Intel Builds 23-Company Korean AI Alliance Around HBM-Free GPU

Intel has signed up 23 South Korean companies for a new AI alliance and used the launch in Seoul to pitch an inference GPU that carries 480GB of LPDDR5X instead of scarce high-bandwidth memory. The K-AI Full Stack Alliance, announced by Intel Korea chief executive Bae Tae-won at the 2026 Intel AI Summit Seoul at COEX, is a direct attempt to build an alternative AI stack on Korean-designed servers while HBM supply remains tight.

The summit drew 52 companies and roughly 1,000 attendees, according to Intel. Samsung Electronics, SK Hynix and LG Electronics were present. Named alliance members include AI chip startups FuriosaAI and DeepX, model developer Upstage, Nota AI, Megazone Cloud, Dinotisia, Panesia and server maker KTNF.

The plan is straightforward: combine Korean server hardware with members' chips, models and software, so that companies can assemble industry-specific and sovereign AI systems. Intel Korea also signed a memorandum of understanding with the Seoul AI Hub to support startups. Sovereign AI is the commercial hook here. Governments and large enterprises want local stacks, and Intel is offering to be the silicon inside them.

What is the D-Series GPU?

Anil Nanduri, who leads AI product go-to-market for Intel's data center group, showed the physical chip: the GPU D270, code-named Crescent Island. The card draws 350 watts, is rated at up to 26,200 teraflops at FP4 precision, and carries up to 480GB of LPDDR5X — the low-power DRAM more commonly found in smartphones. Intel says that capacity is up to five times the 96GB on Nvidia's RTX Pro 6000 Blackwell Server Edition.

The trade-off is bandwidth. LPDDR5X moves data more slowly than HBM, which is why Intel is aiming the card at inference on long documents and many simultaneous requests — workloads where memory capacity matters more than raw speed. In effect, Intel is conceding the throughput race and competing on how much context a single card can hold.

Which performance claims are verified?

None of them, formally. Intel said internal tests run before mass production showed FP64 throughput of 1.92 to 5.9 times that of comparable Nvidia products on four financial simulations, 30% lower latency on Google's Gemma model, and 2.2 times faster AI-agent execution. The company did not release a launch date or a price. Buyers will want third-party numbers before committing fleets; pre-production vendor benchmarks have a habit of flattering their authors.

The CPU roadmap received firmer confirmation. Intel said the Xeon Clearwater Forest server chip, with up to 288 efficiency cores, is now in mass production. Diamond Rapids, with up to 256 performance cores, follows next year.

Why Korea, and why now?

The alliance gives Intel a Korean foothold alongside Nvidia, which has moved aggressively to lock in the market. Nvidia reaffirmed its 260,000-GPU supply plan for Korea on Friday, and earlier this month agreed to pair Korean NPUs with its GPUs. Intel's counter is not a bigger chip but a broader coalition: local chips, local models, local servers, with Intel silicon and ecosystem weight behind them.

For Korean suppliers, HBM scarcity is not an abstraction — SK Hynix sells most of its output before it is made. A GPU that runs inference on commodity LPDDR5X sidesteps that queue entirely, which may prove the most persuasive line in Intel's pitch.

The open questions are schedule and price. Until Crescent Island ships in volume with independent benchmarks attached, the alliance is a framework rather than a product, and Nvidia's 260,000-GPU commitment will remain the benchmark Intel must beat on delivery, not on slides.