Intels Inference Gambit: How AI Agents Are Quietly Reshaping the Chip Business

Intel reported first-quarter earnings Thursday that rewrote the narrative around a company the market had left for dead. Data Center and AI revenue climbed 22% year-over-year to $5.1 billion — but the number is a symptom, not the story. The story is a structural shift in what AI workloads demand from silicon, and whose architecture is positioned to supply it.

The received wisdom of the AI chip era has been simple: Nvidia wins at training, everyone else fights over scraps. GPUs process billions of parameters in parallel, making them the natural engine for building foundational models. Intel's Gaudi accelerators were supposed to compete in that market. They did not.

But AI is changing shape. The chatbots that launched the boom were one-shot inference tasks: a user asks something, the model generates a response, the interaction ends. Agentic AI works differently. Agents loop. They observe, decide, act, and loop again — often hundreds of times per task, often running continuously across an enterprise. That persistent, iterative workload has different compute requirements than a training sprint.

The next wave of AI will bring intelligence closer to the end user, moving from foundational models to inference to agentic, Intel CEO Lip-Bu Tan said on the earnings call Thursday (CNBC).

What that means in practice: CPUs — Intel's historic franchise — are better suited than GPUs for the long-running, latency-sensitive, decision-tree-chasing workloads that define the agentic era. Not for training. Not for the massive batch inference of a model answering a single query. For the continuous, interactive, often boring work of an agent that runs in the background of an enterprise workflow, checking, updating, routing, and flagging.

Two named customers are now building on that assumption. Google committed in March to using multiple generations of Intel's Xeon 6 processors to run AI workloads in its data centers (CNBC). And this month, Intel and SambaNova announced a joint solution pairing Xeon 6 as the host and action CPU with SambaNova's RDU accelerators for high-throughput decode — specifically engineered, according to their joint announcement, for coding agents and other multi-step agentic systems expected in enterprise and sovereign AI deployments in the second half of 2026 (Intel Newsroom).

The architecture logic is technical but the implication is not: as AI workloads fragment into training, batch inference, and persistent agentic loops, the chip that wins each category differs. GPUs won training. Custom ASICs are competing for batch inference. CPUs — Intel's existing franchise — are positioned to win the persistent, latency-sensitive agentic loop. That positioning is what Intel's 22% DCAI growth actually reflects, if you look at the shape of demand rather than just the headline number.

The broader context: Lenovo and Intel published benchmark data this year showing Intel Xeon 6 with Advanced Matrix Extensions running 7-8 billion parameter models for agentic workloads entirely on CPU — production-ready inference without GPU infrastructure (Lenovo Press). Dell and others are selling Intel Gaudi 3 accelerators for inference workloads, and Intel's own Crescent Island data center GPU for inference is in customer sampling. The company is trying to be a full-stack AI silicon vendor, having failed at the one market everyone expected it to win.

Intel is not a simple comeback story. Net loss widened to $4.28 billion in Q1 from $887 million a year earlier (CNBC). The company is still burning cash even as revenue ticks up. The 22% DCAI growth needs to be sustained across multiple quarters before it becomes evidence of a structural shift rather than a data center refresh cycle — the kind of one-time server replacement that inflates one quarter's numbers and then disappears.

Intel's 18A manufacturing node, which powers its newest chips, has had yield problems — defects on wafers reducing usable chips per wafer (CNBC). The 14A node planned for 2028 is what anchor customer commitments like the Terafab partnership with Musk's SpaceX, xAI, and Tesla are supposed to justify. If 14A has similar ramp challenges, the manufacturing dependency that has always been Intel's differentiator becomes a liability.

And then there is the Gaudi question. Intel's GPU efforts have been troubled: Reuters reported in late 2024 that Gaudi uptake was slower than expected, with software issues cited (Reuters). If the inference market develops primarily on CPUs rather than custom AI accelerators, Intel wins by default on the products it already makes. If the market decides it wants purpose-built inference chips, Intel needs Gaudi to work — and it has not yet.

The critical open question is what fraction of Intel's 22% DCAI growth is Xeon CPUs doing agent work versus general server refresh. Intel does not break out revenue by workload type, and Google and SambaNova commitments, while real, are not yet in production — both are expected to materialize in the second half of 2026. The architectural positioning is genuine. The revenue confirmation requires another quarter or two of the same number.

Intel reported Q1 2026 revenue of $13.6 billion, up 7% year-over-year, with non-GAAP EPS of 29 cents (StockTitan). The company guided for Q2 revenue of $13.8 billion to $14.8 billion (StockTitan). Intel stock rose 16% in after-hours trading Thursday (CNBC).