Intel leads with 480 GB of memory for its new AI GPU; its own card carries 160 GB and the stock fell 3.2%

A name, at last, for Intel's AI inference card

At the Open Compute Project Global Summit in San Jose on October 12, Intel▼−4.25% — Intel, down 4.25 percent today named its first commercial data-center AI GPU since Ponte Vecchio in 2021: Intel GPU D-Series, formerly code-named Crescent Island↗ ¹³. The lead card, the D270, is a 350-watt, air-cooled PCIe card with 1.3 petaflops of FP8 compute and 1.5 TB/s of memory bandwidth, sized for cheap, sustained inference: models up to 300 billion parameters on one card, past a trillion on eight ¹. Customer sampling is underway. There is no price and no launch date ¹³.

The 480 GB headline is a partner build

The memory number Intel leads with is not on the card Intel sells. The release's own footnote↗: the Intel-branded PCIe card carries 160 GB of LPDDR5X, and the design "enables partners to build ODM branded cards with flexible options up to 480 GB" ¹. ServeTheHome's spec table shows both configurations; TechPowerUp↗ confirms 160 GB is exclusive to Intel-branded cards ³⁴. The 480 GB figure, the one that beats AMD's▼−3.02% — AMD, down 3.02 percent today 432 GB MI450X and Nvidia▼−2.00% — Nvidia, down 2.00 percent today Rubin's claimed 288 GB on paper, exists only if original design manufacturers build it ⁵.

The rally is CPU money, and the GPU is a fourth attempt

Two successors to Ponte Vecchio were canceled on the way here, Rialto Bridge in 2023 and Falcon Shores in 2025, and Crescent Island was unveiled at last year's OCP as an interim, cost-optimized part ³. ServeTheHome expects server vendors to wait until 2027 for volume ³. Meanwhile the stock's 183% run this year rests on the other product line: revenue hit $16.1 billion in the June 2026 quarter, up 25.4% year over year↗, and Yahoo Finance↗ attributes the rally to CPU demand from the AI build-out, not GPUs ⁸². On launch day the shares fell 3.2% to $101.35 ²⁶.

The open lane against Nvidia's racks

The counterargument is real. High-bandwidth memory is scarce and expensive; LPDDR5X, the cheap memory in laptops, buys capacity at a fraction of the power ⁵. A 350-watt air-cooled card drops into servers already on the floor, while Nvidia sells inference as the GB300 NVL72: a fully liquid-cooled rack of 72 Blackwell Ultra GPUs and 36 Grace CPUs↗, moving up to 576 TB/s across the rack, roughly 8 TB/s per chip against the D270's 1.5 ⁷. AMD's PCIe rival, the MI350P, peaks at 2.3 petaflops of dense FP8 but at 600 watts ³. If the metric is tokens per watt in ordinary servers, Intel has a lane, and ASUS, HPE, Dell, Supermicro and several neoclouds are building systems for it ². Intel is also driving an OCP workgroup to standardize the sub-rack accelerator card, the form factor Nvidia's racks do not use ¹.

No price, no launch date, no benchmarks

The card needs a price, a launch date and independent benchmarks, none of which exist yet ³⁴. Until one arrives, the GPU leg of the turnaround is a spec sheet with sampling status attached, and holders sitting on a 183% run are underwriting it. Next readings: Intel's third-quarter results in the coming weeks, and any D-Series announcement that carries a number.