The artificial intelligence revolution runs on silicon and electricity. For all the attention paid to algorithmic breakthroughs, model architectures, and software capabilities, the fundamental constraint on AI development in 2026 is physical: the availability of specialized computing hardware and the energy required to operate it. The economics of AI compute have become a defining factor in national competitiveness, corporate strategy, and the pace at which AI capabilities advance. Understanding these economics — the supply constraints, the investment trajectories, the energy implications, and the geopolitical dimensions — is essential for anyone attempting to forecast the AI landscape through 2028.
The GPU Economy: Scarcity as Strategy
The global market for AI accelerator chips is dominated by NVIDIA to a degree that has few parallels in modern technology markets. NVIDIA’s data center GPU revenue exceeded $120 billion in fiscal year 2025, representing approximately 80% of the global market for AI training and inference hardware. This dominance has persisted despite aggressive competitive efforts from AMD, Intel, Google, Amazon, Microsoft, and numerous startups, because NVIDIA’s CUDA software ecosystem creates switching costs that hardware specifications alone cannot overcome.
The GPU shortage that constrained AI development from 2023 through mid-2025 has partially eased, but supply-demand imbalances persist at the frontier. The highest-performance chips — currently the Blackwell B200 and its variants — remain allocation-constrained, with major cloud providers and AI laboratories securing supply through long-term contracts that commit billions of dollars years in advance. Organizations without these pre-existing relationships face wait times of six to twelve months for large orders, effectively locking them out of frontier AI development.
This scarcity dynamic has profound implications for the structure of the AI industry. Access to compute has become a competitive moat comparable to data access or talent concentration. The largest technology companies — Microsoft, Google, Amazon, Meta — have invested hundreds of billions of dollars in GPU procurement and data center construction, creating a compute advantage that smaller competitors and academic researchers cannot match. The concentration of compute resources has created what some analysts describe as an “AI oligopoly” — a market structure in which a small number of firms control the physical infrastructure on which all AI development depends.
The pricing of AI compute reflects this concentration. Despite hardware improvements that have reduced the cost per floating-point operation by roughly 40% annually, the total cost of frontier AI training runs has increased dramatically because the scale of training has grown faster than per-unit costs have fallen. The most expensive training runs in 2026 cost upward of $500 million, and industry projections suggest that training runs exceeding $1 billion will occur before 2028. These costs are far beyond the reach of all but the largest corporate and government entities, raising fundamental questions about who controls the development of the most powerful AI systems.
The Data Center Buildout: Geography, Energy, and Capital
The physical infrastructure required to house and power AI computation has become one of the largest capital investment categories in the global economy. Data center construction spending is projected to exceed $500 billion cumulatively between 2024 and 2028, driven primarily by AI workload demand. This buildout is reshaping physical geography, energy markets, and municipal planning in ways that are only beginning to be understood.
In the United States, data center construction is concentrated in three primary corridors: Northern Virginia (which hosts the world’s largest concentration of data center capacity), the Dallas-Fort Worth metropolitan area, and the Pacific Northwest. Secondary clusters are emerging in Phoenix, Columbus (Ohio), and several locations in the Midwest where affordable electricity and available land create favorable economics.
The energy requirements of this buildout have become a major constraint and a source of growing public controversy. A single large-scale AI data center consumes as much electricity as a small city — typically 100-300 megawatts for a campus-scale facility, with some facilities under construction planning for 500 megawatts or more. The aggregate electricity demand from data centers in the United States is projected to increase from approximately 4% of total national electricity consumption in 2025 to 8-10% by 2028.
This demand growth has created a scramble for power supply that is straining electrical grids, delaying interconnection timelines, and forcing utilities to reconsider retired generation assets. In Northern Virginia, the epicenter of data center construction, the local utility has warned that demand growth from data centers may exceed the pace at which new transmission capacity can be built, potentially creating reliability risks for the broader grid. Similar warnings have been issued by utilities in Texas, Georgia, and the Pacific Northwest.
The energy intensity of AI has attracted significant environmental scrutiny. While major technology companies have made commitments to power their operations with renewable energy, the speed of demand growth has outpaced the deployment of new renewable generation in most markets. The result has been increased utilization of natural gas generation and, in some cases, the extension or restart of nuclear and coal facilities that had been slated for retirement. Several major AI companies have signed agreements to purchase power from existing nuclear plants or to support the development of small modular reactors specifically for data center power supply.
Sovereign Compute: The Geopolitics of AI Infrastructure
The recognition that AI capabilities depend on physical infrastructure has transformed AI compute into a matter of national security and sovereign strategy. The concept of “sovereign AI” — the principle that nations must control sufficient domestic AI compute capacity to ensure technological autonomy — has moved from academic discussion to active policy implementation in multiple countries.
The United States has pursued a dual strategy combining domestic infrastructure investment with export controls designed to limit competitors’ access to advanced AI hardware. The semiconductor export restrictions targeting China, first implemented in October 2022 and progressively expanded through 2025, have attempted to maintain a compute capability gap between the United States and its primary strategic competitor. These restrictions have targeted not only finished chips but also the semiconductor manufacturing equipment and electronic design automation software needed to produce advanced processors domestically.
The effectiveness of these controls is increasingly debated. China has responded with a massive domestic semiconductor investment program — the “Big Fund III” alone committed $47 billion in 2024 — and has made notable progress in producing AI accelerator chips using mature process nodes. While these chips do not match the performance of NVIDIA’s latest products, they are increasingly competitive for many AI workloads, and the volume of Chinese AI compute capacity has grown substantially despite export restrictions.
The European Union has pursued sovereign compute through the European Chips Act and the EuroHPC Joint Undertaking, which funds the construction of high-performance computing facilities across member states. However, European sovereign AI compute capacity remains a fraction of American or Chinese levels, and the EU’s dependence on non-European hardware and cloud infrastructure for AI workloads has not meaningfully decreased.
Smaller nations have adopted varied approaches. The United Arab Emirates, Saudi Arabia, and Singapore have invested heavily in AI data center infrastructure, positioning themselves as regional compute hubs. Japan and South Korea have focused on domestic chip design capabilities while maintaining access to advanced manufacturing through alliances with Taiwan’s TSMC and South Korea’s Samsung. India has launched a national AI compute initiative targeting the construction of publicly accessible AI infrastructure, recognizing that private-sector investment alone will not provide sufficient compute access for the country’s AI development ambitions.
The Economics of Inference: Where the Money Actually Flows
While training captures headlines — the massive one-time expenditures required to develop frontier models — the economics of AI are increasingly dominated by inference: the ongoing computational cost of running trained models to serve user requests. As AI applications proliferate, the total compute spent on inference has surpassed training expenditure by a substantial margin and continues to grow exponentially.
The economics of inference are fundamentally different from training. Training requires maximum computational throughput and can tolerate significant latency. Inference requires low latency and high throughput for serving millions of simultaneous user requests. These different requirements have driven hardware diversification, with specialized inference chips offering dramatically better price-performance than general-purpose training GPUs for deployed models.
The cost structure of inference has major implications for AI business models. At current hardware prices, serving a single complex query to a frontier language model costs approximately $0.01-0.05, depending on the model size and query complexity. For applications handling millions of queries per day, these costs aggregate rapidly. Companies deploying AI at scale must achieve sufficient revenue per query — through subscription fees, advertising, or productivity gains — to justify the infrastructure cost, and many current AI applications struggle to achieve unit economics that cover their inference costs at scale.
This inference cost dynamic is driving two important trends. First, significant research and engineering effort is being devoted to model compression, quantization, and distillation — techniques that reduce the computational requirements of deployed models while preserving most of their capability. Second, the separation between model training and model deployment is creating a stratified market structure in which a small number of organizations train frontier models and a much larger number of organizations deploy optimized versions of those models for specific applications.
The 2028 Compute Landscape
Projecting the compute landscape through 2028 requires synthesizing hardware roadmaps, infrastructure construction timelines, and demand growth estimates. Several conclusions are reasonably well-supported by committed capital expenditure and physical projects under construction.
Total global AI compute capacity will approximately triple between early 2026 and late 2028. This growth reflects both hardware improvements (each new chip generation delivers roughly 2x performance improvement at similar power consumption) and infrastructure expansion (the ongoing data center buildout adding physical capacity).
The cost of a given level of AI capability will decline substantially. A training run equivalent in compute to today’s frontier models will cost roughly 70-80% less by 2028, making capabilities that currently require billion-dollar budgets accessible to organizations spending tens of millions. This democratization will not extend to the new frontier, however — the largest training runs of 2028 will cost more than today’s, not less, because the frontier will have advanced.
Energy constraints will become the binding limit on AI infrastructure growth in several key markets. Absent breakthrough improvements in energy efficiency or rapid deployment of new generation capacity, the pace of data center construction will be limited by power availability in the most desirable locations. This constraint will drive geographic diversification of AI infrastructure toward locations with abundant power supply, potentially shifting the geography of AI development.
The geopolitical compute competition will intensify. The gap between US and Chinese AI compute capability — currently estimated at roughly 18-24 months in hardware performance terms — will narrow or widen depending on export control effectiveness and domestic Chinese semiconductor progress. The outcome of this competition will significantly influence the relative AI capabilities of the two nations and the balance of power in AI-intensive domains including military applications, economic competitiveness, and intelligence operations.
For policymakers and planners, the message is clear: AI compute is infrastructure in the same fundamental sense as transportation networks, energy grids, and telecommunications systems. Nations and organizations that control sufficient compute capacity will shape the AI future. Those that do not will be shaped by it.
Compute capacity estimates in this analysis draw on proprietary modeling of announced data center projects, semiconductor production schedules, and energy interconnection timelines. Estimates are updated quarterly. Last update: Q1 2026.