TheChek News Flow

The Takeaway

Nvidia faces growing competition in the AI chip market. But the chip designer has found a way to make money even from its rivals’ chips.

During its annual GTC developer conference last week, the company unveiled a new server rack designed to run chips made either by Nvidia or by its rivals. The new racks include networking technology that connects the chips on each rack, ensuring that they can communicate quickly and reliably, according to two people with knowledge of the project.

The move shows how Nvidia is starting to cater to the growing array of chips from Google and Advanced Micro Devices used by its major customers. It also gives Nvidia a potential new foothold in China, where companies could fill the new rack with domestically produced chips and run them on Nvidia’s software, generating revenue for the company without requiring it to navigate restrictions on chip sales from either the U.S. or China.

The new rack, known as MGX ETL, utilizes Nvidia’s MGX rack design, a format the company introduced in 2023 and has since made the standard for its data center products. It is also the design Nvidia is now relying on for its Groq and Vera chip racks, which can hold up to 256 chips in a single cabinet.

ETL is not Nvidia’s first attempt to bring rival chips into its servers, but it is the easiest one for customers to use. Since last year, Nvidia has offered a separate program called NVLink Fusion, which allows companies to integrate their chips with NVLink, Nvidia’s proprietary high-speed connection technology that allows chips inside a server to communicate with each other very fast. For example, Amazon has committed to building its next Trainium4 AI chip with NVLink compatibility.

Not all companies are willing to embrace NVLink. More than 115 companies including AMD, Google, Meta and Microsoft have backed UALink, an alternative to NVLink that companies can build without licensing or approval from Nvidia.

The ETL rack is different. It runs on Spectrum-X, Nvidia’s networking technology built on Ethernet, the underlying technology that virtually every chip already supports. Getting Spectrum-X’s full performance benefits still requires Nvidia’s own switch chips and network cards, but the barrier for customers to use ETL rack is lower than NVLink. For Nvidia, that may be the point: even as customers increasingly turn to rival chips, the ETL rack keeps Nvidia at the center of AI infrastructure.

An Nvidia spokesperson said MGX is an open rack design and that it does not restrict system partners from using it with alternative components. The company said that was not new, and described the MGX ETL as an “open rack reference architecture”, a modular system design that can support different mixes of compute and networking.

Some Nvidia employees have been pitching the new rack to some customers, according to two people involved in the conversations. For example, Nvidia has presented the rack to some Chinese companies as a way to plug in a mix of domestically made AI chips and chips from companies such as AMD while still running on Nvidia’s Spectrum‑X networking and software, according to the two people.

The new rack also could help Nvidia counter allegations that it forces customers to buy chips and networking equipment together, a practice that has irked large customers such as Microsoft and that previously triggered an investigation by EU competition regulators.

Sales of networking equipment has become an increasingly important part of Nvidia’s business. Networking accounted for more than 15% of its total revenue, or $11 billion, in the quarter that ended in January. Those sales grew 268% compared to the same period a year earlier, a much faster growth rate than for any other business Nvidia breaks out.

Nvidia is responding to growing efforts by other U.S. companies such as Google, Amazon and Meta Platforms, as well as by Chinese firms like Huawei Technologies, to sway customers toward buying or renting alternative chips, especially for running AI products and agents at high speeds.

At GTC last week, CEO Jensen Huang said the AI industry has moved from training AI models to running them at scale as AI agents, which requires a different type of processing with faster and cheaper chips that access a lot of data storage and memory hardware. He said that in the past two years “the amount of computation required for reasoning”—a form of inference computing where a model performs logical processing to solve problems—“has increased by about 10,000 times, and usage has increased by about 100 times.” He said there was a roughly “millionfold” increase in total computing demand over this period.