Modal Just Made Nvidia's Vera CPU the Default for AI Agents — and Took That Seat From AMD and Intel

Nvidia's Vera CPU launched in March 2026 as the host processor inside the Vera Rubin platform — 88 custom Olympus Arm cores, 1.2 TB/s of LPDDR5X memory bandwidth, and a 1.8 TB/s coherent link to Rubin GPUs via second-generation NVLink-C2C. It is engineered for the branch-heavy, memory-bound work that runs around a GPU: Python runtimes, sandboxed code execution, tool calls, and the orchestration logic of long-running AI agents. Vera systems began shipping to partners in the fall of 2026.

Until Monday, Modal's CPU estate was a binary: Intel Xeon (Sapphire Rapids, now Granite Rapids) plus AMD EPYC — and nothing else. Modal said it will run Vera alongside the existing x86 fleet in a mixed architecture, assigning each workload to whichever CPU handles it best, rather than a wholesale cutover.

Modal's own early tests, disclosed Monday, showed Vera doubling the work-per-second of comparable x86 silicon on agent hosting. The third-party picture lines up: DeepInfra's July 2026 benchmark locked the comparison basis at 20 physical cores, one NUMA node, one thread per core, and measured per-core orchestration-step latency on bare-metal runc and Kata microVMs.

At a fixed service-level objective — 99th-percentile response under 6 seconds with zero errors — Vera hosted 256 concurrent agents per partition on identical hardware. The x86 incumbents topped out around 160 (192 on a best-case Granite Rapids day). On the same 20-core partition running full tilt, leftover cores served a gpt-oss-20b CPU inference workload at 118.5 tokens/second decode — better than the 84.5 tok/s measured on a full 72-core Nvidia Grace (GH200) socket doing nothing else. The gap traces to single-thread IPC: Vera sustains 3.62 instructions per clock on branch-heavy agent code, against 2.44 for the best x86 — a 48% advantage that compounds when the workload is embarrassingly parallel.

Modal closed a $750 million Series D on October 15 at a $15.75 billion post-money valuation, with usage tripling in five months and annualized run-rate revenue tracking around $1 billion. Its sandbox product — the substrate that hosts the agents — crossed $300 million ARR, with roughly three billion sandboxes spun up to date and a stated capacity of one million concurrent executions. The customer list reads like a who's-who of agentic AI: Anthropic, Meta, DoorDash, Cognition (Devin), Ramp, Runway, Physical Intelligence, and Suno. Modal is not a hyperscaler; it is the compute layer underneath hundreds of agentic products.

That makes it a high-signal win for Nvidia. The Vera customer roster at launch in March 2026 included Alibaba, ByteDance, Meta, Oracle Cloud Infrastructure, CoreWeave, Crusoe, Lambda, Nebius, and Nscale; the May 31 expansion added Cloudflare, NYSE, Anthropic, OpenAI, and SpaceXAI. Modal is the first pure-play inference neocloud on the list, and the first reference customer whose entire thesis — serverless compute for AI workloads — depends on CPU performance rather than GPU time.

AMD has the stronger claim in the data-center CPU market — Data Center segment revenue hit $6.7 billion in Q2 2026, up 107% year-over-year and 58% of AMD's total revenue, with EPYC reaching a record 46.2% share of x86 server CPU spending in Q1 2026. But that share is built on traditional cloud and enterprise workloads. The agentic AI inference layer — Modal, CoreWeave, Nscale, Crusoe, Cerebras — is a different buying center, and it now has a non-x86 incumbent.

AMD's answer is EPYC \"Venice\" on a 2 nm-class process with up to 256 Zen 6 cores per socket, plus a dedicated 256-core variant for AI infrastructure with 512 threads, and the Helios rack-scale system that pairs Venice CPUs with MI455X GPUs. AMD says Helios delivers up to 30% more inference tokens per dollar than the unnamed competition, and Microsoft is deploying Helios to power frontier-model inference. Venice is slated for a Q3 2026 launch window — the same window Modal just chose Vera.

Intel's data-center group is structurally weaker than both: Intel posted $57.0 billion in trailing-twelve-month revenue against an $11.3 billion net loss, with Q2 2026 revenue of $16.1 billion masking a 41.5% gross margin and a loss-making operating line. Losing agent-class workloads at the inference neocloud layer is a smaller absolute blow than it is for AMD, but it deepens the gap between Intel and the two architectures — Arm and x86 — that are actually shipping product into AI factories today.

Vera is built on TSMC's N3 process — the same leading-edge node that anchors the Rubin GPU — which means Vera pulls incremental wafer demand into a foundry that is already capacity-constrained for AI accelerators. That compounds the Nvidia-TSMC revenue linkage: every Vera CPU shipped is one more die eating into the same allocation that competes with Rubin, Blackwell Ultra, and the next-gen Feynman family in2027.

The CPU cores are custom Olympus designs running on the Arm architecture — Nvidia's first standalone data-center CPU, and the validation of a multi-year bet on Arm-based server silicon that began with Grace. The 1.8 TB/s NVLink-C2C coherent link to Rubin GPUs is what turns Vera from a CPU into a platform component: it lets the host keep agent orchestration in lockstep with GPU work, which is the entire reason a neocloud like Modal can rebalance jobs across a mixed fleet without losing throughput.

On the memory side, Vera's LPDDR5X complement is distinct from the HBM4 that pairs with the Rubin GPU, but the rack economics flow together. SK Hynix is one of the primary HBM suppliers for Nvidia's Rubin generation, which means a Vera-driven AI factory orders pull both a CPU die from TSMC and accelerator memory from SK Hynix at the same time — a single decision at a neocloud is now a multi-supplier revenue event.

Strip out the announcement language and the operating data are unambiguous. Nvidia told investors in March it expects $20 billion in Vera chip sales in the fiscal year ending January 2027, against a $200 billion CPU opportunity it sizes as the addressable market. Nvidia Q2 FY2027, reported August 26, 2026, booked $96.2 billion in revenue at a 75.0% gross margin, with $59.7 billion in net income — the cash-generation machine that funds the Vera ramp.

Modal validates three claims at once: that there is a non-trivial inference neocloud segment that will standardize on Vera rather than the cheapest x86; that the 1.8x-to-2x throughput advantage holds up under a real production workload, not just a synthetic benchmark; and that the platform plays well in a heterogeneous fleet. Each of those claims, individually, is what a Nvidia sales rep needs to close the next neocloud deal.

Days to quarters — watch Q3 and Q4 FY2027 results (reporting November 2026 and February 2027) for any disclosure of Vera-specific revenue against the $20 billion target. Watch AMD's Q3 2026 earnings (October 28, 2026) for any commentary on EPYC Venice pricing power and whether Microsoft Helios adoption is delivering the token-per-dollar advantage AMD claims. Watch CoreWeave Q3 2026 for any color on how Vera is changing the CPU line item in the bill of materials — CoreWeave is one of the largest Vera adopters and the first neocloud in Nvidia's customer list to scale it broadly.

One to three years — Vera is the bridge between Grace and whatever Nvidia ships as the post-Rubin CPU family. If Modal, CoreWeave, Nscale, and the larger Vera cohort lock in, AMD has to defend the inference layer with Venice and the dedicated AI variant at 256 cores, while Intel faces the harder question of whether its Xeon roadmap can credibly compete for agent-class workloads at all. The structural read: AI inference is becoming CPU-bound at the orchestration layer, and the architecture that wins that boundary controls the platform.