NVIDIA Won an AI Customer With a CPU, Not a GPU. The Real Difference Appeared Under Load.

A GPU company wins a customer for its CPU

NVIDIA has won another AI infrastructure customer, but the product at the center of the story is a CPU rather than a GPU. On October 19, 2026, cloud infrastructure startup Modal described plans to add NVIDIA's Arm-based Vera CPU to its Sandboxes for AI agents. The company had previously operated this execution workload on Intel and AMD x86 processors.

NVIDIA delivered a first Vera server to Modal's San Francisco office for evaluation. Modal's engineering team said adopting Arm requires changes across its infrastructure and container-image build process; it has not claimed that the entire production fleet has already migrated.

The headline result is 'up to 2x'. What changed was not how fast a single AI model generates an answer, but how many completed execution tasks a crowded host can process.

One task was not dramatically faster. Crowded hosts were different

Modal created Agentless Agent Bench from 50 Terminal-Bench 2.1 oracle solutions and verifiers, removing network calls and using prewarmed containers. Workloads included compilation, file grepping, git commands and similar CPU-side tasks an agent might execute. LLM thinking and remote network wait time were intentionally excluded.

On a quiet host, with tasks run in sequence, Vera often performed better than the alternative processors—but not always, and usually without a large gap. Individual task latency was not the main surprise.

Under concurrent load, Modal's early tests found Vera completed up to twice as many tasks per unit time as a comparably configured Intel Granite Rapids x86 GNR host. This is task throughput under Modal's chosen benchmark conditions, not proof that every agent runs twice as fast end to end.

A Sandbox is usually idle until it suddenly isn't

An agent does not only ask a GPU-backed language model to reason. It reads and modifies files, installs packages, runs tests and sends results back to the model. Modal isolates these actions in Sandboxes and may schedule hundreds of execution environments on one machine.

According to Modal, CPU activity can suddenly spike roughly tenfold above baseline when a container leaves an idle period for npm install or an intensive grep. Unknown task shapes mean the scheduler cannot predict which tenant will demand more CPU at a given moment.

Allocating enough dedicated capacity for every peak wastes resources. Packing environments according to typical use improves utilization but raises the risk of noisy neighbors. Modal's bursting approach makes the ability to maintain throughput under contention commercially relevant.

What Vera's technical specifications can and cannot explain

NVIDIA describes Vera as an 88-core Olympus processor with up to 1.2TB/s of memory bandwidth and a standalone CPU-server option. Modal's engineering team attributes much of the concurrent-workload advantage to its ability to keep progressing without the early throughput plateau it saw on alternative machines.

That is an interpretation grounded in vendor specifications and an early, company-run benchmark. It does not establish identical gains in a fully deployed, independently audited fleet across all workload mixes, energy prices or total-cost scenarios.

The CPU can keep an expensive GPU waiting

Modal also points to reinforcement learning rollouts. In synchronous training workflows, a GPU may be ready for the next training step while waiting for CPU-side sandbox execution to finish. Reducing that delay could improve utilization of costly GPU capacity.

But the benchmark does not quantify any actual fleet-wide GPU cost reduction. That requires end-to-end measurements of rollout synchronization, accelerator idle time, hardware costs and the architecture of each training job.

Modal is not replacing Intel and AMD

Reuters reported that CEO Erik Bernhardsson intends to retain Intel and AMD while adding Vera. The plan is a heterogeneous fleet where jobs can be routed to whichever hardware performs best for the relevant workload.

This matters because the trade-offs differ for serial latency, throughput and CPU-architecture migration. Arm support brings software and image-building work, and future Intel and AMD designs may narrow or reverse individual benchmark gaps. The initial tests are not a declaration of one permanent CPU winner.

Why the economics now matter

On October 14, Modal announced a $750 million Series D at a $15.75 billion post-money valuation. It said its annualized run rate had exceeded $1 billion and the Sandboxes product alone had reached a roughly $300 million annualized run rate.

Those are company-reported run-rate metrics, not audited revenue for a completed fiscal year. They do show why capacity density and predictable performance could become material decisions for an AI execution provider rather than minor engineering optimizations.