Chip manufacturer Nvidia achieved sales of $30 billion last fiscal quarter, due largely to the AI ​​industry's insatiable demand for GPUs. GPUs are essential for training and running AI models. They contain hundreds of cores that work in parallel to quickly execute the linear algebra equations on which the models are based.
The appetite for AI stays strong and Nvidia's GPUs have develop into the chip of alternative for AI gamers of all sizes. But TensorWavean organization founded at the top of last yr, goes against the grain and introducing a cloud that only offers access to hardware from Nvidia competitor AMD for AI workloads.
“We recognized that an unhealthy monopoly is at work – one which denies end users access to computing and stifles innovation in AI,” Darrick Horton, CEO of TensorWave and considered one of its co-founders, told TechCrunch. “Driven by our desire to democratize AI, we’re committed to providing a viable alternative and restoring competition and alternative.”
Winding paths
Pickleball initially brought Horton along with TensorWave's two other co-founders, Jeff Tatarchuk and Piotr Tomasik – or a minimum of it got the ball rolling (pardon the pun).
One day after a match, Tomasik and Tatarchuk—close friends and longtime pickleball doubles partners—invited Horton, a former colleague of Tatarchuk's, to their favorite bar in Las Vegas.
“As the conversation progressed, we discussed the monopolistic influence over GPU computing capability, which led to produce shortages,” Horton said. “This realization led to the founding of TensorWave.”
The three partners didn't just know one another from pickleball.
Tatarchuk co-founded cloud provider VMAccel with Horton before selling one other startup, customer relationship management (CRM) developer Lets Rolo, to digital identity company LifeKey. Horton, who has bachelor's degrees in mechanical engineering and physics, once worked at Lockheed Martin's Skunk Works research and development division after which co-founded VaultMiner Technologies, a crypto mining company and parent company of VMAccel.
As for Tomasik, he co-initiated Lets Rolo with Tatarchuk. (Tomasik can also be co-founder of influencer marketer website Influential, which French PR firm Publicis acquired for $500 million in July.)
What made three entrepreneurs with little knowledge of the hyperscaler landscape think they might compete with the giants of the AI ​​industry? Basically, persistence.
“We believed we could solve the GPU supply problem,” Horton said.
Vegas, Inc.
TensorWave is headquartered in Las Vegas, an unusual city alternative for a cloud infrastructure startup. But Horton said the team liked the possibilities.
“We thought Vegas had the potential to be a thriving tech and startup ecosystem,” he said.
This prediction is just not entirely incorrect. After According to Dealroom.co, Las Vegas is home to only over 600 startups with greater than 11,000 employees, attracting over $4 billion in investments in 2022.
Energy costs and overhead costs Are lower in Vegas in addition to in lots of major US cities. And each Tomasik and Tatarchuk have close ties to the town's VC community.
Tomasik was previously a general practitioner at Vegas-based seed fund 1864 Fund and now works with nonprofit accelerators StartUp Vegas and Vegas Tech Ventures. (Oddly, Vegas Tech Ventures' website threw a 404 error for the pages listing its partners and portfolio corporations; A spokesman said it was a technical error that was being resolved Christian religious organization.
These connections – together with Horton's – helped TensorWave develop into considered one of the primary clouds to launch AMD Instinct MI300X instances for AI workloads. TensorWave delivers setups with dedicated storage and high-speed connections upon request. TensorWave rents GPU capability by the hour and requires a minimum six-month contract.
“We are in good company across the cloud space,” Horton said. “We see ourselves as a complement and offer additional AI-specific computing power at a competitive price-performance ratio.”
AMD forward
There is a booming marketplace for startups developing low-cost, on-demand, GPU-powered clouds for AI.
CoreWeave, the GPU infrastructure provider that began as a crypto mining company, recently raised $1.1 billion in recent funding (and $7.5 billion in debt) and signed a multi-billion dollar contract signed act with Microsoft. Lambda Labs secured a special financing vehicle value as much as $500 million in early April allegedly I'm aiming for one more $800 million. Non-profit organization Voltage Park, backed by crypto billionaire Jed McCaleb, announced last October that it was investing $500 million in GPU-powered data centers. And Together AI, a cloud GPU host that also conducts generative AI research, raised $106 million in a funding round led by Salesforce in March.
So how does TensorWave plan to compete?
Firstly, the worth. Horton notes that the MI300X is that this significantly cheaper as Nvidia's current hottest GPU for AI workloads, the H100, and that TensorWave can pass on savings to customers consequently. He wouldn't reveal TensorWave's exact instance prices. But to beat the more competitive H100 plans, it might need to be under about $2.50 an hour — a difficult but not unimaginable feat.
“Prices range from about $1 per hour to $10 per hour, depending on individual workload requirements and GPU configurations chosen,” Horton said. “As for the per-instance cost incurred by TensorWave, we cannot share these details because of confidentiality agreements.”
Secondly, about performance. Horton points to benchmarks showing the MI300X superior the H100 with regards to running (but not training) AI models, especially text-generating models like Meta's Llama 2. (Other Evaluations suggest that the profit could also be workload dependent.)
It seems that Horton's claims are somewhat credible, considering the tech industry's movers and shakers are concerned about the MI300X. Meta announced in December that it might use MI300X chips to be used cases comparable to running its Meta AI assistant, while OpenAI, the maker of ChatGPT, plans to support the MI300X in its developer tools.
The competition
Others betting on AMD's AI chips range from startups like Lamini and Nscale to larger, more established cloud providers like Azure and Oracle. (Google Cloud and AWS remain unconvinced AMD's competitiveness.)
At this time it is necessary to work for the good thing about all of those providers ongoing Nvidia GPU shortage and the delay from Nvidia's upcoming Blackwell chip. But the shortage might be alleviated soon with a ramp-up in manufacturing of critical chip components, particularly memory. And that would allow Nvidia to extend shipments of the H200, the successor to the H100, which features significantly improved performance.
Another existential dilemma for cloud newbies betting on AMD hardware is bridging the competitive benefits Nvidia has built around AI chips. Nvidia's development software is taken into account more sophisticated and user-friendly than AMD's – and is widely used. AMD CEO Lisa Su has approved that it “takes work” to introduce AMD.
In the long run, price war could develop into difficult as hyperscalers increase their investments in custom hardware to operate and train models. Google offers its TPUs; Microsoft recently unveiled two custom chips, Azure Maia and Azure Cobalt; and AWS has Trainium, Inferentia and Graviton.
“As developers search for alternatives that may effectively handle their AI workloads, especially within the face of increased storage and performance requirements and ongoing production issues causing delays, AMD will proceed to take care of its dominance and play a key role in democratizing computing in AI. “Play Ages,” Horton said.
Early demand
TensorWave began onboarding customers in preview late this spring. However, Horton says the corporate already generates $3 million in annual recurring revenue. He expects that number to achieve $25 million by the top of the yr – an 8-fold jump – once TensorWave increases capability to twenty,000 MI300Xs.
Provided $15,000 per GPU20,000 MI300Xs would equate to a $300 million investment – yet Horton claims that TensorWave's burn rate is “well throughout the sustainable range.” TensorWave before told The Register said it might use its GPUs as collateral for a big debt financing round, an approach also taken by other data center operators including CoreWeave; Horton says that's still the plan.
“This reflects our strong financial health,” he continued. “We are strategically positioned to weather potential headwinds by delivering value where it is required most.”
I asked Horton how many shoppers TensorWave has today. He declined to reply for “confidentiality reasons,” but emphasized TensorWave’s publicly announced partnerships with a network backbone provider Edgecore networks And MK1an AI inference startup founded by former Neuralink engineers.
“We are rapidly expanding our capability with multiple nodes available, and we’re continually increasing capability to satisfy the growing needs of our pipeline,” Horton said, adding that TensorWave plans to launch AMD’s next-generation MI325X GPUs on the It will likely be released available on the market within the 4th quarter of 2024, online in November/December.
Investors seem like pleased with TensorWave's growth trajectory to this point. Nexus VP announced Wednesday that it led a $43 million round of the corporate that also included participation from Maverick Capital, StartupNV, Translink Capital and AMD Ventures.
The tranche – TensorWave's first – values ​​the startup at $100 million after funding.
“AMD Ventures shares TensorWave’s vision to remodel AI computing infrastructure,” AMD Ventures SVP Mathew Hein said in a press release. “Their use of the AMD Instinct MI300X and talent to supply public instances to AI customers and developers positions them as an early competitor within the AI ​​space, and we’re excited to support their growth with this latest round of funding.”
