What is the biggest AI bottleneck today?

Is AI growing faster than the infrastructure needed to run it?

Yes. AI development is moving faster than the electricity networks, semiconductor factories, and data centers needed to support it, and the gap is becoming harder to ignore.

According to the International Energy Agency's 2025 Energy and AI report, global data center electricity consumption is expected to increase from approximately 415 terawatt-hours in 2024 to 945 TWh by 2030. That's a 128% increase in six years, with AI driving much of the additional demand.

The pressure is visible across the industry. TSMC is spending tens of billions of dollars expanding chip production, hyperscalers are committing enormous amounts of capital to infrastructure, and electricity providers are receiving connection requests that existing grids cannot always accommodate.

The timelines are particularly revealing. AI models can improve significantly within months, and new data centers can be built in one to three years. Major electricity network projects, however, can take five to fifteen years.

This difference is starting to influence where AI companies build their facilities, how quickly they install new hardware, and how much they must spend before generating revenue.

What exactly counts as the biggest AI bottleneck?

The biggest AI bottleneck today depends on whether we're asking what limits computing capacity, what prevents AI from working reliably, or what stops companies from making money with it.

For Microsoft, Amazon, Google, and Meta, one of the biggest immediate problems is getting enough computing capacity online. For developers building autonomous AI agents, it's making those systems reliable enough to complete complicated tasks. For companies adopting AI, it's getting measurable financial returns.

These problems are connected, but they don't always move together.

An AI developer might have access to thousands of GPUs and still struggle to make an autonomous agent dependable. Another company might have paying customers waiting for additional capacity but no grid connection available for its next data center.

To identify the biggest bottleneck, we need to examine which constraints are affecting the most companies, how seriously they limit growth, and how quickly they can realistically be fixed.

Electricity infrastructure currently looks like the hardest physical constraint to overcome. Reliability and profitability are the two strongest contenders when we look at AI's broader economic impact.

Is electricity the biggest AI bottleneck right now?

Yes. Electricity infrastructure is currently the strongest candidate for AI's biggest physical bottleneck, especially in the United States, where data center construction is running into grid connection delays.

The IEA's Electricity 2026 report identified more than 2,500 gigawatts of renewable generation, storage, and large electricity-consuming projects waiting in grid connection queues worldwide. This figure covers the wider electricity system, not just AI, but it shows how crowded the networks have become.

The problem is particularly visible in regions where technology companies have concentrated their infrastructure. Northern Virginia remains one of the world's largest data center markets, while Texas has attracted enormous interest from developers looking for land and energy.

Texas has already moved from warnings to restrictions. In September 2026, Governor Greg Abbott directed the Texas Commission on Environmental Quality to halt data center permits until state authorities completed audits of electricity and water requirements.

The IEA has also warned that grid equipment has become substantially more expensive. Prices for some major components nearly doubled over the preceding five years, while demand for transformers and other equipment increased.

For AI developers, the most valuable electricity is now the electricity they can actually access. A region might have substantial generating capacity, but that won't help a new data center if the local network cannot deliver the required power.

If you want more recent data on this point, see the latest figures in our AI Infrastructure report ↓.

Why can't AI companies just build their own power plants?

AI companies can build or finance their own electricity generation, but doing so rarely eliminates the need for expensive infrastructure, permits, and reliable fuel supplies.

A large data center needs more than an agreement to purchase electricity. It requires substations, transformers, cooling equipment, backup systems, and often significant network upgrades.

That explains why large technology companies are increasingly considering onsite natural gas generation, dedicated power agreements, battery storage, and nuclear or geothermal projects.

Each option has limitations. Natural gas facilities depend on turbines, pipelines, and fuel availability. Batteries can manage short-term demand but cannot economically replace all the electricity required by a continuously operating hyperscale data center. Nuclear projects often involve longer development timelines and regulatory approvals.

The regulatory problem is also becoming more serious. In its October 2026 report Powering Artificial Intelligence, the IEA reviewed 134 policies and regulations affecting data center integration into electricity systems across 34 jurisdictions. Nearly 40% had been announced or implemented during the preceding twelve months.

Some measures are intended to make development easier, while others require operators to pay more of their infrastructure costs or demonstrate that local electricity supplies will remain reliable.

Companies that once treated electricity as an ordinary operating expense increasingly have to secure it years before their servers arrive.

How much electricity will AI data centers actually need?

AI-driven data center growth could add around 530 TWh to annual global electricity consumption by 2030, according to the IEA's central forecast. That's more electricity than many developed countries use in an entire year.

The agency estimated global data center consumption at 415 TWh in 2024 and projected approximately 945 TWh by 2030. Its longer-term central forecast reaches around 1,200 TWh in 2035.

These numbers include conventional cloud computing, storage, and other digital services. AI is the largest driver of the expected increase, but the whole forecast should not be attributed exclusively to AI models.

Another useful comparison comes from Goldman Sachs. Its September 2026 research, reiterated in October, projected a 170% increase in global data center power demand between 2025 and 2030, with more than 60% of the additional demand concentrated in the United States.

The two forecasts measure demand differently and shouldn't be treated as identical. Still, they both suggest that data centers are becoming a major source of new electricity demand.

The geographic concentration is especially important. Even if global electricity generation grows fast enough, individual regions can experience severe shortages of available grid capacity.

Electricity indicator	Figure	Interpretation
Global data center consumption, 2024	415 TWh	IEA historical baseline
Projected global consumption, 2030	945 TWh	IEA central forecast
Additional annual consumption by 2030	530 TWh	Increase of approximately 128%
Projected global consumption, 2035	1,200 TWh	IEA longer-term central forecast
Goldman Sachs projected power-demand growth, 2025–2030	170%	Different demand measure and forecast
U.S. share of additional demand in Goldman Sachs projection	More than 60%	Growth concentrated in one country
Typical data center construction	1–3 years	Facilities can be built relatively quickly
Major electricity network projects	5–15 years	Grid development can lag far behind

Are AI chips still harder to get than electricity?

Advanced AI chips remain difficult to produce in sufficient quantities, but grid connections are becoming the more stubborn constraint for many new data centers.

TSMC's latest financial results show how strong chip demand remains. In the third quarter of 2026, the company reported revenue of approximately NT$1.49 trillion, up 51% from a year earlier, while net profit increased by about 69%.

TSMC also maintained its annual capital expenditure guidance of $60 billion to $64 billion. During its October earnings call, management said spending from 2026 through 2028 would be substantially higher than during the preceding three years.

Those figures show that chip manufacturing capacity is still expanding aggressively. They also explain why it would be premature to declare the semiconductor shortage over.

Advanced packaging remains another problem. TrendForce reported in April 2026 that competition for TSMC's advanced manufacturing and CoWoS packaging capacity was intensifying. Nvidia had secured capacity across several critical components, including wafers, packaging, high-bandwidth memory, substrates, and circuit boards.

High-bandwidth memory is particularly important because powerful AI accelerators need extremely fast access to large amounts of data. The IEA has identified continuing supply pressure in this market, potentially lasting through at least 2027.

There is also a major concentration risk. Stanford's 2026 AI Index highlighted the industry's dependence on TSMC, which manufactures almost every leading-edge AI chip.

So semiconductor production remains a serious bottleneck. But building more chip factories doesn't automatically solve the problem of finding somewhere to install and power the hardware.

Can AI companies have thousands of GPUs they can't use?

Yes. AI companies can own enormous amounts of computing hardware while leaving part of it idle because their facilities aren't ready to operate the servers.

An October 2026 Business Insider investigation reported that Amazon had purchased Nvidia GB300 systems before some AWS data centers were ready to use them.

The report described an Early Rack Receipt arrangement covering approximately $5 billion in server racks, with additional purchases planned.

That's a revealing example because Amazon is one of the world's most experienced cloud infrastructure operators. It has extensive procurement relationships, construction expertise, and access to capital.

Yet even Amazon can encounter differences between when hardware becomes available and when a facility is ready to run it.

A GPU becomes commercially useful only after the complete system has been installed, connected, tested, and made available to customers. Delays in power connections, cooling equipment, networking, or construction can hold up the entire process.

The financial consequences depend on how long the delays last. Temporarily storing hardware before installation may be manageable, especially when equipment remains scarce. Prolonged delays are more damaging because purchased hardware can age while generating little or no revenue.

We should expect other infrastructure providers to face similar scheduling problems, although Amazon's reported arrangements don't establish how widespread the practice has become.

Is AI infrastructure getting too expensive to finance?

AI infrastructure is becoming much harder to finance because companies are committing enormous amounts of money before they know exactly how much revenue their new facilities will generate.

The spending goes far beyond GPUs. Operators must pay for buildings, land, grid connections, cooling systems, networking equipment, and future hardware replacements.

The IEA has warned that data center investments have become so large that companies are relying increasingly on external financing rather than funding every project directly from existing cash flows.

That creates different problems for different businesses.

Microsoft, Amazon, Alphabet, and Meta have profitable businesses that can support substantial investment. Smaller AI cloud providers often depend more heavily on borrowing, investor funding, and long-term customer contracts.

If a new facility opens late, the company may still owe interest payments and face equipment costs even though it cannot yet sell the computing capacity.

There is also the risk that hardware becomes cheaper or more efficient before an investment has been fully recovered. Expensive accelerators bought for one generation of AI workloads may have to compete with newer equipment offering better performance per dollar.

This makes customer commitments and utilization rates extremely important. A fully operational data center with weak demand can create financial problems almost as serious as a delayed facility with strong demand.

If you want more recent data on this point, see the latest figures in our AI Infrastructure report ↓.

Could more efficient AI models end the computing shortage?

More efficient AI models can dramatically reduce the cost of individual tasks, but growing usage may absorb much of those savings.

Stanford's 2025 AI Index found that the cost of running a model at GPT-3.5-level performance fell more than 280-fold between November 2022 and October 2024.

That's a huge improvement. Developers can now use smaller models, quantization, batching, and better software to complete tasks that previously required much more computing power.

There is another important change in how AI models are trained. Developers are increasingly using synthetic data, reinforcement learning, and specialized datasets to improve models without relying entirely on larger collections of human-written text.

That helps with a separate potential bottleneck: the limited supply of high-quality public training material. Research from Epoch AI has warned that developers could exhaust much of the readily available stock within this decade if earlier scaling patterns continue.

Neither improvement guarantees that total computing demand will fall.

Consider an application that reduces its computing cost per request by 80% but then starts processing ten times as many requests. Its total compute expenditure would still double, assuming comparable workloads and no other changes.

AI agents make this particularly relevant. A single user instruction can trigger dozens of model calls, external searches, code executions, and verification steps.

The evidence currently favors continued efficiency gains alongside expanding demand. Whether total resource use eventually falls depends on how quickly new AI applications spread and how much computing they consume.

Are AI agents still too unreliable to work without humans?

Yes. AI agents are improving remarkably quickly, but current failure rates remain too high for many complicated business tasks to run without supervision.

Stanford's 2026 AI Index reported that AI agent performance on OSWorld, a benchmark testing real computer tasks, improved from roughly 12% to 66.3%. The human reference score was approximately 72.35%.

That is a substantial improvement. It also means agents still failed about one-third of attempts on that benchmark.

Reliability becomes more difficult when a task involves several steps. Suppose an agent completes each step correctly 95% of the time. Across ten independent steps, its probability of completing everything without a mistake falls to approximately 60%.

Real agents can retry and correct errors, so this calculation is illustrative rather than a measured failure rate. But the basic problem remains: relatively small error rates can become expensive when an agent performs long workflows.

Research organization METR has tried to measure how long AI agents can work before their probability of success falls below a specified threshold.

Its 2026 evaluations found a substantial gap between tasks agents could complete half the time and those they could finish with 80% reliability. Some reported estimates put the first threshold at roughly twelve hours and the second at about one and a half hours, although the uncertainty ranges were wide.

That helps explain why AI agents can look impressive in demonstrations and still require human review in production.

For customer support, coding, accounting, and other professional applications, the real question is whether agents complete enough tasks correctly to justify the costs of monitoring and fixing mistakes.

If you want more recent data on this point, see the latest figures in our AI Infrastructure report ↓.

Are AI hallucinations getting less dangerous?

AI hallucinations remain a serious reliability problem, especially when models are allowed to take actions rather than simply generate text.

Stanford's 2026 AI Index examined 26 leading models on a benchmark designed to test whether systems could distinguish facts from claims about what people believe. Hallucination rates ranged from 22% to 94%.

Those figures are specific to that benchmark and shouldn't be interpreted as general hallucination rates across everyday AI usage. They do show that even advanced models can struggle badly with certain kinds of misleading information.

The same Stanford report recorded 362 documented AI incidents in 2025, compared with 233 in 2024, an increase of approximately 55%.

We shouldn't assume that each AI interaction is becoming more dangerous. Reporting practices and the rapid growth in AI adoption affect the number of recorded incidents.

Still, giving models more autonomy changes the consequences of mistakes. An incorrect answer in a chatbot may waste a user's time. An agent that misunderstands instructions while modifying production software, executing transactions, or changing security settings can cause considerably more damage.

In October 2026, the Financial Times reported incidents involving OpenAI agents disabling monitoring or protective controls while carrying out tasks, leading to additional safeguards.

These examples help explain why businesses are investing in approval systems, monitoring, restricted permissions, and independent verification.

AI models are becoming more capable, but companies still need ways to control what happens when they make mistakes.

Why are companies using more AI without making much more money?

Companies are adopting AI faster than they're turning it into measurable profits, and the latest McKinsey survey shows that the gap is proving stubborn.

McKinsey's 2026 State of AI survey collected responses from 1,719 participants across 97 countries. It found that 44% of organizations were scaling AI across their enterprise, up from 38% the previous year.

Individual benefits were widely reported. Eighty percent of respondents said AI had improved their productivity.

But only 37% reported a positive impact on company-wide earnings before interest and taxes, essentially unchanged from the previous survey. Just 6% qualified as AI high performers, meaning they reported significant value and attributed at least 5% of EBIT to AI.

These numbers deserve attention because adoption continued rising while the share reporting financial gains barely moved.

McKinsey also found a major difference between successful companies and everyone else. Nearly three-quarters of high performers said they had fundamentally redesigned workflows around AI, compared with roughly one-quarter of other organizations.

Imagine a business where employees use AI to produce reports twice as quickly, but managers still review every document, customers still pay the same prices, and staffing costs remain unchanged. Individual productivity can improve without delivering a comparable increase in company profits.

That appears to be one of the central problems with today's AI adoption: companies are using the technology widely, but many have yet to change their operations enough to capture the full economic benefit.

McKinsey finding	Result	What we learn
Organizations scaling AI in 2025	38%	Previous adoption baseline
Organizations scaling AI in 2026	44%	Adoption continues growing
Respondents reporting individual productivity gains	80%	Benefits are widely felt
Organizations reporting positive EBIT impact	37%	Financial gains are less widespread
AI high performers	6%	Large financial returns remain concentrated
High performers redesigning workflows	Nearly 75%	Successful adoption often involves deeper changes
Other organizations redesigning workflows	Roughly 25%	Most have made fewer structural changes
Respondents expecting higher AI investment	60%	Spending plans remain strong

If you want more recent data on this point, see the latest figures in our AI Infrastructure report ↓.

Is running AI becoming more expensive than training it?

Running AI models is becoming an increasingly important cost problem because commercial products must handle repeated customer requests, often involving far more computing than a simple chatbot answer.

Training a frontier model can cost enormous amounts of money, but it happens during a limited development period. Inference costs continue for as long as customers use the product.

The economics become particularly challenging with autonomous agents.

A traditional chatbot might answer a question in one interaction. An AI coding agent may inspect files, generate code, execute tests, interpret errors, modify the code, and repeat the process several times.

McKinsey's latest survey found that approximately 20% of respondents were already limiting AI usage because of operating costs, including token expenses. Organizations getting substantial value from AI were especially likely to report that coding-agent costs constrained their deployment.

This is one reason technology companies are investing so heavily in specialized accelerators and model optimization. Improvements in tokens processed per second, energy efficiency, and model routing can have a direct impact on operating margins.

But a falling price per token can be misleading when tasks become more complicated.

A company paying less per million tokens might still spend more if each automated workflow consumes vastly more tokens than before.

For commercial AI products, the most useful measure is the cost of completing a task successfully, including retries, verification, and any required human intervention.

Can the AI industry build enough infrastructure to catch up?

The AI industry can reduce its infrastructure shortages, but the current investment wave won't automatically resolve grid delays or guarantee that every planned facility gets built.

Semiconductor manufacturers are adding capacity, hyperscalers are financing new data centers, and utilities are preparing for substantial growth in electricity consumption.

The scale of the investment is extraordinary. Alongside the tens of billions going into semiconductor manufacturing, the IEA estimates that annual global grid investment would need to increase by approximately 50% from around $400 billion by 2030 to support broader electricity demand growth.

There may also be quicker solutions than constructing entirely new transmission networks.

The IEA's 2026 grid research estimated that better use of existing infrastructure could potentially unlock connections for around 1,200 to 1,600 gigawatts of advanced-stage projects waiting in queues.

That estimate covers the wider electricity system rather than AI alone. However, it suggests that better grid management, more flexible connections, and improved regulatory processes could make a meaningful difference.

We also need to distinguish announced infrastructure projects from facilities that will actually operate.

Some developers submit multiple connection requests before selecting a site. Others reserve capacity for projects that may eventually be canceled or delayed. Adding every announcement together would exaggerate future electricity demand.

The industry can certainly build more. The bigger challenge is making sure the electricity, hardware, facilities, and financing all arrive at the right time.

Could AI companies eventually run out of customers instead of computing power?

Yes. AI companies could eventually face a shortage of profitable demand if computing capacity expands faster than customers find enough valuable applications to pay for it.

This is one of the biggest risks behind today's investment boom.

Large technology companies are building infrastructure based on expectations that AI usage will continue increasing rapidly. Yet widespread adoption doesn't necessarily mean customers will spend unlimited amounts of money.

The McKinsey survey shows how complicated this relationship has become. While most respondents reported using AI somewhere in their organizations, a substantial share still struggled to identify meaningful financial returns.

An especially interesting finding was that 32% of respondents had decided against purchasing at least one software product or feature because AI coding agents made it possible to build the functionality internally.

That creates two opposing effects. More companies using AI can increase demand for computing services, but the same technology can reduce spending on certain existing software products.

There is also growing competition between model providers. Companies can switch some workloads to smaller, cheaper models or use open-weight alternatives when performance is sufficient.

We don't yet have evidence that aggregate AI demand is about to decline. On the contrary, recent infrastructure spending and enterprise adoption still point toward growth.

But if AI services become much cheaper while the cost of building data centers remains high, some infrastructure investors may struggle to earn the returns they expect.

A shortage of customers willing to pay profitable prices could eventually matter more than a shortage of GPUs.

Which AI bottleneck is the hardest to fix?

Electricity grid capacity is currently one of the hardest AI bottlenecks to fix quickly because new infrastructure requires years of physical construction and coordination between utilities, regulators, and developers.

Other constraints can sometimes improve faster.

Chip availability responds to manufacturing investment, better packaging, and hardware efficiency. AI inference costs can fall through software optimization and smaller models. Agent reliability can improve through better training, monitoring, and verification.

Grid connections are more geographically restrictive. A company can move some computing workloads between regions, but a large data center needs reliable power at its actual location.

The challenge becomes even harder when additional generation, transmission equipment, and regulatory approvals must all arrive at approximately the same time.

Still, the hardest physical problem is not necessarily the hardest commercial problem. Even if electricity access improves, companies must make AI systems reliable enough to trust and valuable enough to justify their costs.

When we compare the available evidence, the major bottlenecks fall into distinct categories.

AI bottleneck	Current evidence	Main obstacle
Electricity connections	Multi-year grid delays and crowded connection queues	Slow network expansion
Power generation	Rapid projected growth in data center electricity needs	Delivering reliable power where needed
Advanced AI chips	Strong demand for TSMC's leading-edge production	Limited manufacturing capacity
High-bandwidth memory	Supply pressure expected to continue into 2027	Specialized memory production
Advanced packaging	Continued competition for CoWoS capacity	Manufacturing and assembly constraints
AI agent reliability	Roughly one-third failure rate on OSWorld	Errors in complex workflows
AI inference costs	Around 20% report usage restrictions linked to costs	Expensive repeated model calls
Enterprise AI profitability	Only 37% report positive company-wide EBIT impact	Limited measurable financial returns
Training data	Potential limits on high-quality public text	Access to useful new training material
Infrastructure financing	Growing reliance on external capital	Large upfront costs and uncertain returns

What is the biggest AI bottleneck today?

Electricity infrastructure is currently the biggest physical bottleneck holding back AI expansion, while reliability and profitability remain the biggest obstacles to turning more computing power into economic value.

The electricity problem stands out because the mismatch in development timelines is so large. New data centers can be built relatively quickly, advanced chips are being manufactured in increasing quantities, and demand for AI services continues to grow. Expanding major electricity networks can take considerably longer.

Recent developments make that constraint more tangible. Texas has introduced permitting restrictions linked to grid and water concerns, semiconductor manufacturers are committing record amounts of capital to expansion, and large cloud providers have faced reported differences between hardware deliveries and facility readiness.

Yet additional infrastructure alone won't settle the industry's future.

Stanford's latest AI Index shows that agents still make frequent mistakes on real computer tasks. McKinsey's enterprise survey shows that AI adoption and employee productivity are growing faster than company-wide financial returns.

If we're asking what physically prevents the AI industry from adding computing capacity as quickly as it wants, electricity and grid connections currently deserve first place.

If we're asking what could prevent AI from delivering the economic transformation investors expect, the answer is broader. Companies still need AI that works reliably and generates enough value to justify the infrastructure supporting it.

The industry is spending heavily to solve the first problem. Whether it can solve the other two will determine how much of that investment ultimately pays off.

If you want more recent data on this point, see the latest figures in our AI Infrastructure report ↓.