[Partner Insight] The Infrastructure Gap: Why AI Needs Decentralised Compute

AI teams aren't short on ideas. They're short on hardware. That's the pattern we hear from enterprises, research groups and startups every week. They have models ready to run, but the GPUs they need are tied up in long queues, delayed by power upgrades, or sitting idle in clusters they can't access. The gap between demand and available compute is starting to dictate product roadmaps, and not in ways anyone planned for.

Hyperscalers are throwing extraordinary amounts of capital at the problem. Amazon, Microsoft, Google and Meta are on track to spend roughly $650 billion on AI infrastructure in 2026. Yet the queues for GPUs and power connections haven't cleared. A data-centre deal can be signed in weeks; the physical build takes far longer. This is the environment AI teams now operate in: ambitious plans on one side, and compute that arrives on its own timeline.

Why demand looks different now

For much of the past decade, the biggest bursts of AI infrastructure demand came from training models. Teams planned their runs, booked capacity and worked inside predictable windows. The load was heavy, but it was contained.

Running those models inside products creates a different pattern. Once a model sits behind a chatbot, recommendation engine or image generator, usage follows people rather than schedules. Requests arrive across time zones and markets, rising and falling throughout the day as users interact with applications.

That demand rarely grows in a straight line. A feature that gains traction in one market can push infrastructure in another. Teams end up managing bursts of usage that move across regions rather than staying inside a single data‑centre window.

When infrastructure can't follow demand

Infrastructure moves on longer timelines than software. GPUs take time to manufacture, large facilities take years to build and power upgrades often become the slowest step in the process.

Grid connections, cooling systems and electrical equipment all have their own lead times. In some regions, the transformers required for large data‑centre substations can take one to three years to deliver. Those timelines sit well outside the control of the teams trying to deploy new capacity.

At the same time, AI demand rarely stays in one place. A product launch in one market can drive traffic in another. Enterprise deployments spread across regions as new teams adopt the same tools.

Large, centralised infrastructure struggles to adjust quickly when usage moves this way. Capacity can be added over time, but it cannot be repositioned overnight.

Where unused compute already exists

The shortage of AI compute is often described as a lack of hardware. In many cases, the hardware already exists.

Across enterprises, universities and regional data centres, GPUs are frequently provisioned for peak workloads that only appear occasionally. Clusters built for internal research may run at full capacity during training cycles and sit far quieter the rest of the time. Infrastructure designed for one application rarely moves easily to another.

For teams trying to deploy new systems, that capacity might as well not exist. Procurement models, security boundaries and long-term contracts keep most of it tied to the organisations that installed it.

The result is a market where supply and demand sit close together without connecting. Developers wait months for access to compute while GPUs elsewhere run well below full utilisation. This is where decentralised physical infrastructure networks (DePIN) change the equation. Rather than waiting for new capacity to be built, DePIN uses token incentives to bring idle compute online ahead of demand - turning latent supply into accessible infrastructure.

How DePIN stitches distributed capacity together

Traditional cloud platforms concentrate hardware in a few locations and charge accordingly. DePIN networks coordinate machines that already exist - across enterprises, regional data centres and independent operators - and make them available to developers at a fraction of the cost. Operators are compensated with tokens for contributing capacity, creating an economic layer that surfaces supply faster than construction timelines allow.

The cost difference is significant. Teams using decentralised compute report savings of up to 70% compared to traditional cloud rates, largely because the infrastructure doesn't need to be built from scratch - it's already deployed and underutilised.

Workloads can then be scheduled across that distributed supply rather than waiting for capacity to appear in a single location. For developers and infrastructure teams, the experience resembles requesting compute from a cloud provider. The difference is that the underlying hardware may come from many participants rather than a single facility.

This model treats distributed infrastructure as a coordination challenge rather than a construction project. Instead of waiting for new facilities to come online, existing machines can be brought into a shared pool of compute.

The gaps that still need solving

Distributed computing has been tried before, and early versions struggled with reliability. Coordinating large numbers of independent machines introduced variability that enterprises found difficult to absorb. Uptime, predictable performance and clear verification of results all became barriers to adoption.

Those challenges haven't disappeared. Networks that rely on machines they don't own still need to prove that workloads run reliably and that outputs can be trusted. Interoperability between different systems also remains unresolved as new networks appear. New tools are emerging to address uptime guarantees, cryptographic verification of compute and cross-network orchestration, but they haven't been proven at enterprise scale yet.

What has changed is the pressure on infrastructure. AI workloads are expanding faster than new capacity can be built, which is pushing more teams to experiment with ways of organising compute that were previously considered too difficult to manage.

The supply-side answer to a demand problem hyperscalers can't solve alone

AI infrastructure is drawing supply from far more places than the major cloud platforms alone. Enterprises, regional data centres, universities and independent operators all hold pieces of the hardware needed to run modern AI systems.

The question is how easily that hardware can be reached. Many of these machines were installed for internal workloads and rarely appear in the wider compute market. When DePIN networks coordinate that capacity and make it accessible to developers, a much larger pool of infrastructure becomes available - not years from now, but as soon as the incentives align.

Cloud providers will remain central to the ecosystem. What changes is that AI workloads are no longer limited to the hardware inside a handful of platforms. DePIN is the supply-side answer to a demand problem that centralised infrastructure cannot solve on its own. Compute begins to come from many more locations, and from many more operators than the hyperscale cloud market alone can provide - and the data is starting to back that up.