Beirit Harvey

At Web Summit 2025, TensorWave CEO and co-founder Darrick Horton took the stage to talk about how AI teams are frustrated with a world where NVIDIA has all the leverage.

Breaking NVIDIA’s Grip: AMD AI Infrastructure at Web Summit 2025

At Web Summit 2025, TensorWave CEO and co-founder Darrick Horton took the stage to talk about something a lot of AI teams are quietly frustrated with: a world where NVIDIA has all the leverage.

In a session titled “Breaking NVIDIA’s Grip: AI Infrastructure with AMD GPUs,” Horton dug into how openness, hardware diversity, and a new generation of “NeoClouds” are reshaping the AI infrastructure landscape and why TensorWave went all-in on AMD GPUs.

Below is a recap of the key themes from the conve

At Web Summit 2025, TensorWave CEO and co-founder Darrick Horton took the stage to talk about something a lot of AI teams are quietly frustrated with: a world where NVIDIA has all the leverage.

In a session titled “Breaking NVIDIA’s Grip: AI Infrastructure with AMD GPUs,” Horton dug into how openness, hardware diversity, and a new generation of “NeoClouds” are reshaping the AI infrastructure landscape and why TensorWave went all-in on AMD GPUs.

Below is a recap of the key themes from the conversation.


From Data Centers to Dedicated AMD AI Cloud

Horton didn’t come into AI from the model side; he came from the racks.

He’s spent nearly a decade building and operating data centers, long before “AI infrastructure” was a phrase everyone used. So when the demand spike hit after the rise of ChatGPT, his team was already in position.

“We’d been building data centers since before AI was super popular… so when AI infrastructure started to become in high demand, we were in a good position to tackle that market.”

But the decision to found TensorWave wasn’t just about timing. It was about correcting an unhealthy dynamic: a near-monopoly at the GPU layer.

Horton notes, “... We saw huge demand from end customers for additional capacity, additional compute, and they weren't satisfied with only having a single option. Right? People like NVIDIA, They make a good product, but nobody likes to be forced to purchase from only one vendor. And so we saw that as an opportunity, a problem that we could solve in the space to bring a viable alternative to the market so that end users actually had choice.”

TensorWave’s mission grew out of that reality: make high-end AI compute more open, more accessible, and not controlled by a single company.


Betting Early on AMD as the Credible Challenger

When TensorWave launched, the GPU landscape looked very different. If you wanted AI compute, you went to NVIDIA. Full stop.

Other players existed on paper, but not in production:

 * A few startups were working on chips, but couldn’t get traction at scale.
 * Intel had an AI product roadmap that ultimately flopped.
 * AMD and others were talking about future products, but weren’t shipping yet.

Despite that, TensorWave dug into where the market was actually headed.

“What we determined early on was that AMD was going to be the only vendor that could credibly challenge NVIDIA in the GPU marketplace.”

Fast-forward to today, and that thesis is bearing out. AMD’s AI roadmap has accelerated, hardware has become more competitive each generation, and the next wave of AMD GPUs is expected to meet or beat NVIDIA on key performance metrics.

That moment, when AMD not only matches but outperforms NVIDIA, is what Horton called a “turning point” for the industry, where choice at the hardware layer becomes real, not theoretical.


TensorWave’s Competitive Advantage Against Hyperscalers

The moderator raised a question many people have: if AWS, Google, and Microsoft are so massive, why do specialist GPU clouds like TensorWave exist at all?

Horton’s answer boiled down to one word:

“The main answer is speed.”

Hyperscalers have historically planned data center builds on 3–5 year timelines. That doesn’t map to the reality of today’s AI demand:

 * Data centers today must be larger than data centers from just a few years ago.
 * They often need to be deployed in 3–6 months, not in several years.

That requires different muscles:

 * Less bureaucracy and fewer internal dependencies.
 * Teams focused purely on AI cluster design, deployment, and operation.
 * Willingness to optimize specifically around GPUs, power, cooling, and interconnects.

“Neoclouds have the capability to deploy much much faster because Neoclouds in general are focused on just that… doing AI cluster building and operation at scale and nothing else.”

TensorWave narrowed that focus even further: be the best in the world at deploying and operating AMD GPU clusters.

According to Horton, that depth of specialization is why hyperscalers, who have to spread attention across hundreds of services, can’t easily copy what NeoClouds are doing in this niche.


Building the Full Stack: Firmware, Orchestration, and Cooling

Buying GPUs is the easy part. Making them reliable and easy to consume at scale is where things get hard.

Most of TensorWave’s customers don’t want to think about infrastructure at all. They want a cluster that just works; so they can focus on models, data, and product.

That’s why TensorWave invests heavily across every layer:

 * Data center site selection & power delivery.
 * Cooling engineering (including advanced solutions to support dense GPU racks).
 * Firmware, orchestration, and cluster management.

“Our role is essentially to build and provide reliable infrastructure, then make it as easy as possible to consume that infrastructure.”

The larger the cluster, the more non-trivial this becomes. At scale, small mistakes in power planning, thermals, or orchestration can turn into massive reliability problems. TensorWave’s goal is to absorb that complexity so customers don’t have to.


Scaling to the Gigawatt Era

One of the most striking parts of the session was the discussion around scale.

TensorWave expects to have over a gigawatt of data center capacity under contract, and Horton was blunt about why that matters:

Early in the AI boom, doing something meaningful with 10–20 MW of power was possible. Today, that’s no longer true. The new floor for “serious” AI deployments is 100 MW and up, with the industry rapidly moving into the gigawatt data center era.

Horton emphasizes, “in this industry power and scale are the most important thing right, especially moving forward into 2026 and 2027.”

As clusters grow, the constraints are less about the willingness to buy GPUs and more about finding enough power in the right locations.


Opensource, AMD, and an Easier Path Off NVIDIA Lock-In

Breaking NVIDIA’s grip isn’t just about offering another chip. It’s about making it practical to move workloads.

That’s where open ecosystems and open source come in.

“We’re trying to make AI and AI infrastructure as accessible as possible… It’s important to not only focus on the hardware, but also support the ecosystem overall, supporting open source software and development projects.”

For customers, this looks like:

 * Choosing frameworks and tooling that support multiple hardware backends.
 * Avoiding proprietary glue that locks them to one vendor.
 * Using open-source components so they don’t have to “reinvent the wheel” for each new model or hardware type.

When the ecosystem is open and portable, switching from NVIDIA to AMD (or running both) becomes much easier, which is exactly what TensorWave wants to enable.

“Having a unified ecosystem, the open source ecosystem makes that more accessible and makes the market ultimately more competitive.”


Sovereign AI: Why Multi-Vendor Matters Even More for Countries

The conversation then shifted from startups and enterprises to governments.

“Sovereign AI” has become a buzzword, but underneath the hype are two real drivers:

 1. Data sovereignty – keeping sensitive data and models within national borders.
 2. Infrastructure ownership – countries building and operating their own AI infrastructure.

Many nations are now:

 * Planning or deploying in-country AI data centers
 * Training their own models for internal and external use
 * Setting policies around where data can live and how it can be processed

Horton’s key point: sovereign AI is incomplete if it still depends on a single hardware vendor.

“Countries don’t like to give all of their money to one vendor any more than large companies do, they probably like it even less. That’s why having a multi-vendor strategy, and an ecosystem that can support multiple different hardware types, is so important for sovereign AI.”

Multi-vendor strategies and open ecosystems are even more important at the national level.


The Looming Power Crunch

Looking 2–3 years ahead, Horton sees two big forces shaping AI infrastructure:

 1. More support and optimization for AMD GPUs

 * More frameworks and tools will treat AMD as a first-class target.
 * Optimization work will continue to close and then invert the gap with NVIDIA.

 1. A global power constraint

 * The US, in his view, has roughly 18 months of “easy” data center power left.
 * After that, building gigawatt-scale infrastructure in the US becomes extremely challenging until new capacity comes online around 2030.

Horton notes, “There will be this natural demand for large-scale power that’s forcing all providers, ourselves included, to really evaluate other countries and other geos.”

That’s part of why regions like Europe, including countries like Portugal, are becoming more interesting as locations for future AI infrastructure builds.


Advice for AI Startups: Don’t Get Locked In

The session ended with a simple piece of advice for AI startups trying to navigate this shifting hardware landscape:

“Be multi-vended. Don’t get locked into one ecosystem or another.”

Practically, that means:

 * Designing your stack so you can run on more than one GPU vendor
 * Embracing open-source frameworks and tools that support that flexibility
 * Partnering with infrastructure providers who are aligned with openness, not lock-in

It’s better for resilience, better for pricing power, and better for long-term optionality.

If you’re building serious AI workloads and want to explore what an AMD-powered, open, and scalable GPU cloud can look like in practice, this Web Summit conversation made one thing clear: The era of “NVIDIA or nothing” is ending and companies like TensorWave are working hard to make sure the next phase of AI infrastructure is faster, more open, and a lot less monopolized.


About TensorWave

TensorWave is the AI neocloud purpose-built for performance. Powered exclusively by AMD Instinct™ Series GPUs, we deliver high-bandwidth, memory optimized infrastructure that scales with your most demanding models–training or inference. 

Ready to get started? Connect with a sales engineer. 











AMD Instinct™ MI355X GPUs Now Available

Breaking NVIDIA’s Grip: AMD AI Infrastructure at Web Summit 2025

Subscribe to our Blog

Stay ahead of the curve with the latest in AI, AMD accelerators, and all things TensorWave.

Product

Resources

Company

© 2025 TensorWave Inc. - All rights reserved.