Ben

During $IREN ’s earnings presentation, I was surprised to hear @danroberts0101 say, “Bare metal is better. The vast majority of GenAI workloads don’t need virtualization. They need direct access to resources to run training, inference, and experimentation with maximum performance and low latency.” Most experts on X argue that the total addressable market for bare metal is small, and the internet is never wrong about this sort of stuff. Accordingly, I had to ask myself: was Dan just talking his own book? Not really, because the above quote isn't even from IREN; it's from $CRWV. So let's talk about this, starting with an overview of bare metal. ----------------------- Bare metal servers are physical machines that provide direct access to hardware without a virtualization layer. By eliminating the hypervisor, tenants gain predictable performance, better GPU utilization, and full control over the operating system and software stack. This makes bare metal ideal for workloads that demand consistency and efficiency, particularly GPU-intensive AI training. While traditional cloud platforms emphasize virtualization for flexibility and multi-tenancy, certain clients—often hyperscalers, HPC labs, or advanced enterprises—demand bare metal because they already have their own orchestration layers like Kubernetes and want maximum performance without additional abstraction. For IREN, pursuing bare metal positions the company in the highest-value segment of the AI cloud market. By offering infrastructure at the hardware level, IREN can appeal directly to sophisticated buyers who don’t need managed services—just guaranteed access to compute at scale. This approach plays to IREN’s strengths: its experience in procuring power, optimizing data centers, and deploying large fleets of hardware quickly from its Bitcoin mining roots. Unlike a virtualized cloud that requires heavy investment in proprietary software stacks (which IREN admittedly lacks), bare metal can be delivered with relatively little engineering overhead, while still commanding premium demand. Strategically, bare metal also supports IREN’s financing model. Long-term contracts with creditworthy hyperscalers and enterprises provide stable utilization and predictable cash flows. This in turn allows IREN to raise asset-backed financing for GPUs and data centers, fueling its growth flywheel. And because IREN doesn’t bear the expense of building a large managed-services layer, it can undercut peers like CoreWeave or Nebius while still achieving attractive returns. The result is an “Uber/Amazon” style pricing strategy that enables rapid market share capture at a time when GPU demand far outstrips supply. In this context, IREN’s close relationship with Fluidstack is revealing. Fluidstack is an AI-focused GPU cloud provider that specializes in orchestrating large-scale bare metal clusters. Its technical expertise lies in cluster setup: tuning InfiniBand or Ethernet fabrics, optimizing NCCL and MPI libraries, integrating high-performance storage, and managing scheduling and recovery across thousands of nodes. The relationship between IREN and Fluidstack probably is as follows: IREN is highly capable when it comes to data center build-out. They know how to procure power, install substations and switchgear, design cooling systems, and maximize rack density. They handle GPU burn-in, racking, and cabling, and make sure the physical environment can support high-density compute. What they do not specialize in, however, is distributed AI cluster tuning. Their expertise lies in infrastructure engineering, not in the HPC software stacks that are essential for turning thousands of GPUs into a usable training cluster. This is where Fluidstack’s mastery comes in. Fluidstack has built Atlas OS specifically for managing multi-thousand GPU clusters. They know how to configure Kubernetes and SLURM at scale for AI workloads and have extensive experience in benchmarking and performance tuning to achieve more than 95 percent scaling efficiency. They integrate advanced storage systems such as DDN and VAST, use NVMe burst buffers, and deploy custom networking stacks to ensure smooth data flow during training. Their track record with companies like Mistral, Character.AI, Aleph Alpha, and Black Forest Labs demonstrates that they have already set up and optimized these kinds of massive clusters for real-world clients. In practice, IREN racks the GPUs and provides the high-bandwidth networking that ties them together, while Fluidstack turns that raw hardware into a coherent AI cluster capable of scaling training across thousands of nodes. This is why sovereign AI programs, startups, and enterprises rely on Fluidstack: they don’t want to invest the time and expertise required to make 4,000 GPUs behave like a single machine. Fluidstack has the tools and experience to do it for them. The key division of labor, then, is clear. IREN is the infrastructure builder, focused on physical capacity and efficiency. Fluidstack is the cluster builder, stitching that infrastructure into usable, high-performance compute environments with Atlas OS, orchestration, tuning, and reliability. In this technical sense, Fluidstack is the master of large-scale cluster setup. Nonetheless, the partnership with Fluidstack gives IREN more than just a reliable demand channel; it provides access to world-class expertise in cluster setup and orchestration. By working closely together with Fluidstack, IREN benefits from direct knowledge transfer—learning how InfiniBand networks are tuned, how NCCL and MPI libraries are optimized, and how storage and schedulers are configured for distributed AI training at scale. This collaboration effectively accelerates IREN’s climb up the value chain. In the short term, Fluidstack ensures utilization and cash flow, making IREN’s financing flywheel spin faster. Over time, IREN absorbs the operational know-how that will allow it to eventually handle more of the orchestration and customer-facing layers itself. The result is a symbiotic relationship: Fluidstack gains access to high-density infrastructure, and IREN not only monetizes its GPUs but also builds the technical competence needed to compete directly in the AI cloud market. ----------------------- By partnering with leading AI companies like Fluidstack (and Together.ai), IREN is steadily enhancing its credibility and accelerating its evolution from a pure infrastructure builder into a sophisticated AI cloud provider. These collaborations validate IREN’s ability to host some of the most demanding workloads in the industry, while also exposing it to best-in-class practices in cluster orchestration, software integration, and customer delivery. The result is not only immediate utilization of its GPU fleets but also a transfer of knowledge and competencies that strengthens its long-term positioning, enabling IREN to compete more effectively with established CSPs and reinforcing its reputation as a credible partner in the global AI ecosystem.