Q1 2026 Earnings Season Check-In: Data Center CPUs and the $700B Capex Behind Them

An interesting theme emerged from earnings season regarding the AI buildout: the composition has shifted toward CPUs as we move into an agentic inference world. In this article, we explore three data center CPU names (INTC, AMD, ARM) that reported strong Q1 2026 results and even more bullish outlooks, framed against the four hyperscalers’ increased capex and guidance.

A quick note on the workflow. Daloopa’s MCP is the data infrastructure layer Claude uses to query Daloopa’s normalized financials. When asking for a segment number, the server returns the value and the source. Each number is hyperlinked to its filing, so you can audit any line in one click.

The setup: As AI moves from training to inference, CPUs increase in importance

To make sense of why the three data center CPU names all printed strongly this quarter, start with what is changing in the underlying workload mix.

The first wave of AI capex was overwhelmingly about training. Training is a brute-force matrix-multiplication problem: feed a massive static dataset through a model and adjust the weights. The compute is GPU-dominated by structure, because parallel floating-point math is what GPUs do. CPUs in a training cluster handle data loading, checkpointing, and orchestration.

Inference at scale is different, and agentic inference at scale is different in kind. Every user query or agent action is an asynchronous workflow with stages: receive and parse the prompt, manage authentication and security, retrieve context from vector databases or memory, route to the right model, orchestrate inference on the accelerator, post-process and format the response, then handle agent state, tool calls, and follow-on steps. Most of those stages are CPU work. The accelerator does the matrix math. The CPU does everything to get the matrix math the right inputs and turn the outputs back into something useful.

ARM management quantified it this quarter: agentic AI will require more than 4x today’s CPU capacity per gigawatt of data center buildout, with roughly 30 million CPU cores per gigawatt under the new workload mix. AMD said basically the same thing, noting that inference and agentic AI workloads require additional CPU processing for orchestration, head nodes, and pre- and post-processing alongside accelerators. Intel disclosed that Xeon was selected as the host CPU for NVIDIA’s DGX Rubin NVL8 system, a GPU-centric AI rack that needs a heavy CPU layer to function. Microsoft’s CFO Amy Hood summed it up most cleanly on its print, noting that most of Q1 2026 capex went to short-lived assets, primarily GPUs and CPUs, not GPUs alone.

The CPU is not displacing the accelerator. It is scaling alongside it as the workload shifts from one-shot training runs to continuous, agentic inference. That is why three different CPU companies, each with its own competitive positioning, posted very strong quarters.

AMD: Data center segment passes Intel, server CPU TAM doubled

AMD’s data center segment grew 57% year over year to a record $5,775M, up from $3,674M in Q1 2025. Server CPU revenue grew more than 50% year over year for a fourth consecutive record quarter, and Q2 guidance calls for server CPU revenue to grow more than 70% year over year. EPYC-powered cloud instances are now over 1,300, up nearly 50% year over year.

The bigger signal came in the TAM revision. At the November Analyst Day, AMD framed server CPU as an 18% CAGR market reaching $60 billion by 2030. On this call, they roughly doubled both numbers: greater than 35% annual growth and a $120 billion TAM by 2030, driven by agentic AI workloads that need CPU compute for orchestration, head nodes, and pre- and post-processing alongside accelerators. Zen 6 Venice on 2nm ramps in the second half of 2026, with Meta as lead customer alongside a 6 GW Instinct GPU partnership.

Operating margin compressed slightly from 32.6% in Q4 2025 to 27.7% in Q1 2026 even as revenue grew sequentially. Management framed this as product mix tied to the early Instinct ramp. The margin trajectory in H2 will depend on how Helios deploys alongside it.

Intel: DCAI operating income up 168% year over year, but read the share footnote

Intel’s data center and AI segment posted the most significant margin swing in the group. DCAI operating income went from $575M in Q1 2025 to $1,542M, up 168% year over year. Operating margin expanded from roughly 14% to 30.5%, which is actually above AMD’s data center op margin this quarter. Revenue grew 22% year over year to $5,052M.

Management framed demand as running ahead of supply for Xeon server CPUs, with Xeon 6 selected as the host CPU for NVIDIA’s DGX Rubin NVL8 systems. The strategic backdrop matters here: NVIDIA bought a $5.0 billion equity stake in Q4 2025, and the two companies are co-developing custom client and data center products combining x86 with NVIDIA’s accelerated computing.

The footnote to all of this is that Intel’s own 10-K still acknowledges market share losses in both client and data center markets in recent years. AMD’s data center segment revenue at $5,775M is now roughly 14% larger than Intel’s DCAI at $5,052M. The margin story is real, but the share story is also real, and they coexist.

ARM: First in-house silicon launched, data center royalties doubled

ARM’s headline number was a 20% year over year increase in total revenue to $1,490M. Royalty revenue grew 11% to $671M and license and other revenue grew 29% to $819M. Inside the royalty number, management said data center royalty revenue more than doubled year over year, with cloud AI now the fastest growing royalty driver.

The strategic news was bigger than the numbers. ARM launched the Arm AGI CPU, the company’s first in-house designed data center silicon, with Meta as lead customer and Cloudflare, OpenAI, SAP, SK Telecom, and F5 also committed. ARM is now selling chips directly, not just licensing IP. Management’s framing is that data center will become ARM’s largest business, with the TAM growing from about $35 billion today to over $100 billion by FYE31.

There is a structural dynamic to note: by shipping its own silicon, ARM now competes with its licensees that build their own Arm CPUs, including AWS Graviton, Google Axion, Microsoft Cobalt, and NVIDIA Grace. Whether royalty growth from those programs holds as ARM goes direct is a variable to monitor.

The demand pull: Hyperscaler capex up 71% YoY, with $700B+ implied for 2026

Three data center CPU names printing strong results at once is not a coincidence. It is the supply-side response to the four customer prints below. Every major hyperscaler is meaningfully increasing capex, and all four explicitly tie the increase to AI infrastructure.

A few things stand out. First, the combined Q1 2026 capex from the Big Four came to roughly $131 billion in a single quarter, up 71% year over year. Annualized at that run rate, the Big Four alone would spend over $500 billion this year, and forward commitments and CEO letter guidance point to a combined total closer to $700 billion for calendar 2026.

Second, the absolute leader changed. Amazon’s $44.2B Q1 2026 capex is now the highest single-quarter print among the Big Four, ahead of Google’s $35.7B. For most of the last decade, the order ran in the opposite direction.

Third, the language matters. Microsoft’s CFO described the company as supply-constrained at least through 2026 even at $190B of calendar-year capex. Amazon’s CEO letter put it bluntly: “we’re not investing approximately $200 billion in capex in 2026 on a hunch,” pointing to the OpenAI commitment (over $100 billion) as the visible piece of a much larger committed pipeline. Google’s 10-Q says it expects to “significantly increase” capex relative to 2025 levels, on top of a Q4 2025 guidance range already set at $175–$185 billion. Meta raised its FY 2026 guide by $10 billion at both ends just this quarter.

What ties them together

The Big Four hyperscalers are committing the capex. AMD, Intel, and ARM are all reporting revenue growth tied directly to the hyperscaler capex in the table above, not just NVIDIA. The TAM revisions from AMD and ARM are sized to commitments like those in the table above, and the shape of demand is shifting from training-heavy buildouts (GPU-dominant) toward inference and agentic workloads (which require CPU and GPU together).

The variance across the three semis is also informative. AMD is taking data center share. Intel is recovering profitability on a shrinking share base while leaning on a new partnership with NVIDIA. ARM is repositioning its entire business model to sell silicon directly into hyperscaler racks. Three different responses to the same demand wave, each visible in the data above and traceable back to the source in one click.

A note on the data and the workflow

Every figure in this post was retrieved through Daloopa’s MCP, which connects the LLM directly to Daloopa’s comprehensive financial database. Each number you can click takes you to daloopa.com/src/ followed by an identifier that points back to the original filing line. The MCP returns the value and the source link together in the same call, so there is no separate reconciliation step.

During earnings season, the bottleneck is not finding the numbers. It is keeping the right number from the right period attached to the right line in your model when you are moving across seven prints in a sitting. Daloopa’s MCP solves that—pulling from a normalized, structured financial database for more accurate AI reasoning and linking every datapoint back to its original filing. Everything above was pulled and written in a single working session, with fully auditable outputs available in just one click.