AI Compute: The Cloud Chokepoint

Modern AI runs on compute. Training large foundation models requires huge clusters of specialised processors working for long periods. Running those models for millions of users also requires inference capacity: the compute needed every time a person asks a model to write, search, classify, summarise, generate code, analyse an image, or make a recommendation. Training gets attention because it is dramatic. Inference is the daily operating cost.

This creates a structural advantage for organisations with capital, cloud access, chip supply, engineering talent, and data-centre capacity. A small research group can experiment. A startup can fine-tune or build on existing models. But frontier-scale training and mass deployment require infrastructure that most organisations cannot buy outright. They rent it, partner for it, or become dependent on providers that already operate at hyperscale.

The hardware layer is where the AI compute cloud chokepoint becomes visible. Advanced AI depends heavily on accelerator chips, especially GPUs and related AI processors. These chips are not ordinary consumer components. They are specialised, expensive, supply-constrained, and tied to complex manufacturing, packaging, memory, networking, and software ecosystems.

NVIDIA has become the central company in this layer because its chips, systems, software stack, and developer ecosystem are deeply embedded in AI training and deployment. CUDA, networking, server systems, libraries, and data-centre platforms create more than a chip sale. They create an operating environment. Once institutions build workflows around that stack, switching becomes difficult.

The GPU layer also shows why AI is geopolitical. Advanced chips rely on global supply chains involving design firms, foundries, packaging capacity, memory suppliers, export controls, and manufacturing equipment. A disruption in any part of that system can affect model builders, cloud providers, defence contractors, research labs, and national AI plans.

Cloud companies do not need to own every model to shape the market. They can provide the infrastructure underneath many models. They can bundle AI into productivity suites, developer platforms, government contracts, cybersecurity products, analytics systems, and enterprise workflows. They can also impose usage policies, compliance terms, rate limits, security requirements, and commercial restrictions.

This is why cloud providers may become more important than many AI startups. A startup with a model still needs compute. A public agency buying an AI tool often buys it through an enterprise cloud channel. A defence contractor deploying AI needs secure infrastructure. A hospital, bank, insurer, or school district may not want a raw model; it wants a supported cloud service with legal, security, and compliance packaging.

AI is becoming an energy story. Training and serving advanced models requires electricity at a scale that links AI growth to grid planning, generation capacity, transmission lines, power purchase agreements, and energy politics. The more AI becomes embedded into search, office tools, coding, healthcare, finance, defence, education, and government administration, the more inference demand becomes a permanent load rather than a one-off training expense.

This is why major technology companies are pursuing long-term power arrangements, including renewable contracts, nuclear discussions, and direct energy partnerships. The point is not branding. The point is capacity. Compute growth needs power certainty. A company that can secure energy at scale can support AI expansion. A competitor without that access may be limited even if it has talent and models.

Compute creates power because it determines possibility. Without compute, a model cannot be trained. Without serving capacity, a model cannot reach users. Without cloud contracts, a startup cannot scale. Without energy, data centres cannot expand. Without chips, national AI strategies become paper ambitions. In that sense, compute is not merely a technical input. It is the physical condition of AI capability.

This also changes governance. A government can regulate model behaviour, but cloud providers may enforce access faster. A standards body can write risk frameworks, but hyperscalers can decide procurement requirements. A startup can publish research, but the infrastructure owner can decide whether it receives the capacity to compete. The more AI depends on scarce compute, the more infrastructure policy becomes AI policy.

Executive Order 14110 treated advanced AI capability as a safety and security issue, including reporting requirements for certain powerful model-development activities and large-scale compute. The order was revoked in January 2025, so those requirements should be read as a historical policy record rather than current law. Its significance remains: it showed government attention moving from model outputs to development inputs. The state was not only interested in what models said. It was interested in who had the compute to build them.

This is the bridge between AI governance and national infrastructure control. If compute thresholds, advanced chips, export controls, data-centre security, and cloud reporting become part of policy, then AI power will increasingly be mapped through infrastructure access. In practice, compute becomes something like strategic industrial capacity.