The physical constraint on AI has shifted from chip supply to electricity. I am tracking the reported numbers: big tech AI capex, the US grid interconnection queue, AI data centre power consumption. The gap between what is being spent and what can be powered is widening.
$349B
Big technology AI capital expenditure, 2025
Confirmed
2,600 GW
US electricity grid interconnection queue
Confirmed
4 GW
AI data centre power consumption, 2024
Traceable, estimated
Disclaimer: This is personal research, not professional advice. I am a technologist, not an analyst, economist, or forecaster. Nothing here constitutes financial, investment, career, or legal advice. The data comes from third-party sources I believe to be reliable, but I make no representations or warranties as to its accuracy or completeness. My views may be wrong. Past trends do not guarantee future outcomes. If you make decisions based on this, that is your call, not mine. Consult a qualified professional before acting on any of it.
About this report
Why I built it, and why it will get better
I am trying to track what the real measurable effects of AI are being reported. I trace every number to its primary source. Where the data is confirmed, I say so. Where it is directionally accurate but imprecisely sourced, I say that too. Where it cannot be verified, I say that as well. This tracker is the result.
I am not an analyst or forecaster. I trace every number to its primary source. Where a number is confirmed, I say so. Where it is directionally accurate but imprecisely sourced, I say that too. Where it cannot be verified, I say that as well. Where credible sources disagree, I present both positions. This means my dashboard will sometimes show fewer data points than a typical industry report. I think that is a feature, not a bug.
I will update the tracker every month. Each update will add a new data point, so over time you can see whether a signal is accelerating, stabilising, or receding. Historical data will never be overwritten. After three or four updates, the trends will become the most important part of the dashboard. I am tracking five areas (labour market, revenue concentration, compute, regulation, sector disruption), each broken into multiple metrics. This page covers the compute and energy bottleneck, which has four metrics. The other four signals will get the same treatment in subsequent updates.
Methodology
Source selection: I search for primary datasets from government sources (DOE, FERC, LBNL), research institutions (RAND Corporation, IEA), industry publications (DataCenterKnowledge, Data Center Dynamics), and company filings (SEC 10-K, earnings calls). I exclude any source where the underlying dataset cannot be located or where the methodology is undisclosed.
Confidence ratings: VERY HIGH = published government dataset or peer-reviewed study with large sample. HIGH = primary dataset with transparent methodology. MEDIUM-HIGH = primary aggregator with broad scope. MEDIUM = secondary source or different methodology. MODERATE = contextual only, not directly comparable. LOW = qualitative or untraceable.
Limitations: AI data centre power consumption figures are estimates that depend on how you define "AI data centre" and what utilisation rate you assume. The 4 GW figure for 2024 is a broad estimate from multiple sources with different methodologies. Grid interconnection queue data includes all generation types, not just data centre requests. Capex figures are from company earnings reports but represent total capital expenditure allocated to AI, which each company defines differently. Margins of error are not included in the figures shown. Where data points cannot be traced to a single publication, I note this explicitly.
Date of access: All sources were last accessed in June 2026. Source URLs were verified live during research. Where a source may have been updated since, the most recent version at time of writing is cited.
Built with Odokai. The initial research synthesis ran in a single session, with each source then individually verified. The brand styling came from a reusable Skill. The data, charts, and HTML live in a persistent cloud workspace, ready for next month's update with no setup. The platform makes the process repeatable: each cycle is a matter of re-running the Skill against fresh data, not starting from scratch.
Metric 1 of 4
Big technology AI capital expenditure, 2025
The original claim: the largest technology companies collectively committed $349 billion to AI capital expenditure in 2025. This traces to company earnings reports and SEC filings from Microsoft, Google, Amazon, and Meta, aggregated by DataCenterKnowledge and industry analysts.
Verdict: Confirmed
The $349B figure is directly traceable to public earnings reports and SEC filings. The aggregation is straightforward, though each company defines AI capex differently.
The four largest technology companies (Microsoft, Alphabet/Google, Amazon, and Meta) disclosed combined capital expenditure of approximately $349 billion for 2025, with the majority allocated to AI infrastructure including data centres, GPU clusters, and networking equipment. The individual figures are from public filings. The aggregate is a simple sum. The caveat: each company uses its own definition of what counts as "AI" capital expenditure. Some include all data centre spending. Others separate AI-specific from general infrastructure. The $349B figure represents the total capital expenditure of these companies, not a narrowly defined AI-only spend. The AI allocation is estimated by the companies themselves and by analysts, and those estimates vary.
Source
Figure
Scope
Confidence
Primary?
DataCenterKnowledge
$349B
Combined big tech AI capex, 2025
HIGH
YES
Company earnings reports
$349B
Microsoft, Google, Amazon, Meta combined capex
HIGH
YES
GPUnex
$349B
Industry analysis, corroborated aggregate
HIGH
YES
Big technology company AI capital expenditure, 2025
Combined capex for Microsoft, Alphabet, Amazon, and Meta. Each company reports total capital expenditure; the AI allocation is their own classification. The $349B aggregate is a sum of publicly disclosed figures.
Combined AI capex
$349 billion is an extraordinary commitment, but the definition matters. If each company is including all its data centre spending as "AI capex," the figure conflates general infrastructure growth with AI-specific investment. If the AI allocation is narrower, the true AI-specific spend may be lower. The number is confirmed. What it means depends on how each company draws the line.
What I am watching
1. 2026 capex guidance from the same four companies. If the aggregate rises above $400B, the investment cycle is accelerating. If it flattens or falls, the industry may be hitting the limit of what it can physically deploy. 2. The split between "committed" and "deployed" capex. Spending on GPUs that sit in inventory waiting for power is committed but not productive. 3. Company-level capex as a share of revenue. If capex continues to rise while revenue growth decelerates, the investment thesis is being tested.
Metric 2 of 4
US electricity grid interconnection queue
The original claim: the US electricity grid interconnection queue holds 2,600 gigawatts of pending requests, with an average wait time exceeding five years. This traces to the Lawrence Berkeley National Laboratory (LBNL), a Department of Energy research lab that publishes the authoritative dataset on grid interconnection queues.
Verdict: Confirmed
The 2,600 GW figure is directly from LBNL's publicly available dataset. The five-year average wait time is from the same source. This is a government dataset with transparent methodology.
LBNL publishes annual reports on generator interconnection queues across all seven US grid regions. The 2,600 GW figure represents the total capacity of generation projects waiting for grid connection approval. This includes solar, wind, battery storage, and gas projects, not just data centre requests. The queue has grown dramatically as renewable energy projects have proliferated, and data centre power demands have added further pressure. The five-year average wait time reflects the time from application to commercial operation. The queue is a measure of demand for grid access, not a measure of projects that will actually be built. Many projects withdraw before completion. But the queue size signals that the grid approval process is a bottleneck for any large power consumer, including AI data centres.
Source
Figure
Scope
Confidence
Primary?
LBNL / DOE
2,600 GW
Total US grid interconnection queue capacity
VERY HIGH
YES
LBNL / DOE
5+ years
Average wait time from application to operation
VERY HIGH
YES
DataCenterKnowledge
2,600 GW
Industry reporting, corroborated from LBNL
HIGH
YES
GPUnex
2,600 GW
Industry analysis, corroborated
HIGH
YES
US grid interconnection queue: 2,600 GW of pending requests
The queue includes all generation types (solar, wind, battery, gas), not just data centre requests. Average wait time exceeds five years. Data from LBNL/DOE.
Queue
2,600 GW
0 GW~1,250 GW total US installed generation capacity (EIA, 2024)
Pending interconnection requests For reference: total US installed capacity is ~1,250 GW (EIA, 2024)
The queue is larger than the entire existing US generation fleet. Total US installed generation capacity is approximately 1,250 GW (EIA, 2024). The interconnection queue holds roughly 2,600 GW of projects waiting for approval, more than double the entire fleet built to date. Even if only a fraction of these projects are ultimately built (and many withdraw before completion), the queue signals that the grid approval process is the rate-limiting step for new power capacity. AI data centres need gigawatts of power. They cannot wait five years for grid connection.
The queue is not just a data centre problem
Structural grid constraint
The interconnection queue reflects a fundamental mismatch between the speed of AI demand growth and the speed of grid infrastructure. Permitting, environmental review, and physical transmission construction all take years. Even fast-tracked projects face material lead times. The queue is a symptom of a system that cannot scale at the pace AI requires. Microsoft's CEO confirmed that GPUs are "sitting in inventory" because there is not enough power to run them.
Data Center Dynamics, Microsoft earnings call, LBNL
Permitting reform can help
The queue is large partly because the permitting process is slow, not because the grid is physically incapable of expansion. Federal permitting reform, already under discussion in Congress, could significantly reduce wait times. Some data centres are bypassing the queue entirely by building on-site generation (nuclear, natural gas, batteries). The constraint is real but not permanent: policy change and private generation can both accelerate connection.
DOE grid modernisation initiative, industry reports
My assessment: The grid constraint is structural in the near term, addressable in the medium term. No amount of permitting reform will build transmission lines in under two years. But the queue is not destiny either. Data centres are already pursuing on-site nuclear (Three Mile Island restart for Microsoft), behind-the-meter gas, and battery storage. The question is whether these workarounds can scale fast enough to prevent a significant slowdown in AI deployment over the next two to three years. I am watching the gap between what is committed and what is actually energised.
Metric 3 of 4
AI data centre power consumption, 2024
The original claim: AI data centres consumed approximately 4 gigawatts of power in 2024. This traces to the International Energy Agency (IEA), RAND Corporation research, and industry estimates. The figure is an estimate, not a metered reading, because there is no universally agreed definition of an "AI data centre."
Verdict: Traceable, estimated
The 4 GW figure is directionally consistent across multiple credible sources, but the exact number depends on how you define and measure "AI data centre" power.
The IEA and RAND both estimate AI-specific data centre power consumption at roughly 4 GW in 2024. However, this figure requires explanation. Most data centres run a mix of AI and non-AI workloads. Isolating the AI-specific portion requires assumptions about utilisation rates, workload types, and hardware efficiency. Different methodologies produce different estimates. The 4 GW figure represents the central estimate from multiple sources, but the range is likely 3 to 6 GW depending on definition. RAND's research notes that continued exponential growth in demand would "far outpace previous data centre expansion," which is a physical constraint that could bend the AI growth curve regardless of how capable the models become.
Source
Figure
Scope
Confidence
Primary?
RAND Corporation
~4 GW
AI data centre power consumption, 2024
MEDIUM-HIGH
YES
IEA
~4 GW
Data centre power for AI workloads, 2024
MEDIUM-HIGH
YES
GPUnex
~4 GW
Industry estimate, corroborated
MEDIUM
YES
DataCenterKnowledge
~4 GW
Industry reporting, corroborated
MEDIUM
NO
AI data centre power consumption, 2024
Estimated at approximately 4 GW. The exact figure depends on how "AI data centre" is defined and what utilisation rates are assumed. Multiple credible sources converge on this range.
AI data centre power (estimated)
The 4 GW figure is a snapshot, but the trajectory is what matters. If AI data centre power demand is doubling every 18 to 24 months (as many industry participants project), the 4 GW figure becomes 8 GW by 2026, 16 GW by 2028. At some point, the growth rate must slow because the physical infrastructure cannot keep pace. RAND's observation that exponential growth would "far outpace previous data centre expansion" is the key insight: the constraint is not about today's 4 GW. It is about whether 4 GW can become 16 GW in four years without breaking the grid.
Metric 4 of 4
Primary constraint on AI growth
The original claim: the primary constraint on AI growth has shifted from chip supply to electricity. In 2024, the bottleneck was GPU availability. By 2026, it is power. Microsoft's chief executive confirmed that GPUs are "sitting in inventory" because there is not enough power to install them.
Verdict: Directional, well-sourced
The shift from chips to electricity as the primary constraint is supported by multiple credible sources, but "primary constraint" is a qualitative assessment, not a measured quantity.
Microsoft's CEO stated publicly that the company has GPUs it cannot deploy because of power constraints. DataCenterKnowledge and Data Center Dynamics both report that the bottleneck has shifted from silicon to electricity. RAND's research confirms that power infrastructure is the binding constraint on AI data centre expansion. Power management chips and discrete semiconductors remain in shortage throughout 2026, extending the constraint beyond GPUs to the wider chip infrastructure. The convergence of these sources is strong. However, "primary constraint" is inherently qualitative: it depends on which part of the stack you are looking at and when. For a company that cannot get GPUs at all, chips are still the constraint. For a company that has GPUs but no power, electricity is the constraint. The shift is real but uneven across the industry.
Source
Figure
Scope
Confidence
Primary?
Data Center Dynamics
ELECTRICITY
Microsoft CEO: GPUs in inventory, lacking power
HIGH
YES
DataCenterKnowledge
ELECTRICITY
Bottleneck shifted from chips to power, 2024 to 2026
MEDIUM-HIGH
YES
RAND Corporation
ELECTRICITY
Power infrastructure is the binding constraint
MEDIUM-HIGH
YES
GPUnex
ELECTRICITY
Energy crisis constraining AI data centre growth
MEDIUM
YES
The constraint shift: from chips to electricity
In 2024, the primary bottleneck was GPU supply. By 2026, it is electricity. Multiple sources confirm this shift, though it is uneven across the industry.
Constraint intensity (illustrative)
2024
Chips
Power
2026
Chips
Power
Chip supply constraint Electricity constraint
The shift matters because electricity is harder to solve than chips. Chip production can be scaled with more fabs (given time and money). Electricity requires grid connection, permitting, transmission lines, and generation capacity, each of which involves regulatory, environmental, and physical constraints that money alone cannot overcome. Microsoft's admission that it has GPUs sitting idle because there is no power to run them is the clearest signal yet that the AI industry has hit a different kind of wall.
The wider chip constraint
Power is the binding constraint
If you have GPUs but cannot power them, the chip shortage is solved but the deployment bottleneck remains. Microsoft's inventory confession is the clearest evidence. The 2,600 GW grid queue and five-year wait times confirm that power delivery is the rate-limiting step. Even if NVIDIA produces every GPU the industry wants, those GPUs are useless without electrons. The constraint is physical and cannot be solved by software efficiency gains alone.
Data Center Dynamics, LBNL, RAND Corporation
The constraint is broader than power
Power management chips and discrete semiconductors remain in shortage throughout 2026. The constraint extends beyond GPUs to the wider chip infrastructure needed to build and operate data centres. Even if grid connection were instant, the supply chain for power delivery components (inverters, switchgear, cooling systems) has its own bottlenecks. The "electricity" framing may understate the breadth of the constraint. It is an infrastructure constraint, of which power is the largest component.
GPUnex, DataCenterKnowledge
My assessment: The primary constraint is electricity, but the bottleneck is broader. Power is the single largest limiting factor, and the evidence for that is strong. However, the wider infrastructure supply chain (power management chips, switchgear, cooling, construction labour) also has bottlenecks that would persist even if the grid problem were solved tomorrow. The AI industry is learning that software can scale instantly, but the physical world does not. I will be watching whether the power constraint begins to slow NVIDIA's revenue growth, which would be the clearest market signal that deployment, not demand, is the problem.
What I am watching
1. Data centre power purchase announcements. If major technology companies start securing nuclear and renewable agreements aggressively, the constraint is biting. The Three Mile Island restart for Microsoft was the first major signal. 2. NVIDIA earnings and guidance. GPU inventory building without corresponding deployment signals power constraints. 3. US and EU electricity grid investment policy. Permitting reform could unlock or block the next five years of infrastructure buildout.
Signal 03 severity: High (score: 65/100). The physical constraint on AI growth is real and well-documented. The shift from chips to electricity as the primary bottleneck is confirmed by multiple credible sources including Microsoft's own CEO. The 2,600 GW grid queue and five-year wait times represent a structural barrier that cannot be resolved quickly. However, the severity is not yet Critical because workarounds exist (on-site generation, permitting reform, private wire deals) and the constraint may ease over a three to five year horizon. If the constraint tightens further in the next two quarters, or if NVIDIA's revenue growth decelerates because of deployment limits, the severity will rise.
Lewis Barclay · Personal research, not professional advice · Updated June 2026
