Darrick Horton, CEO and co-founder of TensorWave, believes the company has a significant competitive advantage, as the price of AMD MI300X is much cheaper than the current most popular nvidia GPU H100, and MI300X performs better than H100 in running (but not training) AI models. The company valuation has reached 0.1 billion US dollars, with expected annual recurring revenue to reach 25 million US dollars by the end of the year, an 8-fold increase.
The AI boom is in full swing, making Nvidia the undisputed leader, with the company's GPUs becoming the preferred chips for AI companies of all sizes. However, media reports that the company TensorWave, founded at the end of last year, is making waves by launching a cloud computing service that only offers competition to Nvidia's rival, AMD hardware, for processing AI workloads.
CEO and Co-founder of TensorWave, Darrick Horton, stated in an interview with TechCrunch:
"We realize this is an unhealthy monopoly situation, one that prevents end users from accessing computing resources and stifles innovation in the AI field. In our desire to make AI more accessible, we have decided to provide a viable alternative to restore competition and choice."
Established in the gambling city.
According to reports, Horton and TensorWave's other two co-founders, Jeff Tatarchuk and Piotr Tomasik, met through the sport of Pickleball, and subsequently reached a consensus on the issue of GPU computing power monopoly leading to supply restrictions, ultimately prompting TensorWave.
TensorWave is headquartered in Las Vegas, an unusual choice for a cloud infrastructure startup. However, Horton explained that the team saw an opportunity here. "We believe Las Vegas has the potential to become a thriving ecosystem for technology and startups."
This prediction is not entirely unrealistic. According to Dealroom.co data, Las Vegas has over 600 startups, employing over 11,000 employees, and attracted over $4 billion in investments in 2022.
Compared to many major cities in the USA, Las Vegas has lower energy costs and operating expenses. In addition, Tomasik and Tatarchuk have close ties to the local venture capital community.
Tomasik was previously a general partner of the Las Vegas-based seed fund 1864 Fund and now collaborates with the non-profit accelerator StartUp Vegas and Vegas Tech Ventures. Tatarchuk is an angel investor in the incubator Fruition Lab.
These connections helped TensorWave become one of the first cloud computing services in the market to offer AMD Instinct MI300X instances for processing AI workloads. TensorWave offers GPU computing resource rental services by the hour and requires a minimum contract term of at least six months.
Centered around AMD
Media reports indicate that the low-cost, on-demand GPU cloud market is thriving. Horton states that TensorWave has a clear competitive advantage.
First is price. Horton points out that the price of AMD MI300X is much cheaper than the current popular Nvidia GPU H100, allowing TensorWave to pass on cost savings to customers. He did not disclose specific pricing for TensorWave instances, but to beat the H100, pricing must be below $2.50 per hour.
Second is performance. Horton cites benchmark test data showing that the MI300X outperforms the H100 in running (but not training) AI models, especially in text generation models like Meta's Llama 2.
In addition to TensorWave, other companies betting on AMD AI chips include startups Lamini and Nscale, as well as larger, more established cloud providers like Azure and Oracle.
Analysis shows that the current favorable factors for users of AMD chips are the continuous shortage of Nvidia GPUs, as well as the delay of the upcoming Blackwell chip from Nvidia. However, with the accelerated manufacturing of key chip components (especially memory), the shortage may soon ease. This will enable Nvidia to increase the shipment volume of H200, and the performance of this new product has been significantly improved.
For emerging cloud computing service providers relying on AMD hardware, another existing challenge is how to overcome the competitive barriers built around AI chips by Nvidia. It is widely believed in the industry that Nvidia's development software is more mature, easier to use, and has been widely deployed. Even AMD's CEO, Lisa Su, admits that adopting AMD chips "requires effort".
Looking further ahead, with large-scale companies increasing their investments in custom hardware, the challenge of focusing on price competition may become increasingly difficult. For example, Google offers its TPU; Microsoft recently launched two custom chips, Azure Maia and Azure Cobalt; and Amazon AWS has Trainium, Inferentia, and Graviton.
But Horton said:
"As developers seek alternative solutions that can effectively handle their AI workloads, especially as memory and performance requirements increase, coupled with delays caused by production issues, AMD will maintain a leading position for a longer time in the future and play a key role in the democratization of computing in the AI era."
Expected annual recurring revenue to increase 8 times, valuation has reached 0.1 billion USD
TensorWave currently has annual recurring revenue of 3 million USD. Horton stated that once TensorWave increases the capacity of its MI300X to 20,000, the recurring revenue is expected to reach 25 million USD by the end of the year, an 8-fold increase.
Assuming a cost of 15,000 USD per GPU, then 20,000 MI300Xs would mean an investment of 0.3 billion USD. However, Horton claims that TensorWave's cash burn rate is "completely sustainable." TensorWave had previously stated that it would use its GPUs as collateral for large-scale debt financing, a method adopted by other data center operators, which Horton says is still part of the company's plan.
He continued:
"This reflects our strong financial health. By strategically positioning ourselves to provide value where most needed, we are preparing to face potential headwinds."
However, when asked about the current number of customers TensorWave has, Horton refused to answer due to "confidential reasons," but emphasized the publicly announced partnerships of TensorWave with networking backbone supplier Edgecore Networks and AI inference startup MK1 founded by former Neuralink engineers.
Horton stated that the company is rapidly expanding its computing power to meet the growing demand. He also added that TensorWave plans to use the next generation MI325X GPU from AMD, which may be released in the fourth quarter of this year.
Investors are pleased with the growth trajectory of TensorWave. Nexus VP revealed on Wednesday that the firm led a $43 million round of financing, with participants including Maverick Capital, StartupNV, Translink Capital, and AMD Ventures. This marks TensorWave's first financing round, with the company valuation reaching $0.1 billion after the funding.