Everyone talks about AI. But nobody talks about what actually runs it. The reality is, artificial intelligence does not exist without compute. And for most of the past decade, compute meant one company: Nvidia. Full stop.
But a startup born out of a Las Vegas coworking space is doing something most people said was impossible. TensorWave is building the AMD-powered alternative to Nvidia-dominated cloud infrastructure. And it is working.
How TensorWave Started: Problem, Solution, and Target Audience
The problem was hiding in plain sight.
Between 2022 and 2023, as AI adoption exploded, one bottleneck kept showing up everywhere: GPU access. Nvidia had captured so much of the AI chip market that prices shot up, wait times stretched out, and smaller companies simply could not afford the compute they needed to build. It was not just inconvenient. It was choking innovation.
Darrick Horton saw it clearly. A Forbes 30 Under 30 alum who once worked on plasma physics at Lockheed Martin’s Skunk Works, Horton was not the typical cloud startup founder. He told TechCrunch directly: “We recognized an unhealthy monopoly at work, one that was starving end-users of compute access and stifling innovation in the AI space. Motivated by our desire to democratize AI, we set out to provide a viable alternative and restore competition and choice.“
So in December 2023, Horton co-founded TensorWave with Piotr Tomasik as COO and Jeff Tatarchuk as chief growth officer. Their solution was sharp and specific: an AMD-exclusive cloud platform for AI workloads. Instead of fighting Nvidia on its home turf, they bet on AMD’s Instinct GPU lineup, specifically the MI300X, MI325X, and MI355X chips, which carry significant memory bandwidth advantages for training and running AI models.
Who were they building for? Three groups, mostly. AI-native startups that needed serious GPU compute but could not stomach hyperscaler pricing. Enterprises running complex AI systems that kept getting underserved. And research institutions building large models who needed memory-intensive infrastructure at a reasonable cost. Within a month of founding, StartUpNV wrote them a $2.2 million check. The rest, as they say, followed fast.
Competitive Advantage
Here is the thing about competitive advantage in infrastructure: it is rarely one big thing. It is usually several smaller things that stack up.
AMD-Exclusive Hardware with Real Memory Benefits. CoreWeave and Lambda Labs run almost entirely on Nvidia GPUs. TensorWave does not touch Nvidia. Their entire stack runs on AMD’s Instinct series, chips built with large HBM3E memory and high bandwidth. For large language model training and AI inference, memory speed is often the constraint. TensorWave built around that reality.
Better Pricing. AMD hardware costs less than Nvidia’s H100 and H200 lines. That gap is not trivial. It lets TensorWave offer more competitive per-GPU rates to enterprise customers. And while bigger rivals are burning through multibillion-dollar debt facilities, TensorWave keeps its capital efficiency sharp.
AMD Is Both Supplier and Investor. This is the part most people miss. AMD Ventures is a lead investor in TensorWave. So AMD is not just selling them chips. They are financially invested in TensorWave’s success. That means better access to next-generation silicon and a very public signal that AMD sees TensorWave as the proving ground for its AI hardware at scale.
Open Software Through ROCm. Nvidia’s CUDA ecosystem is its biggest moat. It is deeply familiar to developers and hard to abandon. AMD’s ROCm platform is open-source and more flexible, which appeals to enterprise teams that do not want to be locked into proprietary tooling forever.
Scale That Doubles Every Year. TensorWave has grown its infrastructure capacity tenfold annually since founding. They have secured over 2 gigawatts of long-term data center capacity and deployed the world’s largest liquid-cooled AMD-native AI training cluster, with 8,192 Instinct MI325X GPUs running simultaneously.
Marketing Techniques
Let’s be honest: most infrastructure companies market like they are filing quarterly reports. TensorWave takes a different approach.
The “Anti-Nvidia” Frame. From day one, they positioned themselves as the opposition. The Wall Street Journal called them an “anti-Nvidia data center.” That is not a line TensorWave had to pay for. It came from media picking up on a genuinely differentiated story. Horton regularly appears in press, reinforcing the democratization angle. It costs almost nothing and generates constant coverage.
Partnerships That Talk. In February 2026, TensorWave partnered with Credo, a networking technology company, to improve reliability across their GPU clusters. Earlier, they paired with MK1, a startup founded by Neuralink engineers, to bring fast AI inference to AMD cloud infrastructure. Each of these partnerships does two things at once: it builds technical capability and it gets written about. That is smart marketing disguised as operations.
Funding as a Brand Play. Every time TensorWave closes a round, it becomes a news story. The $43 million seed was the largest in Nevada startup history. The $100 million Series A followed. Then the $350 million Series B. Press releases go out through Business Wire, tech publications pick them up, and suddenly a cloud infrastructure company is front page news. That is brand awareness most startups would kill for.
Enterprise and Research as the GTM Core. The actual go-to-market is direct. They target cloud-native firms and research institutions through sales and referral channels, building long-term contracts that create steady revenue while also functioning as case studies for future customers.
Nevada as a Home Advantage. Headquartered in Las Vegas, TensorWave has leaned into the state’s economic development incentives and local tech community. It is a smaller play, but it supports hiring, earns local media, and keeps operating costs manageable.
How TensorWave Makes Money
The model is not complicated, but the execution is everything.
TensorWave charges customers for access to its AMD GPU clusters. On-demand access is billed by the GPU-hour. Larger customers lock in dedicated training clusters under longer-term contracts, which gives those customers price certainty and gives TensorWave predictable recurring revenue. It is the same basic model as other cloud providers, but with AMD hardware and a much sharper focus on AI-specific workloads.
The numbers show it working. In 2024, TensorWave’s annual revenue run rate sat at around $5 million. By 2025, they were projecting north of $100 million. That is a 20x jump in a single year. The upcoming rollout of AMD MI355X clusters across multiple North American regions is expected to push that even further as capacity grows to meet demand.
Market Share of TensorWave
TensorWave sits in what the industry now calls the “neocloud” market: specialized cloud providers built around GPU-dense infrastructure for AI workloads.
And it is crowded. The two clear frontrunners right now are CoreWeave and Lambda Labs. Behind them sit Nebius, Fluidstack, and TensorWave. CoreWeave, as a public company, holds the biggest slice. TensorWave’s share is still relatively small in raw terms.
But here is the kicker: the whole market is growing fast. The neocloud sector is projected to exceed $400 billion by 2027. TensorWave’s $1.55 billion post-Series B valuation shows investors believe it can carve out a meaningful piece of that growth, specifically as the AMD-powered option in a market where most players are Nvidia-dependent. With over 2 gigawatts of capacity secured and infrastructure growing tenfold each year, the trajectory is pointed in one direction.
Business Model Canvas of TensorWave
Key Partners: AMD (chip supplier and lead investor through AMD Ventures), Credo (networking technology), MK1 (inference optimization), Magnetar Capital (financial backing), Edgecore Networks, Aviz Networks, and the data center operators that provide the physical real estate.
Key Activities: Deploying and managing large-scale AMD GPU clusters. Building and maintaining fast networking across those clusters. Onboarding enterprise and research customers. Expanding the data center footprint across North America.
Key Resources: AMD Instinct GPU clusters across three generations (MI300X, MI325X, MI355X). Around 117 employees across engineering, operations, and sales. Over $493 million in total capital raised. More than 2 gigawatts of secured data center capacity.
Value Propositions: Cost-competitive GPU compute for AI training and inference. High memory bandwidth that suits large model workloads. A genuine alternative to Nvidia-dominated hyperscalers. Open ROCm software ecosystem. Both dedicated and on-demand cluster options for different customer needs.
Customer Relationships: Long-term cluster contracts for enterprise customers. On-demand access for smaller AI teams. Technical support and infrastructure services bundled in.
Channels: Direct enterprise sales. Partnerships with cloud-native firms and research institutions. Earned media through press coverage and funding announcements. Developer community engagement.
Customer Segments: Enterprise AI teams running serious model training. AI-native startups. Research institutions. Any company priced out or underserved by Nvidia-based cloud options.
Cost Structure: GPU hardware procurement is the biggest cost by far. Data center leases and operations. Networking and liquid cooling infrastructure. Employee payroll across technical and commercial teams. Power and bandwidth costs at scale.
Revenue Streams: GPU-hour billing for on-demand compute. Reserved cluster contracts that provide long-term, recurring revenue. Dedicated training cluster rentals for larger enterprise customers.
Conclusion: Is TensorWave a Viable Business?
Short answer: yes. But let’s not pretend it is without risk.
In roughly two years, TensorWave has raised over $493 million, grown revenue 20x, locked in AMD as both a supplier and a financial backer, and hit a $1.55 billion valuation. That is not luck. That is execution. The story is real, the differentiation is real, and the demand pulling it forward is very real.
But the risks matter too. The neocloud model is brutally capital-intensive. You are constantly pouring money into hardware before revenue catches up. CoreWeave’s enormous debt load shows how fragile the economics can get if AI demand ever softens. And AMD’s ROCm software ecosystem, improving as it is, still has ground to cover before it matches Nvidia’s CUDA in developer familiarity. That gap creates friction when trying to pull customers away from what they already know.
So what keeps TensorWave in the game? The mutual dependency with AMD. AMD needs TensorWave to prove its silicon can compete in real production AI workloads. TensorWave needs AMD to keep shipping competitive chips. Neither can afford to let the other fail. That is a structural moat that money alone cannot buy.
It is lonely building something nobody has done before. It is hard betting on the underdog chip when the whole world is running on the other one. But TensorWave is doing it. And right now, it is working.
Read about – Startup business models
Read in – Startup Directory
Read about Solo businesses
Hi Friends, This is Swapnil; I love reading and sharing knowledge. Currently working as a content writer at startupsunion.com. You all can hang out with me here.