Last month, Intel held its 4Q24 earnings call, hosted by the company’s interim co-CEOs. We’ll cover Q4 results after Nvidia reports. Here we’ll discuss what the call revealed about Intel’s roadmap.
18A Mixed Messages—Products
- Focused on products, co-CEO Michelle Johnston Holthaus stated that Intel Foundry Service (IFS) must earn her business and expressed confidence in the manufacturing organization.
- She reiterated that the next laptop processor, Panther Lake, will employ the IFS 18A process. (Of course, past CEO Pat Gelsinger said IFS would fabricate Arrow Lake.) In context, her comments could mean all Panther compute tiles (chiplets) will be made in house but others (e.g., the GPU tile) will not.
- However, for the subsequent Nova Lake processor, Intel will outsource some compute-tile production.
- Moreover, Intel will “be doing the same look” [considering outsourcing] for its data-center portfolio.
- To summarize: she’s confident in the IFS team but not completely.
18A Mixed Messages—IFS
- Co-CEO David Zinsner acknowledged that in addition to delivering technology, IFS must win customers’ trust.
- He stated that IFS has a good RFQ pipeline but acknowledged it won’t win every deal and will be selective. (Probably like how I selectively don’t date supermodels.)
- He targets the end of 2027 for IFS breaking even, driven by greater internal demand and fabricating more wafers in an EUV-based process (i.e., 18A).
- Curtailed capex will also help the company’s cash flow. Reduced capex could be either a cause or an effect of the product group outsourcing some Nova CPU tiles.
- Overall, he’s expressing more realism about 18A’s prospects than the company has previously but can’t escape from it being a make-or-break technology for IFS.
Charging the AI Market
- In Q4, the company again took charges against Gaudi inventory.
- Sensing a similar disaster with Falcon Shores, Holthaus has relegated it to a test chip.
- Intel is turning its AI focus to rack-scale solutions and will deliver Falcon’s successor, Jaguar Shores, as part of a system.
- Given the company’s adjusted free cash flow of negative $2.2 billion in Q4 and $22 billion in the bank, we expect Intel to cancel Jaguar as well.
Muddy Waters
- Holthaus indicated the P-core Xeon code-named Granite Rapids and its successor, Diamond Rapids, are garnering positive customer feedback. At the same time, she acknowledged share loss and called Granite only a first step toward bringing back customers.
- Clearwater Forest, the next E-core Xeon, was to be a leading 18A product but has slipped to 1H26.
- The reasons for Clearwater’s delays are muddy. One possibility is 18A yield, but Holthaus alluded to packaging issues and characterized E-core server chips as niche products.
- We think Intel will nix Clearwater also.
- The first E-core Xeon, Sierra Forest, and Clearwater exist to fend off Arm processors. Originally, Arm-based server chips delivered similar performance to big-core Xeons only when comparing one dual-thread Intel (or AMD) CPU to two physical Arm cores. Except for the unpopular AmpereOne processor, newer Arm-compatible server chips have delivered similar performance as big-core x86 processors when comparing a like number of physical CPUs.
- Clearwater, however, was also to be an answer to AMD’s Epyc processors with V-Cache—separate SRAM dice bonded to Epyc’s compute chiplets. Clearwater was to pioneer Intel’s Foveros Direct 3D technology, which bonds dice to an active base die—that is, compute chiplets to an SRAM die.
- Foveros Direct 3D could be the packaging technology Holthaus alluded to in explaining Clearwater’s delay. As AMD has shown, bonding SRAM to compute dice can yield an appealing product. It’s an odd fit for an E-core chip, however.
Bottom Line
Although Intel’s co-CEOs are conveying more realism and less optimism than their predecessor, the company’s turnaround is as dependent as ever on 18A. External foundry customers will not contribute much to 18A demand compared with Intel’s product group, and the company is reducing its planned capital spending.
The product group, meanwhile, is hedging its bets, outsourcing some manufacturing. In the data-center market, Intel understands that its high-performance server processors must regain market share and mindshare. The E-core Xeons aren’t nearly as relevant to its future, other than to prove SRM bonding. Meanwhile, Intel is still digging for the AI pony.