Nvidia's Vera CPU doubles throughput in Modal's sandbox test, but the test is Modal's own and the chip is one node deep
Modal says 'up to 2x' the task throughput of an equivalent Intel Granite Rapids box, and Vera kept scaling where x86 plateaued. The catch: ~50 tasks, one Vera server, early runs.
preliminary evaluations showed that Vera could deliver twice the throughput of its existing AMD and Intel processors, although the company does not intend to cease using CPUs from its current suppliers.
Comprehensive take
Vera's pitch is not peak speed, it is not slowing down when a host is crowded, and Modal's chart is the first public evidence of that. It is one customer, one node and a self-built test, so read it as a reason to put Vera on your evaluation list, not as a verdict on x86.
Use cases
Agent sandboxes and RL rollouts
Bursty, high-concurrency CPU work where per-task latency has to hold under contention is exactly what Modal tested.
Mixed Arm/x86 fleets
Modal is keeping its x86 CPUs and choosing per workload, a template for operators who can't migrate everything at once.
CPU buyers benchmarking agent workloads
Use Modal's method (fixed-window completions at rising concurrency) on your own jobs rather than relying on single-thread scores.
Specs
| Core architecture | Arm v9.2 'Olympus' cores, up to 10 instructions decoded/executed per cycle source |
|---|---|
| Memory bandwidth | 1.2 TB/s source |
| Modal throughput claim | Up to 2x task throughput vs an equivalent x86 Granite Rapids machine source |
| Test scope | ~50 Terminal-Bench-derived tasks, 60-second completion window, one Vera node (up to 176 logical CPUs) source |
| Serial latency | Vera often faster, 'not by a particularly large margin' source |
Insights
- Modal's post claims the doubling against Intel Granite Rapids specifically; the 'Intel and AMD' framing comes from the press summary.
- The gain shows up under contention. In serial, unloaded runs the lead is small.
- Modal remains an x86 shop for now: Arm support requires infrastructure and image-build changes.
Axion Scorecard
Overall: 6.5
Not yet scored — no cited data
Not yet scored — no cited data
Olympus cores and 1.2 TB/s of memory bandwidth in a standalone server CPU, per Tom's Hardware. source
Modal says Arm support needs infrastructure and image-build changes; it keeps x86 in the fleet. source
Not yet scored — no cited data
Not yet scored — no cited data
Not yet scored — no cited data
Nvidia has spent a decade selling the accelerator that sits next to the CPU. Now it wants to sell the CPU too, and the first public customer story is more interesting than the headline number. Modal, a serverless AI cloud, published early benchmark results on October 19, 2026 after Nvidia hand-delivered its first Vera server to a San Francisco office the week before. Tom's Hardware, citing Reuters, reports that Modal will deploy Vera alongside its existing x86 fleet rather than replace it.
What Modal actually measured
The workload matters more than the chip. Modal Sandboxes run hundreds of concurrent containers and VMs per host, and their CPU use is spiky: an npm install or a grep can push a sandbox to roughly 10x its baseline. Modal sells bursting, so it provisions for typical use rather than worst case. That makes sustained throughput under contention the metric that decides margins.
Modal built an internal suite called Agentless Agent Bench from about 50 Terminal-Bench oracle solutions, with network calls stripped out and tasks prewarmed. For throughput it raised the number of concurrent containers and counted tasks finished in a fixed 60-second window. Machines tested were Vera (up to 176 logical CPUs), Intel Granite Rapids bare metal (up to 256) and as a VM, AWS Graviton 4 and 5, and AMD Zen 5 (up to 192).
The headline claim in Modal's own post is narrower than the press version: up to double the task throughput versus an equivalent x86 Granite Rapids machine. Tom's Hardware phrases it as twice the throughput of Modal's existing AMD and Intel processors. Read the primary source, not the summary. In serial latency on an unloaded host, Modal says Vera often won but not by a large margin. The advantage appears under load: the x86 and Graviton systems hit throughput plateaus as core counts rose, while Vera kept improving and its per-task latency held.
Why it might be real
Tom's Hardware lists Vera as an Arm v9.2 design with Olympus cores that can decode and execute up to 10 instructions per cycle, with 1.2 TB/s of memory bandwidth. Modal attributes the gain to that bandwidth letting it pack more work per core without throttling. That is a plausible mechanism for build-and-grep workloads that stream through memory, but Modal published no bandwidth or throttling measurements, so treat it as a hypothesis.
Why to stay skeptical
The benchmark is vendor-adjacent in the wrong direction: it was designed by the customer for its own workload and run on the one Vera node Modal could get. Modal calls these early runs. Tom's Hardware notes that Modal has not disclosed the specific AMD and Intel models, configurations or performance per watt. Nothing was compared against AMD's or Intel's upcoming generations. And Modal says supporting Arm needs infrastructure and image-build changes, so the post says it is excited to roll out Vera, not that it has.
The business context
Tom's Hardware reports Nvidia targeting at least $20 billion in Vera sales this fiscal year against a roughly $200 billion addressable market. Modal, which raised $750 million at a $15.75 billion valuation per the same report, is a credible design win for agent-heavy workloads, which is where CPU demand is growing.