AMD and Intel Just Put CPUs Back at the Center of the AI Buildout

AMD has sharply raised its server CPU outlook, while Intel has reported its strongest revenue growth in more than 15 years. Together, those signals challenge the assumption that AI infrastructure spending belongs almost entirely to GPUs.

The change is visible in actual orders, capacity plans, and product schedules. Intel’s data center business grew 59% year over year in the second quarter of 2026. AMD now expects the server CPU market to exceed $120 billion by 2030.

This is not a story about CPUs replacing Nvidia GPUs. It is about CPUs becoming more valuable as AI systems expand beyond model training. Inference, agents, data movement, networking, and storage all create work that accelerators cannot handle alone.

The central contest is now AMD versus Intel for the host-compute layer surrounding those accelerators. Memory availability and manufacturing execution will determine whether either company can convert demand into sustained growth.

Intel’s Earnings Turn AI Demand Into Measurable CPU Growth

Intel’s second-quarter results provide the clearest evidence that AI spending is reaching the CPU layer.

Intel reported second-quarter revenue of $16.1 billion, a 25% increase from the same period last year. Chief executive Lip-Bu Tan called it the company’s strongest revenue growth in more than 15 years.

The company’s Data Center and AI division produced $6.3 billion in revenue. That represented 59% annual growth and made the division Intel’s fastest-growing major product business.

Those figures matter because Intel does not dominate the accelerator market. Nvidia remains the central supplier of GPUs used for training and serving large models. Intel’s growth therefore points toward a broader infrastructure cycle.

Every large AI cluster requires host CPUs. These processors initialize accelerators, run operating systems, schedule jobs, manage storage, and handle data before it reaches a GPU.

Agentic AI adds another layer of demand. An AI agent is software that can plan and execute multistep tasks using models, tools, and external data.

Agents repeatedly call models, databases, search systems, and business applications. Those interactions create orchestration workloads that often run on conventional server processors.

Intel’s quarterly results also showed that the growth was not limited to one product category. Client and physical AI revenue reached $8.9 billion, up 13% year over year.

Intel Foundry reported $5.8 billion in segment revenue, up 31%. However, much of that revenue came from Intel’s own product groups rather than outside chip designers.

The company’s overall results also require careful interpretation. Intel reported an $11 billion GAAP net loss, despite posting $2.2 billion in non-GAAP net income.

Much of that gap reflected charges outside ordinary operating performance. Still, it shows why revenue growth alone cannot settle the company’s turnaround case.

Intel expects third-quarter revenue between $15.8 billion and $16.8 billion. It also plans to increase investments in manufacturing equipment, clean-room capacity, and substrates.

That response is significant. Management is not treating the second-quarter increase as a temporary rush for existing inventory. It is preparing for higher product and foundry demand into 2027.

Intel launched Xeon 6+, its first server processor manufactured with the Intel 18A process. The company also said its factories exceeded internal volume targets across several process technologies.

The tension begins here. Intel has demand, a large installed base, and improving production output. Yet it must spend heavily before customers prove that current AI infrastructure plans will translate into lasting utilization.

Why AMD Sees a Much Larger Server CPU Market

AMD believes AI agents will require more CPU capacity for every deployed accelerator, not less.

During its first-quarter earnings discussion, AMD said it had materially increased its expectations for the server processor market. The company previously expected that market to grow about 18% annually.

AMD now projects growth above 35% annually, with total addressable demand exceeding $120 billion by 2030. Total addressable market, or TAM, estimates the revenue available across a defined product category.

That forecast comes from AMD rather than an independent market measurement. Investors should treat it as a strategic target shaped by customer discussions, product plans, and management assumptions.

The underlying mechanism is credible, even if the exact figure remains uncertain. AI systems still need CPUs to coordinate accelerators and connect them with the rest of a data center.

Training clusters use CPUs for job scheduling, data preprocessing, checkpoint management, and infrastructure monitoring. Inference systems add routing, retrieval, security, and application logic.

Agentic systems intensify that pattern. A single user request can trigger several model calls, database searches, permission checks, and tool executions.

Each step consumes accelerator time, but it also generates conventional compute work. As enterprises connect agents to internal applications, the CPU workload grows alongside model usage.

AMD said these requirements include orchestration, data movement, parallel execution, and head-node services. A head node is the server that coordinates a larger collection of accelerator nodes.

The company expects server CPU revenue to grow more than 70% year over year in its second quarter. AMD is scheduled to report those results on August 4, creating an immediate test for the forecast.

AMD also needs products capable of serving that demand. Its answer is the sixth-generation EPYC family, code-named Venice, based on the Zen 6 architecture.

The company began the Venice production ramp on TSMC’s 2-nanometer manufacturing process. AMD says customer validation is broader at this stage than for any previous EPYC generation.

Venice includes Verano, a processor designed specifically for AI infrastructure. AMD expects the family to support cloud, enterprise, and accelerated computing deployments.

The distinction between a general server CPU and an AI-focused host CPU is becoming more important. Data center buyers increasingly evaluate the complete rack rather than individual chips.

A rack-scale system combines processors, accelerators, memory, networking, cooling, and software into one deployable unit. Performance depends on how those components interact under production workloads.

AMD can supply several parts of that system. Its portfolio includes EPYC processors, Instinct accelerators, Pensando networking products, and the ROCm software stack.

That breadth gives AMD an opportunity to optimize across components. It also increases the number of schedules that must align before a complete system reaches customers.

AMD’s Helios platform is expected to pair EPYC processors with MI450-series accelerators. Production deployments are planned to begin during the second half of 2026.

Microsoft has already said it intends to deploy next-generation EPYC and Instinct products under an expanded partnership. Meta has separately agreed to deploy up to six gigawatts of AMD accelerators across several generations.

Those commitments give AMD more visibility than a speculative TAM estimate alone. They do not guarantee the pace, margins, or utilization of future installations.

AMD Versus Intel Is Becoming an AI Infrastructure Contest

AMD and Intel are competing for control of the systems that feed, coordinate, and monetize AI accelerators.

Intel still has the larger installed base across enterprise servers. Its processors are deeply integrated with corporate software, cloud services, and data center operating procedures.

That position creates inertia in Intel’s favor. Replacing a server processor can require application testing, firmware changes, performance validation, and new purchasing agreements.

AMD has spent several EPYC generations reducing those barriers. It has gained adoption by offering higher core density, energy efficiency, and competitive total ownership costs.

The current AI cycle changes the terms of that competition. Buyers are building new clusters instead of simply refreshing familiar enterprise servers.

New construction gives AMD opportunities to win deployments without first displacing an existing Intel system. It also lets customers design around a complete accelerator platform.

Intel approaches the contest from the opposite direction. It can use existing Xeon relationships to retain the host-processor role even when customers select Nvidia or AMD accelerators.

Its second-quarter business highlights included a system combining Intel Xeon processors, SambaNova accelerators, and Nvidia Blackwell GPUs. That design shows Intel positioning Xeon as neutral infrastructure around several accelerator types.

AMD has stronger incentives to promote tightly integrated systems using its own components. Intel benefits when customers preserve flexibility across accelerator vendors.

This makes the competition larger than a benchmark comparison. It involves supply commitments, platform qualification, networking, memory access, software support, and deployment timing.

Both companies also face Arm-based server processors. Amazon, Google, Microsoft, and other cloud operators have developed custom CPUs for selected workloads.

These processors can reduce dependence on merchant suppliers. They also allow cloud companies to tune hardware around their own software and data center designs.

AMD claims Venice can deliver more than twice the throughput per socket of leading Arm-based AI solutions. That statement has not yet received broad, independent production validation.

Arm competition remains strategically important because hyperscalers purchase processors at enormous scale. A successful internal chip can remove a large block of demand from AMD or Intel.

However, custom processors do not eliminate the merchant CPU market. Enterprises still want broad software compatibility, multiple cloud options, and systems available from established server vendors.

The likely outcome is a fragmented infrastructure layer. Hyperscalers will mix custom processors with AMD and Intel products according to workload, availability, and cost.

This environment rewards execution more than architectural purity. A processor with excellent benchmark results creates little value if memory, boards, or complete racks arrive late.

AMD must coordinate its roadmap with TSMC and packaging partners. Intel must raise manufacturing output while improving yields and managing its capital needs.

The rivalry therefore rests on two different promises. AMD argues that its product cadence and integrated portfolio can continue taking share.

Intel argues that better execution can reactivate its scale, customer relationships, and manufacturing assets. Its latest revenue figures make that case more credible than it appeared one year ago.

Neither promise has been fully settled. AMD’s upcoming results must confirm that strong demand is becoming revenue. Intel must show that growth can survive beyond a favorable comparison period.

The Real Bottleneck Is Moving From Accelerators to Memory

More CPU demand does not remove the AI supply problem; it spreads that problem across memory, packaging, networking, and power.

Intel’s Tan described memory as a major supply constraint during the company’s earnings discussion. He said the component had become a bottleneck and a customer pain point.

This constraint includes high-bandwidth memory, or HBM. HBM places stacked memory close to an accelerator, providing the data throughput required by large AI workloads.

Conventional server memory also matters. CPUs require DRAM to run databases, retrieval systems, virtual machines, and orchestration services surrounding AI models.

Higher CPU shipments therefore compete for capacity within an already constrained memory supply chain. Demand can move between product categories, but fabrication and packaging capacity cannot change instantly.

SK Group and Nvidia have responded with a much larger commitment. The companies announced letters of intent for an initiative valued above $500 billion.

The plan covers an AI factory of up to two gigawatts and a long-term memory relationship. The first factory is scheduled to begin operating in 2027.

Under the SK partnership, SK hynix and Nvidia plan to co-develop memory products, including HBM. Nvidia also expects the agreement to secure future supply.

The arrangement expands work announced in June. That earlier memory agreement connected SK hynix products with Nvidia’s Vera Rubin systems, Vera processors, personal AI computers, and robotics platforms.

Letters of intent do not carry the same certainty as completed purchases or delivered infrastructure. The announced total also spans several activities and years.

It should not be interpreted as a single immediate order for memory chips. The figure describes a broad infrastructure initiative involving construction, compute systems, and long-term supply.

Even with that qualification, the partnership shows where leading suppliers see the risk. Nvidia wants memory development and allocation aligned with its system roadmap before demand arrives.

That alignment puts pressure on AMD and Intel. Both need memory partners to prioritize capacity for their products while Nvidia remains the largest customer for AI accelerators.

AMD’s challenge is especially complex because Helios requires processors, accelerators, networking, HBM, packaging, and software to arrive together. One delayed component can restrict the complete rack.

Intel faces a different constraint. It plans to increase equipment and substrate investment while its foundry business still consumes substantial capital.

Intel Foundry generated $5.8 billion in segment revenue during the quarter. Yet it recorded a $2.1 billion operating loss, according to the company’s detailed results.

The foundry loss complicates the optimistic CPU narrative. Intel must fund manufacturing improvements while competing against AMD products made by TSMC.

Its 18A output was approximately 25% above target and more than 50% higher than the previous quarter. Those figures indicate progress, but external foundry revenue remained only $293 million.

Most foundry growth therefore came from Intel’s internal production network. Meaningful external adoption remains a separate test.

The AI hardware cycle can support several suppliers, but it cannot erase production economics. Expanding capacity too slowly loses orders, while expanding too quickly can damage returns.

That tradeoff will become sharper if data center operators delay projects because of power access or financing. Chips cannot generate revenue inside facilities that remain unbuilt.

What the CPU Growth Story Does Not Prove

Strong quarterly demand does not establish that every announced AI data center will be completed or profitably used.

The first uncertainty concerns the demand comparison. Intel’s 25% revenue increase followed a difficult period and benefited from improved factory execution.

A strong annual growth rate can therefore reflect recovery as well as structural expansion. Investors need several quarters of data center growth before separating those effects.

Intel’s GAAP loss also shows that operating improvements do not automatically produce clean shareholder returns. Accounting charges can be temporary, but capital requirements are persistent.

Management plans to invest more in equipment, manufacturing space, and substrates. Those investments assume that product and foundry demand will remain elevated.

The second uncertainty concerns AMD’s market forecast. A server CPU TAM above $120 billion depends on sustained agent adoption and substantial infrastructure construction.

Agents can increase orchestration work, yet software efficiency can offset some hardware demand. Better scheduling, caching, quantization, and model routing can reduce compute used per task.

AI providers also have strong incentives to raise utilization. An accelerator or CPU sitting idle still consumes capital and occupies valuable data center capacity.

The relationship between application usage and chip demand is therefore not linear. More AI users can create more compute demand, but each generation of infrastructure can also process work more efficiently.

The third uncertainty involves customer concentration. A small group of hyperscalers accounts for much of the industry’s planned AI spending.

Large commitments from Microsoft, Meta, and other operators give suppliers visibility. They also leave suppliers exposed to changes in a few capital budgets.

A delayed campus or revised model strategy can affect several hardware vendors simultaneously. Long supply agreements reduce uncertainty, but they do not remove construction and utilization risk.

The fourth uncertainty is competitive response. Nvidia continues to expand beyond standalone GPUs through rack systems, networking, software, and its own Arm-based processors.

Cloud operators are developing custom accelerators and CPUs. These products can redirect workloads that would otherwise run on AMD or Intel hardware.

AMD’s software progress remains another important variable. ROCm is the company’s open software environment for programming and operating its accelerators.

AMD said ROCm downloads increased tenfold year over year at its November 2025 analyst event. Download counts do not measure production usage, developer satisfaction, or workload reliability.

Nvidia’s CUDA environment retains a deep base of libraries, tools, and trained developers. AMD needs software adoption to grow alongside hardware deliveries.

Intel also needs a clearer position in accelerated computing. Strong Xeon demand helps, but customers increasingly purchase architectures assembled around accelerators and networking.

The company can remain a valuable host-processor supplier without leading the accelerator market. That role may still limit its influence over complete system design and margins.

These risks do not invalidate the CPU recovery. They define what must happen for it to become a durable AI infrastructure cycle.

The strongest interpretation is not that every chip company wins. It is that AI spending now reaches more components, creating more opportunities and more points of failure.

Three Signals Will Decide Whether the CPU Cycle Lasts

AMD’s next earnings report, Intel’s manufacturing economics, and memory delivery schedules will test the new CPU thesis in that order.

The first signal arrives on August 4, when AMD reports second-quarter results. Management previously forecast server CPU revenue growth above 70% year over year.

Meeting that expectation would support AMD’s claim that AI is increasing demand for host processors. It would also validate customer discussions behind the company’s larger market forecast.

Investors should focus on data center revenue, EPYC growth, supply availability, and the Venice schedule. Commentary about 2027 capacity commitments will matter as much as one quarter’s sales.

A weaker result would not eliminate long-term demand. It would suggest that customer interest, qualification, and delivered revenue are moving at different speeds.

The second signal is Intel’s foundry and capital-spending trajectory. Its internal production improved during the second quarter, but external foundry revenue remains limited.

Watch 18A yields, output, external customer commitments, and foundry operating losses. Progress across all four would strengthen Intel’s claim that manufacturing scale can support its CPU recovery.

Higher spending without lower foundry losses would weaken that case. Intel cannot rely indefinitely on product demand to absorb manufacturing costs.

The third signal is the delivery schedule for next-generation memory. SK hynix and Nvidia have outlined extensive co-development and supply plans, with larger infrastructure expected in 2027.

Watch HBM availability, memory qualification, and whether competing platforms receive enough supply for scheduled deployments. Memory shortages would limit AMD and Intel even when CPU demand remains strong.

Those three signals connect the entire hardware chain. Orders must become processors, processors must enter complete racks, and racks must reach powered data centers.

For developers, the outcome affects which platforms receive optimization work and cloud availability. Hardware competition can influence model costs, deployment choices, and the portability of AI software.

Enterprise buyers should watch the same signals before committing to long-lived infrastructure. A compelling processor roadmap matters only when memory, networking, support, and deployment capacity arrive on schedule.

Knowledge workers will encounter the effects indirectly. More available inference capacity can support faster agents, larger business workflows, and wider use of private organizational data.

Teams evaluating those systems should preserve the decisions, benchmarks, and source material behind each infrastructure choice. A searchable engineering knowledge base can keep that evidence connected as products and schedules change.

The practical question is no longer whether GPUs remain central to AI. They do. The question is how much additional CPU, memory, and networking each accelerator deployment pulls into the market.

Follow AMD’s August results first, then Intel’s 18A economics, and finally the memory schedules behind 2027 racks. If all three advance together, the CPU revival has structural support. If shipments separate from announcements, the current optimism is running ahead of deployment. Either outcome will shape the cost and availability of the AI services businesses plan to use next.