Some figures are impressive on their own, while others distort the debate. The estimate of around $930 billion in cumulative investment in data centers between 2020 and 2026 for the group composed of Amazon, Microsoft, Alphabet, Meta, and Oracle clearly falls into that second category. Looked at without context, it seems like a figure that invites talk of an immediate bubble, excess capacity, and another tech boom doomed to end badly.
But that reflection might be looking in the wrong place. The gross volume of investment is enormous, yes, but economic history shows that transformative megaprojects are not usually well judged solely by the size of the bill. It also matters how much that bill impacts the economy and, above all, how quickly the investment is expected to be recovered. And that’s where the story changes.
Recent official data from the big tech companies show that the acceleration is real. Amazon closed 2025 with $128.32 billion in net property and equipment purchases; Microsoft, $64.55 billion in its fiscal year 2025; Alphabet, $91.4 billion in 2025; Meta, $72.22 billion in 2025, including major payments on financial leases; and Oracle, $21.22 billion in its fiscal year 2025. Just these five companies totaled about $377.71 billion in capex in their latest reported fiscal year. Additionally, Alphabet has guided between $175 billion and $185 billion for 2026, Meta between $115 billion and $135 billion, and Oracle around $50 billion for its fiscal year 2026. Microsoft already spent $37.5 billion on capex in a single quarter—the second of its fiscal year 2026—and Amazon has not published an equivalent annual guidance in the reviewed documentation.
The numbers are intimidating, but context shifts the perspective
This surge in investment is not an accounting illusion. Epoch AI calculated in February 2026 that the combined capex of Alphabet, Amazon, Meta, Microsoft, and Oracle had quadrupled since the launch of GPT-4, and if recent trends continued, it could reach $770 billion in 2026. This isn’t an official forecast from the companies but an extrapolation of the trend, providing insight into how rapid this change is happening.
Table 1. Recent Capex of the Hyper-Scale Block
| Company | Latest reported annual capex | Guidance for 2026 | Notes |
|---|---|---|---|
| Amazon | $128.32 billion (2025) | No formal annual guidance in reviewed documentation | Largest absolute expenditure in the last reported fiscal year |
| Microsoft | $64.55 billion (FY2025) | $37.5 billion in a single quarter (FY2026 Q2) | Approximately two-thirds of that quarter’s capex was short-lived assets, mainly GPUs and CPUs |
| Alphabet | $91.4 billion (2025) | $175-185 billion (2026) | Official capex guidance well above the previous year |
| Meta | $72.22 billion (2025) | $115-135 billion (2026) | Includes principal payments on finance leases |
| Oracle | $21.22 billion (FY2025) | $50 billion (FY2026) | The company that has scaled up the most in the shortest time |
Note: Capex definitions vary somewhat across companies, so comparisons are not perfect. Nonetheless, the overall trend is clear.
Yet, one thing is to state that spending is enormous, and quite another is to conclude that, solely for that reason, a bubble is imminent. Historical comparisons help temper the drama. The comparative estimate included in the charts provided by the user—built from corporate reports and references like FHWA, NASA, CRS, GAO, Brookings, and Epoch AI—suggests that the current deployment of data centers, even though gigantic, does not surpass several major historical programs when measured as a percentage of US GDP accumulated over time.
Not every massive project is a bubble
Table 2. Megaprojects vs. current data center deployment
Comparative estimate based on charts provided by the user
| Project | Inflation-adjusted cost | Duration | Spending as % of US GDP |
|---|---|---|---|
| Data centers of big hyperscalers (2020-2026 estimate) | ~$930 billion | 6 years | ~3.3% |
| Manhattan Project | ~$36 billion | 5 years | ~1.3% |
| Apollo Program | ~$257 billion | 14 years | ~4.8% |
| Marshall Plan | ~$170 billion | 4 years | ~6.4% |
| Interstate Highway System | ~$620 billion | 37 years | ~9.0% |
| F-35 Program (to date) | ~$400 billion | 25 years | ~1.8% |
| International Space Station | ~$150 billion | 27 years | ~1.0% |
| US Railroads | ~$550 billion | 71 years | Out of scale in original chart |
The uncomfortable conclusion is that the debate over the “AI bubble” is probably using the wrong metric. The issue doesn’t seem to be the economy’s capacity to absorb large-scale investments. The US has absorbed more intense and longer-lasting programs before. The real delicacy is another factor: timing.
The real risk lies in the gap between expenditure and monetization
Railway expansion took decades to mature. The Interstate Highway System also. The Apollo Program was extremely expensive but had geopolitical and state-driven logic that did not depend on justifying each dollar with a quarterly operating margin. In contrast, current AI capex is hitting the books of publicly traded companies, which must convince the market that monetization will come on time.
This is where the concern becomes more relevant. Microsoft acknowledged in January 2026 that $37.5 billion of its quarterly capex had already been largely allocated to short-lived assets, mainly GPUs and CPUs. That significantly alters the conversation. An interstate highway can be amortized over several generations; a cluster of accelerators cannot. In AI infrastructure, technological obsolescence progresses much faster than in traditional public works.
Therefore, the real risk is not “overspending,” but “spending too fast for the rate of genuine value capture.” If business demand for inference, copilots, agents, and AI software consolidates and maintains margins, this cycle will be remembered as the foundation of a new industrial layer. If not, the blow won’t come from the absolute number being obscene, but because the expected payback was too close, and revenue curves too far.
In other words: an economy can sustain enormous investments if it has time. The market’s fear isn’t just the scale of the wager but the shortening of the timeline. And in that sense, there is a truly fragile element. Capital-intensive projects destroy value not only when they are misdirected but also when they demand results before the infrastructure has found its definitive business model.
Industrial history offers a useful reminder: many infrastructures that later appeared inevitable first caused trouble for those who funded the initial wave. Not because they were useless, but because they arrived before the market was ready to return the investment at the pace their promoters demanded.
Something similar could happen with AI. The world probably will need more data centers, more networks, more energy, and more computing power. The question isn’t whether the infrastructure will make sense in the end; it’s how many balances can withstand the early phase until it truly does.
Frequently Asked Questions
Are the $930 billion an official figure from the companies?
No. It is a comparative estimate of data center capex between 2020 and 2026, based on the provided charts and a methodology combining corporate reports and assumptions about what portion of total capex actually relates to data center infrastructure.
Why isn’t it enough to just look at total capex to talk about a bubble?
Because a bubble depends not only on how much is spent but also on whether that spending can be monetized timely and with sufficient returns. The timing matters as much as the amount.
Which company is currently spending the most on infrastructure?
In the latest reported year, Amazon registered the highest absolute capex, with $128.32 billion in 2025. However, Alphabet, Meta, and Oracle have guidance for 2026 pointing to another significant scale jump.
Are the figures from Amazon, Microsoft, Meta, Alphabet, and Oracle directly comparable?
Only with some nuances. Each company defines capex somewhat differently: some include financial leases, others use net property and equipment purchases, and additionally, Oracle reports by fiscal year, not calendar year.