Power before performance

Artificial intelligence is no longer constrained primarily by algorithms or silicon, but by electricity. The competition now turns on which nations can build and power the infrastructure required to run AI at an industrial scale. The societies that can combine abundant firm power with dense layers of local compute embedded in everyday infrastructure will shape the economic, military and political balance of the century ahead.

The new basis of power

A profound shift is underway in the global race for leadership in artificial intelligence. Progress in AI has come from better models, more data and exponential gains in accelerator hardware, but the constraint is now moving from mathematical sophistication to physical capacity: AI progress is running up against the availability, location and quality of power, not the cleverness of algorithms.

Data centres in the United States consumed roughly 4.5% of national electricity in 2023 and, in some scenarios, could reach the low teens by the late 2020s. Northern Virginia’s “Data Centre Alley,” where Amazon, Microsoft, Google, and Meta continue to expand, is forecast to exceed 20 gigawatts of demand by 2028. The local grid operator, Dominion Energy, has already announced multi-year delays for new interconnections because of transmission and transformer constraints. Around Cheyenne, Wyoming, a proposed AI campus is targeting an eventual load of up to 10 gigawatts, comparable to the state’s existing generation capacity and far above its residential consumption. None of this is a software problem.

In Europe and the UK, climate mandates, decarbonisation commitments, and opposition to large transmission projects have slowed growth in firm power capacity just as governments push to electrify transport and heating. In the United States, fragmented planning and permitting add years to transmission timelines. Grid connection queues in the UK now extend into the 2030s, with some major projects pushed into the mid-2030s. Energy scarcity is becoming a structural feature, not a temporary shock.

China faces the opposite imbalance. It is barred from the most advanced sub-5nm GPUs, high-bandwidth memory and packaging technologies essential for frontier training, but it does not face the same immediate shortage of firm power. Total installed generation capacity is on track to reach roughly 3,000–3,300 gigawatts by 2030, underpinned by rapid expansion in renewables alongside nuclear, hydro, gas and coal for grid stabilisation. That build-out gives Beijing room to grow compute even under tight silicon restrictions.

Rising workload intensity is outpacing efficiency gains. Token counts per query, context window sizes and embedded software demands are increasing faster than hardware and optimisation can offset. Every enterprise rollout adds sustained, always-on load. Nothing about this wave of demand is occasional or bursty. The last wave of digital transformation was defined by cloud elasticity; this one is bounded by the physical limits of generation, transmission and cooling, so advanced AI now depends more on construction equipment, civil engineering and grid reinforcement than on algorithmic advances.

The energy bottleneck

Electricity has become the principal limiting factor in the deployment of AI across the UK, Europe and the United States. Demand from hyperscale data centres, electrified transport, industrial reshoring and domestic heat-pump programmes is rising faster than generation and transmission capacity. Ambitions to decarbonise, electrify large parts of the economy and accommodate exponential digital growth are colliding with systems never designed for this convergence.

The grid stack is stressed at every level: substations and urban feeders were not built for clusters of 100–500 megawatt data centres and electrified vehicle fleets; ageing high-voltage lines struggle to move large blocks of power from resource-rich regions to load centres; national interconnectors and system-operator tools are stretched to keep frequency and voltage within tight tolerances as inverter-based renewables take a larger share. Without reinforcement at each layer, additional generation cannot be turned into reliable power where AI workloads sit.

Europe’s loss of Russian pipeline gas from 2022 triggered a scramble for LNG imports and record wholesale prices, while green policies accelerated the closure of coal and nuclear capacity in some member states. Governments cushioned consumers and industry in the short term but postponed decisions on firm generation, leaving an energy system that is less competitive and more exposed to global gas markets, just as data centre load and electrification targets rise.

In the United States, electricity demand is growing at the fastest rate since the 1970s. PJM Interconnection, the country’s largest grid region, has repeatedly raised its demand forecast and now expects an increase of around 32 gigawatts in peak load by 2030, largely driven by data centres and electrified industry, and has warned that planned additions to generation are inadequate. Analysts expect US data centre electricity consumption to rise sharply by 2028, creating a substantial shortfall without accelerated investment. Electric utilities are forecast to spend around $220–$230 billion a year on generation, transmission and substations by 2027, reversing a decade of under-investment, with further increases planned through 2030 as companies replace ageing transformers and lines and build new corridors into data centre and manufacturing clusters.

To escape these constraints, hyperscalers are starting to secure their own power: xAI is building dedicated gas-fired capacity to supply its Colossus campuses in Tennessee and Mississippi, with hundreds of megawatts of on-site or adjacent gas-turbine capacity already committed and proposals that would push the complex towards gigawatt scale. This bypasses local grid bottlenecks and accelerates deployment but effectively moves AI infrastructure outside national planning frameworks.

Microsoft has applied for up to 960 megawatts of gas generation in Virginia to support its infrastructure, while Amazon has signed multibillion-dollar power purchase agreements to lock in long-term renewable energy supply. These moves protect continuity for individual operators but do not resolve systemic scarcity and can deepen regional inequities if the public grid must still fund reinforcement without fully sharing in the benefits of the new load.

Major infrastructure projects move slowly: the SunZia transmission line linking New Mexico and Arizona took 18 years to secure approval, and the Grain Belt Express has been stalled for more than a decade. In the UK, queues for large grid connections extend into the 2030s. Without significant reinforcement, renewable projects cannot be integrated at pace, and new data centres cannot be sited where demand is highest.

Policymakers are starting to adjust in Europe; the Commission is shifting towards a more top-down approach to grid planning, built around a limited number of “energy highways” that map cross-border needs, identify bottlenecks, and propose projects to relieve them, after congestion and curtailment costs reached billions of euros annually. The goal is to treat the grid as a genuine internal market backbone rather than a patchwork of national systems, and to compress permitting timelines that currently stretch over many years.

In Great Britain, Ofgem has approved an initial £28 billion investment in gas and electricity networks over the next five years, the first stage of a roughly £90 billion programme to 2031 focused on building new lines and substations to move power from northern and Scottish generation zones to demand centres in England and Wales, to try to limit further increases in already high electricity prices.

The generation mix is in flux, with wind and solar now the cheapest new sources of power in many markets but remaining variable and largely non-synchronous; they do not naturally provide inertia, frequency support or dynamic voltage control. Nuclear, gas, hydro and residual coal still supply most of the firm, dispatchable capacity that can be called on at will to meet sudden swings in demand or renewable output and provide the system services that keep the grid stable. Studies of high-renewable systems point to the same conclusion: some combination of firm low-carbon capacity (nuclear, hydro, gas with carbon capture) and storage is required to maintain reliability as renewable penetration rises.

The blackout across Spain and Portugal in April 2025 illustrates what happens when these pieces fall out of balance: high levels of solar and wind, low system inertia and limited synchronous backup meant that when a disturbance triggered oscillations and a sudden loss of generation, frequency dropped quickly, and protection systems disconnected large parts of the network. Tens of millions lost power, and restoration required a rare black-start sequence and imports from neighbouring systems. The episode showed that grids dominated by inverter-based resources become fragile without sufficient firm capacity, modern control systems and robust interconnection.

Large-scale training and inference loads are subject to the same physics as the rest of the grid and are highly sensitive to voltage and frequency deviations. They require ample generation and robust local and regional networks capable of maintaining quality of supply under stress and rerouting power around faults. That favours architectures where compute is distributed across many sites, closer to both load and generation, rather than concentrated in a handful of hyperscale hubs that amplify local failures. As grids strain, it will often be cheaper and more resilient to push intelligence into factories, vehicles and substations than to haul every watt and every byte back to a few central campuses.

If baseload nuclear, hydro, gas and, in some countries, coal capacity is withdrawn faster than alternative firm resources are built, system operators have fewer tools to balance demand and supply in real time. Storage and demand response can help, but current deployments are far short of the levels needed to substitute for synchronous machines fully. Maintaining or replacing firm, dispatchable generation is essential for both net-zero goals and the stability required for nationwide AI deployment.

China is moving aggressively on this front, planning to reach around 2,000–2,500 gigawatts of renewable and other non-fossil capacity by 2030 and adding well over 100 gigawatts of new nuclear, hydro and pumped-storage capacity this decade, additions that rival, and in some cases exceed, the combined net increases expected across North America, Europe and other advanced economies. On current trajectories, China could add more than 300 gigawatts of new coal- and gas-fired capacity by 2030, reinforcing a system in which expanding renewables and nuclear sit atop a still-growing fossil baseline.

State Grid and other operators are matching this with record grid budgets, committing tens of billions of dollars a year to new lines and substations, much of it directed into ultra-high-voltage corridors that move power thousands of kilometres from western wind and solar bases to eastern industrial belts and enable rapid integration of new supply and more effective regional balancing.

Access to advanced GPUs and high-bandwidth memory remains constrained, but China does not face an immediate energy shortfall on the same scale as many Western economies and can scale compute capacity aggressively. The strategic comparison is blunt: either Western countries fix their energy and grid constraints first, or China solves its silicon problem first.

The cost of compute

Leadership in frontier semiconductor design and advanced model development remains concentrated in the United States and Europe. NVIDIA’s H100 and Blackwell platforms are the reference standards for large-scale training clusters, while AMD’s MI300 series is gaining traction among hyperscalers seeking diversification. ASML holds an effective monopoly on extreme ultraviolet lithography, without which sub-5nm fabrication is impossible. TSMC remains at the centre of global production, with 3nm output scaling and 2nm entering commercial availability from 2026. Western hyperscalers such as Google, Meta, Microsoft, and Amazon still operate the largest training systems in the world, with clusters of more than 100,000 accelerators now common.

As AI moves from research and prototypes into commercial deployment, the decisive economic metric is shifting from peak training capability to inference cost. More than 90% of lifetime compute expenditure occurs after a model has been trained, and competitive pressure is concentrating on throughput, utilisation and marginal cost per query, not just headline benchmarks.

Hardware bottlenecks are shifting in parallel, with capability increasingly constrained by memory bandwidth rather than raw floating-point performance. High-bandwidth memory has become the limiting factor on accelerator utilisation. Without adequate HBM, performance claims remain theoretical. Shortages have already forced NVIDIA and AMD to prioritise production for select customers, extend lead times for large orders beyond 2025, and push memory production and packaging into the foreground as strategic assets alongside chip fabrication.

Huawei sits at the centre of China’s response. It is no longer defined by telecommunications alone; the company underpins energy networks, transport systems, manufacturing platforms, healthcare operations, financial services, and civic infrastructure, providing the connective tissue between operational systems and AI platforms. Its identity rests on the ability to integrate compute, software and network intelligence across the national economy.

Huawei invests on the order of $24–$25 billion a year in R&D, more than twice NVIDIA’s, and greater than the combined R&D of Ericsson and Nokia, with more than half its workforce in R&D roles. Sell-side estimates suggest that by 2026, Huawei could supply close to half of China’s AI accelerators, while Nvidia’s share may fall into single digits as export controls tighten and domestic preference policies deepen. Improving yields on the Ascend 910C series and expanding capacity at SMIC are turning Huawei’s advanced AI chips into a financially sustainable product line rather than a patriotic science project.

SMIC’s advanced nodes still trail the global cutting edge by at least a generation, and yields on 7nm-class processes remain below best practice. Roadmaps and analyst estimates indicate that yields and packaging throughput, not demand, will determine how quickly domestic high-performance compute can scale. In the near term, the binding constraint for China is physics: how much useful silicon can be produced at acceptable cost and how fast it can be assembled into systems.

Targeted tariffs and subsidies for strategic data centre and AI projects mean that electricity pricing is not yet the primary bottleneck. Subsidised power in designated industrial zones can partly offset the weaker power efficiency of domestic accelerators and reduce effective data centre costs. Every new generation of clusters still requires substations, transmission capacity and cooling, but the ability to align industrial policy, grid expansion and domestic semiconductor development at speed differentiates China from Western markets, where electricity has already become the dominant constraint.

Other Chinese accelerator designers, such as Cambricon, Biren, and Iluvatar, are building platforms around inference economics rather than peak training performance, trading some absolute capability for efficiency, lower power draw, and large-scale deployability. For inference workloads where cost sensitivity is greatest, this can be a competitive advantage. Lower effective electricity pricing and centralised power planning reinforce this structural cost base, which Western hyperscalers will struggle to match.

Open or semi-open ecosystems such as Qwen, Yi, ERNIE, InternLM, and DeepSeek have advanced quickly and closed much of the gap with commercial Western systems, offering customisation and sovereign deployment without dependence on proprietary cloud pricing. In parts of the Global South where budgets are tight and data must remain onshore, that combination can matter more than nominal access to the “best” model.

Quantisation, mixture-of-experts routing, and distillation reduce computational overhead, while moving inference to edge devices can relieve some pressure on central infrastructure; however, the aggregate load still rises as model scale, context length, and enterprise demand increase faster than efficiency gains.

The next phase of the AI race will be determined less by the fastest chip or largest model and more by control over fabrication, memory, packaging, and electricity. Industrial and energy economics, not benchmark scores alone, are becoming the foundation of competitive power. Western firms still hold a technological lead, but without comparable progress in cost structures, that lead will be eroded by nations able to scale infrastructure faster and more cheaply.

Shifting influence

Nations able to deliver energy and compute infrastructure at scale and at competitive cost will shape the balance of power in artificial intelligence. As accelerator supply expands and pricing pressure increases, the economics of inference are likely to converge toward the marginal cost of electricity, making abundant, low-cost power the decisive variable. Breakthroughs in efficiency, storage or distributed computation could reshuffle positions, but the direction is clear.

Data centre development is already shifting toward regions with surplus baseload electricity, reliable cooling and supportive regulation. Iceland, Norway, and Quebec position themselves as destinations for energy-intensive compute on the back of stable hydropower and cold climates. Texas has attracted significant investment despite grid reliability concerns because power pricing remains favourable and industrial land is available. Regions facing high costs and congestion, such as parts of California and southeast England, are seeing projects cancelled or deferred.

Advantage in the AI era rests not only on who can generate the cheapest kilowatt-hour, but on who can move and balance that power. Nations with dense, modern transmission grids, strong regional interconnectors, digitalised distribution networks and advanced system-operator tools will be able to support AI workloads across many locations, not just a few privileged hubs. Those that neglect this infrastructure will find that compute, like heavy industry before it, clusters in a handful of regions with stable grids and firm power, leaving the rest of the economy constrained.

China is using its energy and infrastructure position to extend geopolitical influence. The Belt and Road Initiative, initially focused on ports, rail and logistics, is evolving into a digital-and-energy infrastructure programme in which state-backed financing bundles power generation, ultra-high-voltage transmission, data centre construction and AI hardware. Several African and Southeast Asian governments have signed agreements that integrate renewable projects with compute facilities, allowing local AI deployment without dependence on Western cloud providers. Trials of Chinese accelerator systems paired with open model ecosystems such as Qwen and ERNIE are underway across multiple emerging markets.

Distributed AI embedded inside critical infrastructure is a central feature of this strategy and increasingly a template for any state with constrained grids and rising digital loads. National-scale intelligence is conceived not as a single, centralised supercomputer but as millions of inference engines embedded in grids, factories, ports, vehicles, logistics networks and civic systems. Huawei's architecture, spanning custom processors, interconnects, cloud platforms and edge devices, enables computation at the point of data generation, with central clusters focused on coordination and retraining rather than handling every inference.

By pushing inference closer to where data and power are generated, distributed compute can smooth demand peaks, optimise local power flows, and reduce strain on long-distance transmission and hyperscale hubs. The same logic will increasingly apply in advanced economies trying to reconcile exponential digital demand with finite grids.

This model reduces dependence on central nodes, improves latency and resilience, and strengthens oversight of digital operations that underpin national security and economic stability. Substations, vehicles, port equipment, clinical systems and industrial machinery can all act as distributed inference points within a national fabric of connectivity and control. Huawei has become a load-bearing pillar in a broader state-led push toward technological sovereignty and secure national AI capability.

The United States is promoting “trusted infrastructure” partnerships focused on digital sovereignty and cybersecurity as alternatives to Chinese financing, while the European Union is developing frameworks around regulatory trust, data governance and supply-chain security. These offer meaningful differentiation but do not change the primary constraint for many emerging economies: the availability and affordability of power. For countries where per-capita electricity consumption is below roughly 1,500 kilowatt-hours per year, values cannot override cost. A cheap, reliable megawatt still wins most arguments.

Over the next decade, several paths are possible: China may achieve enough domestic capability in fabrication, packaging and memory to serve its internal market and much of the Global South, using energy pricing and industrial financing to outcompete Western offerings; Western nations may accelerate small modular nuclear deployment, grid reinforcement and permitting reform, stabilising power prices and retaining leadership in advanced AI research; the world may split into parallel technology blocs defined by incompatible hardware, regulatory frameworks and security norms. Some messy combination of all three is likely.

AI-enabled defence applications, from autonomous systems to simulation and targeting, depend on sustained compute availability. The nation able to train and iterate at a higher tempo gains an operational advantage, much as oil determined reach and speed in the twentieth century. The contest for AI leadership is shifting from a race to build the most advanced models to a race to build the most capable industrial base: compute determines productivity and military readiness; energy determines compute; and influence will follow those who can generate and deliver the lowest-cost kilowatt-hour and convert it into capability at scale.

illuminem Voices is a democratic space presenting the thoughts and opinions of leading Sustainability & Energy writers, their opinions do not necessarily represent those of illuminem.

Curious how major companies measure up on climate? On illuminem’s Data Hub™, explore verified emissions data, net‑zero targets, and sustainability performance of thousands of firms — from industry leaders to emerging innovators.