Nvidia Vera Doubled Throughput — Modal's First Step Off x86
By TechPulse USA Newsroom | October 20, 2026
Nvidia is not content selling you the GPU anymore. It wants the CPU under it, the memory around it, and — if this week is any indication — the loyalty of the clouds that run on top of it. On Monday, Modal, the $15.75 billion AI infrastructure startup, revealed that it measured up to double the task throughput from Nvidia's new Vera CPU against a comparable Intel machine in its own sandbox benchmark, and is now planning to do something it has never done before: put an Arm chip into its fleet.
Modal's entire history has been x86 — Intel and AMD, nothing else. Adding a third architecture means retooling its infrastructure and image-build system, the kind of thankless engineering that nobody volunteers for. Which makes the decision itself more interesting than the benchmark. A company with a $300-million-a-year sandbox business just told the world it believes an Nvidia CPU is worth the pain.
A benchmark built for chaos
To understand why Modal cares, you have to understand what its Sandboxes product actually does. Sandboxes are isolated compute environments where AI agents run code they write themselves — compiling, grepping through files, cloning repos, emulating devices. Hundreds of these land on each host machine, and their behavior is gloriously unpredictable. An agent can sit idle for minutes waiting on a language model, then suddenly spike CPU usage tenfold because of an npm install or an expensive search command. Modal calls this the "noisy neighbor" problem, and it is the defining headache of the sandbox business: one customer's worst-case moment degrades everyone else's experience.
Rather than force users to reserve capacity for their peak usage, Modal lets sandboxes burst — you pay for your typical load and trust the padding inside the host machine to absorb the spikes. The company says the approach has saved customers millions of dollars. The trade-off is that it leaves the fleet exposed to exactly the kind of contention that makes benchmarks boring and real life miserable.
So Modal built a benchmark that tries to be miserable on purpose. It published the design in its Monday blog post: a collection of 50 Terminal-Bench "oracle" solutions and their verifiers, with all network calls stripped out, prewarmed inside concurrent Docker containers. The project, called Agentless Agent Bench, is public on GitHub — it assembles a deliberately heterogeneous mess of the things agents and reinforcement-learning rollouts actually do on a CPU, then measures both straight-line per-task latency and how many verified task completions the machine can loop through in a fixed time under load.
The quiet-machine surprise
Here is the part of the story that deserves more attention than it has gotten. Run serially on an unloaded machine — measuring pure, quiet per-task speed — Vera was not particularly special. It beat the tested Intel GNR (Granite Rapids), AMD Zen 5, and AWS Graviton 4 and 5 chips on many of the 50 tasks, the company wrote, "but not always, and not by a particularly large margin." The post's author admitted he was "admittedly skeptical" when he saw those numbers first.
Then they added load, and the story flipped. "Counterintuitively," the company wrote, "Vera gets better under heavy load." In the throughput test — ramping up concurrent containers and counting verified task completions — early runs showed up to double the task throughput of an equivalent x86 GNR machine, while individual task latency measured equal to, or sometimes even lower than, the x86 box. "During one of our onsites at NVIDIA HQ," the post reads, "I vividly remember first seeing these results, and being unable to believe how they could be true." Then, with admirable honesty: "It was at this moment I became a Vera believer."
The mechanism, as Modal explains it, is memory bandwidth. Vera's high-bandwidth memory subsystem lets it pack significantly more container workloads onto each core without throttling, so it does not slow down as jobs pile up — it keeps ripping through each one at its original pace. The tested machines' physical core boundaries were 128 for Intel GNR, 96 for Graviton and the GNR virtual machine, and 88 for Vera. As container counts rose, the x86 and Graviton alternatives hit early throughput plateaus. Vera kept climbing.
For Modal's scheduler, that translates into denser packing and more forgiving burst protection. For sandbox users, the pitch is higher absolute scale without sacrificing latency. And there is a second use case hiding in the post that matters more than agents: reinforcement-learning rollouts, where expensive training GPUs sit idle waiting for CPU-bound work to finish. "Time saved on CPU is time saved on GPU," as Modal puts it — and neoclouds live and die on GPU utilization.
The number Modal didn't headline: over 6x per-CPU vs AMD
Buried in Modal's own published benchmark charts is a number the writeup never headlined: at equal 124-logical-CPU configurations, Vera recorded 6,360 verified task completions per 60 seconds against AMD Zen 5's 961 — roughly 6.6x the per-CPU task throughput. The AMD number comes from Modal's own chart data, so it carries the same caveat as the Intel comparison: Modal's internal measurement, not an independent evaluation. But it sharpens the story in one useful way — Vera's advantage widens, not narrows, against the other x86 incumbent, which makes the 'why rebuild for Arm' decision easier to read.
What Vera actually is
Vera is the processor half of Nvidia's Vera Rubin platform, unveiled on March 16 at this year's GTC conference. It is an Arm-compatible CPU built around 88 custom Nvidia-designed "Olympus" cores, each able to run two threads through a feature Nvidia calls Spatial Multithreading — dedicated worker and I/O threads, so background operations can't starve a core's primary work. The memory subsystem is second-generation low-power LPDDR5X delivering up to 1.2 TB/s of bandwidth. Nvidia says Vera is in full production, with partner availability in the second half of 2026, and has named Alibaba Cloud, ByteDance, CoreWeave, Lambda, Nebius, Nscale and Oracle Cloud Infrastructure among the clouds planning to deploy it.
The chip was purpose-built, in Nvidia's telling, for agentic AI and reinforcement learning — the workloads that happen between model calls, when an agent uses tools, processes data, runs code and checks results. Nvidia claims up to twice the performance of leading x86 processors on such workloads and 1.5x the performance of its own Grace CPU. Those are Nvidia's numbers, presented at a keynote, and should be read the way keynote numbers always should be read. Independent corroboration is thin so far: third-party hardware site Phoronix did verify roughly a 20% efficiency uplift from Vera's spatial multithreading in its own testing, according to Nvidia's Vera team — one genuine independent datapoint on one specific feature, not on the headline claims.
The caveat, in large letters
Now the part that matters for honest coverage. The 2x throughput figure is Modal's internal number from early benchmarking runs on a single Vera node — not an independent evaluation, and Modal does not pretend otherwise. It is 50 curated tasks with network calls stripped out, which is a reasonable approximation of sandbox CPU work but not the same as production traffic. The company itself notes it only had access to one benchmark node, and that the results "suggest we could push it even further than the node we had access to" — phrasing that tells you as much about what wasn't tested as what was. Rival vendors will note, correctly, that they did not choose the benchmark, the tasks, or the machine configurations.
This is not a reason to dismiss the result. It is a reason to read it precisely: an infrastructure company with every incentive to avoid rebuilding its stack for Arm has measured enough of an advantage under the exact contention conditions its business depends on that it is willing to do the rebuild anyway. The benchmark is self-serving; the decision is the signal.
Why a $15.75 billion company is rebuilding for Arm anyway
Modal is not some scrappy upstart taking a flyer. Six days before the Vera post, it announced a $$750 million Series D led by Accel at a $15.75 billion post-money valuation — more than triple the $4.65 billion tag on its $355 million Series C just five months earlier. The company says usage has tripled since May, revenue is up more than 15x year over year, and it has crossed $1 billion in annualized run rate. Its Sandboxes product alone is doing a $300 million annualized run rate, with nearly 3 billion sandboxes spun up and the system scaling toward a million concurrent executions. Customers include DoorDash, Cognition, Ramp, Runway and Physical Intelligence; its infrastructure spans more than 20 cloud providers.
At that scale, a CPU that doubles throughput under contention is not a nice-to-have. It is a direct lever on cost of goods sold — and Modal said exactly that in the post, naming background agents (citing Ramp Inspect and agents on Niteshift) and RL rollouts as the two prevalent sandbox use cases where faster CPUs cut real dollars. Modal's founder and chief executive, Erik Bernhardsson, was blunt with Reuters about the strategy: running a mixed fleet means the company can "always deploy whatever hardware performs best." Asked about future Nvidia parts, he was equally direct: "We'll test them when they're available." Intel and AMD stay in the fleet; Vera becomes an additional lane, not a replacement. Nobody is burning bridges. Everyone is reading the benchmark.
The wider war for the CPU
Modal is not Vera's first believer, and the pattern of who else is lining up tells you what Nvidia is really after. Perplexity — the AI search company now handling more than 400 million queries a month — said earlier this year it would run agent work on Vera after its vice-president of enterprise infrastructure, Nate Kupp, found the chip ran agentic coding tasks about 1.5 times faster than the traditional CPUs the company had been using. Nvidia has also named Anthropic, OpenAI and Oracle among Vera's early users. The chipmaker has told investors it expects at least $20 billion in Vera sales this fiscal year, in a CPU market it values at roughly $200 billion.
That $200 billion figure is the actual story. Nvidia is coming for the ordinary server processor — the long, boring, enormously profitable market Intel and AMD have divided between themselves for decades. The pitch is that the agentic era has moved the bottleneck: when AI systems chain model calls, tool uses and code runs into long loops, the general-purpose processor coordinating all of it matters as much as the accelerator doing the heavy maths. It is the same argument Broadcom's Hock Tan has been making from the other direction — that AI's biggest problem isn't chips in the way most people think — and it reframes the CPU from commodity plumbing into the pacing item for the whole AI factory.
There is an irony worth savoring. Nvidia hand-delivered Modal's first Vera server to its San Francisco office — the blog post lingers on the detail with the pride of a kid showing off a new bike — while Intel and AMD, the incumbents Modal still runs, delivered nothing of the sort into this story. The optics are juvenile and the substance is real: the world's most valuable company is doing white-glove delivery for a single server to a startup, because every logo on the Vera customer slide is worth more than the silicon.
Our judgment: can Nvidia actually break the x86 hold?
Here is the honest version. Modal's 2x is a company number, measured on company hardware, on a company benchmark — it should move your priors by inches, not miles. Intel and AMD are not going to wake up obsolete; they have roadmaps, relationships, and a century of combined enterprise trust, and most data-center buyers will wait for independent, production-scale numbers before touching their architectures. The Graviton precedent also counsels patience: AWS spent the better part of a decade making Arm respectable in the cloud, and it took genuine, sustained price-performance wins to do it.
But the direction of travel is now unmistakable, and it favors Nvidia more than the headline numbers suggest. The moat was never just the chip — it is CUDA, the software stack, and now the Vera Rubin platform story that lets a buyer standardize CPU and GPU from one vendor with one throat to choke. Modal volunteering to rebuild its image system for Arm is the tell: clouds do not do that kind of work for a 10% benchmark win. They do it when the economics of their fastest-growing product line depend on it. If Vera lands at scale in the CoreWeave, Nscale and OCI fleets slated for the second half of this year, Intel's data-center narrative — already under siege from every direction — gets genuinely uncomfortable. Nvidia does not need to win the whole $200 billion CPU market to change the industry. It needs to win the agentic slice, and this week it collected its most credible witness yet.
Frequently asked questions
What is Nvidia Vera?
Vera is Nvidia's Arm-compatible CPU, unveiled at GTC on March 16, 2026. It is the processor half of the Vera Rubin AI platform but is also sold as a standalone CPU server. Nvidia positions it as a chip purpose-built for agentic AI and reinforcement-learning workloads — the CPU-side work that happens between AI model calls.
Is Nvidia Vera an Arm chip?
Yes. Unlike Nvidia's earlier Grace CPU, which was also Arm-based but sold mainly as a companion to GPUs, Vera uses 88 custom Nvidia-designed "Olympus" cores and is explicitly marketed as a standalone server CPU competing with Intel and AMD parts.
How fast is Vera compared with x86?
Nvidia claims up to 2x the performance of leading x86 processors on agentic workloads (a company claim, not independently verified). Modal's own early benchmark found up to double the task throughput versus a comparable Intel Granite Rapids machine under heavy concurrent load — but per-task speed on an unloaded machine was only modestly better, and Modal's number is internal, not independent.
Who is using Nvidia Vera?
Named early users and testers include Anthropic, OpenAI, Oracle, Perplexity and now Modal. Cloud providers planning deployments include Alibaba Cloud, ByteDance, CoreWeave, Lambda, Nebius, Nscale and Oracle Cloud Infrastructure, with availability in the second half of 2026.
What is Modal's Agentless Agent Bench?
It is the open-source benchmark Modal built to test Vera: 50 Terminal-Bench oracle solutions and verifiers, stripped of network calls and prewarmed in concurrent Docker containers, measuring the CPU tasks AI agents and RL rollouts actually perform — compilation, file grepping, git operations, device emulation. The code is public on GitHub.
How does Vera compare with AMD, not just Intel?
Modal's chart data shows roughly 6.6x per-CPU task throughput vs AMD Zen 5 at matched 124-logical-CPU configurations (6,360 vs 961 completions/60s); internal Modal measurement, not independent.
Will Modal drop Intel and AMD for Vera?
No. Mixed fleet: Intel and AMD stay; Vera is an additional lane. CEO Erik Bernhardsson told Reuters the company can “always deploy whatever hardware performs best” and that future Nvidia parts get the same treatment: “We’ll test them when they’re available.”
We checked the live blog feed and the full published log for any prior Modal or Vera CPU coverage before drafting this piece, and curl-verified every internal link in it against the live site — each returned a live page. If you want the longer explainer on what AI agents actually do inside those sandboxes, we have that too.
Sources: Modal company blog post "NVIDIA Vera CPUs double throughput for Modal Sandboxes" (Oct 19, 2026); Reuters original reporting via AI Weekly (original headline: "Modal Picks Nvidia Vera as First CPU for Hosting AI Agents, Joining Its Intel and AMD Mix"); Unite.AI; Bitcomme; RuntimeWire; PYMNTS, TechStartups and Oossa on Modal's $750M Series D; Futurum Group on the $300M Sandboxes run rate; Wccftech's exclusive with Nvidia's Vera team; Creative Collaboration's July 2026 interview with Perplexity's Nate Kupp.
TechPulse USA is a US-first technology newsroom covering AI infrastructure, chips, and the companies racing to build them. We verify every fact against primary sources and named reporting — and we say so when the numbers come from the company itself.
Comments
Post a Comment