Intel Partners with 23 Korean Companies to Form AI Alliance, Launches HBM-Free Inference GPU
nashnova research
Intel launched the K-AI Full-Stack Alliance in Seoul with 23 Korean partners and unveiled its GPU D-series — an inference card that swaps HBM for low-power LPDDR5X memory. This means → Intel is betting that big capacity, low power, and no HBM can carve a niche in Nvidia's AI accelerator stronghold.
What is this alliance actually for?
Intel Korea president Taewon Bae announced the K-AI Full-Stack Alliance with 23 founding members spanning hardware, software, and industry AI services.
Members include NPU makers FuriosaAI and DeepX, model developer Upstage, plus Nota AI, Panmnesia, KTNF, and Hancom. Target markets: sovereign AI and physical AI.
This means → Intel is not just selling chips. It is assembling a full stack — chip to application — with Korean partners, pitching an "open ecosystem" against Nvidia's closed CUDA moat.
A GPU with no HBM — what is the play?
The GPU D-series (codename Crescent Island) uses LPDDR5X memory instead of HBM, drawing 350W — just over half the 600W of Nvidia's RTX Pro 6000.
In plain terms = HBM — high-bandwidth memory, a fast but expensive and power-hungry technology that stacks memory chips vertically — is the default for AI GPUs today. Intel's card skips it entirely, opting for cheaper, cooler LPDDR5X.
Memory capacity reaches up to 480GB, as much as 5× Nvidia's comparable card. Intel's own version ships with 160GB; partner cards scale to 480GB.
This means → the strategy is "trade bandwidth for capacity" — accept somewhat slower data movement in return for loading larger models at a lower deployment cost.
The performance numbers look good — but are they reliable?
Intel claims the D-series delivers 30% lower latency than Nvidia on the Gemma model and up to 5.9× better performance in financial-simulation workloads.
Yet Intel VP Anil Nanduri admitted after the demo: "The hardware and software are not yet production-ready." The card is not expected to ship until next year.
Jon Peddie Research warned that the card's 1.5 TB/s bandwidth "will limit decode speed on dense models" and said independent cost-per-token data and pricing are needed before any judgment can be made.
This reflects a key tension: Intel's numbers come from select scenarios, not broad benchmarks. Until independent verification arrives, they are reference points, not proof.
Why is SK Hynix involved?
SK Hynix VP Youngpyo Joo said the company is collaborating with Intel from the system-design stage and proposed a shared "memory rack" architecture across servers.
Test data: using CXL — Compute Express Link, a high-speed interface connecting CPUs to remote memory — to keep key-value caches in remote DRAM cut first-token generation time from 80.76 seconds (when the cache is evicted and recomputed) to 15.45 seconds.
This means → SK Hynix is not betting on this GPU per se. It is betting on the broader "move away from HBM" trend — if LPDDR5X can replace HBM in some inference scenarios, SK Hynix's DRAM and CXL product lines get a wider addressable market.
What is the Korean government betting on?
Korea's Ministry of Science and ICT said it will build open AI infrastructure through Intel's partnerships with Korean chip, server, and software firms.
The Ministry of SMEs and Startups pledged over ₩5 trillion (roughly $3.73 billion) by 2030 to bring physical AI to small and mid-sized manufacturers.
In plain terms = Seoul's calculus: rather than let all AI compute run on Nvidia's closed platform, back an open chain where Korean companies fill the middle layers — Intel supplies chips and foundry capacity, local firms build the rest.
Can this strategy actually work?
Intel custom-chip lead Marcus Lawrence stressed that two Korean companies (Samsung and SK Hynix) produce two-thirds of the world's memory chips; Intel can layer chip design, advanced packaging, and foundry services on top.
But Tom's Hardware analyst Jeff Kampman noted that whether the D-series can establish a foothold for Intel remains to be seen — its previous data-center GPU, Ponte Vecchio, ended in failure.
This reflects Intel's core risk: the direction may be right (low-power inference, open ecosystem), but execution is unproven. The real test comes next year, when mass production and independent cost verification either validate or sink the pitch.
市场有风险，内容仅供研究参考，不构成投资建议。