Most people who build a company in AI infrastructure start by asking what they can build. Darrick Horton started by asking what was broken. The answer, he decided, was the market structure itself. One supplier. One software stack. No real choice. So in December 2023, the 28-year-old former plasma physicist from Lockheed Martin’s Skunk Works co-founded TensorWave in Las Vegas with Jeff Tatarchuk and Piotr Tomasik, with a single stubborn goal: restore competition to AI compute.
That bet is now worth $1.55 billion.
What is TensorWave? The AI Cloud Built on AMD GPUs
TensorWave is an AI cloud platform built exclusively on AMD hardware. Not mostly AMD. Not AMD with some NVIDIA sprinkled in. Entirely AMD, top to bottom, every data center, every chip, every line of software.
The company runs AMD Instinct series GPUs, including the MI300X, MI325X, and the next-generation MI355X, across all of its infrastructure. It pairs that hardware with AMD’s open-source ROCm software platform and a direct-to-chip liquid cooling system that lets those GPUs run at higher sustained clock speeds than air-cooled setups allow. The result is a cloud built for one thing: the most demanding AI workloads you can throw at it.
Here is the kicker. Horton did not stumble into AMD. He chose it on purpose, after a deliberate decision to go all-in on an alternative to a market he saw as dangerously concentrated. “I don’t like buying things from monopolies,” he told the Wall Street Journal. “You don’t have a lot of leverage.” So TensorWave was created, as Horton puts it, “to restore competition to the market.”
That is a founder’s conviction, not a marketing line. And it shaped every technical choice the company has made since.
TensorWave’s $350 Million Funding Round and $1.55B Valuation
Let’s be honest: the funding trajectory here is almost hard to believe.
TensorWave received $2.2 million from StartUpNV within its first month of existence. Then a $43 million seed round in October 2024, which set the record for the largest seed funding in Nevada history. Then a $100 million Series A in May 2025, which broke its own record. And now, on June 10, 2026, a $350 million Series B co-led by Magnetar Capital and AMD Ventures, with continued participation from Maverick Silicon, Nexus Venture Partners, and Western Frontier.
Total funding since founding: approximately $493 million. Post-money valuation: $1.55 billion. That is nearly four times the company’s value from just a year ago.
The notable part is not just the size of the round. It is who is writing the checks. AMD is both TensorWave’s chip supplier and now a lead investor, using its own balance sheet to fund a company that serves as a high-profile proof point that AMD silicon can compete at scale in production AI. That is a strategic relationship, not just a financial one. AMD needs TensorWave to succeed almost as much as TensorWave needs AMD.
The new capital goes toward expanding data center capacity globally and deploying next-generation AMD Instinct MI355X GPU clusters built for memory-intensive work: large language model training, high-throughput inference, and generative AI at production scale.
How TensorWave Competes with NVIDIA Without Using a Single NVIDIA Chip
This is the question everyone asks. And the answer sits at two levels: hardware and software.
On hardware, TensorWave has already deployed a cluster of 8,192 AMD Instinct MI325X GPUs, the largest all-AMD AI training cluster in North America. That cluster produces over 2 petabytes per second of aggregate memory bandwidth and an estimated 21 exaFLOPS of FP8 throughput. The MI325X features 256GB of HBM3E memory per chip. And the incoming MI355X brings 2,300GB of VRAM and 8,000GB per second of memory bandwidth, making it a genuine competitor to NVIDIA’s Blackwell generation on memory-heavy tasks.
But hardware alone is not the whole story.
The real friction with AMD historically was software. NVIDIA’s CUDA platform has a years-long head start and a massive developer ecosystem. AMD’s ROCm, the open-source alternative, was considered buggy and hard to work with. TensorWave spent the last two years working directly alongside AMD to fix that. Horton has said publicly that ROCm is now “pretty much plug-and-play” for most production workloads, including PyTorch and Hugging Face pipelines.
And then there is supply. NVIDIA chips have been brutally difficult to get. Long wait times. Premium pricing. Take it or leave it. TensorWave offers something that NVIDIA-based clouds structurally cannot: actual availability, with no vendor lock-in and no dependency on a single company’s production schedule.
That combination, capable hardware plus improving software plus real supply access, is what makes TensorWave’s position real rather than theoretical.
TensorWave’s Data Centers: Where Are They Located and How Big Are They?
Right now, TensorWave operates three AI data centers in the United States: Pennsylvania, Arizona, and Florida. Each facility holds approximately 10,000 AMD Instinct GPUs and delivers computing capacity equivalent to around 14 megawatts of electrical power.
The model is intentional. TensorWave leases physical data center space from developers and fills those shells with its own AMD-designed hardware, liquid cooling infrastructure, and networking equipment. It is capital-efficient and fast to deploy.
But it is scaling quickly. In January 2026, TensorWave signed an additional 20-megawatt capacity agreement with TecFusions for its Pennsylvania and Arizona sites. The company has now secured more than two gigawatts of long-term data center capacity in total. That is not a vanity number. In this market, power commitments, grid access, and cooling capacity are what actually determine whether a provider can deliver at scale. Chips generate headlines. Power determines what is actually possible.
In February 2026, TensorWave also partnered with Credo, a networking technology firm, specifically to improve network reliability across its large-scale GPU clusters. The company currently employs around 117 people and has plans to expand across engineering, infrastructure, operations, sales, and customer success.
What Can You Run on TensorWave? AI Training, LLMs, and Inference Explained
Short answer: if your workload needs serious memory, TensorWave is worth looking at.
AI training is the process of building a model by exposing it to enormous datasets over many iterations. TensorWave’s AMD MI325X and MI355X GPUs are built for this. The MI300X carries 192GB of HBM3 memory per chip. The MI355X goes further, with 2,300GB of VRAM, meaning teams running 70-billion-parameter-plus models can often fit them onto far fewer GPUs than competing setups require. Fewer GPUs means lower cost and less parallelism complexity.
AI inference is different. It is the process of running a trained model to answer questions, generate images, or write text. Every interaction with a production AI product is inference. AMD GPUs have been specifically praised for efficiency here. TensorWave’s managed inference service supports autoscaling, burst capacity, and batch processing for exactly these workloads.
The platform supports PyTorch, TensorFlow, JAX, DeepSpeed, and Hugging Face Transformers, all on ROCm. The MI355X has been validated for frontier open-weight models including DeepSeek V3.2, Kimi K2.6, and GLM 5.1 at long context lengths with high throughput.
The reality is: if you are running large models and cannot afford memory bottlenecks, TensorWave was built for your specific problem.
Who Uses TensorWave? Real Companies Running AI on AMD Infrastructure
The question that separates real infrastructure from promising infrastructure is always the same: who actually uses it?
TensorWave has real answers here. Fireworks AI, a platform for deploying open-source models at production scale, is running large-scale generative AI workloads on TensorWave’s AMD infrastructure. Luma AI, known for its video generation technology, is using TensorWave for production inference systems.
And then there is Zyphra. In May 2026, Zyphra launched Zyphra Cloud, a full-stack AI platform built entirely on AMD Instinct MI355X GPUs running on TensorWave’s infrastructure. Zyphra Cloud launched with Zyphra Inference, a serverless inference service for frontier open-weight models including DeepSeek V3.2, Kimi K2.6, and GLM 5.1. It is built for long-horizon agentic workloads and was designed, from the start, on AMD rather than NVIDIA hardware.
The pattern is clear. The companies building the next generation of AI products are not waiting for NVIDIA capacity. They are finding TensorWave, running their workloads, and staying. Customer references in this market carry more weight than benchmarks. Fireworks, Luma, and Zyphra are not small experiments. They are production systems.
Is TensorWave the Future of AI Cloud Computing?
Here is what the numbers actually say. Since its founding in December 2023, TensorWave has grown its overall capacity tenfold, every single year. Horton has said publicly that he intends to do it again. That kind of growth does not happen in a market where customers are dissatisfied.
But let’s be real about what TensorWave is and is not. It is not trying to be AWS. It is not a general-purpose cloud. It is a specialized AI infrastructure company, purpose-built for organizations that need massive memory capacity and cannot afford to wait for NVIDIA’s supply chain.
The broader market trend is moving in TensorWave’s direction. Enterprise demand for AI compute is accelerating faster than any single chip manufacturer can satisfy. NVIDIA’s pricing remains elevated. Its supply constraints are real. And its software ecosystem, while mature, creates the kind of dependency that makes CTOs nervous. TensorWave is built around open-source software, a chip supplier actively investing in the platform, and a growing roster of production customers who have already made the switch.
The company has $493 million in total funding, a $1.55 billion valuation, more than two gigawatts of secured data center capacity, and a 28-year-old CEO who left nuclear fusion research at Lockheed Martin because he thought the AI compute market was a bigger problem to solve.
The market needed a genuine alternative. TensorWave decided to build one.
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.