Intel’s Falcon Shores future looks bleak as it concedes AI training to GPU rivals On Monday, Intel sent a letter to employees detailing its comeback plan after an abysmal second-quarter earnings report with critics calling for CEO Pat Gelsinger to resign.

The letter detailed the company’s plan to center its products around the x86 architecture. That could be bad news for GPU codenamed Falcon Shores, as Gelsinger previously said he will not compete with Nvidia and AMD in the AI training space.

Gelsinger cleared the air in a Monday letter to employees, starting with, “There has been no shortage of rumors and speculation about the company … so I’m writing today to provide some updates and outline what comes next.”

The next phase of Intel’s restructuring includes cutting another $10 billion in costs and laying off 15,000 employees, of which 7,500 have already agreed to voluntary buyouts.

Gelsinger stated a third pillar: “We must refocus on our strong x86 franchise as we drive our AI strategy while streamlining our product portfolio in service to Intel customers and partners.”

The company is exiting the AI training market and focusing on inferencing, in which x86 chips help Intel operate from a position of strength, Gelsinger said last month at the Deutsche Bank analyst conference.

“That’s an area where our strengths are much more substantial, where the role of the CPU… even as you think about some of the evolutions that we’re seeing in the architecture of AI systems,” Gelsinger said.

The desire to cut costs and axe irrelevant products could impact Falcon Shores, a discrete GPU designed for data centers that have already been delayed until recently. It is the successor to a GPU codenamed Ponte Vecchio, which was discontinued after the release of the Aurora supercomputer.

Intel didn’t respond to a request for comment on the future of Falcon Shores, which is due for release next year.

The major architects behind it — Jason McVeigh and Raja Koduri — are either gone or reassigned, and Falcon Shores looks like an orphan.

Gelsinger acknowledged that Intel was way behind its competitors on GPUs and AI training chips.

“As I view it… in the four-horse race on this side of the page, Nvidia, (AWS’s) Trainium and Inferentia, Google Cloud’s TPU, and AMD, and Intel’s number four… that’s hard,” Gelsinger said.

Gelsinger estimated that 70% of computing is done in the cloud, while 80% of the data remains on-premise. Intel has a good installed base of x86 chips on-premise where inferencing is done, so the company wants to build on that.

“So our position in the CPU… and we’re starting to see businesses ask those questions about AI. It’s on this side of the page where our strengths become stronger, our Xeons are stronger, and our system platform of CPU plus GPU are stronger,” Gelsinger said.

Intel is not competitive with Nvidia, AMD, and Google in AI training chips.

“That race, you know, they are so far ahead, some of the other challenges that we have. We’re just not going to be competing anytime soon for high-end training,” Gelsinger said at the analyst meeting.

Intel’s AI strategy is best represented by the Lunar Lake PC chip, which was announced earlier this month. The chip includes a neural processor and a GPU that the company claims can deliver more than 100 TOPS of AI performance.

Gelsinger’s Monday note also made a reference to chiplets being a key part of their strategy.

Intel is separating its chip design business as it reinvents itself into a manufacturing company. The company’s market value has crashed to $89 billion, while Nvidia’s is at $2.86 trillion, and its Blackwell GPU will be out soon.

Intel’s only legitimate training chip is Gaudi 3, which started shipping recently. Last month, IBM announced it would offer Gaudi 3 through its cloud service to customers.

“Our AI investments—including continued leadership of the AI PC category, our strong position with AI in data center, and our accelerator portfolio—will leverage and complement our x86 franchise with a focus on enterprise, cost-efficient inferencing,” Gelsinger said in the Monday letter.

Intel has faith in its PC roadmap but not as much in the server chips.

“Where we still haven’t completely gotten the business to a good place is on the data center side of CPU,” said Intel chief financial officer Dave Zinsner this month during an interview at Citi Global’s analyst day.

Intel was expecting a stronger server cycle, but the “AI build” market is depressing the server CPU market, he said.

“That portion of the business hasn’t been as robust as we expected,” Zinsner said. “That area has been a bit disappointing that we haven’t seen more uptick there, but with good products, we’re starting to compete more effectively.”

The current Emerald Rapids server CPU has been a dud, and the next-generation Granite Rapids is due early next year. The successor, Diamond Rapids, will put Intel at the top.

“I think Granite [Rapids] is a meaningful step forward for us in terms of making us competitive. Diamond Rapids will definitely put us in a good place competitively. It’s just important we work our way through the roadmap to get us to where we want to be,” Zinsner said.

Intel was expecting Granite Rapids to be its major breakthrough server product that would separate it from AMD’s Epyc chips. AMD plans to ship its next-gen server chip codenamed Turin, which is based on the Zen 5 architecture, later this year.

However, a lot of server competition is coming from ARM-based chips. Microsoft and Google have made their ARM-based CPUs.

Intel’s final goal is to get to its 18A manufacturing node, which includes new chip technologies such as RibbonFET and PowerVia. Intel’s dense server chip codenamed Clearwater Forest will be released on that node.

Intel also made some foundry announcements.

Intel signed a deal with a major cloud provider to customize x86 chips based on the Xeon 6 architecture. The cloud provider makes its own ARM-based CPUs that do not work with GPUs. Custom x86 chips will allow the cloud provider to offer more power-efficient chips that work with GPUs.

The U.S. government also awarded Intel $3 billion to make chips for U.S. agencies.

“As the only American company that both designs and manufactures leading-edge logic chips, we will help secure the domestic chip supply chain,” Gelsinger said in the letter.

Intel also delayed its plans to establish a factory in Germany by two years, given its financial struggles.