AI Capex at $725 Billion and the Distance to Revenue── Placing AI capital expenditure alongside telecom-bubble numbers reveals the assumptions embedded in current valuations

01Capital expenditure hit $725 billion while AI revenue remains a fraction of that

In 2024, Sequoia Capital partner David Cahn published an analysis titled "AI's $600 Billion Question," arguing that the gap between AI infrastructure spending and AI-derived revenue was widening. He initially estimated the shortfall at $125 billion; by 2026 the figure had grown to $500–600 billion.

Interpretation depends on where you stand. Optimists point out that building infrastructure ahead of demand is standard practice for technology investment cycles — railroads, electricity grids, and fiber-optic networks all required years of front-loaded spending before utilization caught up. Skeptics note that spending nine to ten dollars on equipment for every dollar of current AI revenue brings the ratio close to the telecom bubble of 2000. The purpose of this article is not to settle the debate but to decompose the assumptions on each side so that readers can form their own assessment.

02The telecom bubble peaked at a 32 percent capex-to-revenue ratio — AI has passed 34 percent

At the height of the US telecom buildout in 2000, aggregate capital expenditure as a share of revenue reached roughly 32 percent. Capex grew at a compound annual rate of 33 percent between 1996 and 2000, while revenue grew at only 10 percent over the same period. After the crash, 85–95 percent of the fiber-optic cable laid during the boom sat unused for four years.

AI-related capex-to-revenue ratios reached 34 percent in 2026, with projections of 37 percent by 2028. At the individual-company level, the dispersion is wider: Oracle's capex-to-revenue ratio stands at 86 percent and Meta's at 54 percent, meaning these firms are spending more on infrastructure than they earn in a given year. The aggregate numbers are now comparable to the telecom peak in structural terms.

One important difference separates the two episodes. The telecom buildout rested partly on a fabricated demand signal: WorldCom claimed internet traffic doubled every 100 days, when the actual rate was roughly once a year. AI demand, by contrast, is reflected in real sales — NVIDIA's revenue grew 3.3 times in two fiscal years. The question is not whether demand exists, but whether it translates into downstream earnings for the companies making the investments. Demand and profitability are separate questions, and conflating them is one of the most common errors in reading AI valuations.

03NVIDIA's revenue tripled in two years, but that is infrastructure spend, not end-user earnings

NVIDIA reported annual revenue of $60.9 billion in fiscal 2024, $130.5 billion in fiscal 2025, and $215.9 billion in fiscal 2026. Data-center revenue alone reached $62.3 billion in the fourth quarter of fiscal 2026. The demand for AI compute hardware is real and measurable.

However, NVIDIA's revenue is, in substance, the hyperscalers' capital expenditure. When Microsoft or Amazon buys GPUs, that purchase appears as NVIDIA revenue and as capex on the buyer's balance sheet. The next stage of the chain is the cloud provider's service revenue — Google Cloud grew 63 percent year on year in Q1 2026, AWS 28 percent, and Azure 40 percent. Cloud revenue is rising, but not as fast as the capex feeding it. Each hyperscaler exceeded its own initial 2025 capex guidance — Google announced $75 billion and spent $91.4 billion, Amazon guided roughly $100 billion and spent $131.8 billion, Meta guided $60–65 billion and landed at $72.2 billion — suggesting that competitive pressure, not just demand, is driving the spending pace.

The revenue chain therefore has at least three stages: GPU maker revenue (realized), cloud service revenue (growing but lagging capex), and end-user enterprise earnings from AI (largely unrealized). Any valuation depends on which stage it assumes as its revenue base, and how quickly value travels down the chain.

04Only 37 percent of AI-adopting firms report any impact on operating profit

McKinsey's 2026 State of AI survey found that 37 percent of organizations using AI reported any effect on EBIT, unchanged from 2025. Only 6 percent qualified as "AI high performers." At the individual level, 80 percent of users reported higher personal productivity, but that has not translated into measurable financial returns at the organizational level at the same rate.

Goldman Sachs revised its consensus estimate for 2026 hyperscaler AI capex upward to $527 billion in January 2026. The same research unit projected a 24-fold increase in global AI token consumption between 2026 and 2030. The distance between investment-side projections and realized earnings remains wide. Enterprise AI adoption is broad — Goldman Sachs Research reported that 88 percent of organizations used AI in at least one function in 2025, up from 78 percent the year before — but breadth of adoption has not converted into depth of financial impact at the same pace.

05Pharma AI investment is smaller in scale but follows the same structure

AI investment in drug discovery is orders of magnitude smaller than hyperscaler capex. The AI-in-drug-discovery market was estimated at $1.9 billion in 2024, projected to reach $6.9 billion by 2029 at a compound annual growth rate of roughly 30 percent. Cumulative AI investment in life sciences from 2022 through 2026 exceeded $100 billion.

The gap between investment and revenue follows the same pattern. As of late 2024, more than 75 AI-designed molecules had reached clinical stages, but none had received regulatory approval. AI-designed molecules have reported Phase I success rates of 80–90 percent, compared with approximately 40 percent for conventionally designed candidates. Phase I, however, tests safety; demonstrating efficacy through Phase III and reaching approval typically takes several more years.

Evaluating pharma AI investment requires the same decomposition: what exactly has been shortened, and when does that time saving convert into revenue? Even if AI compresses the discovery phase by two to three years, the full development timeline from molecule to approved product spans 10 to 15 years, and the investment payback period remains long. The AI-in-biopharmaceuticals market was valued at $1.55 billion in 2024, with long-range projections reaching $24.5 billion by 2034 — a compound annual growth rate of about 32 percent. These projections assume that clinical success rates hold and that regulatory frameworks accommodate AI-derived evidence, both of which are untested at scale.

06Three lenses for decomposing valuation assumptions

When reading the valuation of an AI-linked company, at least three assumptions can be isolated.

Revenue stage. Identify which of the three stages the valuation treats as its revenue base. GPU-maker revenue has been realized. Cloud service revenue is growing but trails capex growth. Enterprise-level earnings from AI remain largely unrealized. A valuation built on Stage 3 revenue carries more uncertainty than one built on Stage 1.

Payback horizon. Capital expenditure is depreciated over the useful life of the asset. GPUs have an effective operating life of three to five years. Whether the revenue generated within that window covers the initial outlay is a key variable. In pharma, the payback horizon stretches to 7–12 years because clinical development must follow discovery.

Concentration risk. Approximately 75 percent of AI capex targets data centers, GPUs, and networking equipment. When spending is concentrated among a small number of buyers and sellers, a single revision in investment plans can cascade. NVIDIA's revenue depends heavily on the capex decisions of four hyperscalers; if those plans are cut, the effects propagate through the supply chain. Goldman Sachs projects cumulative capex of $5.3 trillion across the four largest hyperscalers from fiscal 2025 through 2030. The scale of this commitment means that even a modest downward revision — say, 10 percent — would remove over $500 billion in expected demand from the ecosystem.

07Building the habit of checking numbers yourself is the most durable skill

Capital-expenditure figures are available in each company's quarterly filings (10-Q, 10-K). The capex-to-revenue ratio can be computed by dividing capital expenditure by revenue. Lining up five years of this ratio reveals the trajectory of investment intensity.

Comparing revenue growth rates with capex growth rates shows whether the gap is widening or narrowing. A widening gap means the market is pricing in future revenue growth that has not yet materialized. A narrowing gap indicates that investment is beginning to convert into sales.

Analyst consensus estimates are available through terminals such as Bloomberg and FactSet. Tracking revisions — upward or downward — signals shifts in market expectations. During the Q4 2025 earnings cycle, Wall Street consensus for 2026 hyperscaler AI capex was revised upward to $527 billion, indicating that analysts were raising, not lowering, their spending assumptions. The FDA's count of AI-enabled medical device authorizations (1,451 cumulative through 2025, with 295 in 2025 alone) provides an additional gauge of how quickly AI is reaching practical deployment.

This article does not recommend any security or financial product and makes no price forecast. The ability to read the assumptions behind a valuation — by checking the numbers yourself — is the point.