Intel's Data Center Revenue Grew 78 Percent. Memory Prices Still Cost It a Quarter of Its Market Value.

The quarter Intel reported on Thursday afternoon was, by the company's own measure, the best data center quarter it has had in 15 years. Revenue came in at $16.5 billion, up 21 percent. GAAP gross margin was 43.0 percent against 38.2 percent a year earlier. Non-GAAP earnings per share of $0.52 beat the $0.38 the company had guided to. Data Center and AI revenue reached $7.3 billion, up 78 percent.

By the end of the after-hours session the shares were at $70.86, against a $92.07 close. That is about 23 percent of Intel's market value gone in an afternoon, and The Register's Simon Sharwood put the fall at almost a quarter.

The results do not explain that. The guidance does.

A quarter that beat every number it set

The quarter ended 26 September. Intel generated $5.1 billion in cash from operations and completed a $23 billion equity offering, which is part of why GAAP earnings per share fell 33 percent to $0.60 while revenue rose: the share count grew. GAAP net income was $3.3 billion, down 18 percent, while the non-GAAP figure was $2.8 billion, up 169 percent. Much of the gap between those two lines is a tax rate that went from 6.6 percent to 11.6 percent.

Client Computing and Physical AI, the group that sells PC processors, reported $8.1 billion. That was down 5 percent year over year and 9 percent sequentially, and David Zinsner, the CFO, told the call it came in below the company's own expectations. The segment still made $2.3 billion of operating profit, 28 percent of its revenue, roughly flat on the prior quarter. AI PC revenue grew 35 percent inside a shrinking segment, and Wildcat Lake, the cost-optimized 18A client part, shipped about 70 percent more units than in Q2 and is sold out.

Then the guidance.

The guide is where it broke

Intel guided fourth-quarter revenue to $15.0 billion to $16.0 billion, a midpoint of $15.5 billion, with non-GAAP gross margin of 38 percent and non-GAAP earnings per share of $0.25. Ben Reitzes of Melius, on the call, described the guide as roughly $1.8 billion below the Street. Gross margin drops five points from the 44.1 percent Intel just delivered.

Two things sit inside that number. The first is an accident. "Early this month, an unfortunate and tragic accident led to an extended shutdown in our Israel factory," Lip-Bu Tan said in his prepared remarks. The plant, Fab 28, was down about two weeks and was due to restart production that week. Intel's early assessment puts the cost at roughly $300 million of fourth-quarter revenue and 200 basis points of gross margin, almost all of it in 10-nanometer server products, and Zinsner expects a similar revenue hit in the first quarter.

The second is the PC business. Intel expects Client Computing revenue to fall more than 20 percent sequentially in Q4, as customers "move quickly to reduce weeks of inventory and align purchases with their expectations for 2027 consumption."

A note on the arithmetic, because two trade outlets read this guidance differently and both readings are wrong. The Register reported that Tan predicted the PC chip market "will shrink by 12 percent next year." Thurrott reported a 26 percent decline in 2026 followed by another 10 percent in 2027, and read Intel's foundry line as a 58 percent decline where the earnings release's segment table says foundry revenue rose 58 percent to $6.7 billion. The prepared remarks and the call transcript say the same thing as each other: the 2026 PC unit TAM is expected to decline approximately 12 percent, and 2027 is planned at roughly 10 percent down.

What memory actually did this quarter

Intel's problem is not that PCs got worse. It is that the parts inside them got more expensive faster than anyone could reprice a finished product.

TrendForce's 30 September forecast put conventional DRAM contract prices up 10 to 15 percent quarter over quarter in Q4, with NAND up 15 to 20 percent. A 20 October TrendForce item citing Commercial Times put the DRAM figure closer to 20 percent and gave the trajectory: about 80 percent in the second quarter, about 40 percent in the third, now 10 to 20. Slowing, in other words, from a very high base, and still rising. TrendForce also expects PC DRAM supply to decline in 2027 as capacity shifts to server.

That is the mechanism behind the shipment numbers IDC and Omdia published on 9 October. IDC counted 62.7 million PCs shipped in Q3, down 20.1 percent from 78.5 million a year earlier and 9.1 percent from Q2, the sharpest third-quarter decline it has recorded. Omdia counted 58.1 million units, down 21.2 percent, with desktops down 23.5 percent. Omdia's Ben Yeh put memory and storage at nearly 40 percent of a PC's bill of materials, against a typical 15 percent, and Omdia expects another 24 percent decline in Q4 before a shallower 7 percent drop in 2027.

IDC's June forecast had the full year at minus 11.3 percent. The third quarter alone came in at minus 20.1. Forecasts made before the pricing reset have been overtaken by it.

The price notices followed. MCPC, a device lifecycle management company that tracks manufacturer and distribution price actions for its customers, logged HP's 10 to 15 percent increase on Personal Systems effective 15 September, Lenovo's configure-to-order and Top Seller increases averaging about 6 percent, Panasonic's 10 to 15 percent on TOUGHBOOK from 1 October, and an expected Dell increase of 10 to 20 percent in early November. MCPC is a channel partner working from partner communications rather than published price lists, and says so. The pattern across four vendors is still the pattern. HP told partners that technology prices are up more than 50 percent year over year in 2026 and that it does not expect the trend to slow in 2027.

Microsoft reported the same break from the other end

Microsoft's quarterly report, out the same week, described the demand break from the software side. "Windows OEM revenue was below expectations and decreased 27% and 28% in constant currency, driven by the Windows 10 end of support demand in the prior year, inventory drawdowns by OEM and channel partners and weaker PC market demand," Amy Hood, the CFO, said on the call. The broader Windows OEM and devices category fell 22 percent. Hood expects "these trends to continue through the rest of the fiscal year," which for Microsoft runs to the end of June 2027.

Intel's client revenue fell 5 percent in the same quarter. Microsoft's Windows OEM revenue fell 27. One is a component sold to OEMs, the other a licence sold with the finished machine, and both numbers moved in the same direction at the same time. Microsoft had a good quarter otherwise: $92.3 billion in revenue, Azure up 48 percent, and Microsoft Cloud up 30 percent, now roughly 70 percent of the total.

The capacity Intel is choosing to move

Intel's answer to the client problem is, in part, to stop feeding it quite so much. Zinsner said the company is taking 18A wafers offline to bring Intel 3 into Oregon, which frees capacity to build more data center wafers. Industry server CPU units are expected to grow more than 30 percent in both 2026 and 2027, and Intel expects to be short of that demand. "We will likely not fully meet what they require, but I think we are going to be in a more challenging situation in the first half of the year than we are the second half," Zinsner said. He later put it in supply terms: the first half of 2027 will be tighter on wafers and substrates than the second, with 40 percent more substrate capacity arriving mostly in the back half.

Tan's argument for why that capacity is worth chasing is the most interesting thing on the call. He argued that agentic workloads shift the computational bottleneck toward retrieval and orchestration, work CPUs do cheaply. "While GPU reasoning query run as high as multiple watt-hours of energy, a single CPU query could cost only joules of energy," he said. On that basis Intel estimates the server CPU market at $40 billion today and more than $200 billion in five years.

The tension is that Intel is making this bet while its largest business contracts, and the improvement it has scheduled for that business is a comparison effect rather than a recovery. Client gets better in the second half of 2027 "against an easier comparison," with Nova Lake, the first 18A-P client part, arriving into it. On memory, from the same transcript: "Make no mistake, we are not optimistic about next year in terms of memory. We think that it is going to impact the TAM. We think TAM will be down 10% on a units basis next year. Then we will have to see in 2028 whether things improve. But our expectation is that it does not improve in 2027."

Intel plans to sell into a PC market it expects to be 10 percent smaller, with a gross margin five points lower than the one it just reported, while moving fab capacity toward a server market it cannot fully supply. SK Hynix said this week that memory shortages will persist until 2030, as The Register reported. Nobody in this chain is forecasting a quick return to earlier pricing.

What this means if you have a refresh on the books

Read it as a procurement story rather than an Intel story. Intel's client business could be run perfectly and the arithmetic would not change: memory is close to 40 percent of the bill of materials, those prices are set by a market allocating capacity to AI servers, and a PC quoted in January costs materially more to build in October.

That produces two constraints landing in different budget lines. Endpoints cost more and quotes expire faster, with Dell's 30-day validity, per MCPC's tracking, now the longest lock in the market. Servers are allocated rather than merely priced, and Intel has said the first half of 2027 will be tighter than the second.

The planning question is not whether to refresh. It is which parts of a refresh can wait for a memory market that may not normalize this decade, and which parts cannot, and what you specify while you wait. Memory and storage configurations are now the largest lever on unit cost, which makes the low-memory standard build a strategic choice rather than a corner cut. On the server side, allocation is a relationship and a forecasting problem, and the teams that committed volumes early are the ones getting wafers in the first half.

At DMC we work on that boundary with hardware companies and the organisations buying from them: BOM cost modelling when component prices move quarterly, sourcing strategy when allocation rather than price is the constraint, and production and refresh planning that holds up when the market reprices mid-cycle. If your 2027 plan is still built on January's prices, let's talk.