TheChek News Flow

TensorWave will use fresh $350 million to fill more data centers with chips from AMD, an investor

In the artificial-intelligence world, pretty much everyone wants to lay hands on chips and other computing hardware made by the market leader, Nvidia. Everyone except Darrick Horton, that is.

Horton is the 28-year-old co-founder and chief executive of TensorWave. The Las Vegas-based cloud-computing startup refuses to use Nvidia’s graphics processing units—or GPUs, its signature chips—or its other products, out of concern that Nvidia controls too much of the AI infrastructure market. Horton says that situation is bad for competition.

Instead, TensorWave exclusively uses hardware and software made by Advanced Micro Devices AMD -3.02%decrease; red down pointing triangle, Nvidia’s smaller rival in the GPU space. The startup has closed its Series B funding round, led by AMD and the hedge fund Magnetar Capital, raising $350 million at a post-money valuation of $1.55 billion.

The investment nearly quadruples TensorWave’s value. About a year ago, the startup raised $100 million in a round that was led by Magnetar and AMD and valued the company at around $400 million.

“We wanted to figure out how we can solve problems for customers and restore competition to the market,” Horton says. “I don’t like buying things from monopolies. You don’t have a lot of leverage.” Nvidia didn’t respond to a request for comment.

Startups and larger companies are seeking to provide AI labs and large enterprise customers with a longer menu of computing alternatives besides Nvidia. That includes Cerebras, which makes platter-size chips used for running AI models quickly and which went public last month; Majestic Labs AI, which produces chips with vast amounts of memory; and Decart, whose software makes it easier to switch between computing technologies.

Horton says that in 2023, when TensorWave started, Nvidia was “a monopoly by default”—rather than one built on anticompetitive practices—and that customers were eager to diversify into alternative suppliers.

As the AI boom accelerated and computing shortages worsened, exacerbated by bottlenecks in everything from power supply to memory-chip production, TensorWave expanded rapidly.

“The race to build AI infrastructure has created urgent demand for providers who can deliver at speed without sacrificing reliability,” says Ross Laser, Magnetar’s president and co-founder.

Currently, TensorWave has three operational data centers, in Pennsylvania, Arizona and Florida, with computing capacity at the equivalent of 10,000 megawatts of electrical power, which is a relatively small amount.

TensorWave leases data-center shells from developers and fills them with AMD-designed chips and other equipment. The startup says it has signed leases for 500 megawatts of capacity in total, but most of it is in data centers that are still being built. In the coming year, the company hopes to raise that number to 2 gigawatts.

TensorWave plans to use the fresh round of capital to scale up its operations by adding more data centers and buying generators, chillers, critical power-supply gear, transformers and other equipment.

“As many data centers as we can sign, we can sell instantly,” Horton says.

TensorWave has worked with AMD over the past year to help improve ROCm, AMD’s custom-software platform, which has been criticized for being buggy and hard to use compared with Nvidia’s rival programming library, known as CUDA.

The software has improved to the point that “these days, it’s pretty much plug-and-play,” Horton says, making AMD’s chips well-positioned to capitalize on surging demand for what is known as inference computing, which allows AI models to respond to queries. AMD has marketed its GPUs as especially powerful when handling inference workloads. The influential analyst shop SemiAnalysis has identified AMD’s Instinct series as one of the best processors for inferencing.