits a 350 W air-cooled envelope and is positioned to undercut Nvidia and AMD on cost per token for enterprise agentic workloads.
A different memory bet
By leaning on LPDDR5X rather than scarce HBM, Intel sidesteps the supply bottleneck that has constrained Nvidia's H200 and B200 shipments. The reference card starts at 160 GB and tops out at 480 GB across the LPDDR5X stack – enough capacity to keep long-context agent traces, retrieval-augmented memory and multiple model replicas resident on the card. Bandwidth is lower than HBM3e, but Intel argues that for inference – especially agentic chains that issue many small generations – capacity matters more than peak throughput.
Air-cooled, rack-friendly
Crescent Island is designed to drop into standard 2U servers without liquid cooling, easing deployment for colocation and on-prem buyers who are not ready for the immersion or direct-to-chip plumbing that GB200 and MI355X racks require. Intel is targeting customer sampling in the second half of 2026, with broader availability and revenue expected to ramp in 2027.
Built for agentic AI
Intel described the part as "built for agentic AI," and the spec sheet backs that up: the large memory pool is sized for long-context inference, mixture-of-experts routing and multi-tenant serving. The pitch goes after Nvidia's tightening grip on AI infrastructure budgets, which has driven hyperscalers to publish their own silicon and made room for a credible third source.
Where it fits
The launch lands alongside Nvidia's 550B Nemotron 3 Ultra and Qualcomm's Dragonwing Q-7790/Q-8750 announcements, marking Computex 2026 as the moment inference silicon decisively broke out as its own category.
Reporting based on coverage from Tom's Hardware, TechTimes, Startup Fortune and Intel's Computex briefing.