How Intel’s AI Chip Strategy Is Coming Into Shape After Years Of Struggles
CRN explains how Intel’s AI chip strategy under CEO Lip-Bu Tan is coming into shape around the idea of GPUs and other processors powering inference and agentic workloads after several previous attempts by the chipmaker failed to catch traction in the data center market.
The chipmaker has generated excitement around its latest push in this area under Intel CEO Lip-Bu Tan. This has been boosted recently by the hiring of Qualcomm’s longtime GPU leader and an AI chip partnership. It is also experiencing a resurgence in CPU demand due to the rise of agentic AI. But the company still has much to prove.
[Related: Intel-Nvidia Deal Will Create ‘New Class Of Integrated Graphics Laptops,’ Huang Says]
Intel has spent billions of dollars over the past 15 or so years to build accelerator chips to take on Nvidia’s GPUs and open a new profit center beyond its traditional CPU business, but the company has had to pivot several times as various efforts failed to catch traction.
This has been a sticking point for some in the channel, including Dominic Daninger, vice president of engineering at Intel systems integration partner Nor-Tech, based in Burnsville, Minn. He told CRN on Friday that his company has invested in new Intel products before—such as the Xeon Phi processors, which competed with Nvidia and AMD GPUs—only for the chipmaker to retrench from the category a few years later.
“One of the things with Intel we’ve seen so many times [was] that they were out there a year or two with a particular product line, and if it’s not sticking, it’s gone,” he said.
With hardware demand growing due to increased focus on inference workloads, Daninger said he thinks Intel could see some traction but only if their next solution is competitive.
“[It would] probably have to have a heck of a lot more performance than anything the competitor would have to make people interested—and a reasonable price,” he said.
Intel’s GPU Push Goes Back Several Years Under Different CEO
While Tan is homing in on a strategy that relies on GPUs and other chips working in parallel as Intel’s challenge to Nvidia, the push stems from several years of development that started under former Intel CEO Brian Krzanich, who preceded Tan by three tenures.
But in trying to build competitive accelerator chips, the company has also veered into ASICs, short for application-specific integrated circuits, in recent years, mostly recently with the Gaudi chips, which were originally designed for AI training workloads but fell short of modest sales expectations, even after Intel pivoted the focus to inference.
“We did not meet the needs of the frontier AI training market, and we didn’t meet the market needs,” said Anil Nanduri, Intel’s vice president of product management and go-to-market for data center AI accelerators, during the CES 2026 event in January.
With Intel’s new strategy, Nanduri said, the chipmaker has “a plan and a road map” to meet growing needs for inference and agentic AI workloads “with our CPUs, our GPUs, from the client, to the edge, to the data center.”
“We really want to go focus on this new world of AI inferencing, start leading with AI agents, and we want to be here to support you in that journey,” he said.
Tan Makes AI Top Priority As Intel Preps Next GPU
When Tan came on board as Intel’s CEO last year, he made it clear that AI, particularly AI accelerator chips for data centers, is a top priority for the company. He also stated plainly that he wasn’t “happy with our current position” in the market, vowing to learn “past mistakes and work towards a competitive system.”
“It won’t happen overnight, but I know we can get there,” Tan said.
While he didn’t explicitly state GPUs as the processor of choice for these efforts in his first public remarks last April, Intel had been in the process of moving from ASICs to GPUs as the lynchpin of its AI strategy over the previous two years.
The company elaborated on its strategy several months later, in October, when Intel revealed a 160-GB, energy-efficient data center GPU code-named “Crescent Island,” which it said is part of a new road map with an annual release cadence. This followed similar pushes by Nvidia and AMD over the past two years to move to annual release strategies.
Intel said at the time that the company expected to start sampling Crescent Island with customers in the second half of 2026, foreshadowing a launch the following year.
Positioning Crescent Island as “power- and cost-optimized” for inference workloads running on air-cooled enterprise servers, the chipmaker said the GPU is expected to feature its Xe3P microarchitecture that is optimized for performance-per-watt, 160 GB of LPDDR5X memory and support for a broad range of data types.
The Xe3P microarchitecture stems from the Xe GPU architecture the company first disclosed back in 2019 when it revealed its “Ponte Vecchio” GPU for high-performance computing and AI workloads in data centers. The company has released different implementations of the Xe architecture, including for PCs, over the past several years.
Intel Veers Into Heterogeneous Computing Again With New Developer Pitch
When Intel revealed Crescent Island in a briefing with journalists last September, Sachin Katti, who was the AI leader at the time, laid out its vision for AI hardware. Like previous eras at Intel, this vision centers around heterogeneous computing, which refers to systems that use multiple, distinct types of processors to boost performance and efficiency.
Katti—who Tan appointed a year ago to lead Intel’s AI strategy before he unexpectedly departed several months later last November for a job at OpenAI—said this vision will be based on open systems and software architecture that will involve delivering the “right-sized” and “right-priced” compute needed to power future agentic AI workloads.
This, he said, will translate into cost-effective systems with multiple processor components, potentially from different vendors, designed to address different aspects of agentic AI workloads—namely the “pre-fill” and “decode” stages.
Noting that Chinese AI startup DeepSeek popularized the concept of separating the pre-fill and decode stages in a large language model for inference, Katti said compute-optimized GPUs are better suited for the former while the latter gets more benefit from GPUs that have the “highest memory bandwidth possible.”
A similar approach was touted by Nvidia when it announced last September the upcoming Rubin CPX as a “new class of GPU” that is meant to work alongside the vanilla Rubin GPU in a single system, with the former handling pre-fill and the latter handling decode. Nvidia then paused development of Rubin CPX and announced in March that it would launch the Groq 3 LPU to handle prefill in tandem with the vanilla Rubin GPU.
“If we can build such a heterogeneous infrastructure, then we can optimize that performance-per-dollar by making sure that the right part of that agentic workload runs on the right-priced hardware with the right performance and delivers that overall, system-level performance per dollar that customers need,” Katti said last September.
He said this will be made possible by an “open software approach” that supports multiple infrastructure vendors and won’t require developers to “change any of their habits.”
As an example, Katti said Intel tested systems that run the pre-fill stage of a large language model on an Nvidia GPU and the model’s decode stage on an Intel accelerator chip. This allowed the company to achieve a 70 percent improvement in performance per dollar “compared to the homogenous systems out there today,” he added.
“That’s the strategy: We will be building scalable heterogeneous systems that deliver that zero-friction experience to agentic AI workloads and can deliver on the best performance-per-dollar for these workloads by leveraging this open heterogeneous architecture,” he said.
Intel Sees Rack-Scale Solutions As Important Part of AI Strategy
Shortly before Intel announced Tan as its new CEO last year, the chipmaker disclosed that it had cancelled an upcoming data center GPU to focus on a successor chip it was building for rack-scale AI solutions to challenge Nvidia.
Up until its cancellation, the “Falcon Shores” GPU was set for a 2025 release, and it was supposed to mark Intel’s first product to merge the road maps for its data center GPUs and Gaudi accelerator chips, which were previously developed on separate tracks.
Now that convergence was set to happen with the next generation, code-named “Jaguar Shores,” which Michelle Johnston Holthaus, then Intel’s interim co-CEO, said would allow the company to deliver a “rack-scale solution.”
Nvidia introduced the first rack-scale solution for AI workloads in 2024 with its Grace Blackwell GB200 NVL72 platform, which packs 72 Blackwell GPUs connected over its NVLink chip-to-chip interconnect. Such products have driven a significant amount of revenue growth for Nvidia.
When Tan came on board last March, he called rack-scale solutions a top priority.
“There’s no question we need to strengthen our position in the cloud-based Al data center market by developing competitive rack-scale system solutions, which will be a key priority for me and the team,” he wrote in Intel’s annual report that month.
While the company discussed Jaguar Shores during its first-quarter earnings call in April of last year, the chipmaker has only made a few references to the product line since then.
When Intel CFO David Zinsner was asked about the status of Jaguar Shores at an investor meeting last September, the executive responded, “Jaguar Shores is the product that is where we want to end up. But I think there will be milestones along the way and we do have a lot of technology within the company that we can leverage into the AI space. So I would just say, next few months, look for Lip-Bu to unveil that and talk about it.”
A few months later at CES 2026 in January, Nanduri, Intel’s AI go-to-market executive, showed a new road map promising an “annual predictable GPU cadence” that referenced a next-generation product line to follow the Xe3P-based Crescent Island GPU.
Nanduri called this the “next-generation inference-optimized GPUs and the Shores product line” without mentioning the “Jaguar” part of the code name Intel has previously used. (An Intel spokesperson doubled down on this nomenclature in a statement to CRN from earlier this year, referring it to as the “Shores product line.”)
In Intel’s latest earnings call last month, much of the focus was on high demand for server CPUs due to the growing role such processors play in agentic AI workloads, but Tan still made time to reaffirm the company’s continued work in GPUs and accelerator chips.
“The accelerator remains central to frontier AI, and we will continue to participate, innovate and partner in that category,” he said.
Tan Folds AI Group Back Into Data Center Unit, Hires Chief GPU Architect
When Tan appointed Katti to lead Intel’s AI strategy and road map last April, the CEO had split the AI group out from the company’s data center business unit.
But a few months after Katti left for OpenAI, Tan reversed that decision in January by putting the AI group back into the data center business unit under the former division’s new leader, former Arm executive Kevork Kechichian.
In Intel’s fourth-quarter earnings call in January, Tan, who oversaw the company’s AI efforts directly following Katti’s departure, said he merged the two groups again to ensure “tight coordination across CPUs, GPUs and platform strategy.”
“To support our AI objectives, I believe that our traditional server and accelerator roadmaps must advance together,” he said at the time.
In explaining Tan’s decision to move Intel’s AI accelerator chip team back into the Data Center Group, Kechichian said in a memo to employees the day before that “AI and the modern data center are fundamentally linked.”
“Customers are standardizing on complete AI platforms spanning compute, networking and software as the industry shifts toward inference and agentic solutions that run across the infrastructure stack. Xeon remains central across head nodes, edge deployments and many inference workloads,” he wrote in the memo, which Intel provided to CRN.
In the same memo, Kechichian confirmed that Intel in January hired Eric Demers, Qualcomm’s longtime GPU leader, to head up GPU engineering and the development of data center GPU solutions, as CRN reported earlier that month.
The data center leader also announced the hiring of Nicolas Dubé, a former HPE executive who “led the design, execution and delivery” of the server vendor’s Frontier exascale supercomputer for the U.S. Department of Energy.
Dubé, most recently a senior vice president at Arm, was named the leader of Intel’s data center systems and solutions. This role, according to Kechichian, will make him responsible for driving the “technical architecture and strategy for [the Data Center Group] toward full-stack systems and solutions, from chips to applications, ensuring integrated designs across compute, storage and networking.”
Intel Partners With Tan-Backed SambaNova For Heterogeneous AI
A month after confirming the hiring of Qualcomm’s longtime GPU leader, Intel announced in late February that it was entering a “multiyear strategic collaboration” with AI chip startup SambaNova Systems, for which Tan is the chairman and an investor.
As part of the partnership, Intel’s venture capital arm, Intel Capital, participated in SambaNova’s newly announced $350 million Series E funding round.
The announcement was made after Bloomberg reported in January that discussions for Intel to acquire the startup stalled. The publication first reported on acquisition talks between the two companies last October.
Then in early April, Intel and SambaNova shed more details about its partnership, saying that they are developing a reference architecture that combines GPUs, Intel’s Xeon 6 CPUs and the startup’s RDUs to handle different components of agentic AI models.
Set for release in the second half of the year for enterprises, cloud providers and sovereign AI programs, the design will make use of GPUs to handle a model’s prefill phase, RDUs to handle the decode phase and CPUs for tool calling and system orchestration.
A SambaNova spokesperson told CRN that OEMs will be able to use the reference architecture to “sell their own implementations.” No OEMs have been confirmed yet.
The startup positioned the pairing of GPUs, CPUs and RDUs as a solution that would require far less energy, enable higher utilization and fit into existing data centers in contrast to Nvidia’s Vera Rubin platform that consists of Vera CPUs, Rubin GPUs and Groq LPUs.
““Together with Intel, we’re giving customers a blueprint they can deploy in existing air‑cooled data centers, with broad x86 coverage for the coding agents and tools they already use today,” said SambaNova CEO and co-founder Rodrigo Liang in a statement.
Intel Makes Play For AI Workstations With Arc Pro
Outside of data centers, Intel is building upon several years of development of discrete GPUs with the Xe architecture to build products for the emerging category of AI workstation PCs.
The semiconductor giant put this into focus in March with the expansion of its Arc Pro B-series GPUs, which it said provides “optimized performance for multi-user and multi-agentic AI workloads” with “strong price-to-performance for inference.” This announcement built upon Intel’s initial launch of Arc Pro-B series GPUs last year.
In an interview with CRN, Nanduri, Intel’s AI go-to-market executive, said he thinks the Arc Pro B-series GPUs will play well with the rise of autonomous AI agents, which has become a focal point in the industry thanks to the popular OpenClaw AI assistant software.
“We are looking at the cusp of the OpenClaw moment and agentic AI inference, and you start to realize that we have a very nice product that fits the right price point, fits the right capabilities and can give you a multi-card scaling that takes the graphics workstation to the next level,” Nanduri said in March.