One Trillion Dollars and Still Stuck: The Physical Bottlenecks Throttling the AI Build-Out

For two years, the AI story has been told in one number: hyperscaler capital expenditure. Now even that number has outgrown its own frame. In an August 7 report, Goldman Sachs Research published the most comprehensive attempt yet to measure how much the world is actually spending on artificial intelligence — and concluded that global AI-related investment will exceed $1 trillion in 2026, including $581 billion in the United States alone.

Yet the more consequential story may be what that money is running into. With earnings season complete, the forecasts keep climbing — JPMorgan sees $697 billion of US spending this year, Bank of America sees “a path toward ~$1.2 trillion” by 2027 — but a growing body of evidence suggests the binding constraint on AI is no longer capital. It is transformers, switchgear, fiber-optic crews, and the time it takes to bring electricity to a building that does not exist yet.

The $1 trillion, measured properly

The commonly cited figure for AI spending is the consensus forecast of roughly $794–800 billion in 2026 capex by the big US hyperscalers — Microsoft, Alphabet, Amazon, and Meta. Goldman’s economists, led by Joseph Briggs, argue that number misleads in both directions.

It understates global investment by around $200 billion, because it leaves out AI spending by private companies (OpenAI, Anthropic, and the neoclouds among them), by public US companies outside the hyperscaler cohort, and by AI-exposed firms abroad — especially in Asia. And it overstates US investment by roughly $200 billion, because American tech giants build globally: a large share of that capex lands in Ireland, Singapore, Japan, and elsewhere.

Goldman’s adjustments produce the $1 trillion global / $581 billion US split. Methodologically, the team also subtracted 2022 capex levels from hyperscaler totals, on the logic that only the increment above pre-AI-boom baselines is genuinely AI-driven — a conservative touch that makes the headline more credible, not less.

Zoom out and the sums get stranger. Goldman’s companion “Tracking Trillions” analysis sketches a baseline of roughly $7.6 trillion of AI capital investment between 2026 and 2031 across compute, data centers, and power — with some scenarios running to $8 trillion. For comparison, that is on the order of building the entire US interstate highway system several times over, compressed into five years.

Why the money isn’t the problem

A widely shared August 14 analysis from Yahoo Finance, framed around those very forecasts, made the sharper point: the bottleneck isn’t cash. The physical supply chain is the throttle.

Electrical equipment first. Nearly half of US AI data centers planned for 2026 are reportedly facing delays or cancellation because of shortages of transformers, switchgear, and other basic electrical gear. Large power transformers — the house-sized units that step grid voltage down to usable levels — have lead times that have stretched toward multi-year waits. You cannot expedite a transformer with a purchase order; the global manufacturing capacity for core electrical steel and finished units is the constraint, and it expands on its own schedule.

Labor second. Construction contractors keep flagging the shortage of skilled trades — electricians, pipefitters, linemen — needed to deliver projects on clients’ desired timelines. One striking datapoint from NBC’s reporting in early August: by some estimates the US is 58,000 people short of the workforce needed to install the fiber-optic cable that connects data centers to the internet. That is not chip design talent; that is people pulling cable through conduit, a trade that takes months to train and years to master.

Regulation third. Public backlash against data centers is hardening into policy. New York State imposed a one-year moratorium on certain data center development; power hookups in some jurisdictions now face formal audits before approval. Communities that once courted hyperscaler campuses are increasingly skeptical of the water, noise, and grid strain that come with them. Permitting timelines that once ran months now run years.

Chips, still. Memory prices have been soaring, and Nvidia — which can essentially command whatever price it sets for its newest GPUs — remains supply-constrained despite new fabrication capacity coming online. The compute layer of the stack is expensive and scarce at the same time.

What this means for the race

The economics of AI have always been a race between demand curves and supply curves. What is changing in 2026 is which supply curve binds first. Through 2024 and 2025, the story was GPU allocation — who got Hopper and Blackwell-class silicon, and how much. In 2026, the binding constraints are moving down the stack into domains with far less elasticity: utility interconnection queues, transformer foundries, and vocational labor pipelines.

This has three practical implications.

First, the timeline debate is miscalibrated. Analysts arguing about whether AGI arrives in 2027 or 2035 are implicitly assuming compute arrives when purchased. If nearly half of planned 2026 US capacity is delayed, effective compute in 2027 will lag the capex-derived forecasts that modelers feed into their scaling predictions. Bottlenecks in the physical layer quietly become bottlenecks in capability.

Second, capital intensity will keep rising even as efficiency improves. Microsoft alone expects to spend roughly $175 billion on AI infrastructure in calendar 2026, with more than $50 billion in the current quarter. Bank of America’s analysts estimate hyperscaler capex could reach roughly $1.1 trillion in 2027. Better chips and better models do not shrink these budgets; they simply buy more constraint relief per dollar. The race to secure power and equipment is becoming a competitive moat in itself — whoever energizes a campus first ships intelligence first.

Third, the geography of AI is being decided by grid operators, not just CEOs. If US investment is $581 billion of the global trillion, the remaining $419 billion+ follows electricity, land, and permissive permitting. That is why AI campus announcements cluster in Texas, the Gulf states, and increasingly Asia — and why “sovereign AI” budgets in Japan, Korea, and the Gulf states are growing into a distinct capex category. Power availability is now industrial policy.

The trillion-dollar question

None of this means the AI build-out is in trouble — capex forecasts keep rising, not falling. It means the build-out is becoming a physical-economy story rather than a purely financial one. The interesting question for the next eighteen months is not “how much will be spent?” — that answer keeps being revised up — but “how much of what is spent actually converts into energized, cooled, connected racks?”

A trillion dollars buys a great deal of silicon. It cannot conjure a three-year transformer queue, a 58,000-person fiber shortfall, or a substation that the county commission has not yet approved. In 2026, the scarcest asset in AI isn’t GPUs or even capital. It is time — measured in construction schedules, interconnection studies, and the training pipelines for the people who physically build the future’s compute.

The industry that promised to compress the world’s thinking into milliseconds is discovering that its own foundations move at the speed of concrete, copper, and human hands.