PostHeaderSection

Symbol

Global NavBar

PostContentSection

Global Footer

Blog Detail Page

Beirit Harvey

Why did TensorWave choose an AMD-exclusive strategy from day one and stay that way? Inside the leadership team’s thinking on performance, openness, and long-term strategy.


TensorWave on AMD: All-In on Red

When TensorWave launched, choosing an AMD-exclusive strategy looked like a contrarian move.

NVIDIA owned mindshare, most clouds treated AMD as a side option, and demand for GPUs far outstripped supply. But in early strategy sessions, one theme kept coming up: if AI was going to be truly accessible, the market needed a real second option.

TensorWave co-founder and CEO Darrick Horton often frames it this way: the team saw an unhealthy level of dependence on a single vendor and an opportunity to build a different kind of AI cloud around AMD.

This post distills conversations with Horton and other TensorWave leaders on why we chose AMD Instinct™ GPUs from day one, and why we’ve doubled down on that choice as we’ve scaled.


1. The Starting Point: “Where’s the Second Option?”

Before TensorWave, our founders were in the same position as many of our customers are now: trying to secure GPUs for real workloads and running into the same issues:

 * Long waitlists and chronic scarcity.
 * Opaque pricing and unpredictable costs.
 * A single dominant vendor controlling both hardware and software.

AMD was the only realistic candidate for that role; serious accelerator roadmap, a rapidly improving software stack, and a willingness to work closely with early partners. TensorWave was created to turn that potential into an actual, production-grade platform.


2. Why AMD was the Right Fit

Going AMD-only was never about ideology. It was a bet that AMD could deliver competitive hardware with a different set of strengths.


Better Unit Economics

For large training jobs and always-on inference, small percentage differences in cost per GPU-hour add up fast.

AMD’s accelerators gave us room to offer:

 * Strong performance on modern AI workloads.
 * More favorable price–performance than comparable alternatives.
 * A path to sustainable, predictable pricing for customers.

That combination is what makes high-performance AI accessible. 


Big-Memory GPUs for Modern Workloads

Most cutting-edge AI work today is memory-hungry: long-context LLMs, multi-modal models, retrieval-heavy systems, and complex fine-tuning.

AMD leaned into high-memory designs, which means customers can often:

 * Fit larger models per GPU.
 * Run longer sequences without extreme model parallelism.
 * Simplify cluster design by using fewer, more capable nodes.

In practice, this lets teams spend less time wrestling with sharding tricks and more time shipping features.


Practical Advantage: You Can Actually Get the GPUs

One of the most pragmatic reasons our leadership cites: supply.

Early on, TensorWave was able to secure meaningful allocations of AMD accelerators while much of the market was stuck on waitlists for competing hardware. For a company built around democratizing compute, being able to say “yes, we have capacity” is a big deal.


3. Why We Stayed Exclusive: Focus, Depth, and Alignment

Choosing AMD at launch is one thing. Sticking to AMD-only as you grow is a deliberate strategy.


One Hardware Family, Deep Tuning

Running a serious AI cloud is about far more than racking GPUs. It’s about:

 * Cooling and power at rack and cluster scale.
 * Networking and storage tuned for training and inference.
 * Scheduling, partitioning, and observability for multi-tenant environments.

By standardizing on AMD Instinct, we can tune that entire stack around one set of characteristics instead of spreading effort across multiple architectures.

The result: fewer edge cases, faster iteration, and a platform that gets more efficient and capable with every generation of AMD hardware.

In an AMD Tech Talk with Jim Greene, CEO Darrick Horton notes, “The hyperscalers are good at a lot of things. That’s their role. They are good at most of the activities that a general company would do in the cloud. AI is not a general thing. It is not as simple as a lot of traditional cloud workloads.” 


Building on an Open, ROCm-Based Stack

The second pillar is software.

AMD’s ROCm stack gives us an open, evolving platform for:

 * PyTorch, JAX, TensorFlow, vLLM, and other AI frameworks.
 * CUDA-interop paths for teams bringing existing code over.
 * Deep performance tuning on Instinct GPUs without locking customers into a proprietary ecosystem.

For customers, “AMD-only” translates into a cleaner story: an AI cloud built around an open stack, with a clear path away from single-vendor dependency.

“We founded TensorWave to democratize access to high-performance AI computing through open ecosystems,” says founder and CEO Darrick Horton. That commitment to openness is why TensorWave has gone all-in on AMD’s ROCm-based stack instead of a closed, single-vendor ecosystem. 

Paired with what co-founder and President Piotr Tomasik calls “deep specialization in AMD technology”, the result is an AMD-only cloud that’s optimized for next-gen AI workloads while keeping customers out of vendor lock-in.


Strategic Partnership

The third reason we’ve stayed exclusive is alignment.

AMD isn’t just a supplier; they’re a strategic partner. That means:

 * Early visibility into their roadmap.
 * Close collaboration on performance and ROCm features.
 * Confidence to design very large clusters specifically around AMD’s trajectory.

If you’re going to be opinionated about hardware, you want your hardware partner just as invested in your success as you are in theirs.


4. What AMD Exclusivity Means for Customers

From a customer’s perspective, “AMD-only” should show up as concrete benefits; not just branding.


Better Economics End-to-End

Because of the combination of hardware cost, high memory per GPU, and cluster-level efficiency, customers can often:

 * Run bigger models on fewer GPUs.
 * Reduce the number of clusters/nodes they need to manage.
 * Cut power, cooling, and management overhead for the same work.

Ultimately, that means more useful work per dollar and per watt.


A Real Path Away from Lock-In

By centering the platform on ROCm and open frameworks, teams get:

 * A realistic path to move off single-vendor stacks.
 * The ability to use mainstream open-source tools without heavy rewrites.
 * Less concentration risk around one chip vendor’s pricing and roadmap.

For CISO / CIO profiles, this is as much about resilience and control as it is about technology.


Faster Access to New AMD Generations

Because we’re all-in on AMD, we’re also a natural landing zone for new generations of Instinct GPUs and ROCm releases.

For customers, that looks like:

 * Early access to newer SKUs for training and inference.
 * Rapid support for updated ROCm features and performance improvements.
 * A platform that keeps pace with AMD’s roadmap by design.


5. MI455X and the Next Wave of AMD AI

As AMD moves from the current MI3xx family into the MI400 series, our bet on exclusivity only gets more interesting; especially with MI455X, AMD’s next flagship for large-scale AI training and inference.

MI455X-class GPUs are designed for exactly the workloads our customers care about:

 * Massive on-GPU memory – Enough HBM to keep larger models and longer context windows resident in memory, reducing the need for extreme sharding tricks and cutting communication overhead.
 * Higher bandwidth and throughput at rack scale – Architected for dense, liquid-cooled racks where GPUs stay fed with data, instead of stalling on I/O bottlenecks.
 * Scale-out AI as a first-class goal – The MI400 family is built with cloud deployments and large AI services in mind, not just single-node benchmarks.

For our internal roadmap, that translates into a few consistent themes:

 * The world needs more than one high-end AI stack. MI455X and the broader MI400 roadmap are AMD’s answer to that, and we’re aligned with it.
 * Memory capacity and efficiency will matter as much as raw FLOPs, especially for long-context, multi-modal, and retrieval-heavy systems.
 * Open, flexible software stacks like ROCm will continue to beat closed ecosystems for teams that want control and portability.

TensorWave’s answer is just as clear: stay focused on AMD, be ready for MI455X-class hardware as it becomes available, and keep turning that choice into concrete advantages; lower cost, larger models, and more open AI; for the teams building on our cloud.


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. 

Why did TensorWave choose an AMD-exclusive strategy from day one and stay that way? Inside the leadership team’s thinking on performance, openness, and long-term strategy.

TensorWave on AMD: All-In on Red

Product

Solutions

Resources

Company

Product

Resources

Solutions

Company