Nvidia wants to own the entire AI data center stack, from chips to networking and storage

At GTC 2026, Nvidia showed that it no longer wants to be just a GPU supplier. The company is offering customers an entire AI data center as a single product, from compute and networking to context storage and management software.

A bet on vertical integration

The main signal from Jensen Huang's presentation was not a single new chip but the architecture of the business. Nvidia lined up a range of racks on stage and effectively told the market that AI infrastructure will be cheaper, faster and more profitable if you buy the whole stack from one vendor rather than individual components. This is no longer the model of "we sell accelerators, you assemble the rest yourself" but an attempt to turn the data center into a system fully designed by Nvidia.

Huang has long promoted the idea of the AI factory: a plant that produces tokens and intelligence the way a factory produces parts. Now the idea has become even more rigid. Nvidia wants to control not only compute but also networking, memory, storage of intermediate data, the CPU layer and the software that ties it all together. Amid the boom in agentic systems, the approach looks logical: the bottleneck is no longer just the GPU but everything that moves data between GPUs.

What is in the stack

In the new configuration, Nvidia has assembled several racks, each addressing a separate problem in the AI data center. Together they form an almost complete kit for companies building large clusters for training and inference that do not want to stitch together dozens of disparate suppliers into one design.

• Vera Rubin NVL72 — the flagship rack-scale system with 72 Rubin GPUs and 36 Vera CPUs.
• Vera CPU Rack — a standalone rack with 256 CPUs for agentic AI workloads, where tool calling, SQL and code execution matter.
• BlueField-4 STX — a layer for storing and quickly serving the KV cache that large language models need during inference.
• Spectrum-6 SPX — a new Ethernet network for connecting racks and scaling the cluster.
• Groq 3 LPX — an inference rack with 256 LPU accelerators, focused on low latency and large context.

The point of this lineup is that Nvidia now sells more than the "horsepower" of GPUs. It covers the parts of the system where milliseconds, watts and money are usually lost: moving data between chips, handling context, network latency, CPU tasks for agents and overall load balancing. The larger the model and the longer its context, the more noticeable this overhead becomes.

Where Nvidia sees the payoff

Nvidia's most prominent argument is inference economics. The company says the combination of Vera Rubin and Groq 3 LPX reduces calls to external DRAM thanks to the large amount of SRAM in the LPU, which lowers latency and speeds up token output. According to Nvidia, this setup can deliver up to 35 times more throughput per megawatt for trillion-parameter models and up to 10 times more revenue per watt in scenarios with expensive "premium" tokens.

"What used to require a whole day of queries will now be done in less

than an hour."

Nvidia is also betting on the CPU layer. It says plainly that even in the GPU era, agents constantly run into ordinary computing tasks: tool calls, SQL queries, compilation and sandboxed code execution. So the company puts its own Vera CPUs at the center and adds DPUs and specialized context storage alongside. This extends Nvidia's ambitions far beyond accelerators and shows that it wants to capture more and more margin in AI infrastructure.

What it means

For the market, this is another step toward vertically assembled AI data centers in which one vendor is responsible for nearly everything. For customers, this model may bring simpler deployment and better efficiency, but at the price of greater dependence on Nvidia. For competitors, from CPU and networking players to storage vendors, it is a signal that Nvidia no longer plays only in the GPU field.