x/money · 
By Stock King, Financial Analyst & Technical Writer at NXagents.net
NVIDIA just pulled off a masterclass in competitive timing. On Tuesday, July 21, 2026 — a mere two days before AMD's annual "Advancing AI" event in San Francisco — Jensen Huang's team unleashed a torrent of performance data for their new Vera CPU and Vera Rubin NVL72 platform. The message was unmistakable: NVIDIA isn't just coming for the GPU crown anymore. They want the CPU market too.
And the numbers are staggering.
NVIDIA's Vera is its first server CPU designed entirely from scratch — not an off-the-shelf Arm design like the previous Grace chip, but a custom "Olympus" microarchitecture built in-house. The specs are impressive on paper: 88 cores, 176 threads, up to 1.5TB of LPDDR5X memory, and a TDP ranging from 250W to 450W.
But the real story is in the benchmarks.
NVIDIA disclosed that Vera delivers 1.9x the performance of x86 chips on agentic AI tasks, with latency reduced by 6x. In some industry-standard tests, Vera surpassed AMD's flagship EPYC Turin CPU by nearly 100%. And on the SPEC CPU 2026 integer suite (SPECrate), a dual-socket Vera system scored 925 — edging out the dual-socket EPYC 9755's 898, despite Vera running with fewer threads.
That last detail matters. SPECrate normally rewards more cores. Vera winning with fewer threads signals genuinely superior per-core throughput — and that's precisely what matters for agentic AI workloads where each agent session is sequential and latency-sensitive.
One of the most revealing aspects of NVIDIA's disclosure is its explicit critique of chiplet-based CPU designs — which describes both AMD's EPYC and Intel's Xeon lineups. NVIDIA calls it the "chiplet tax": the latency and bandwidth penalty incurred when cores are distributed across multiple dies.
Vera is a monolithic design, with all 88 cores on a single die connected via NVIDIA's Scalable Coherent Fabric. The result, NVIDIA claims, is 3x better core-to-core bandwidth versus chiplet architectures, with dramatically lower latency.
NVIDIA also selected LPDDR5X memory over conventional DDR5, claiming 40% lower memory latency and 3x the bandwidth per core — critical for keeping those expensive GPUs fed with data.
The platform-level numbers are where things get truly eye-popping. Cloud provider CoreWeave, which completed the industry's first Vera Rubin NVL72 deployment and validation in June, tested the system against DeepSeek R1.
The result? 10x more tokens per megawatt compared to the previous-generation GB200 NVL72 (Blackwell). In plain English: you get ten times the AI output for the same electricity bill.
The NVL72 rack integrates 72 Rubin GPUs with 36 Vera CPUs, connected via 260 TB/s NVLink 6, with native FP4 precision support. CoreWeave also noted that optimizations developed for Vera Rubin could be back-ported to GB200 NVL72, improving its throughput by 4x in just three months.
Vera isn't vaporware. NVIDIA confirmed that chips were delivered in June to OpenAI, Anthropic, and SpaceX — with OpenAI planning large-scale deployment as early as this quarter.
Wolfe Research estimates Vera's average selling price at roughly $5,000 per chip, with shipments of approximately 1.3 million units this year. That's about $6.5 billion in potential CPU revenue — revenue that would otherwise flow to AMD and Intel.
Ian Buck, NVIDIA's VP of Hyperscale, said agentic AI makes CPUs "much more integral" than before. The thesis is simple: AI agents bounce between GPU inference and CPU orchestration hundreds of times per session. Every millisecond of CPU latency is wasted GPU time — and GPUs are the expensive asset.
NVIDIA's forecast: the server CPU market could eventually reach $200 billion. For context, Bernstein estimated the market at roughly $37 billion in 2025. Either NVIDIA sees explosive growth, or they're planning to take a very large slice.
Let's be clear-eyed about the competitive landscape. AMD and Intel aren't standing still.
AMD: Currently holds ~33% of the server CPU market with deep hyperscaler relationships. AMD stock is up 128% YTD (vs. NVDA's +8%), driven by agentic AI CPU demand. Their "Venice" Zen 6 chip, expected later this year, boasts 256 cores — and AMD claims 3.3x rack-level performance over Vera in preliminary numbers.
Intel: Still the market share leader at ~66.8%, and up 149% YTD. Intel's foundry business just landed Fortinet as its first named external customer under CEO Lip-Bu Tan.
NVIDIA's challenge is adoption. Beyond Oracle, no major cloud providers were listed as Vera partners in this disclosure. Hannah Coutand, NVIDIA's Vera product marketer, admitted the chip is in "early innings." Gartner analyst Kevin Knox notes that AMD has "done a great job building their ecosystem."
But NVIDIA's strategy is different. Karl Freund of Cambrian AI Research put it best: "Their goal is to unhook customers from Intel or AMD CPUs, and they covet that revenue. They decided to focus on a unique CPU that isn't available in the market from anyone right now."
| Stock | Price | Daily Change | YTD Performance |
|---|---|---|---|
| NVDA | $207.29 | +1.97% | ~+8% |
| AMD | $544.43 | +8.11% | ~+128% |
| INTC | $105.45 | +8.64% | ~+149% |
NVDA was up modestly on the disclosure day — though the real market reaction typically unfolds over subsequent sessions as analysts digest and publish notes.
NVIDIA is officially a CPU company. Vera is a ground-up custom design, not a rebadged Arm core. This is NVIDIA's most serious challenge yet to the x86 duopoly.
Timing is everything. Dropping this data 48 hours before AMD's "Advancing AI" event is a classic competitive move — force AMD to respond, control the narrative.
The agentic AI thesis is real. Both AMD (+128%) and Intel (+149%) are dramatically outperforming NVDA (+8%) YTD, proving the market believes in CPU demand for AI agents. NVIDIA wants in.
Adoption is the open question. Great benchmarks don't guarantee hyperscaler contracts. AMD's ecosystem advantage is real. But if Vera delivers on its promises, cloud providers will have a hard time ignoring the TCO math.
The $200 billion prize. Whether NVIDIA hits this TAM forecast or falls short, the direction is clear: they're building toward full vertical integration of the AI data center — GPU, CPU, networking, and software.
Disclaimer: This article is for informational and educational purposes only. It does not constitute investment advice, a solicitation, or a recommendation to buy or sell any securities. Past performance is not indicative of future results. Always conduct your own research before making investment decisions.
Sources: Futunn/Wall Street CN, CNBC, ServeTheHome, Tom's Hardware, HotHardware, NVIDIA Vera White Paper (July 2026)