Every quarter we look at earnings from an illustrative set of processor companies targeting data-center, PC, and embedded markets.
Data Center
AMD
AMD’s reported data center revenue surpassed Intel’s. Epyc sales set another record, but it was MI300X revenue that propelled AMD past its x86 rival. Growing rapidly from a near-zero base, the company’s data-center GPU sales almost matched those for server processors. AMD raised its full-year GPU estimate to $5 billion.
Customers employ the AMD MI300X mainly for inference, not training. Although AMD is improving its GPU ROCm software needed to achieve the MI300X’s performance and features, this effort to catch up with Nvidia’s Cuda is but one required to gain meaningful training share.
Epyc’s share gains, however, have stalled out but should creep upward. AMD is winning enterprise designs, and OEMs offer 50% more models than a year ago. Incremental enterprise progress even after several Epyc generations illustrates how slowly the industry moves to a new vendor.
Intel
This past quarter Intel launched the Granite Rapids Xeon 6, finally achieving competitive server-processor core counts. The company’s focus, however, is operational efficiency, and a priority is its “x86 franchise.” Although the company isn’t clear what this term means, it implies to us cutting non-x86 projects. However, the company still articulates dreams of supplying data-center AI accelerators (GPUs/NPUs). Intel cannot have it both ways. We expect it to exit the latter.
It won’t be missed. Whereas AMD raised its 2024 AI-accelerator estimates by $500 million, Intel disclosed it won’t hit its estimate—which was only $500 million total. It’s probably not a small miss, either. The company wrote down $300 million in accelerator inventory. Intel blames the shortfall on the transition from Gaudi 2 to Gaudi 3. Having announced it was abandoning the Gaudi architecture, Intel shouldn’t be surprised its few customers aren’t upgrading and new ones aren’t coming on board. Although Intel recognized the AI market potential early, it has failed to capitalize on it. The company’s operating cash flow is positive, free cash flow is negative—a situation that will force the company to move on from data-center AI.
Nvidia
Data-center AI, however, is propelling Nvidia, but growth has come at a cost. Company-average gross margins dove to 75%. The sickening plunge should continue, dropping to a measly 73% next quarter. How will the company survive? The H200 Hopper midlife kicker generated double-digit billions in revenue. Blackwell is in full production, and the quarter’s shipments topped 13,000. Given Elon Musk’s xAI company is raising $6 billion to buy 100,000 chips, we can estimate that each Blackwell costs $60,000 (counting the chip itself and every other system aspect).
Overall, Nvidia’s data-center revenue more than doubled compared with the year-ago quarter, as Figure 1 shows. Even sales to China grew, despite export restrictions. The company reports demand exceeds supply and is raising its Blackwell sales estimates. One data-center area that fared relatively poorly was networking. Although Spectrum-X (Ethernet switching) revenue tripled, overall networking revenue increased only 20% year over year and fell sequentially. This speaks to how keen customers are to avoid Nvidia’s proprietary InfiniBand networking. Recent Ethernet-based AI clusters show the standardized technology has improved enough to replace the specialized interconnect.
Figure 1. AMD, Intel, and Nvidia data-center revenue. (Data source: companies, except for estimates.)
PC
AMD
AMD’s PC-processor business also had a strong quarter. At $1.9 billion, sales were up 29% from a year ago, and AMD’s x86 unit share approached 24%. The company has long been strong in the retail desktop-processor segment (i.e., to gamers), and this business was also up despite a flubbed Zen 5 (Ryzen 9000) launch. The recent 9800X3D launch, however, went well. We expect that product to bolster Q4 revenue, and the company expects further growth. The company’s challenge is to transcend the consumer-PC niche and gain enterprise share, particularly in mobile PCs.
Intel
Intel’s Arrow Lake debut was at least as disastrous as the initial Ryzen 9000 launch, and the 9800X3D destroys it on the gaming benchmarks that matter to retail buyers. Intel Lunar Lake, however, has done well. The company has moved this niche backstop product to the forefront of its mobile-PC efforts. It’s a costly, low-margin product, however, because it employs chiplets on a silicon substrate, integrates DRAM, and relies on TSMC-fabricated dice. Nonetheless, Intel dominates PC-processor sales. Due next year or early 2026, the next-generation Panther Lake PC processor will employ an Intel-fabricated computing die and won’t integrate DRAM but will still employ at least one TSMC-made die (presumably for the GPU) and still rely on Intel’s costly chiplet packaging.
As an aside, Intel’s stated x86 prioritization could imply demand for Arm-based PC processors. Qualcomm recently said it expects to sell $4 billion in PC processors annually by 2029. That would make it half the size of AMD now, which realistically isn’t a sustainable share. Other companies (e.g., AMD, MediaTek, Nvidia, and Samsung) could enter the Arm-compatible PC processor business this year, which could trigger Windows flipping entirely to Arm.
Nvidia
For now, Nvidia’s PC business is selling GPUs. Revenue climbed 15% from a year ago, as Figure 2 shows. We believe strong sales of AMD desktop processors helped Nvidia grow. The GPU company projects Q4 revenue to fall owing to supply constraints. We expect the company to prioritize H200 and Blackwell production over that of consumer GPUs for the obvious reason.
Figure 2. AMD, Intel, and Nvidia PC revenue. (Data source: companies, except for estimates)
Embedded
The embedded-processor market has been in turmoil. We track a company group that, while small, illustrates market dynamics. Mainly coming from FPGAs, AMD’s embedded revenue is in recovery, down year over year but up sequentially. The company states design wins are up 20% in 2024, however, boding well for the future. Intel’s NEX revenue has been steady, having arrested its decline earlier than the other companies. This reporting unit, however, includes a combination of Ethernet components, processors for networking and telecommunications, and processors for “edge” (i.e., other embedded) applications. The company will start folding results from the lattermost segment into client PCs and is rumored to be selling off the networking/telecom piece.
Among these companies, Microchip is the most conventional embedded supplier. Its microcontroller (MCU) sales have continued to fall and should decline again next quarter, as Figure 3 shows. The company sees strength in the military/aerospace and data-center AI segments but weakness in all other end markets regardless of region. Visibility is weak, but the company usually has a lot of turns business (orders for same-quarter shipments). It believes that it is shipping considerably under consumption owing to the inventory bullwhip effect as order curtailments from end customers ripple up through the supply chain.
Figure 3. AMD, Intel, and Microchip embedded revenue. (Data source: companies, except for estimates)
Bottom Line
The lackluster embedded market indicates that a broad swathe of the economy is doing well but not so well as to digest oversize semiconductor inventories. Until levels normalize, the semiconductor industry won’t bounce back, assuming a demand-side downturn doesn’t occur first. Meanwhile, the PC market is humming along fine, and the data-center market is rocketing upward. Although AI investments may be diverting spending from conventional servers to AI systems, server processors are selling well.
The biggest issues are company specific, and the specific company in question is Intel. It’s easy to fault its product development, but its manufacturing problems overshadow everything. Racing to regain process-technology preeminence (or at least parity), the company also must invest in capacity even as it takes charges for underutilization at older nodes. Therefore, even as its operations generate cash, free cash flow is negative $2.7 billion. This position constrains Intel’s flexibility and will force difficult decisions.