Intel’s new inference card is advertised with up to 480 GB (gigabytes) of memory on partner cards and no high bandwidth memory (HBM) at all [1]. It won’t skip your memory queue. The D270 was announced in October, after Intel argued that the limit had moved off bandwidth and onto capacity for long prompts and many sessions at once [1]. Its headline numbers are memory and watts: 350 watts, air cooling, and a standard PCIe slot [1][2]. Sixteen days later the memory maker (SK hynix) reported an 80% operating margin, and said more than half of its 2027 and 2028 output is already sold [3].

The big picture:

I keep returning to the gap between these two releases. Intel’s page argues the bottleneck has moved, because agents read far more than they write [1]. Long prompts dominate, and the card is built to keep more of a model resident, with what Intel describes as “no high bandwidth memory (HBM)” at all [1]. The page says nothing about where the alternative memory comes from. The memory maker’s quarter shows what HBM has become: “the mainstream product from the fourth quarter” [3]. More than half of the next two years’ output is already signed, on deals that run “typically five years”, at an operating margin of 80% [3].

Put the two together and the swap changes shape. The low-power memory Intel relies on is not outside that queue. The same release that lifts HBM4 output also plans to sell more server memory for AI systems, as part of a range built on high-value parts [3]. A buyer is not choosing between an expensive memory market and a cheap one. A buyer is choosing which queue to stand in, and the queue this card shortens is the one for HBM stacks rather than memory wafers.

Price direction points the same way. TrendForce expects standard dynamic random access memory (DRAM) contract prices to rise 10 to 15% this quarter, with server DRAM still short and graphics DRAM “on an upward trajectory” [4]. Nothing on that list is falling. So the reason to buy capacity in a standard slot is not a lower memory bill. It is what a bigger card does to where models sit, and to the rack.

By the numbers

  • 480 GB — Partner-card ceiling: Intel’s own card carries 160 GB from twenty 8 GB chips; partner cards reach 480 GB using 24 GB chips, on an estimated 1,280-bit path at 1.5 terabytes a second [2].
  • 1.5 against 3.6 to 4 terabytes a second — Bandwidth given up: AMD (the chip designer) runs HBM3E in its MI350P at about 3.6 to 4 terabytes a second. NVIDIA (the card maker) pairs 96 GB of GDDR7 with 1.6 to 1.8 terabytes a second [2].
  • 80% — Memory maker’s margin: on revenue of 104.6155 trillion won and operating profit of 83.3338 trillion won, with HBM4 set to pass HBM3E in volume and revenue this quarter [3].
  • 10 to 15% — Memory prices: the expected rise in standard DRAM contract prices this quarter, with flash memory up 15 to 20% and server DRAM still short [4].

What I’d watch:

What I am watching is how workloads split between prompt processing and output generation. Prompt processing builds the cache and leans on raw compute. Output generation leans on memory bandwidth, which explains why the two phases fail in different ways. Intel built the card around the first phase and says so, presenting speculative decoding that retires 2.9 to 4.9 tokens per check [2].

A firmware unlock published this month shows the same split from the other end. Doubling the half-precision math issue rate on consumer cards lifted prompt processing by 14 to 36% and left output, which is limited by memory bandwidth, about the same [5]. More compute does not fix a bandwidth wall. Neither does more memory.

  • Prompt share: If most of the clock goes on prompts and cache building, capacity is the lever, making a large card the cheaper way to hold a model [2].
  • Placement: A 480 GB partner card holds weights that would otherwise split across several cards, which cuts traffic between them [2].
  • Watts and air: 350 watts with passive cooling fits racks that cannot take a liquid retrofit, and that decides more rollouts than a bandwidth table does [1][2]. The cap-tuning arithmetic in Why Your 80% Cap Wins is the other half of that budget.
  • The next supply deal: Whether low-power server memory joins HBM in the sold-ahead category, which its maker now lists among its high-value lines [3].

The catch

The card is not on sale. Intel is sampling the hardware with no price and no ship date, and the analyst house that tore it down expects volume availability in 2027 [2]. Its costs are tied directly to the memory market. The same analysis notes that low-power memory prices have been unstable, and that “the economics of the 480GB configuration depend heavily on DRAM pricing” [2]. Intel’s record in this market offers another reason to wait, since two earlier programmes were cancelled or pulled [2].

Where I’ve landed: the business case looks narrower than the launch text. The D270 does not cut the memory bill, because no price has been published for the card or its memory, and every memory type TrendForce tracks is rising [2][4]. What it does instead is turn a bandwidth budget into a capacity and power budget, which suits long-prompt agent work on air-cooled racks and fits badly any service whose users sit waiting on output tokens.

At a glance

  • The Big Shift: Intel’s D270 sells memory capacity instead of bandwidth, putting up to 480 GB of low-power memory on partner cards in a 350-watt slot, and markets the result as inference without HBM [1][2].
  • Why It Matters: The HBM it avoids is now a sold-ahead product, with more than half of the supplier’s 2027 and 2028 output signed and an 80% margin on the quarter [3]. The swap changes which queue a buyer joins rather than what memory costs.
  • What I’d Watch: How much of the workload is prompt processing, and whether the rack can take the power.
  • Prompt share: The time spent reading prompts and building the cache, where extra compute and extra capacity both help [2][5].
  • Output speed: The time spent producing tokens, set by memory bandwidth, where this card trails HBM parts by well over half [2].
  • Supply terms: Whether low-power server memory gets contracted the way HBM now is, which decides whether the cheaper option stays available [3].
  • The Catch: No price, no ship date, sampling only, and a cost case the analyst says depends heavily on DRAM pricing [2].

Related reading

Sources

[1] Intel, “Intel GPU D270 — Inference-Optimized Data Center GPU,” product page, D-Series announced at the Open Compute Project summit, 12 October 2026. https://www.intel.com/content/www/us/en/products/details/discrete-gpus/data-center-gpu/d-series.html [2] Steve McDowell, “Intel GPU D270: An Inference-First GPU for the Data Center,” NAND Research, 15 October 2026. https://nand-research.com/intel-gpu-d270-an-inference-first-gpu-for-the-data-center [3] SK hynix, “SK hynix Announces 3Q26 Financial Results,” press release, 28 October 2026. https://news.skhynix.com/en/q3-2026-business-results [4] TrendForce, “AI Server Demand Sustains Memory Contract Price Increases in 4Q26, While Consumer-Side Pressure Persists,” press release, 30 September 2026. https://www.trendforce.com/presscenter/news/20260930-13258.html [5] Zagitta, “GPU unlock 1.0.0 — unlock FP/BF16 instruction issue rate on NVIDIA 3000-series cards,” GitHub repository, read 24 October 2026. https://github.com/Zagitta/gpu-unlock