One More Thing in AI · Deep Dive

The AI Spending Boom: Bubble, Build-out, or Both?

AI is producing real revenue. It is also producing the largest corporate infrastructure build-out in modern history, with financing structures that increasingly resemble the telecom boom.

Research current to 18 July 2026

Scott Galloway recently argued that the first stages of an AI-bubble unravelling are visible in the same places that warned of the dot-com crash: extravagant valuations, circular financing and infrastructure suppliers dependent on customers that burn cash.

The comparison is useful, but Pets.com is not the right analogue. The more revealing precedent is the late-1990s telecom build-out: real demand, sound technology and enormous long-term value, but also too much capacity, poorly matched financing and unrealistic expectations about how quickly customers would arrive.

First, define “AI spending”

There is no single honest number. Current estimates measure different things and should not be added together:

Measure	2026 estimate	What it includes
Worldwide AI-related spending	$2.59tn	AI-enabled devices, infrastructure, software and services
Dedicated AI infrastructure	$487bn	Servers, accelerators, storage and associated infrastructure
Five large U.S. cloud providers	About $750bn	Total company capex; substantial, but not entirely AI

Sources: Gartner’s broad spending forecast, ¹ IDC’s infrastructure forecast, ² and S&P Global’s estimate for Alphabet, Amazon, Meta, Microsoft and Oracle. ³

The Federal Reserve calculated that capex at those five companies reached $412 billion in 2025, or roughly 1.31% of U.S. GDP. Current guidance implies another extraordinary acceleration in 2026. ⁴

What the biggest spenders are actually saying

Company	Spending marker	Demand evidence	Watch-out
Amazon	About $200bn 2026 total capex	AWS Q1 sales +28% to $37.6bn	TTM free cash flow fell to $1.2bn; capex also covers logistics, satellites and robotics
Microsoft	About $190bn calendar-2026 capex	AI revenue run-rate above $37bn; Azure +40%	Large-customer concentration influences demand and backlog
Alphabet	$180bn–$190bn 2026 capex	Cloud +63% to over $20bn; backlog above $460bn	Capex is expected to rise again in 2027
Meta	$125bn–$145bn 2026 capex	Total revenue +33%; AI supports advertising performance	Standalone AI monetisation remains difficult to isolate
Oracle	Roughly $55.7bn implied FY2026 investment	OCI +93%; RPO reached $638bn	Free cash flow was negative $23.7bn; financing and customer concentration are materially higher

The periods and accounting definitions differ. These are disclosed company-wide capex figures or an implied figure in Oracle’s case. They are not a clean estimate of AI-only cash expenditure. Sources: company results and earnings commentary. ^{5 6 7 8 9}

Why this is not simply Pets.com 2.0

The companies writing the largest cheques are not newly listed retailers with Super Bowl advertisements and no customers. They are incumbents with profitable advertising, cloud, commerce and software franchises.

And monetisation is visible. Nvidia’s latest quarterly data-centre revenue reached $75.2 billion, up 92% year over year. Broadcom reported $10.8 billion of quarterly AI semiconductor revenue, up 143%. ^{10 11}

The real danger: connected balance sheets

Cloud companies invest in frontier AI labs. Those labs use the capital to buy cloud capacity. Long-term compute contracts support financing for data-centre developers. Developers buy chips from suppliers that may also hold equity in their customers.

None of this automatically makes the revenue unreal. But it means apparently independent demand signals may originate from a much smaller number of economic customers.

Oracle’s disclosure is especially revealing: much of its recent backlog growth came from large AI contracts, while customers prepaid for, or supplied, $75 billion of GPUs. The arrangement lowers Oracle’s immediate funding requirement, but also illustrates how unusual the financing chain has become. ⁹

The Bank for International Settlements describes some data-centre structures as “shadow borrowing”: debt-like obligations placed in special vehicles and serviced by leases or hyperscaler guarantees. These structures connect technology companies with private-credit funds, insurers and bank funding lines. ¹²

The enterprise reality check

The industry’s upstream signals remain strong, but downstream customers are becoming less tolerant of uncontrolled consumption. Uber reportedly exhausted its annual AI budget in four months and later introduced employee usage limits. ¹³

That does not prove demand is collapsing. It shows the market moving from experimentation to procurement discipline.

PwC found that the top 20% of surveyed companies captured 74% of measured AI returns. ¹⁴ The Federal Reserve similarly found consistent productivity gains in controlled studies but little evidence so far that those gains have lifted aggregate sector productivity. ¹⁵

AI is delivering value. The problem is that value remains concentrated among companies capable of redesigning workflows, data and incentives, not merely buying more tokens.

The new companies entering the spending chain

Frontier model companies

OpenAI, Anthropic and SpaceX/xAI are now major infrastructure customers as well as technology companies. Their funding rounds support enormous compute commitments, but funding-round valuations and self-reported revenue run-rates are not substitutes for audited economics.

OpenAI announced $122 billion of committed capital at an $852 billion post-money valuation. ¹⁶ Anthropic announced a $65 billion round at a $965 billion valuation and said its annualised revenue run-rate had crossed $47 billion. ¹⁷

Neoclouds and AI-factory developers

• CoreWeave reported $99.4 billion of backlog, but its latest quarter also included a $740 million net loss, $536 million of interest expense and $7.7 billion of equipment purchases. ¹⁸
• Nebius signed a Meta agreement worth up to approximately $27 billion over five years. ¹⁹
• Crusoe says it has 4.9 gigawatts of contracted infrastructure and a development pipeline exceeding 40 gigawatts. ²⁰
• Nscale and Lambda are expanding through combinations of equity, secured credit and long-term infrastructure agreements. ^{21 22}

These companies are the closest modern equivalent to the telecom infrastructure layer. Their central risk is the mismatch between long-lived financing and short-lived, rapidly depreciating GPUs.

The wider supply chain

Spending is spreading into Nvidia, AMD, Broadcom, Micron, SK hynix, Samsung, TSMC, Arista, Vertiv, Eaton, Schneider Electric, GE Vernova, Caterpillar, Bloom Energy, power producers and data-centre landlords. This can extend the industrial cycle, but it also moves bottlenecks into electricity, memory, turbines, transformers and construction.

China and the Gulf

China’s parallel stack includes Alibaba, Tencent, ByteDance, Baidu and Huawei. In the Gulf, the central participants include MGX, G42, Core42, Khazna and Mubadala in the UAE, and HUMAIN, PIF and stc/center3 in Saudi Arabia.

Microsoft has announced $15.2 billion of investment in the UAE, including more than $7.9 billion scheduled from 2026 through 2029. ²³ These sovereign projects add a strategic motive to the cycle: capacity may be built for national influence and security even when near-term commercial returns are uncertain.

Three ways the cycle could end

The following probabilities are judgement calls, not forecasts. Their purpose is to make the thesis falsifiable.

35%	Productive build-out Inference demand grows faster than prices fall. Utilisation remains high, enterprise renewals broaden and infrastructure returns justify the spending.
45%	Digestion and consolidation Capex growth slows. Weaker labs and neoclouds merge or restructure, while hyperscalers keep building and acquire useful assets more cheaply.
20%	Severe infrastructure bust Enterprise demand slows, frontier funding weakens and refinancing problems force contract cancellations, impairments and unfinished projects.

What to watch next

• Utilisation: Are installed GPUs generating revenue or waiting for workloads?
• Unit economics: Is workload growth outrunning declines in price per token?
• Backlog quality: Are contracts take-or-pay, cancellable or dependent on continued fundraising?
• Concentration: How much demand ultimately comes from OpenAI, Anthropic and a handful of hyperscalers?
• Free cash flow: Does incremental AI gross profit begin to catch up with capex and depreciation?
• Credit: Do spreads, refinancing costs and private-credit structures begin to deteriorate?
• Enterprise renewals: Do experimental deployments become repeatable production workloads?
• Power: Can announced campuses obtain electricity on time and at commercially viable prices?

The view from here

AI is not a fiction. Neither is the risk of overbuilding it.

The current evidence does not show an aggregate collapse in demand. It does show that capital commitments are growing faster than independently verifiable end-customer returns, while financing is becoming more leveraged and interconnected.

The decisive period is likely to be 2027–2029, when today’s contracted capacity becomes operational. If utilisation, enterprise renewals and AI gross profit rise fast enough, this spending will look prescient. If token prices fall faster than workload volume grows, the industry will discover that correct technological predictions can still produce terrible investments.

That was the lesson of the internet boom. It may also be the lesson of AI.

This article is general research and commentary, not investment advice. Forward-looking estimates and scenario probabilities are inherently uncertain. Company guidance, backlog and funding announcements should not be treated as audited future revenue.