Four companies — Amazon, Microsoft, Alphabet, and Meta — currently plan to spend somewhere in the range of $700–750 billion on capital expenditures in 2026, most of it on AI infrastructure. That is up from roughly $410 billion in 2025, which was itself a record. At the same time, a handful of AI-linked stocks have come to dominate the S&P 500: the Magnificent Seven alone make up roughly a third of the index by market cap, and broader definitions of the AI trade put the figure above 40%.
By the end of this post you will know what “AI capex” concretely buys and why it flipped the hyperscalers from asset-light to asset-heavy businesses; the verified scale of the buildout as of August 2026; where the money actually flows; what index concentration means if you own a plain S&P 500 fund; and the honest versions of both the bull case and the bear case, plus the historical episodes — railroads, 1990s fiber — that everyone keeps reaching for.
Two caveats up front. Every figure here was checked against earnings reports and primary reporting as of August 2026, and these numbers move fast — guidance has been revised upward repeatedly this year alone. And this is an educational explainer written by an engineer, not investment advice.
What “AI capex” actually buys
Capital expenditure is money spent on long-lived assets rather than day-to-day operations. For a hyperscaler in 2026, that means three things stacked on top of each other: buildings, silicon, and power. The building is the data center shell — land, concrete, cooling, substations. The silicon is the expensive part: GPU and custom accelerator clusters, plus the high-bandwidth memory, storage, and networking gear that connects them. Power is the constraint everyone underestimated — grid interconnects, on-site generation, and long-term electricity contracts that increasingly show up as capital commitments.
This is a structural change in what these businesses are. Software companies were historically asset-light: write code once, serve it nearly for free, enjoy gross margins that make industrial firms weep. Training and serving large models broke that model. Every token a model generates has a hardware and electricity cost behind it — the same economics that drive per-token API pricing — so growing AI revenue means growing physical plant, more like a utility or a railroad than like the software businesses these companies grew up as.
The scale of the buildout, in verified numbers
Here is where guidance stood after the July 2026 earnings round, per company reports and CNBC’s coverage of the quarter:
| Company | 2026 capex guidance | Direction |
|---|---|---|
| Amazon | ≈$220 billion | Raised from ≈$200 billion in July, citing memory costs |
| Alphabet | ≈$195–205 billion | Raised twice during 2026 |
| Microsoft | ≈$175 billion (calendar-year basis) | $41 billion in the June quarter alone |
| Meta | $130–145 billion | Low end raised again in July |
A few details from the primary sources are worth flagging. Meta’s Q2 2026 report showed $31.1 billion of capex in a single quarter, nearly double the $17 billion of a year earlier — and free cash flow compressed to under $1 billion. Microsoft’s fiscal Q4 call disclosed that roughly two-thirds of its capex goes to short-lived assets — primarily CPUs and GPUs — and guided to more than $50 billion for the September quarter. And Amazon’s July raise to about $220 billion was attributed partly to rising memory prices: the buildout is now large enough to move the prices of its own inputs.
Add the four together and you get roughly $720–745 billion of planned 2026 spending — before counting Oracle, xAI, CoreWeave and other so-called neoclouds, or the frontier labs’ own commitments. Several banks already model the group crossing $1 trillion in 2027. Whether that happens is a forecast, not a fact; the 2026 guidance is on the record.
Where the money flows
Capex is one company’s expense and another’s revenue, so the buildout has a supply chain worth tracing:
- Semiconductors. The largest slice goes to accelerators — Nvidia above all, with AMD and each hyperscaler’s custom chips (TPU, Trainium and kin) behind it — and to the high-bandwidth memory and advanced packaging those chips require, which is why memory suppliers and TSMC sit directly in the flow.
- Power and utilities. Data centers need firm, around-the-clock electricity at gigawatt scale. That money reaches regulated utilities, gas-turbine makers, nuclear operators signing long-term supply deals, and grid-equipment vendors with multi-year backlogs.
- Construction and cooling. Shells, substations, liquid-cooling plants, and the engineering firms that build them — the unglamorous majority of the physical work.
- Networking. Optical transceivers, switches, and the fiber connecting clusters both inside and between data centers.
This is why the stock-market footprint of AI is much wider than the model vendors themselves. A utility with a data-center pipeline or an optics supplier is an “AI stock” in 2026, which matters for the concentration story below.
A handful of names now drive the index
As of late July 2026, Nvidia’s market cap was around $4.7–4.9 trillion, making it the largest weight in the S&P 500 at roughly 7.5–8% — bigger than entire sectors. The Magnificent Seven together sat at roughly 31–33% of the index, depending on the day you measure, versus about 13% in 2018. Broader tallies are larger still: analyst counts of AI-exposed names put the cluster at 40–45% of index market cap, and Yahoo Finance’s chart-of-the-day coverage of one such basket found AI-linked stocks responsible for the large majority — by some sell-side counts more than 80% — of the index’s gains in 2026 so far.
The precise percentage depends on who counts as “AI-linked,” which is genuinely fuzzy — is a utility with data-center contracts in or out? But the direction survives any reasonable definition: strip out the AI cluster and the index’s recent returns shrink dramatically. The S&P 500’s performance and the AI trade are, for now, close to the same trade.
What concentration means for a passive index investor
The point of an index fund is diversification: own everything, care about nothing in particular. Concentration quietly erodes that deal. If the top seven names are a third of the index, then a third of every dollar you contribute buys those seven stocks; roughly seven to eight cents of each dollar buys Nvidia alone. You did not choose that bet, but you are making it.
Cap-weighting does this by construction — winners grow into larger weights — and it is worth being precise about what the risk is and is not. It is not that indexing is broken; cap-weighted indexes have ridden concentrated leadership before and recovered. The risk is correlation: the top weights are no longer a bank, an oil major, and a retailer with independent fortunes, but seven companies exposed to the same question — will AI monetization justify the capex? If you also work in tech, hold employer equity, and own growth funds that overweight the same names, your true exposure to that one question is larger than any single account suggests. Equal-weight index variants, international allocations, and plain awareness are the standard responses; which is right for you is a personal-finance question this post deliberately does not answer.
The bull case and the bear case
The bull framing: cash-funded and demand-constrained
The strongest version of the optimistic case starts with how the buildout is financed. The late-1990s telecom bubble was built on debt issued by companies with thin profits; today’s spenders are among the most profitable firms in history, funding capex largely from operating cash flow while remaining profitable. Second, the companies report demand ahead of supply: cloud backlogs keep growing, and executives spent the July calls describing capacity constraints, not demand shortfalls. Third, the workloads are real — model API consumption keeps climbing as agentic systems that burn far more tokens than chatbots move into production. On this view, underbuilding is the costlier mistake, and management teams say so explicitly.
The bear framing: depreciation, monetization, circularity
The skeptical case attacks each leg. Depreciation first: Microsoft itself says roughly two-thirds of its capex is short-lived assets, and critics — most loudly investor Michael Burry in late 2025 — argue that stretched useful-life assumptions understate the true annual cost of GPU fleets that may be economically obsolete in a few years. Second, the monetization gap: direct AI revenue across the industry, while growing fast, remains far smaller than the annualized spend, and the gap is bridged by projections. Meta’s sub-$1 billion free-cash-flow quarter and the group’s growing use of debt and lease financing show the “all cash-funded” story already softening at the edges. Third, circular financing: Bloomberg has mapped the web of deals in which Nvidia invests in or guarantees financing for customers — including a reported package of up to $600 billion tied to OpenAI — who then buy Nvidia hardware. Supporters call it lining up a supply chain; skeptics note it makes demand signals harder to read, which is exactly what vendor financing did for telecom equipment in 1999.
Historical rhymes: railroads and dark fiber
Two precedents come up constantly, and both carry the same uncomfortable lesson. Railroads in the nineteenth century were genuinely transformative infrastructure — and serial destroyers of shareholder capital, with waves of bankruptcies in the panics of 1873 and 1893 even as the track kept carrying freight for a century. The 1990s fiber buildout rhymes even harder: carriers laid vastly more fiber than near-term demand justified, WorldCom and Global Crossing collapsed, and the resulting glut of cheap dark fiber later became the substrate for broadband video and the modern internet. Society captured enormous surplus; the builders’ shareholders largely did not.
The rhyme is not a proof. One real difference this time: the hyperscalers are their own anchor tenants, consuming the compute they build rather than wholesaling it into a commodity market, and compute — unlike lit fiber — depreciates before a glut can linger for a decade. Whether that makes overbuild less likely or just makes its cost show up faster on income statements is the actual open question, and reasonable analysts land on both sides.
Bottom line
The verified facts, as of August 2026: the four biggest hyperscalers plan roughly $700–750 billion of 2026 capex, nearly double 2025; the spending has made them asset-heavy businesses whose depreciation now matters as much as their code; AI-linked names are around 40% or more of the S&P 500 and most of its recent gains; and an index fund is therefore a bigger bet on AI monetization than its label suggests. The bull and bear cases disagree not about these numbers but about what comes next — and history says infrastructure booms can be right about the technology and still wrong for the builders’ shareholders. If you are deciding which models to build on rather than which stocks to hold, start with how to choose an LLM instead.
This post is educational content about market structure and infrastructure economics. It is not investment advice, and nothing here is a recommendation to buy or sell anything.