NextFin News - Artificial intelligence has stopped being a software story and become the largest physical infrastructure build-out since the internet, and this time the financing, the demand, and the bottleneck are all different from the dot-com era. That is the argument Lopez set out in a television interview on August 27, 2026, framing the current AI capital-spending wave as the biggest infrastructure upgrade since the internet was first wired.
The numbers behind the claim are large enough to test the comparison. The four largest cloud builders - Microsoft, Amazon, Alphabet, and Meta - are on track to spend roughly $700 billion to $750 billion on capital projects in 2026, up about 77% from $410 billion in 2025 and nearly triple the $226 billion deployed in 2024. Goldman Sachs now expects those four companies alone to commit $5.3 trillion in capital spending between fiscal 2025 and fiscal 2030, up from $4.5 trillion before first-quarter earnings. The question is not whether the money is real. It is whether the economy on the other end of the fiber can pay for it.
The Scale of the Build-Out
The AI infrastructure cycle has moved faster than most investors expected. In 2024, combined capital expenditure by the four hyperscalers was $226 billion. By 2026, that figure has roughly tripled, according to company filings and guidance. Alphabet is guiding toward roughly $185 billion to $200 billion of 2026 capital spending; Amazon is pointing to about $200 billion; Meta has lifted its range to $125 billion-$145 billion; and Microsoft is tracking toward $120 billion or more. Add Oracle and the broader data-center developer universe, and researchers tracking data-center construction put 2026 spending by the largest data-center firms near $750 billion, with more than 23 gigawatts of data-center IT capacity under construction.
What makes this an infrastructure story rather than a chip story is where the money goes. The spending is not just graphics processors. It is the shell around them: purpose-built data centers, high-density power distribution, liquid cooling, long-haul fiber, transformers, and the electrical equipment that turns a warehouse into an AI factory. Nvidia's own trajectory shows the shift. The chip designer's data-center revenue grew from $15 billion in fiscal 2023 to $115 billion in fiscal 2025, with $35 billion in the fourth quarter alone - but the hyperscalers buying those chips are now, in effect, massive construction and utilities companies.
"The largest infrastructure buildout in human history," Nvidia founder and CEO Jensen Huang said of artificial intelligence at the World Economic Forum in Davos, framing AI as a five-layer stack running from energy and chips through cloud data centers, models, and the application layer.
The comparison to the internet build-out of the late 1990s is unavoidable - and that is exactly where the investment debate begins. In the five years after the Telecommunications Act of 1996, telecom companies poured more than $500 billion, mostly financed with debt, into fiber-optic cable, switches, and wireless networks. Four years after the bubble burst, an estimated 85% to 95% of that fiber sat unused, earning the label "dark fiber." The NASDAQ Composite fell 78% from its peak. Telecom capital expenditure peaked at roughly 1.0% to 1.2% of U.S. GDP in 2000, and the poster children of the bust were not just Pets.com and eToys but the equipment makers underneath: Corning's stock fell from nearly $100 in 2000 to around $1 by 2002, and Ciena's revenue dropped from $1.6 billion to $300 million as its shares plunged 98%.
Here is the first tension: AI capital spending has already exceeded that mark. Measured on an annualized basis, capital expenditure by the Magnificent Seven reached about 1.28% of U.S. GDP in the second quarter of 2025 - above the dot-com telecom peak. The scale is not in doubt. The question is whether this time the overbuild is smaller, or the consequences are.
Why the Financing Is Different - and Why That Matters
The most important difference between this cycle and the dot-com overbuild is not the technology. It is the balance sheet. The 1990s fiber boom was debt-financed: telecoms borrowed against traffic projections that never arrived, and when revenue failed to materialize, bankruptcy followed. The AI build-out is being financed largely with equity and operating cash flow from companies with fortress balance sheets.
This distinction changes the failure mode. A debt-financed overbuild ends in defaults, fire sales, and a credit crunch that amplifies the downturn. An equity-financed overbuild ends in write-downs, depressed returns on capital, and multiple compression - painful for shareholders, but not systemic. The 1990s telecoms had to service debt regardless of whether the fiber lit up. Microsoft, Amazon, Alphabet, and Meta can simply slow spending, and their cost of capital does not blow out.
There is a second difference: demand is contracted before construction. Hyperscalers report that their markets are supply-constrained rather than demand-constrained. Microsoft's commercial remaining performance obligations - contracted but not yet delivered services - reached $678 billion by the end of its fiscal 2026, an 84% year-over-year increase, and the company has disclosed an $80 billion backlog of Azure orders it cannot fulfill because of power constraints. A backlog is not revenue, and the dot-com era also had contracts: long-term indefeasible rights of use on fiber that went dark when the telecom customers themselves went bankrupt. But a $678 billion contracted base across the hyperscalers is a stronger signal than the "if you build it, they will come" logic that drove the fiber boom.
Even so, the strain is showing. Amazon is projected to run negative free cash flow of $17 billion to $28 billion in 2026 as capex consumes cash, and Alphabet turned cash-flow negative in the second quarter for the first time as its long-term debt rose 111% in the first half of the year. The builders are not leveraged like 1990s telecoms, but they are no longer printing free cash flow either. The margin for error is thinner than the balance sheets suggest.
The Real Constraint Is Power, Not Chips
The transmission mechanism of this cycle runs through the electrical grid, not the semiconductor fab. For two years the market's bottleneck question was "can TSMC make enough chips?" That question is being replaced by "can anyone power the racks?"
U.S. data-center power demand is forecast to more than double from 31 gigawatts in 2025 to 41 GW in 2026 and 66 GW in 2027, according to Goldman Sachs Research, which based its forecast on facility-level development schedules and a 70% capacity-utilization assumption. The acceleration is steep: year-over-year capacity additions are scheduled to reach 13.6 GW in 2026 and 36.3 GW in 2027, compared with realized additions of 6.4 GW in 2024 and 8.5 GW in 2025. Data centers' share of total U.S. peak summer power demand is projected to rise from 4.1% in 2025 to 8.5% in 2027.
That doubling is arriving against a grid that moves slowly. Interconnection queues across most U.S. grid regions are running four to six years on average. Transformer lead times that once averaged six months now stretch past two years, according to industry reporting, and specialized high-voltage units capable of handling AI data-center loads have become the single biggest cause of project delays, rippling through cooling systems and skilled labor. Microsoft's chief executive, Satya Nadella, has acknowledged that power grid constraints are the main reason the company cannot clear its Azure backlog - and that GPUs are sitting idle in inventory because there is no electricity to install them.
This is the mechanism that separates a cyclical spending wave from a structural one. A cyclical shortage - too few GPUs for six quarters - corrects when fabs add capacity. A grid constraint corrects on the timeline of permitting, transmission construction, and generation build-out: years, not quarters. That is why the AI infrastructure cycle is likely to be longer and less mean-reverting than a normal hardware cycle. The bottleneck is governed by physics and regulation, not by code.
It also flips the value chain. In a chip-constrained world, the GPU designer captures most of the margin. In a power-constrained world, value migrates to the assets that can site, energize, and operate the campuses: utilities with interconnection priority, independent power producers, natural-gas peakers, nuclear developers, and the electrical-equipment suppliers. The market has begun to price this - electrical equipment and power names have rerated - but the rotation is early.
The Return Question: Who Actually Gets Paid
The bear case against the AI build-out is simple and it is getting stronger: the revenue on the other end is not keeping up with the capital. OpenAI, the largest consumer of this infrastructure, has an annual recurring revenue run rate of roughly $20 billion - impressive for a company that barely had consumer products three years ago, but only about 3% of the projected 2026 hyperscaler capex total. Anthropic's run rate, while growing fast, occupies a similar position. The pure-play AI vendors are growing rapidly but from modest bases relative to the capital being deployed.
There is, however, an early counter-signal. Research from Exponential View estimates that AI revenue is now roughly covering estimated infrastructure depreciation - an important milestone, though not proof that the entire build-out has paid for itself. Depreciation is an accounting charge, not a cash cost of replacement, and it lags the actual spending: today's depreciation reflects yesterday's smaller capex base. If capex keeps compounding at 70% a year, depreciation will catch up to revenue quickly unless revenue compounds just as fast.
This is where the second-order thinking matters. The consensus worry is "AI is a bubble." The more precise question is: what kind of bubble, if it is one? The dot-com bubble had two layers - the application layer (Pets.com) and the infrastructure layer (fiber). The application layer went to zero. The infrastructure layer, despite 85% to 95% dark fiber and a 78% equity drawdown, became the backbone of the modern internet. The $500 billion fiber overbuild was, in the long run, not wasted. It was just financed badly and priced expensively.
The same pattern is plausible here. Even if AI application revenue disappoints, the compute, power, and networking capacity being built will exist and will be used - at prices that clear the market. The infrastructure endures; the equity returns do not necessarily follow. This is the distinction investors are failing to make: a build-out can be economically necessary and still be a bad stock investment if you pay too much for the builders.
The Strongest Counter-Thesis - and What Would Prove It Wrong
The strongest case against the structural view is the dark-fiber analogy itself. The 1990s telecoms also believed demand would compound forever. They were right about demand growth and wrong about pricing power: bandwidth costs fell 90% as supply glutted the market, and the equity holders were wiped out anyway. If AI inference costs fall at a similar pace - and there is early evidence of price compression in commoditized inference workloads - then the revenue per watt of data-center capacity could collapse even as utilization stays high. The builders would own full racks and shrinking margins.
This counter-thesis attacks the core of the structural argument: it concedes the build-out is real but denies that the builders capture the value. It is backed by the historical record of every infrastructure overbuild from railroads to fiber - capacity endures, pricing power does not.
The falsifying signal is specific. This article's threshold: if AI-related revenue at the four hyperscalers does not reach at least 15% of their combined capital expenditure within the next four quarters, the "value migrates to infrastructure" thesis weakens materially and the dark-fiber comparison strengthens. Conversely, if AI revenue holds above that threshold while gross margins on inference remain stable, the structural call is confirmed. That is the number to watch - not utilization, not gigawatts, but the revenue-to-capex ratio.
What to Watch
The cyclical-versus-structural call, separated by horizon: the GPU shortage is cyclical and will revert as supply catches up. The power and transmission constraint is structural and will not revert on its own. The equity-financed balance sheets mean the downside is return compression, not systemic default. The build-out itself is a regime shift; the returns to the builders are not guaranteed.
Short term (6-12 months): sentiment and liquidity dominate. Any sign of capex discipline from a hyperscaler - a guided reduction, a delay - would trigger a sharp de-rating across the AI infrastructure complex, because the trade is crowded and the valuation embeds uninterrupted compounding.
Medium term (1-3 years): fundamentals decide. The base case is continued high spending with moderating growth as power constraints bind. The upside case is that AI revenue compounds fast enough to cover depreciation and then some, validated by a revenue-to-capex ratio above 15%. The downside case is that inference price compression accelerates, the ratio stalls below the threshold, and the market re-rates the builders from growth to utility multiples.
Long term (5+ years): the infrastructure will exist regardless. Like the fiber glut of the 2000s, today's data centers, power lines, and cooling plants will become the foundation of the next productivity layer - even if the companies that build them do not all survive as good investments.
Who benefits and who is exposed: the beneficiaries are the bottleneck owners - electrical-equipment suppliers, utilities and independent power producers with interconnection priority, liquid-cooling specialists, and the chip designers with pricing power in training workloads. The exposed are the pure-play builders whose returns depend on sustained hyperscaler generosity, and any supplier competing on price in commoditized inference hardware.
Watch these signals: the quarterly revenue-to-capex ratio at the four hyperscalers; transformer lead times and interconnection-queue lengths; and any shift in guidance from Alphabet, the only hyperscaler to explicitly signal spending increases beyond 2026. If core AI revenue fails to keep pace with the spending, the infrastructure will still be built - but the market will learn, again, that being right about the build-out is not the same as being right about the stock.
The dot-com era left two lessons: the infrastructure always outlasts the bubble, and the equity holders do not always share in the surplus. The AI build-out is the biggest infrastructure upgrade since the internet. Whether it is the best investment since the internet is a different question - and the next four quarters of revenue data will answer it.
Explore more exclusive insights at nextfin.ai.
