NVIDIA Vera CPUs in Europe: Why Agent Sandboxes Are Moving to Them (Early-Testing Claims)

NVIDIA published a roster of European and global companies on October 20, 2026 that it says are building on its Vera CPU, and the list is more about AI agents than about models. Cloud providers Nebius, Nscale and Verda, server makers from Dell to Supermicro, data-processing vendors, and a long list of agent sandbox providers including E2B, Modal, Daytona, Vercel and Prime Intellect all appear in NVIDIA's European Vera announcement.

The pitch is simple: agents are CPU-heavy. This post explains why that is true, who is named, and what is a vendor claim versus independent evidence. It builds on our earlier coverage of the Chips and Cheese measurements of Vera and the critique of NVIDIA's Vera whitepaper.

TL;DR

Why agents make CPUs matter again

For two years the AI hardware story has been GPUs. NVIDIA's post argues the next bottleneck is the work around the model. An agent that books a flight, fixes a bug or analyzes a spreadsheet loops through the same pattern: the model proposes an action, a tool runs it, the result is parsed, and the model decides what to do next. Only the first step is GPU work.

NVIDIA's description: "Agents do more than generate answers: they retrieve information, run software tools, execute code and evaluate results. Each step depends on fast CPU processing." If that is right, a data center with perfect GPUs and slow CPUs leaves expensive accelerators waiting for tool results.

The effect is sharpest for sandboxes. Each agent needs an isolated environment, often a lightweight virtual machine or container, to run code safely. NVIDIA says sandboxes "spin up millions of times a day," so the startup time of each one compounds. We covered the safety side of this layer in Anthropic's red-team of coding agent monitors: the sandbox is the control that holds when a classifier fails.

Who NVIDIA names

Two entries deserve a closer look. Nous Research uses Vera, per NVIDIA, "to authenticate agent identities and validate every action before it executes", which is a security workload rather than a throughput one. And Inherent is described as having tested Vera and "looking to adopt" it, a softer statement than deployment. The wording across the list ranges from "building with" to "looking to," so read it as an ecosystem signal, not a customer count.

What NVIDIA actually claims about performance

The post credits three hardware features: Vera's per-core performance, LPDDR5X memory bandwidth and the NVIDIA Scalable Coherency Fabric. It says data-processing companies "observed strong results" for analytical, ETL and semantic search workloads in early testing, and that Vera is designed to sustain speed under thousands of simultaneous tasks.

What it does not do is publish any figures: no speedups, no cores or clock speeds, no power, no price, no comparison to x86 or Arm rivals. That is why we treat the claims as unverified. For measured data, the independent Chips and Cheese analysis in our earlier write-up found high memory bandwidth but also notable cross-socket latency, which matters for workloads that span sockets. Sandbox startup is mostly single-socket work, which is plausibly where Vera looks best, but that is our inference, not a measured result.

What happens inside an agent sandbox

It helps to picture one agent task end to end, because the CPU cost hides in places people forget.

1. Provision. A request arrives and the provider creates an isolated environment: allocate memory, attach a filesystem snapshot, start a micro-VM or container, boot a runtime. This is pure CPU and memory-bandwidth work, repeated for every task.
2. Load. Dependencies, repositories and credentials scoped to the task are mounted. Package resolution and file indexing dominate here.
3. Execute. The agent runs shell commands, compiles code, runs tests, or drives a headless browser. Compile and test steps are heavy, parallel CPU jobs.
4. Observe. Output is captured, truncated, parsed and returned to the model. Logging and serialization add overhead at scale.
5. Tear down. The sandbox is destroyed or snapshotted so the next task starts clean, which matters for security.

The model call between steps may take a second or two on a GPU. A cold sandbox start that takes a second longer, multiplied across millions of tasks a day, can cost more in aggregate than the savings from a faster model. That is the economic logic behind NVIDIA's list: if you sell sandboxes, shaving startup and execution time is your margin.

The data-processing half of the story

The second large group in the announcement is not agents at all. ClickHouse, Cloudera, Elastic, Polars, Redpanda, SAP and Starburst are named for analytical queries, extract-transform-load jobs and semantic search. Atlassian specifically plans to use Vera for the data processing that builds semantic search indexes for its AI products.

This is the quieter point of the announcement: retrieval-augmented generation and agent memory depend on indexes that are built and refreshed on CPUs. When a company says it is "AI-first," a large share of its AI bill may be data pipelines rather than model inference. NVIDIA is signaling that it wants that spend too.

How to judge a CPU claim for agents

Because the post publishes no numbers, a buyer needs their own test plan. A reasonable one has four parts.

Run each on your current instances and on any Vera instance a provider offers, using the same container image and workload, and compare cost per completed task rather than raw speed. Include power and cooling only if you run your own hardware.

Why this matters for builders

1. Agent cost is not only tokens. If you run agents at volume, sandbox compute is a line item. Faster, denser CPUs reduce it, which is why sandbox vendors are on the list.
2. Sovereignty is a selling point. The post frames Vera in terms of sovereign AI in Europe: infrastructure that keeps data and operations under customers' control. Expect European providers to advertise Vera capacity next to GPU capacity.
3. Competition for the CPU slot. NVIDIA entering general-purpose server CPUs pressures x86 and Arm incumbents. Whether it wins depends on price and software compatibility, neither disclosed here.
4. Security shifts to the host. More agents on shared hosts raise the stakes of isolation. Agent security tools such as AgentBeam, from the explainx.ai team, focus on stopping agent actions before they are dangerous, a complement to hardware isolation.

What to do now

• Measure your own sandbox economics. Track cold-start time, p95 latency and cost per sandbox-hour for your workload on current instances.
• Ask your provider about Vera. E2B, Modal, Daytona, Vercel and the clouds above are named; ask for availability, regions and pricing before you plan around it.
• Do not rewrite for Vera, but test. Our earlier coverage discusses Vera's custom core design, which is not x86, so check that your container images, native dependencies and language runtimes have builds for it before you plan a migration.
• Watch independent benchmarks. Wait for third-party sandbox and database benchmarks before committing capacity.

Open questions

• What are the per-core and per-dollar gains versus x86 on sandbox startup?
• When do Nebius, Nscale and Verda offer Vera instances, and at what price?
• How do the early-testing results hold on real, noisy multi-tenant workloads?
• Does the cross-socket latency measured by independent testers affect sandbox density?

Related reading

• Chips and Cheese measures NVIDIA Vera
• NVIDIA Vera whitepaper critique
• GTC Berlin 2026 keynote recap
• NVIDIA confidential bare-metal for European AI clouds
• Google Cloud A5X Vera Rubin availability
• Anthropic red-teaming Auto Mode
• Official: NVIDIA: European leaders turn to Vera CPUs · NVIDIA blog

Based on NVIDIA's October 20, 2026 post; adopter statements are vendor claims and performance figures were not published.