Google Cloud's Lohmeyer says agentic AI will kick off a CPU comeback. That's the claim, straight from a named executive at one of the three largest cloud providers on the planet. It lands in a market that has spent two years treating the CPU as collateral damage in the GPU gold rush.

The argument, as Fierce Network reports it, is simple. Agentic AI — systems that plan, call tools, and act rather than just generate text — doesn't look like the training workloads that made NVIDIA's data center GPUs the hottest silicon on earth. Agents run lean inference steps, orchestrate calls, wait on APIs, and fan out across millions of small tasks. That's a workload profile with CPU written all over it.

Why would agents favor CPUs over GPUs?

The GPU boom was built on training. Large models chew through trillions of tokens in massive parallel batches, and that math maps neatly onto tensor cores. Inference changed the picture somewhat. Agentic AI changes it further.

An agent spends much of its life doing anything but matrix multiplication. It parses context, routes requests, hits databases, executes logic, and sleeps while a downstream service responds. Bursty, branchy, latency-sensitive, lightly parallel — the opposite of a training run. General-purpose cores handle that shape well, and cloud providers already own enormous fleets of them.

If Lohmeyer is right, the commercial implications run straight through the silicon supply chain.

Who gains if the CPU thesis holds?

A CPU renaissance would touch the usual suspects — Intel, AMD, Arm and its licensees, and the cloud vendors' own in-house arm designs, Google's Axion among them. Every hyperscaler has spent the past decade designing its own server chips to escape x86 pricing. A wave of agentic workloads that runs better on those cores would validate that strategy in revenue, not slideware.

It would also relieve pressure on the most constrained resource in the industry: advanced packaging and HBM capacity tied up in accelerator production. GPUs don't just cost money. They crowd out everything else at the fab and the substrate supplier.

What's confirmed and what's a bet?

Here the reader deserves precision, and the source gives little. Lohmeyer's statement is an executive's forward-looking assertion, not a shipment figure, not a capacity number, not a benchmark. Fierce Network's headline carries the claim; it does not attach datapoints. Treat it as a roadmap-level signal from a buyer with enormous purchasing power, not as a measured trend.

That distinction matters because vendors talk up CPU-centric AI regularly, and the ledger still tells a GPU story. NVIDIA's data center revenue has dwarfed server CPU revenue growth for several quarters running. For the comeback thesis to count, CPU shipments into AI-specific deployments would need to show up in reported numbers from Intel, AMD or the ODMs — none of which the source provides.

The context that changes the picture

Still, this isn't a random pundit. When a Google Cloud executive frames agentic AI as a CPU driver, it's a signal about where one of the largest AI infrastructure builders expects its customers' workloads to land. Google designs its own chips, operates its own network, and sees demand patterns before they reach the financial reports of chipmakers. Executives at that altitude rarely speculate idly in trade press.

The timing helps the argument too. Enterprises are moving agents from demos into production, and the cost math of running millions of always-on agent loops on premium accelerators looks ugly. CPU capacity, by contrast, is available, cheaper, and already amortized inside existing server fleets.

What to watch next

The proof will arrive in the usual places: server CPU shipment and average-selling-price data from Intel and AMD, Google's own Axion deployment disclosures, and any vendor earnings commentary that names agentic workloads as a demand driver. Until then, Lohmeyer's call stands as the most prominent executive endorsement yet of the idea that the next phase of AI belongs, at least partly, to the humble general-purpose core.