The AI buildout has become a systems problem
The artificial-intelligence infrastructure race is usually described as a contest for GPUs. That description is now incomplete.
NVIDIA’s latest disclosures show why. The company is expanding commitments across memory, manufacturing, substrates, advanced packaging, testing, land, power, and data-centre capacity. In a July filing, NVIDIA said its supply and capacity commitments had risen from US$119 billion in the previous quarter to US$279 billion. The same filing warned that customers may delay purchases if they cannot secure the land, power, shell, and capital required to deploy the systems.
That is a change in the nature of the bottleneck. The question is no longer only whether the industry can manufacture a faster accelerator. It is whether several infrastructure systems can be expanded and connected on compatible timelines.
This matters because those systems move at different speeds. A chip design can be revised in months. A data-centre building can be constructed in a few years. Transmission lines, substations, transformers, generation, and regulatory approvals often require longer planning cycles. A mismatch can leave an expensive accelerator without a powered site, or a utility with new capacity whose expected customer load arrives late—or not at all.
The result is an industry increasingly shaped by coordination risk rather than by silicon performance alone.
The demand numbers are large, but location matters more than the headline
The International Energy Agency estimates that data centres consumed about 415 terawatt-hours of electricity globally in 2024, roughly 1.5% of global electricity use. In its base case, that figure doubles to about 945 TWh by 2030. Electricity consumption from accelerated servers, the category most closely associated with AI, is projected to grow by about 30% a year.
Those numbers do not mean AI will consume a dominant share of all electricity. The IEA expects data centres to account for just under 3% of global electricity demand in 2030 in its base case, and less than 10% of total global electricity-demand growth from 2024 to 2030. Electrification, industry, air conditioning, and electric vehicles remain larger sources of aggregate growth.
The more important point is concentration. Data centres cluster where networks, land, tax policy, and power are available. A concentrated load can stress a local transmission system even when its global share looks modest. The IEA explicitly notes that data centres are more difficult to integrate than diffuse loads such as electric vehicles because they are concentrated in specific locations.
US forecasts illustrate the uncertainty. A Lawrence Berkeley National Laboratory estimate cited by the US Department of Energy puts data centres at 11.8% of total US electricity use by the end of the decade, with a range from 9.5% to 15.3%. The estimate models equipment shipments and energy use. It does not directly model whether the grid can supply every proposed project on schedule.
That distinction is easy to miss. A forecast of potential electricity consumption is not a forecast of connected load. Projects can be delayed by interconnection queues, transformer shortages, financing conditions, permitting, or weaker-than-expected AI demand. Conversely, a shortage of grid capacity can encourage operators to build private generation that changes the emissions and cost profile of the project.
Chips are still a bottleneck—but they are becoming one layer of a stack
Advanced packaging shows how the hardware constraint is broadening beyond wafer fabrication.
At its April 2026 technology symposium, TSMC said it was already producing 5.5-reticle-size CoWoS packages and planned a 14-reticle version for 2028. The company describes these packages as a way to integrate more compute dies and high-bandwidth memory in one system. TSMC also said a 40-reticle system-on-wafer product is expected in 2029.
NVIDIA’s November supply-chain announcement points in the same direction. It described new or expanded commitments involving TSMC, Amkor, SPIL, Foxconn, Kinsus, and Samsung Electro-Mechanics. Kinsus plans to expand advanced package-substrate capacity dedicated to NVIDIA by more than 3.5 times current levels, with production expected to begin in the second half of 2027 and additional capacity targeted for the third quarter of 2028. NVIDIA also said its agreements with Amkor for US and South Korean advanced-packaging expansion total nearly US$3 billion.
These announcements are evidence of demand and supply planning, not proof that every AI forecast will be realized. Some commitments are multi-year, some are adjustable, and company announcements naturally emphasize growth. But their timing is revealing. The industry is making capacity decisions years ahead of deployment because the components cannot be treated as instantly interchangeable.
This creates a more complicated competitive landscape. NVIDIA’s advantage is not only its accelerator architecture. It also includes relationships, prepayments, manufacturing access, packaging capacity, software, networking, and the ability to absorb or guarantee risk. That does not make competition impossible. Google, Amazon, Microsoft, AMD, and newer AI-chip companies can pursue different designs and supply arrangements. It does mean that a challenger must solve an ecosystem problem, not simply produce a credible chip.
Power is becoming a procurement and finance problem
The physical power constraint is already changing corporate behaviour.
In June, the Federal Energy Regulatory Commission ordered all six US regional grid operators to justify or reform tariffs for data centres and other large energy users. The order identified five areas for reform: faster transmission-service studies, prevention of cost shifting, rules for co-located generation, services for flexible large loads, and processes for facilities that serve nearby large customers.
The unusually broad action reflects a basic conflict. Data-centre developers want speed and certainty. Utilities and regulators want evidence that new assets will be used and paid for. Existing customers want protection from being assigned costs for infrastructure built primarily for a new industrial load.
NVIDIA’s own filing makes the financing dimension explicit. In August, the company entered guarantees connected to a planned Ohio campus with roughly 4.25 gigawatts of IT load and leases to OpenAI. NVIDIA said its obligation could reach US$105 billion in aggregate, subject to conditions, and that the site is expected to begin placing data centres in service from 2029. The disclosure does not mean NVIDIA will spend that amount immediately or bear the entire project cost. It does show how chip suppliers, cloud companies, and infrastructure developers are increasingly connected through credit and capacity arrangements.
The public debate is moving from “Can we build enough?” to “Who should pay if the forecast is wrong?” The Harvard Kennedy School’s Belfer Center notes that US data-centre electricity projections for 2030 have ranged from roughly 200 TWh to more than 1,000 TWh across institutions. That spread is not a minor modelling difference. It changes the risk of overbuilding generation and transmission.
Recent reporting also shows why rate design matters. Reuters reported that a large Texas data centre could avoid US$33.4 million in transmission costs by reducing grid use for one peak hour, according to NRG Energy. The article also cited an independent monitor’s estimate that data-centre growth had increased power-system costs across PJM by US$29.4 billion. Large users argue that peak shaving lowers the amount of new capacity the grid needs. Critics argue that sophisticated customers can avoid charges that households cannot easily avoid.
Both claims can be true. Flexible load can improve reliability and reduce marginal investment. But the benefit depends on whether the flexibility is measurable, dispatchable, and reflected in planning rules. If a data centre runs backup generators only to reduce its tariff exposure while the wider system still funds new infrastructure, the accounting may be efficient for the customer but not necessarily fair for other ratepayers.
The practical breakthrough may be flexibility, not just more generation
The simplest response to rising demand is to build more power plants. That is also the slowest and most expensive response if the system is designed around a constant, inflexible load.
AI cloud growth and workloads have a potential advantage: some training and inference jobs can be moved in time or location. Batch workloads can run when power is less scarce. Models can be routed between regions. Smaller models, quantization, caching, and improved utilization can reduce the amount of compute required for a given task. Cooling systems and other facility loads may also provide limited flexibility.
Not every workload can move. Real-time inference, regulated data, low-latency applications, and critical services have harder constraints. Demand response also cannot be treated as free capacity. Operators need clear performance requirements, telemetry, compensation, and rules for what happens when a facility cannot curtail.
Still, flexible computing is one of the few levers that can connect digital optimisation to physical infrastructure. It can reduce peak demand without pretending that total energy use has disappeared. For cloud providers, that could become a product feature: not merely a region and a price, but a workload contract that specifies carbon intensity, curtailment conditions, latency, and resilience.
What this means for businesses and IT teams
Businesses should be more sceptical of AI infrastructure promises that discuss model capability but not deployment conditions. A model may be available through an API while the economics of sustained, high-volume use remain exposed to power, capacity, and regional pricing.
When evaluating AI services, organisations should ask:
- Which regions can actually serve the workload, and what happens during a capacity or power constraint?
- Can non-urgent jobs move across regions or time windows?
- How are rate changes, capacity reservations, and infrastructure surcharges handled?
- What availability assumptions depend on a single data-centre campus or power market?
- Are data residency, privacy, and regulatory commitments preserved when workloads move?
- Does the provider disclose meaningful information about energy use and demand-response practices, or only a broad sustainability target?
IT architects should design for graceful degradation. That means keeping smaller models available, separating interactive requests from batch work, caching expensive results, and maintaining a fallback path that does not depend on one provider or one region. These are not merely cost-optimisation techniques. They are resilience controls for an infrastructure market whose physical constraints are becoming more visible.
For investors and policymakers, the key metric is not announced AI capacity. It is delivered capacity with a credible path to power, network connectivity, financing, and customer demand. Announcements can create optionality; they do not eliminate execution risk.
The likely outcome is slower, more selective expansion—not an end to AI growth
The evidence does not support the claim that power constraints will stop AI development. It supports a narrower conclusion: infrastructure will increasingly determine where, when, and at what cost AI expands.
The IEA’s scenarios show a wide range of outcomes. Its base case reaches about 945 TWh of data-centre electricity use in 2030. A higher-growth case exceeds 1,700 TWh in 2035, while a headwinds case plateaus around 700 TWh. Efficiency improvements, software adoption, hardware supply, grid bottlenecks, and the pace of AI use all influence the result.
That uncertainty should change how the industry communicates. Hardware roadmaps and capital-spending plans are useful signals, but they are not demand guarantees. A new data centre is not productive capacity until it has power, cooling, networking, software, customers, and a viable operating cost.
The next phase of the AI race will therefore reward companies that coordinate the full stack. Chip design still matters. So do packaging, memory, cloud software, and model quality. But the decisive advantage may belong to firms that can align those assets with land, electricity, transmission, financing, and flexible operations without shifting unacceptable costs onto everyone else.
AI’s next infrastructure bottleneck is not one missing chip. It is the ability to make an entire industrial system arrive at the same time.
Sources and further reading
- International Energy Agency: Energy demand from AI
- US Department of Energy: Powering America’s AI Future
- Federal Energy Regulatory Commission: Large-load integration action
- US Energy Information Administration: Electricity-demand forecast
- NVIDIA: 2026 supply-chain commitments
- NVIDIA Form 10-Q, July 26, 2026
- TSMC: 2026 North America Technology Symposium
- Belfer Center: AI, Data Centers, and the US Electric Grid
- Reuters: Data-centre peak shaving and power costs
Editorial note: Company announcements describe planned investments and capabilities. They are not independent forecasts. Forecast ranges vary materially because data-centre demand, AI adoption, hardware efficiency, grid access, and financing remain uncertain.
Feature image: AI-generated conceptual illustration for Guava Insights.


