Hello, this is Hamamoto from TIMEWELL.
Not long ago, software companies were the very definition of a nimble business. No factories, no inventory, no power plants. With a team of talented engineers and a server bill, you could earn margins on a scale no other industry could match. That asset-light quality was exactly why investors awarded technology stocks such rich valuations. And yet AI is quietly, but thoroughly, overturning that conventional wisdom. The giant IT firms are now hoarding semiconductors, building data centers the size of several baseball stadiums, and even reviving nuclear power plants to cover the electricity they lack. What they are doing looks less like software and more like steelmaking or power generation. AI has become a heavy industry.
Every time I talk with executives in the field through WARP, I feel this shift in my bones. What used to be a conversation about a tool ("how should we use AI?") has turned into a heavy management decision about capital and power ("which AI infrastructure should nations and companies ride on?"). In this article, I want to trace the structure of AI's transformation into a capital-intensive infrastructure industry, grounded in verifiable primary sources, and then go further into what companies should be thinking about along the way.
Software Was Supposed to Be Capital-Light
The first thing to grasp is that the flow of money has changed completely. AI-related capital expenditure by hyperscalers (the cloud giants, meaning the likes of Alphabet, Amazon, Meta, and Microsoft), the so-called CapEx, has, on a reported and aggregated basis, swelled from roughly 162 billion dollars in 2022 to around 400 billion dollars in 2025[^11][^17]. The figures vary depending on how you tally them, but that is roughly a 2.5x increase in three years, an annual average of about 40 percent. You almost never see that kind of pace in manufacturing.
Even more telling is CapEx as a share of revenue. According to analysis by the asset manager Breckinridge, this ratio rose from around 10 percent in late 2023 to more than 20 percent in the third quarter of 2025[^11]. The firm described this as a departure from the asset-light model that supported a decade of premium valuations. A company pouring 20 percent of its revenue into equipment is no longer a nimble software company. Looking at each firm's 2026 investment guidance, reports put Alphabet at 175 to 185 billion dollars, Microsoft on the order of 190 billion dollars, and Amazon on the order of 200 billion dollars, with the four major hyperscalers expected to reach nearly 700 billion dollars combined in 2026[^1u].
Owning equipment brings depreciation (the accounting practice of expensing the value of purchased equipment a little each year). Analyst estimates put the depreciation of AI-related assets at a brisk annual rate of around 20 percent, with hyperscalers' annual depreciation expense on the order of 400 billion dollars, a level that could exceed their combined 2025 profits[^11]. These are estimates rather than confirmed figures from any single company, but the direction matters. Profits get swallowed by the weight of equipment, and the shortfall is plugged with debt. Just as a steel mill sinks enormous capital into a blast furnace and sees its earnings tossed about by the business cycle, the profits of AI companies are moving toward a highly cyclical structure (cyclicality, the property of earnings rising and falling with booms and busts) that is sensitive to utilization rates, pricing, and technological obsolescence.
Looking for AI training and consulting?
Learn about WARP training programs and consulting services in our materials.
Compute Has Become the New "Strategic Resource"
Why has investment ballooned this far? The reason is simple: the performance of generative AI scales roughly with how much computation you have done. If you want to build smarter AI, you have no choice but to line up large numbers of high-end GPUs and secure the electricity to run them. That is why compute itself has become a strategic resource, like oil or rare metals. Whoever gets it wins, and the nation that corners it gains the advantage. That is the kind of resource it has become.
The numbers make the force of it clear. TSMC, which holds the cutting edge of the world's contract chip manufacturing, posted full-year 2025 revenue of 122.4 billion dollars (up roughly 36 percent year over year), with capital expenditure of about 40.9 billion dollars, and AI-related HPC (high-performance computing) grew to account for 58 percent of its revenue[^14]. In March 2025 the company announced an additional 100 billion dollar investment in the United States, and AI demand pulled its Arizona plant schedule forward by a year, bringing its total U.S. investment to 165 billion dollars[^14]. NVIDIA, the king of GPUs, saw its data center segment revenue swell to about 197 billion dollars for fiscal 2026 and 62.3 billion dollars in the fourth quarter alone (up roughly 75 percent year over year)[^9].
The symbol of it all has to be Stargate. Backed by OpenAI, SoftBank, Oracle, and MGX, this mega-project was announced at the White House in January 2025 and aims for 500 billion dollars and a target capacity of 10 GW over four years[^5]. According to SoftBank's official announcement, as of September 2025 the planned capacity was nearly 7 GW, with more than 400 billion dollars committed over three years and local employment of more than 25,000[^5]. OpenAI's first facility began operating in Texas[^6]. This is no longer software development; it is the world of civil engineering, of building dams and power plants. In fact, nations are rushing to build "sovereign AI" (national AI infrastructure operated while retaining sovereignty over a country's own language, data, and legal system), and NVIDIA's sovereign AI revenue topped 30 billion dollars for fiscal 2026, reported to have more than tripled from the prior year[^9][^10]. Compute is becoming, at the same time, a company's competitiveness and a nation's very security.
The Physical Wall of Electricity, and Why Companies Are Turning to Nuclear
The last obstacle standing in the way of securing compute is electricity. No matter how many GPUs you buy, without the power to run them they are just lumps of metal. In a report published in April 2025, the International Energy Agency (IEA) projected that global data center electricity consumption would more than double from about 415 TWh in 2024 to about 945 TWh in 2030, accounting for roughly 3 percent of world electricity demand by 2030[^1][^2]. Electricity for data centers is set to grow at an annual rate of about 15 percent from 2024 to 2030, said to be more than four times the combined growth of all other sectors[^1].
The skew in that growth is impossible to overlook. According to the IEA, of the increase from 2024 to 2030, the United States accounts for about 240 TWh (roughly 130 percent growth) and China for about 175 TWh (roughly 170 percent growth), with these two countries alone making up roughly 80 percent of the global increment[^1]. In a "Lift-Off" case where AI adoption accelerates further, the IEA also offers an estimate of more than 1,700 TWh in 2035, reaching about 4.4 percent of world electricity demand[^1]. AI computation has begun to consume electricity on a national scale.
That is precisely why the giant IT firms are taking remarkable measures. Nuclear power. Microsoft has teamed up with Constellation on a plan to restart Unit 1 of Three Mile Island, the site known for that accident, signing a 20-year, 835 MW power purchase agreement (PPA), with a reported investment of about 1.6 billion dollars and a target restart in 2028[^7]. Google has partnered with Kairos Power on small modular reactors (SMRs), targeting 500 MWe by 2035 with a first unit expected around 2030[^8]. Amazon has invested 500 million dollars in the nuclear startup X-energy and is procuring several hundred MW over the long term from Talen Energy's Susquehanna nuclear plant[^8]. Companies are corralling electricity, a piece of social infrastructure, for themselves. Just how unusual this is, and what kind of ripple effects it creates, I trace in detail in the companion article, AI, Power, Data Centers, and Space. What I want to emphasize here is that the cost and environmental burden of securing power could ultimately be externalized to society in the form of higher electricity rates and grid congestion.
An "Infrastructure Industrialization" That Echoes the Rail and Power-Grid Eras
The more you know history, the more déjà vu you should feel from everything described so far. The railroads of the 19th century, the electrification and communications networks of the early 20th. Every time a new foundational technology appeared, society poured enormous capital into building physical infrastructure, and in that process came giant corporations, giant debts, and the pain of overinvestment. I see the infrastructure industrialization of AI as the latest entry in that lineage. Just as the railroads laid track, we are now laying optical fiber, data centers, and transmission lines. The difference is that the pace of construction is orders of magnitude faster.
Stanford HAI's AI Index captures the concentration of investment well. The 2025 edition (covering 2024 data) showed corporate AI investment of 252.3 billion dollars, up 44.5 percent year over year, of which private investment in generative AI was 33.9 billion dollars, up 18.7 percent[^3]. The regional gap was stark too: U.S. private AI investment was 109.1 billion dollars, roughly 12 times China's 9.3 billion[^3]. The following 2026 edition (covering 2025 data) showed U.S. private AI investment expanding to 285.9 billion dollars, widening to more than 23 times China's 12.4 billion[^4]. Capital is concentrating in a handful of countries and companies at a furious pace. Structurally, this looks a lot like the era that produced railroad barons and electric-power trusts.
From a company's perspective, what matters here is the choice of which infrastructure to ride. In the railroad era, not every merchant laid their own track. Most competed on how to load their goods and which lines to use. AI is the same. Most companies are not on the side that builds data centers; they will be asked how to make full use of the AI infrastructure that has already been built. This is exactly what our AI consulting practice WARP works through together with executives. Which model to bet on, which of your data and operations to put on top of it, what to build in-house and what to outsource. The more the heavy, large-scale infrastructure gets built out, the more the winner is decided by the design of what you load onto it. As with the sovereign AI mentioned earlier, in an age where even nations choose with their sovereignty at stake, it is dangerous for a company to drift along unaware. Choosing the infrastructure you ride on with intent becomes the first source of competitiveness.
The Risks Lurking in Infrastructure-ization: Circular Deals and Bubble Fears
There is always a shadow behind the euphoria that deserves caution. What is most debated right now is concern over circular financing (a structure in which the parties cycle investments and orders among themselves to make demand look bigger than it is). According to reports and analysis, the structure pointed to is one where NVIDIA invests up to 100 billion dollars in OpenAI, that OpenAI buys NVIDIA's chips, and OpenAI and Oracle then sign a five-year, 300 billion dollar cloud contract[^12][^13]. OpenAI's total chip and cloud commitments are said to potentially exceed 1 trillion dollars[^13]. The worry is that money circles around among the parties, making demand look much larger than it really is. Some commentators draw a parallel with the dot-com era, when Cisco lent money to telecom startups so they could buy Cisco's own products[^12].
That said, this is a place where I want to avoid firm conclusions. In February 2026 the WSJ reported that NVIDIA's plan to invest 100 billion dollars in OpenAI had "stalled," and OpenAI was reported to be projecting a loss of about 14 billion dollars in 2026 while aiming for 100 billion dollars in revenue by 2029. All of this is on a reported basis, uncertain information that requires further tracking. Likewise, the observation that Oracle's debt-to-equity ratio reached around 6x is a media estimate and needs confirmation against the primary financial statements. U.S. investor-owned utilities (IOUs) raised their capital expenditure plans through 2030 by more than 27 percent from the prior year, to a total of 1.4 trillion dollars. More than 30 of the 51 companies surveyed cited data centers as a major growth driver, and the cost of grid upgrades has begun to feed back into upward pressure on electricity rates for households and businesses[^18]. What is certain is that an enormous amount of capital, debt included, is piling up rapidly, and the voices warning of overinvestment and self-reinforcing valuations exist on a scale that cannot be ignored.
On top of that, because compute has become a strategic resource, geopolitical risk rides along with it too. Sam Altman was reported to have said that AI would be sold like electricity and water[^15], and an "AI-as-utility" view, treating AI as a metered public good like electricity or running water, has spread. Meanwhile, the United States has wavered over where to draw the line on export controls for AI capabilities. Under the Biden administration, the Commerce Department's Bureau of Industry and Security (BIS) published the "AI Diffusion Rule" on January 15, 2025, attempting to classify most of the world's countries into three tiers and cap exports of high-end GPUs. But before its scheduled effective date of May 15, 2025, BIS under the Trump administration formally rescinded the rule, effective May 13 of that year. BIS has stated its intent to provide interim export-control guidance and then draft a replacement rule, but as of the time of writing the new rule has not been finalized[^16]. Compute is no longer a pure commodity; it has become an object over which nations fight for control. The more deeply it embeds itself in society as infrastructure, the more companies are forced to factor in the risk of these policy shifts.
The Management Decisions Companies Must Make in the Age of AI Infrastructure
So what should the vast majority of companies, the ones that are not building data centers, do in this age of heavy-industrialization? My answer is simple. Do not chase compute itself; differentiate through the "content" you load on top of it. It is the same as how a utility may control generation, but what you make with that electricity depends on the skill of the company using it. AI infrastructure is going to become a commodity you can procure like electricity or water (a generic product that is hard to differentiate). If so, the source of competitiveness shifts to the data unique to your company, the operational know-how accumulated on the ground, and the design of the workflows that put it all onto AI and keep it running.
Concretely, I recommend holding three axes of judgment. The first is "which model and infrastructure to bet on." Over-relying on a single vendor leaves you fully exposed to swings in pricing and policy. Compare multiple options and keep your design switchable. The second is "what to build in-house and what to outsource." There is little point in carrying a general-purpose compute foundation yourself; what deserves your focus is organizing your own data and designing prompts and agents. The third is "how to design governance." Since compute is a strategic resource, where you process which data becomes a matter not only of cost but of security and compliance.
To be honest, I sense that few companies can yet work all three of these out on their own. Because AI has become a heavy, large-scale infrastructure industry, the reach of management decisions has expanded from technology selection to capital strategy and geopolitical risk. Our AI consulting offerings, WARP, WARP NEXT, and WARP BASIC, exist precisely to take on the role of "drawing up the management strategy for the age of AI infrastructure together with you." Which AI to ride on, and how to design your competitiveness. We can start from drawing that map together. If you are wrestling with how to turn a heavy-industrialized AI into your own weapon, reach out through an individual consultation, even just to organize where you stand today. In an age where capital and power decide competitiveness, stand on the side that chooses rather than drifts. We will help you take that first step.
[^1]: AI is set to drive surging electricity demand from data centres (related to Energy and AI) — IEA — 2025-04-11 — https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works [^2]: IEA: Data center energy consumption set to double by 2030 to 945TWh — DataCenterDynamics — 2025-04 — https://www.datacenterdynamics.com/en/news/iea-data-center-energy-consumption-set-to-double-by-2030-to-945twh/ [^3]: The 2025 AI Index Report (Economy) — Stanford HAI — 2025-04 — https://hai.stanford.edu/ai-index/2025-ai-index-report/economy [^4]: The 2026 AI Index Report (Economy) — Stanford HAI — 2026 — https://hai.stanford.edu/ai-index/2026-ai-index-report/economy [^5]: OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites — SoftBank Group Corp. — 2025-09-24 — https://group.softbank/en/news/press/20250924 [^6]: OpenAI's first data center in $500 billion Stargate project is open in Texas — CNBC — 2025-09-23 — https://www.cnbc.com/2025/09/23/openai-first-data-center-in-500-billion-stargate-project-up-in-texas.html [^7]: Three Mile Island nuclear power plant to return as Microsoft signs 20-year, 835MW AI data center PPA — DataCenterDynamics — 2024-09 — https://www.datacenterdynamics.com/en/news/three-mile-island-nuclear-power-plant-to-return-as-microsoft-signs-20-year-835mw-ai-data-center-ppa/ [^8]: Amazon, Google, Meta and Microsoft go nuclear — Trellis — 2024-2025 — https://trellis.net/article/amazon-google-meta-and-microsoft-go-nuclear/ [^9]: NVIDIA Q4 FY2026 Earnings Highlight Durable AI Infrastructure Demand — Futurum Group — 2026-02 — https://futurumgroup.com/insights/nvidia-q4-fy-2026-earnings-highlight-durable-ai-infrastructure-demand/ [^10]: Nvidia's sovereign AI revenue tripled to $30B — Dealroom.co — 2026-02 — https://app.dealroom.co/news/feed/nvidia-s-sovereign-ai-revenue-tripled-to-30b-as-governments-fuel-overlooked-growth-opportunity [^11]: The Price of AI: How Capex Is Rewriting Tech Balance Sheets — Breckinridge Capital Advisors — 2025-2026 — https://www.breckinridge.com/insights/the-price-of-ai-how-capex-is-rewriting-tech-balance-sheets [^12]: AI Roundtripping: NVIDIA, OpenAI, Oracle and the Circular Financing Debate — Ventures Edge — 2025-2026 — https://www.venturesedge.io/articles/ai-roundtripping-nvidia-openai-oracle-and-the-circular-financing-debate [^13]: AI Circular Deals: How Microsoft, OpenAI and Nvidia Keep Paying Each Other — Bloomberg — 2026 — https://www.bloomberg.com/graphics/2026-ai-circular-deals/ [^14]: TSMC ramps up Arizona production as AI demand drove 2025 revenue to $122B — Yahoo Finance / Reuters — 2026-01 — https://finance.yahoo.com/news/tsmc-ramps-arizona-production-ai-100500045.html [^15]: Sam Altman said AI will be sold like electricity and water — moneywise — 2025-2026 — https://moneywise.com/news/top-stories/ai-infrastructure-spending-data-centers-power-bottleneck [^16]: Department of Commerce Announces Rescission of Biden-Era Artificial Intelligence Diffusion Rule (publication of the AI Diffusion Rule and its formal rescission on May 13, 2025) — U.S. Bureau of Industry and Security (BIS) — 2025-05 — https://www.bis.gov/press-release/department-commerce-announces-rescission-biden-era-artificial-intelligence-diffusion-rule-strengthens [^17]: Big Tech AI Capex in 2025 — ValueAdd VC — 2026 — https://valueaddvc.com/blog/big-tech-ai-capex-in-2025-microsoft-google-meta-amazon-and-the-spending-race [^1u]: Tech AI spending may approach $700 billion this year, but the blow to cash raises red flags (summary of each company's 2026 CapEx guidance) — CNBC — 2026-02 — https://www.cnbc.com/2026/02/06/google-microsoft-meta-amazon-ai-cash.html [^18]: U.S. utilities plan $1.4 trillion spending spree, up more than 27%, for next 5 years amid AI construction boom (per PowerLines) — Fortune — 2026-04 — https://fortune.com/2026/04/14/us-utility-spending-jumps-to-1-4-trillion-amid-ai-construction-boom/