TensorWave Inc., an artificial intelligence GPU cloud delivery platform aimed at delivering an alternative to Nvidia Corp.’s market dominance using Advanced Micro Devices Inc. chips, said today it has raised $43 million in a so-called “simple agreement for future equity” investment round.
The round was led by Nexus VP with participation from Maverick Capital, Translink Capital, Javelin Venture Partners, StartupNV and AMD Ventures. SAFE agreements, conceived in 2013 by the startup incubator Y Combinator, provide potential future equity in a company for the investor in exchange for immediate cash to the company.
Founded in 2023, TensorWave sought to help solve key challenges in the AI and machine learning market by providing greater availability and optionality to GPUs for AI training and inference. Currently, it’s estimated that Nvidia controls between 70% and 95% of the chip market for training and deploying AI models.
Speaking to SiliconANGLE in an interview, TensorWave Chief Executive Darrick Horton said that before launching the company, even if people could get equivalent or better compute and capability delivered compared with Nvidia cards, if it wasn’t simpler or faster to implement, then they wouldn’t want to switch. That’s when his team took notice of AMD’s Instinct MI300x series data center AI accelerators, released last year, which fit the bill.
As for the funding, Horton said it will be used to scale up the company’s primary data center with AMD Instinct accelerators to tackle AI workloads and also prepare its infrastructure for the next generation of AMD chips. The chipmaker intends to roll out a new generation of AI accelerators, the MI325X, in the fourth quarter of this year that appear to put Nvidia in the company’s sights.
“The MI300 for context is, you know, around the same as the H100, sometimes it’s better, sometimes it’s worse. It depends on the workload,” said Horton. “The MI325 is going to be significantly better than the H200. So, it will have the most memory of any chip on the market and the most memory bandwidth of any chip on the market. It will dominate on inference workloads.”
The H100 and H200 are Nvidia’s most powerful Tensor Core GPUs based on Nvidia Hopper architecture, making them capable of providing powerful accelerated AI compute. Horton explained that AMD’s chip offerings have been helping compete with Nvidia on price-performance ratios for the customers of its cloud compute solution.
TensorWave said the company intends to add MI325X instances to its offering as early as the end of the year after AMD launches its new silicon.
The global AI market is projected to grow from $196.6 billion in 2023 to $1.81 trillion by 2030, according to Grand View Research. Much of this will be driven by the accelerating rate of AI research and innovation and the continuous expansion of AI product offerings and services across every industry, including large language models from companies such as OpenAI and Google LLC. All of these models require massive amounts of compute for both training and deployment.
Horton added that the funding will also help the company launch Manifest, a specialized AI deployment platform.
“Manifest is a purpose-built, highly optimized inference stack for AMD that is accessible only at TensorWave,” said Horton. “We’re building in a lot of unique features that take advantage of AMD and offer solutions for actual enterprise use cases that are missing in the marketplace right now.”
The platform will support larger context windows, which is the amount of data that an AI system can ingest at a time to work on, and offer lower response times. Manifest will provide efficient analysis of complex documents for enterprise customers, provide accelerated reasoning and a secure platform for private data storage. It will launch during the fourth quarter of 2024.