Intel D270 Launched: 480GB of VRAM on a Single Card, a 300-Billion-Parameter LLM Runs on Air Cooling

Alternative titles (final pick by Uncle Peng):

"Not About Compute, About Capacity! Intel D270 Launched, Up to 480GB VRAM, Runs a 300-Billion-Parameter Model on a Single Card"

"Intel D270 Launched: A Single 480GB Air-Cooled Card Fits a 300-Billion-Parameter Model, No Data Center Retrofitting Needed"

💡 What You Will Learn

Intel D270 Launched: 480GB of VRAM on a Single Card, a 300-Billion-Parameter LLM Runs on Air Cooling Alternative titles (final pick by Uncle Peng): "Not About Compute, About Capacity! Intel D270 Lau

📜 Table of Contents

• 1. The Main Course First
• 2. Three "No-Holds-Barred" Design Choices
• 3. But Hold Your Excitement: Three Traps
• 4. The Real Trump Card: An Open-Source Software Stack
• 5. What This Means for You
• Conclusion
• Image Notes for the Main Text
• Source List

Intel D270 Launches: 480GB of VRAM on a Single Card, a 300-Billion-Parameter LLM Runs on Air Cooling

Alternative titles (final pick by Uncle Peng): 1. "Forget Compute, Count Capacity! Intel D270 Launches with Up to 480GB VRAM, Running a 300B-Parameter Model on a Single Card" 2. "Intel D270 Launches: A 480GB Air-Cooled Card Fits a 300B-Parameter Model — No Data Center Retrofit Needed" 3. "NVIDIA Pushes Liquid Cooling, Intel Pushes Capacity — Intel D270 Launches, 480GB on One Card for 300B Parameters"

If you could run a 300-billion-parameter large model just by plugging a single GPU into your server — would you still need a liquid-cooled data center?

On October 12, Intel gave its answer at the OCP Global Summit: the AI inference accelerator codenamed "Crescent Island" is officially named the GPU D series, debuting with the D270 — air-cooled, 350W, up to 480GB of VRAM, running a 300-billion-parameter model on a single card.

The press releases all recite specs. This piece covers "where it's strong and where the traps are" in one go.

1. The Main Course First

The most eye-catching thing isn't the compute — it's the VRAM capacity.

2. Three "No-Holds-Barred" Design Choices

1. Betting on capacity. While Rubin (288GB) and MI450X (432GB) are still fighting over HBM4, Intel went ahead and stacked the LPDDR5X used in laptops onto the card — HBM is expensive and hard to source, while LPDDR5X is cheap and easy to stack. For inference, "can it fit" is more fatal than "how fast does it run."

2. Cutting 3D and ray tracing. The entire rendering pipeline is gutted, with the die area all going to matrix computation. This chip was never meant to render a single frame of a game from birth.

3. 350W air cooling. A standard PCIe card that runs the moment you plug it into an existing server. For enterprises, the money spent on redoing cooling is often more than the cost of the cards themselves.

3. But Hold Your Excitement: Three Traps

1. 480GB is the ceiling, not the standard. Intel's own card is actually 160GB; 480GB is the upper limit for ODM-customized versions — a footnote many press releases conveniently ignored.

2. Big capacity ≠ fast. The 1.5TB/s bandwidth is only about one-fourteenth of Rubin's, and the decode phase is precisely where bandwidth matters most; FP8 compute is also roughly 40% lower than the similarly-shaped AMD MI350P. Its positioning isn't a "performance replacement" — it's a "cost replacement."

3. A product with no price, no date, no benchmarks — plus a troubled history. No pricing, no launch date, no third-party benchmarks. This is Intel's third charge into data center GPUs since 2021, and the first two were both cancelled. This time the strategy is the clearest yet, but clarity doesn't equal success.

4. The Real Trump Card: An Open-Source Software Stack

Full support for vLLM, SGLang, and PyTorch — even NVIDIA's own Dynamo inference framework is supported; Meta's newly open-sourced 30-billion-parameter model Muse Glimmer gets day-zero support; Arc Pro code can be ported without rewriting kernels.

The intent is written all over its face: trade open source for an ecosystem niche, trade "no code changes" for developers.

5. What This Means for You

Enterprises: if the "cost per token per watt" holds up in real testing, this is a new option that doesn't require retrofitting your data center; everyone also knows full well what the PCIe interface positioning is about.

Developers: native vLLM support means low trial-and-error cost — but wait for third-party benchmarks before going to production.

Gamers: the easter egg is that Xe3P will also trickle down to Nova Lake and Arc C discrete GPUs — server architecture. The PCIe + air-cooling combo fits easily into a personal workstation, making it broadly adaptable.

Conclusion

NVIDIA is competing over who can pack a liquid-cooled data center fuller; Intel is competing over who can plug a card into more old servers. This isn't a compute showdown — it's a deployment cost showdown.

Whether the D270 succeeds isn't for anyone to say right now, but it proves one thing: in the inference era, "being able to fit" is itself a competitive advantage.

Do you think Intel can pull it off this time? See you in the comments.

Image Notes for the Main Text

• Figures 1 / 2 / 3 / 4 are all from Intel's official public news materials for Hot Chips 2026 / OCP 2026 (reposted via TechPowerUp), copyright Intel Corporation; per standard tech media practice, "Source: Intel" must be credited, which has been noted in the captions;
• Two backup images (backup-compute-array-panorama.jpg, backup-Xe-core-detail.jpg) can be added as needed for platform layout;
• Before publishing, please note Toutiao's rules on citing third-party copyrighted images; if you need to avoid this, you can keep only the official Intel press release images and retain the captions.

Source List

• Intel Newsroom: Intel Accelerates Open Infrastructure for the Agentic AI Era (2026-10-12)
• Kuaikeji: 480GB VRAM on a Single Card! Intel's Most Powerful Inference Card Launches (2026-10-13)
• WccfTech / VideoCardz / ServeTheHome / Tom's Hardware / StorageReview (2026-10-12)