Finally, an Intel GPU With $8 Billion in Revenue Potential

At SC26 Intel gave Crescent Island its product name, the D270, and showed what it can do. An air-cooled card that fits ordinary data centers, with a real path to $8 billion a year.

By

·

Intel GPU D270 at SC26: 5.9 times option pricing throughput, 350W air-cooled, up to 480GB LPDDR5X, full-rate FP64, and a Blue Team estimate of $2B, $4B and $8B GPU revenue for 2027 to 2029.

For years Intel had no data center GPU that customers were lining up for. That changed this week at SC26 in Chicago. Intel gave Crescent Island its product name, the GPU D270, published its first benchmarks and showed it running in a partner system. It is the first Intel GPU with a believable path to billions in revenue.

What the D270 Is

  • Architecture: Xe3P, with 32 Xe cores and 256 XMX matrix engines.
  • Memory: 160 GB of LPDDR5X at 1.5 TB/s on Intel’s own card, and up to 480 GB on partner boards. No scarce HBM needed.
  • Precision: everything from FP4 for AI up to full-rate FP64 for science and engineering.
  • Power: 350 watts, air-cooled, on a standard PCIe 5.0 card.

What Intel Measured

Intel tested early silicon against Nvidia’s RTX PRO 6000 Blackwell, a 600 watt card. In financial option pricing the D270 delivered up to 5.9 times the throughput on Monte Carlo, 5.6 times on Black-Scholes and 4.6 times on binomial models. On AI inference it delivered 1.12 times the throughput on a Qwen3 30B model and 29 percent lower latency on a decision model. It does this at a little more than half the power.

Full-rate FP64 matters. Simulation, seismic imaging and risk models need double precision, and the newest AI-focused GPUs have moved away from it. The D270 gives these customers a modern card that does both AI and classic HPC.

It Fits the Data Centers That Already Exist

Most enterprise and colocation data centers are air-cooled. The newest liquid-cooled AI racks cannot go into them without a rebuild. The D270 slides into the same servers that hold today’s 300 to 400 watt PCIe cards. With up to 480 GB of memory, a large model fits on one or two cards instead of a stack of older ones.

Partners Are Already Building With It

  • SambaNova: at SC26 four D270 cards handle prompt processing while SambaNova’s SN50 rack handles token generation. Intel Xeon 6 and Xeon 6+ processors run the orchestration. All of it is air-cooled and runs on open software over standard Ethernet.
  • Veridien: moved its seismic imaging code to Intel’s open SYCL stack. Intel says 90 percent of the kernels performed well with no Intel-specific tuning, and a seismic mini-application was ported in less than a month.
  • Software: oneAPI, PyTorch, vLLM and SGLang all run on it from day one.

A Smart Way to Sell It

Intel makes its own 160 GB card and also sells the chip to board partners, who build the larger 480 GB versions and buy the memory themselves. Intel does not carry memory costs or board inventory. Every dollar of GPU revenue earns a better margin, and many partners can scale production at once.

Where $8 Billion Comes From

AI accelerator spending is heading toward $500 to $600 billion a year by 2028, and inference is becoming the larger half. We estimate that air-cooled PCIe inference for enterprises, colocation, sovereign clouds and the edge will be a $40 to $80 billion a year market by 2028. A 10 percent share of that is $4 to $8 billion.

Our model has Intel data center GPU revenue at about $2 billion in 2027, $4 billion in 2028 and $8 billion in 2029. Intel has described an annual GPU cadence, with Xe-Next and Jaguar Shores following the D270. In a market short of AI chips, the limit is how many Intel can supply, not whether customers want them.

The Bottom Line

The D270 is the GPU Intel needed: cheap memory, air cooling, full FP64, open software and partners already shipping systems around it. It runs next to Xeon, so every rack sells both. After years of false starts, Intel finally has a GPU that can grow into an $8 billion business.

Sources: Intel newsroom, November 16, 2026. SambaNova announcement at SC26, November 15, 2026. Benchmarks are Intel’s own tests on pre-production silicon. Market size and revenue figures are Blue Team estimates.

Get the Intel bull case in your inbox

Not investment advice. The author may hold positions in the securities discussed.

One response

  1. […] We covered the D270 in depth yesterday: Finally, an Intel GPU With $8 Billion in Revenue Potential. […]

    Like

Leave a Reply