NVIDIA’s in-house Vera CPU has emerged as a new variable in the server processor market, but a channel survey cited by ABMedia says it is unlikely to shake the x86 camp in the near term. The report argues that the bigger forces behind Intel’s competition with Advanced Micro Devices, or AMD, are the yield trajectory of Intel’s 18A process, new “outside-the-rack” CPU demand driven by AI agent workloads, and whether server market growth can hold up in 2027.
The survey estimates Intel’s server CPU share at about 57% in 2026, down from 65% in 2025. AMD is put at roughly 30%, while hyperscale cloud providers’ in-house chips account for about 13%. Intel is using price increases and a richer product mix to push up ASP, or average selling price, but the report says yield issues and ongoing share loss remain its main risks.
Higher 18A pricing and mix shift lifted server CPU ASP
According to the channel survey, Intel has already carried out three rounds of price increases in 2026 for high-end SKUs built on the 18A process, while basic versions were left unchanged. The report expects no major additional increase for the rest of this year. The next adjustment is seen in the second quarter of 2027, when the enhanced 18A-P enters volume production at Fab 62. At that point, average pricing is expected to rise by about 10%, alongside a conservative 10% to 15% gain in performance.
Intel’s second-quarter server CPU ASP rose about 48% from a year earlier, while shipments increased only about 10%. The survey attributes that gap mainly to a change in product mix: midrange and high-end products climbed from about 10% of total shipments to about 30%, on top of 20% to 30% price hikes for higher-end models.
On large customers, the report says Amazon Web Services, Google, Microsoft and Lenovo operate under annual volume framework agreements. Existing SKUs under those contracts have price protection, and together they account for about 50% to 60% of Intel’s server CPU orders. New high-end 18A products, however, are excluded from older price-protection terms, and discount tiers in 2027 are set to tighten. The survey also names X, Elon Musk’s company, as a new major customer, with total orders of around 800,000 to 900,000 units scheduled across three quarters.
AI agents could create new CPU demand outside GPU racks
The most closely watched shift in the survey is demand for “outside-the-rack” CPUs tied to AI agent workloads. This is not about packing more GPUs into server racks for training or inference. The report describes it as demand for large numbers of mid- to high-end server CPUs running agent tasks outside those rack-based GPU systems. Orders accelerate noticeably in the fourth quarter of 2026, though this year’s scale is still limited to several hundred thousand units, mainly for proof-of-concept, or POC, work and small deployments.
Under an optimistic scenario, demand in 2027 for mid- to high-end server CPUs used in outside-the-rack agent infrastructure could reach 1 million to 2 million units. That would represent roughly 16% to 20% of total server CPU shipments.
The survey says hardware priorities for this workload differ from past patterns. In virtual machine environments, with traffic swings between peak and off-peak periods taken into account, one physical CPU core can typically serve four to five users. Each user needs about 8 GB of DRAM, and DDR5 is considered sufficient. HBM does not offer a clear advantage for this type of workload. Hardware priorities are listed, in order, as core count, storage capacity, single-thread performance and memory bandwidth.
That points to demand favoring high-core-count x86 CPUs with better cost efficiency, rather than tightly coupled CPU-GPU designs.
18A yields and Fab 62 are central to Intel’s margin and share outlook
The survey says Intel’s 18A story still comes down to yields. It puts the breakeven yield threshold at about 83%, while the share of dies with more than 100 usable cores is below 10%. Fab 52 has a year-end yield target of 90%, a level the report says is needed before healthier gross margin can be released. Further improvement would then depend on Fab 62, which is being built specifically for 18A-P. Current 18A good-die output is about 1 million units per quarter.
Pricing also reflects how scarce higher-core-count products are. A 64-core product is priced at about $3,000 to $5,000, while products with more than 100 cores are around $10,000. Lead times have stretched as well: strategic customers face waits of roughly three to four months, while general customers are looking at five to six months, mainly because more complex, higher-end SKUs make up a larger share of supply.
The survey’s conclusion is that Intel’s ability to stabilize and recover market share depends on whether Fab 62 can deliver enough high-core-count products. Until then, AMD and hyperscalers’ in-house chips are expected to keep taking share.
Vera is not seen as the biggest near-term threat
On NVIDIA Vera CPU specifically, the survey takes a restrained view: 「NVIDIA’s in-house CPU is unlikely to replace major x86 server CPU customers in the near term. Its strength lies in tight coupling with GPUs and handling heavy AI compute, but for outside-the-rack workloads related to agents, the cost-performance profile is not ideal.」
The report points to software as another practical barrier. Most software stacks in existing data centers are built on x86. Moving to Arm architecture requires meaningful migration spending, and the survey says that remains a concrete concern for both cloud operators and enterprise customers.
2027 shipment outlook is cautious, while 2028 remains hard to read
The survey says Intel internally estimates total server CPU shipments at about 13 million units in 2027, including 7 million AI server units and 6 million traditional server units. That is materially below the double-digit growth narrative management has stressed publicly. If outside-the-rack agent demand proves strong, the total could rise to 15 million to 16 million units.
Visibility into 2028 remains limited, according to the report. Profitability across the AI industry is still unclear, and server replacement cycles are not settled. Book depreciation runs for three years, but actual service life could extend to five years.
The consumer side is also described as weak. The survey estimates more than 170 million consumer CPU orders for full-year 2026, with only limited annual growth and clearer pressure in the second half. End-product prices have risen sharply, by about 30% to 35% for laptops and about 50% to 60% for desktops, which the report says has hurt buying interest. Channel inventory adjustment is expected to weigh on consumer CPU orders in the first quarter of 2027.