The Sovereign Compute Stack: Why Forge.io is Betting on Open Hardware

The Tax You Didn't Know You Were Paying

Every organisation running workloads on AWS, Azure, or GCP pays three taxes they rarely see on an invoice.

The first is the complexity tax. What should be simple—deploy an application, store some data, run inference—requires navigating hundreds of services, each with its own pricing model, security configuration, and compliance posture. Your engineers spend more time understanding the cloud than building on it.

The second is the sovereignty tax. Your data lives in infrastructure governed by foreign laws, operated by companies whose interests may not align with yours, and shared with millions of other tenants. For healthcare organisations, this isn't a philosophical concern—it's a compliance liability measured in audit findings and regulatory risk.

The third is the dependency tax. The hyperscalers have made themselves indispensable by design. Migration is painful. Multi-cloud is a myth sold by consultants. And every year, you're more locked in than the last.

For consumer applications and startups optimising for speed, these taxes are often worth paying. For healthcare—where data sovereignty is a regulatory requirement, not a preference—they're becoming untenable.

The Cracks in the Foundation

The hyperscaler model was built on three assumptions that are now being challenged.

Assumption one: General-purpose compute is sufficient. For a decade, the answer to every workload was "more GPUs" or "bigger instances." But AI inference at scale, real-time medical imaging, and edge diagnostics don't fit neatly into the hyperscaler's menu of instance types. The workloads are outgrowing the infrastructure.

Assumption two: Software abstraction can solve hardware problems. The hyperscalers responded to security concerns with more software—more firewalls, more IAM policies, more compliance checkboxes. But software controls can be misconfigured. They can be bypassed. They create the illusion of isolation without the physics of it.

Assumption three: Scale equals efficiency. The hyperscalers achieved remarkable economies of scale, then passed almost none of those savings to customers. Margins expanded while prices stayed high. The efficiency gains went to shareholders, not users.

A new generation of hardware companies is challenging each of these assumptions—not with incremental improvements, but with fundamental rearchitecture.

The Hardware Spectrum

Forge.io is hardware-agnostic by design. We believe the control plane should be independent of the silicon beneath it. This means our platform runs equally well on enterprise-standard infrastructure and frontier compute architectures.

Enterprise Standard: The Proven Foundation

Not every workload needs exotic silicon. Most healthcare applications—EMRs, patient portals, clinical workflows—run perfectly well on enterprise hardware.

Dell PowerEdge, HPE ProLiant, Supermicro—these are the workhorses of healthcare IT. Billions of patient records already live on this infrastructure. The hardware isn't the problem. The control plane is the problem.

Forge.io brings sovereign orchestration to traditional hardware. The same IsoCell isolation, the same ForgeMesh networking, the same compliance-by-architecture guarantees—running on the servers your team already knows how to operate.

For organisations with existing data centre investments, this is the path of least resistance: upgrade your control plane without replacing your hardware.

Frontier Compute: The Post-Hyperscaler Era

For organisations ready to move beyond traditional architectures, a new generation of hardware is emerging. These aren't incremental improvements—they're fundamental rethinks of how compute should work.

Tenstorrent: The Case for Open AI Silicon

NVIDIA's dominance in AI compute is so complete it's almost invisible. Every major cloud provider, every AI startup, every research lab runs on CUDA. This isn't because NVIDIA's architecture is optimal for every workload—it's because the ecosystem lock-in is total.

Tenstorrent, founded by legendary chip architect Jim Keller, is building AI accelerators on RISC-V—the open-source instruction set that's doing to processors what Linux did to operating systems. Their Wormhole architecture isn't just an alternative to NVIDIA; it's a bet that AI inference doesn't need the complexity (and cost) of training-optimised GPUs.

For healthcare AI—where models are typically deployed for inference rather than trained from scratch—this matters. Diagnostic AI, clinical decision support, and imaging analysis don't need the full weight of an H100. They need efficient, predictable, cost-effective inference at scale.

We're optimising Forge.io's kernel layer for Tenstorrent's architecture because we believe regulated industries shouldn't be forced to pay the NVIDIA tax for workloads that don't require it.

Cerebras: When Moore's Law Isn't Enough

Conventional chip design accepts a constraint so fundamental it's rarely questioned: chips must be small enough to fit on a silicon wafer, then networked together for large workloads. Cerebras rejected this premise entirely.

Their wafer-scale engine—a single chip the size of a dinner plate—eliminates the interconnect bottlenecks that plague distributed training and inference. The CS-3 contains 4 trillion transistors, 900,000 cores, and 44GB of on-chip memory. It's not a faster chip; it's a different category of compute.

For healthcare organisations with sovereign requirements, Cerebras systems represent something rare: hyperscaler-class AI capability that can be deployed in controlled environments. The challenge is providing enterprise-grade access controls, audit logging, and compliance monitoring for hardware this powerful.

Forge.io is building the control plane to make Cerebras systems accessible to regulated industries—sovereign AI compute with the governance healthcare requires.

Oxide Computer: The Cloud You Can Own

The most radical challenge to the hyperscaler model isn't faster chips—it's the argument that you shouldn't need a hyperscaler at all.

Oxide Computer, founded by Bryan Cantrill (former Sun Microsystems CTO) and a team of infrastructure veterans, has built what they call "the first commercial cloud computer." It's a complete rack-scale system—compute, storage, networking, and control plane—delivered as a single integrated product. No three-year AWS contract. No data residency ambiguity. No shared infrastructure with unknown tenants.

For healthcare organisations with strict data sovereignty requirements—Australian patient data that cannot leave Australian jurisdiction, European health records subject to GDPR—Oxide represents infrastructure that's sovereign by construction, not configuration.

We're exploring integration between Forge.io's orchestration layer and Oxide's Sidecar interface to bring our compliance-native control plane to on-premise deployments. Same developer experience as Forge cloud. Same audit trail. Same sovereignty guarantees. Your rack.

The Sovereign Compute Thesis

These companies—from Dell and HPE to Tenstorrent, Cerebras, and Oxide—represent a spectrum of choice that the hyperscalers don't offer. Traditional or frontier. Cloud or on-premise. Leased or owned.

What they share is a common need: a control plane designed for sovereignty, not bolted on as an afterthought.

Forge.io's thesis is simple: the infrastructure layer and the control plane layer should evolve together, but they shouldn't be welded together. Open hardware architectures deserve control planes designed for them—not retrofitted enterprise software or hyperscaler consoles awkwardly adapted to on-premise deployment.

This is what we're building.

A control plane that treats compliance as architecture, not afterthought. That enforces data sovereignty through physics—hardware isolation, region locking, air-gapped deployments—not just policy. That gives healthcare organisations access to the full spectrum of compute without forcing them to accept hyperscaler terms.

What This Means for Healthcare

The convergence of hardware choice and sovereign infrastructure creates possibilities that didn't exist five years ago.

Run on what you have. Existing Dell and HPE infrastructure, upgraded with Forge.io's control plane. No forklift replacement required.

Diagnostic AI on open silicon. Radiology AI, pathology analysis, and clinical decision support running on Tenstorrent accelerators—without NVIDIA licensing complexity or CUDA lock-in.

Sovereign foundation models. Large language models fine-tuned on clinical data, running on Cerebras systems in controlled environments, with full audit trails and access controls.

True hybrid architecture. Burst to Forge cloud for development and testing, deploy to Oxide racks in your own data centre for production—same control plane, same compliance posture, same developer experience.

The hyperscalers won't offer this. Their business model depends on keeping you dependent. Their compliance story is checkboxes, not architecture.

We think healthcare deserves better. We're building for that future.