Intel will ship its Crescent Island AI chip in limited quantities by the end of this year, pricing it below Nvidia and AMD's data center GPUs by skipping high-bandwidth memory and liquid cooling. The part uses LPDDR5 memory and an air-cooled design, two choices that cut both the bill of materials and the infrastructure cost customers carry when they rack the chip. Intel shares are up more than 200 percent since the start of this year, and Crescent Island is the first concrete AI product to land under CEO Lip-Bu Tan, who took over last year.
Kevork Kechichian, who runs Intel's data center group, told the Financial Times the company is aiming squarely at inference — the runtime stage when a user makes a request of a model — rather than training. That distinction matters. Training is where Nvidia's Blackwell and its predecessors dominate, and where HBM and liquid cooling are effectively mandatory. Inference is a more fragmented market with looser memory-bandwidth constraints, which is where LPDDR5 becomes viable.
Kechichian joined Intel last year from Arm and inherited the wreckage of Gaudi, Intel's earlier training-focused accelerator that sold poorly and whose planned successor was cancelled. Crescent Island took 18 months to develop and was first unveiled in October as part of Tan's broader effort to rebuild a competitive product line in AI silicon.
“We decided to start rebuilding our muscles in AI… [but] we are not particularly aiming for [the training market] based on past experience”— Kevork Kechichian, Head of Intel's data center group
Key facts
- 01Intel's Crescent Island GPU will begin shipping in limited quantities by the end of this year after an 18-month development cycle.
- 02The chip uses LPDDR5 memory and air cooling, avoiding the HBM and liquid-cooling costs in Nvidia's Blackwell and AMD's AI parts.
- 03Intel shares are up more than 200 percent since the start of this year, after Lip-Bu Tan replaced Pat Gelsinger as CEO.
- 04The US government announced in August it would take a 10 percent stake in Intel to keep its foundry business intact.
- 05Crescent Island targets inference workloads, not training — a deliberate retreat from the segment Nvidia dominates.
The economics of the bet are straightforward. HBM is in short supply and expensive, and liquid-cooled racks require data center operators to retrofit power and plumbing. By targeting customers who want to serve inference at scale without rebuilding their facilities, Intel is going after a price-sensitive tier of buyers that neither Nvidia's Blackwell nor AMD's MI-series chips serve cleanly.
Kechichian said there are tiers of the chip that could potentially be sold in China in compliance with US export controls, an opening Nvidia and AMD have been largely shut out of because of trade tensions between Washington and Beijing. He framed it as a demand-and-price-point question rather than a policy commitment, but the China line is the kind of revenue Nvidia has been forced to write off.
The leadership backdrop matters for reading Intel's seriousness here. Pat Gelsinger was ousted last year amid concerns that his turnaround strategy was failing, and Tan moved quickly to cut costs and rein in manufacturing spending. In August, the US government announced it would take a 10 percent stake in Intel over time, a move the Trump administration tied to keeping Intel's foundry business under American ownership.
Tan has also pushed Intel to build advanced PC and server chips in its own factories this year, after a long stretch of relying on TSMC. Crescent Island is expected to follow the same in-house path, which would compound the cost advantage if the foundry yields hold up.
The risk in the strategy is real. Inference workloads are migrating toward larger context windows and reasoning models that punish low memory bandwidth, and LPDDR5 is meaningfully slower than HBM. If the next 18 months of model design lean further into long-context inference, Crescent Island's price advantage could erode against parts that pay the HBM premium and deliver the bandwidth. Intel also has to convince hyperscalers and neoclouds that have standardized on CUDA tooling to integrate a new accelerator — a software lift that Gaudi never solved.
The Crescent Island bet is the first Intel data center product in years that has a coherent answer to the question of who would actually buy it. Nvidia's pricing power at the top of the stack has created room underneath, and a cheaper, air-cooled inference part that runs in existing facilities is exactly the kind of wedge a distant third place needs. Whether it closes the gap depends on yield, software, and how fast the inference market splits between frontier reasoning workloads and the long tail of cheaper serving — and Intel only needs to win the second one to make the chip pay.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
