Nvidia Just Attacked the One Market It Didn’t Already Own
GPU company started selling CPUs, Google shipped everything except the model people wanted, and Intel reports tonight into the worst possible setup.
Nvidia released full specifications and benchmarks for its Vera data center CPU on Tuesday, the kind of architectural detail prospective customers need to actually evaluate a chip. That is not a teaser. That is a sales document.
The first Vera chips were delivered in June to OpenAI, Anthropic, and SpaceX. Shipping already happened. The announcement was just the paperwork.
⚙️ The pitch is agents, not servers
Nvidia says Vera was built specifically to address bottlenecks in AI agents, claiming 50% better agent performance than x86 chips, the silicon Intel and AMD make.
The chip packs 88 custom Olympus cores and 176 threads, with a memory subsystem hitting 1.2 TB/s of aggregate bandwidth on SOCAMM2 LPDDR5X, 3.4 TB/s of on-die fabric bandwidth, and a 164 MB unified L3 cache. Nvidia built it from the ground up instead of using an existing Arm design, prioritizing single-core speed, low latency, and memory bandwidth.
The benchmark is the interesting part. In SPECrate integer, Vera posted a base score of 925 against 898 for a dual-socket Epyc 9755, roughly a 3% margin, and it got there with fewer threads.
Three percent is not a blowout. Three percent with fewer threads on a first-generation part against AMD’s flagship is a different sentence entirely.
💵 The market size claim is doing heavy lifting
Nvidia says the server CPU market could eventually be worth $200 billion. A Bernstein estimate from earlier this year pegged the mature server CPU market at about $37 billion in 2025.
Read those two numbers next to each other. Nvidia is not forecasting share gains in an existing market. It is arguing the market becomes five times bigger because agents need CPUs the way training needed GPUs.
Wolfe Research expects average selling price around $5,000 per Vera unit. It sells standalone and as part of the Vera Rubin platform. HPE already announced a Vera-powered server, the ProLiant Compute DL394 Gen12, targeting agentic AI orchestration and reinforcement learning, shipping fall 2026.
🎯 Who this actually hurts
Not who you’d guess. AMD controls roughly a third of the server CPU market, Intel still holds close to two-thirds, and AMD keeps gaining ground. Gartner analyst Kevin Knox told CNBC that AMD currently leads innovation in enterprise AI server processors and has deep hyperscaler relationships.
Cambrian AI Research founder Karl Freund framed Vera as built for intensive AI workloads rather than conventional enterprise computing like hosting websites or business applications. The niche read is the correct read, at least for now.
Here’s the part the market has not resolved. AMD is up 128% and Intel is up 149% in 2026, the two best-performing chip stocks of the year, while Nvidia has gained 8%.
That’s not a company losing. That’s a company whose competitors got repriced on a thesis it just showed up to compete in.
🔵 Google shipped the wrong models
Google released three new Gemini models on July 21: Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber, a security-tuned variant restricted to governments and trusted partners. Gemini 3.5 Pro, the flagship, has now missed its target multiple times and still is not shipping.
In the same update, Google announced it has begun its most ambitious pretraining run yet for Gemini 4.
Translation: look past the flagship we couldn’t deliver, at the flagship after that.
Flash Cyber slots directly against Anthropic’s approval-gated Mythos security models and Microsoft’s Project Perception. Restricted-access security models are becoming their own product category, and every major lab now has one.
🧮 Intel reports into all of this tonight
Intel’s Q2 call is at 5:00 PM ET. Consensus is roughly $14.42 billion in revenue on 22 cents per share, about 12% revenue growth.
The stock is up more than 160% this year and down roughly 13% in the past week. Analysts want progress on the Terafab project and the Google deals.
Now add Vera to the question list. Intel spent 2026 getting repriced as an AI hardware winner. Nvidia just published benchmarks arguing part of that thesis belongs to someone else.
This is the quarter where Intel has to explain what it owns that Nvidia can’t build.
— The Bandicoots 🔧💻⚡