Intel killed Rialto Bridge in 2023.
Then Falcon Shores in 2025.
Two flagship data center GPU programs, gone before they shipped.
This week the company went the other direction and put final numbers on Crescent Island, the inference chip it now sells as the GPU D-Series.
The D270 runs 32 Xe cores and 256 XMX matrix engines at roughly 2.5GHz, reaching 1.3 PFLOPS of peak FP8 matrix throughput.
Memory bandwidth is 1.5TB/s, capacity goes up to 480GB of LPDDR5X on ODM cards, and the whole card stays inside 350W on standard air cooling with ECC on.
It is not built to beat Blackwell Ultra or the MI455X.
Squeezing cost and power was the point.
Intel calls it an interim part aimed straight at the AI inference market, smaller and cheaper than the flagships.
That is the detail worth reading twice.
The interesting number in AI infrastructure is no longer peak throughput, it is cost per token and watts per rack, because those decide which projects get to run at all.
Most small teams in this position are not blocked by ambition.
They are blocked by the monthly bill for capacity that takes months to arrive.
RTX 3090 capacity at $0.15/hr is available today, and Buy Now Pay Later means zero upfront with payment after one month.
If you run inference in production, what hits the wall first on your side: memory, throughput, or the bill?
https://ai.bhkcloud.com/?utm_source=linkedin&utm_medium=social&utm_campaign=bhk_social_2026w42&utm_content=morning
Originally posted on linkedin_personal