Beirit Harvey

At a recent on-stage discussion, our CEO, Darrick Horton, joined leaders from Vultr and Crusoe, to share how AI infrastructure is shifting away from the GPU allocation race toward building efficient, production-ready AI factories at scale.

Not Bigger AI Factories, but More Efficient Ones: What Three NeoClouds Agree On

At a recent on-stage discussion, our CEO, Darrick Horton, joined leaders from Vultr and Crusoe, to share how AI infrastructure is shifting away from the GPU allocation race toward building efficient, production-ready AI factories at scale.

A year and a half ago, success meant securing hardware. Today, it is about optimizing power, networking, software, and operations to deliver real workloads reliably.

This blog breaks down the main takeaways of this panel.


Sustainability Is Becoming a Core 

At a recent on-stage discussion, our CEO, Darrick Horton, joined leaders from Vultr and Crusoe, to share how AI infrastructure is shifting away from the GPU allocation race toward building efficient, production-ready AI factories at scale.

A year and a half ago, success meant securing hardware. Today, it is about optimizing power, networking, software, and operations to deliver real workloads reliably.

This blog breaks down the main takeaways of this panel.


Sustainability Is Becoming a Core Design Constraint

Kyle Sosnowski, VP of Engineering at Crusoe, emphasized building data centers around available energy, using a mix of hydro, wind, and solar, along with modular deployments near stranded energy sources. This approach supports high-density compute while staying closely tied to power availability.

The message was clear: energy strategy now shapes infrastructure design.


Performance Comes From Co-Design

Horton highlighted the value of working closely with customers to tune infrastructure around specific workloads. While training, inference, and fine-tuning share similarities, real gains come from optimizing network design and cluster architecture for each use case.

Performance today is less about a single GPU and more about how the entire system works together.


Global Scale Depends On Repeatability

Duncan Ng, VP Solutions Engineering at Vultr, stressed a “design once, replicate everywhere” model. By standardizing the control plane and embedding optimization across every layer, they can deliver consistent performance worldwide.

Reliability at scale comes from process and replication, not one-off builds.


Production Readiness Now Means Validation and Self-Healing

Sosnowski described rigorous cluster validation and a growing focus on self-healing infrastructure that automatically responds to hardware and operational issues.

The goal is to minimize how much time customers spend managing clusters while ensuring systems stay stable in production.


ROCm Is Maturing, But Hands-On Onboarding Still Matters

Horton noted that many customers are transitioning from NVIDIA environments to AMD-based infrastructure. Because TensorWave is AMD-exclusive that “... means we also need to be experts at helping customers get going on ROCm”. While AMD has made ROCm more seamless, successful production deployments still require close collaboration to validate workloads and optimize performance beyond benchmarks.


Multi-Cloud Works When Each Provider Has a Role

Rather than replacing hyperscalers, the panel agreed that most teams will operate in multi-cloud environments. Specialized providers excel at certain layers, while hyperscalers maintain strengths like storage.

Making this seamless comes down largely to networking and connectivity.


Speed Beats Perfect Optimization

Horton summed it up simply: customers would rather run a cluster at strong performance today than wait a year for theoretical perfection. Optimization continues after deployment, but speed to production delivers the most value.


Power Is the Next Big Constraint

All three agreed that power availability is becoming the primary bottleneck in AI infrastructure growth.

Sosnowski pointed to smarter workload orchestration around grid limits. Horton warned that the U.S. may hit power constraints before demand slows, with major relief closer to 2030. Vultr added that rising power costs will drive major efficiency gains on the software side.


Final Takeaway

The future of AI infrastructure is not about building bigger data centers.

It is about building more efficient ones that turn power into predictable, production-grade performance at scale. The next winners will be the clouds that optimize utilization, reliability, and energy, not just hardware supply.

Watch the full length discussion:









TensorWave Cloud: Engineered for AI Velocity

TensorWave is the industry’s premier AMD-exclusive cloud provider, delivering memory optimized AI infrastructure for training, fine-tuning and inference. 

Our platform combines the raw computational power of AMD's latest GPU architectures with innovative liquid cooling technology and ultra-high-speed networking, creating an optimized environment for the most demanding AI workloads. From large language model training to real-time inference, TensorWave provides the performance, scalability, and cost-efficiency that modern AI initiatives demand.

Ready to get started? Connect with a sales engineer. 

AMD Instinct™ MI355X GPUs Now Available

Not Bigger AI Factories, but More Efficient Ones: What Three NeoClouds Agree On

Subscribe to our Blog

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

Product

Resources

Company

© 2026 TensorWave Inc. - All rights reserved.