AI Infrastructure Investment 2026: Stargate, the AI Bubble Debate, the ROI Gap, and the Power Bottleneck, Read Through Primary Sources

Hello, this is Ryuta Hamamoto from TIMEWELL.

Over the past year, the conversation around AI infrastructure has shifted, from being astonished at the sheer size of the investments to asking whether that investment can actually be recouped. Heading into 2026, the combined data-center-related spending of the major tech companies is estimated at roughly $650 billion¹. Yet in the very same period, the Bank of England and the IMF explicitly warned of the risk of a market correction, and MIT reported that around 95% of enterprise generative-AI pilots are not yet profitable. Investment on an extraordinary scale, and genuine uncertainty about the payoff. Those two things are running side by side, and that is what AI infrastructure looks like in 2026.

This article isn't about lining up the digits and calling it a day. As a set of lenses that Japanese B2B executives can actually use in decision-making, I've reorganized the picture around four bottlenecks, the AI bubble debate, power, financing, and geopolitics, drawing on primary sources wherever possible. If you want to gauge whether your own AI adoption is overheated or, conversely, falling behind, it helps to first check where you stand with the AI Literacy Check before reading on. The rest of the piece will connect more easily.

The Order of Magnitude Changed in 2026

First, let's calibrate a sense of the numbers. When I wrote the first version of this article back in January 2026, I estimated the capital expenditure of the five hyperscalers at $602 billion, a mid-2025 forecast. Six months later, the actuals and guidance came in higher, and 2026 AI data center spending is now estimated at roughly $650 billion across the major players combined¹. Add up the capex of Amazon, Microsoft, Google, Meta, and Oracle, and the picture is unchanged: it has climbed sharply from the prior year, driven largely by AI. The precise figure for each company is the kind of thing you should confirm in the quarterly earnings, so here I'll treat it at the combined-scale level.

The fastest way to feel the scale is probably to look at the stock market. NVIDIA's market capitalization crossed $4 trillion in July 2025 and $5 trillion in October, reaching the point where a single company accounts for about 7.3% of the S&P 500. Roughly 80% of the 2025 gains in US equities are attributed to AI-related companies. As of the end of 2025, the top five companies made up about 30% of the S&P 500, and the cyclically adjusted price-to-earnings ratio (the Shiller PE) topped 40 for the first time since the dot-com era. You may think you're buying the index, but in substance you're making a concentrated bet on AI-infrastructure-related names. That's the market we're in.

What I want to pause on here is whether this concentration is a risk or an opportunity. Before answering, let's look at the most emblematic project of the moment.

Stargate: One of the Largest AI Infrastructure Plans in History

On January 21, 2025, OpenAI announced a large-scale infrastructure plan called Stargate². OpenAI and SoftBank each hold a 40% stake, joined by Oracle and MGX (an Abu Dhabi investment firm). The initial investment is $100 billion, with up to $500 billion committed through 2029. The US president attended the announcement, and the project was treated almost like a national initiative.

It has only grown since. In September 2025, five additional US sites (in Texas, New Mexico, Ohio, and elsewhere) were announced, swelling the figures to roughly 7 GW and more than $400 billion over three years. A site in the United Arab Emirates (UAE) is slated to open in 2026, and in Argentina an investment of up to $25 billion has been floated. The fact that SoftBank, a Japanese company, is a core backer means this is not someone else's story for readers in Japan.

OpenAI itself has reportedly committed to roughly $1.4 trillion in data center contracts over the next eight years. Note the gap in the numbers here. The company's annual revenue is said to be around $13 billion, and one estimate projects an operating loss of about $74 billion in 2028. Deutsche Bank's Jim Reid estimated cumulative losses of $140 billion from 2024 through 2029. Signing infrastructure contracts worth more than a hundred times revenue, while continuing to run at a loss. How you read that setup is the entry point to the bubble debate that follows.

Can It Be Recouped? The AI Bubble Debate and the Reality of ROI

The tone shifted from celebrating the investment figures because public institutions issued warnings one after another. In its Financial Stability Report of October 8, 2025, the Bank of England pointed to an "increased risk of a global market correction," with AI-related equity valuations in mind³. IMF Managing Director Kristalina Georgieva likened the current situation to the 2001 dot-com bubble, warning that a correction would damage even emerging-market growth⁴. It is unusual for a central bank and an international institution to face the same direction and sound the same caution.

What complicates the debate is the circular flow of money. In September 2025, NVIDIA announced a $100 billion investment in OpenAI, and OpenAI, in turn, buys NVIDIA GPUs in large volumes. In October 2025, OpenAI and AMD formed a partnership, reportedly with OpenAI becoming a major shareholder. Between OpenAI and Oracle there is a contract on the order of $300 billion. The seller invests in the buyer, and with that money its own products get bought again. This "circular financing" is what makes real demand hard to see, critics argue. Prominent investors and executives such as Ray Dalio and Jamie Dimon have sounded the alarm, while Sam Altman acknowledges the bubble-like aspects but argues for the value of the underlying technology.

But what really matters to executives isn't the stock-price story. It's the question, "Can our own AI investment be recouped?" Here are some cold numbers. MIT reported that even though corporate AI investment exceeded $60 billion in 2025, about 95% of generative-AI pilots are not profitable. A study released by the National Bureau of Economic Research (NBER) in February 2026 likewise found that about 90% of companies could not confirm workplace productivity gains.

I don't read these figures as grounds for pessimism. Rather, I see them as the result of many companies making "introducing AI" the goal in itself, skipping the design work of deciding which parts of the business to change and how. The more the hyperscalers build out infrastructure with orders-of-magnitude investment, the more the contest moves to "the design capability of the side that uses it." This is where opinions diverge, but I believe redesigning your own operations produces far more reproducible return on investment than fighting over infrastructure. I dig into how to embed AI into which operations in Three Strategic Options for AI-Agent Management as well.

Power: The Hardest Bottleneck

It may sound odd to say after going on about money, but the biggest constraint on AI infrastructure isn't money, it's power. Goldman Sachs estimates that global data center power demand will rise 165%, from 55 GW in 2023 to 84 GW in 2027 and 122 GW in 2030⁵. An AI server rack consumes about 60 kW each, six times the roughly 10 kW of a general-purpose server. The more you pack in, the more heat and electricity it draws: a wall of physics stands in the way.

Building out the supply side isn't easy either. Goldman Sachs estimates that grid investment will require a cumulative $720 billion through 2030, and that the workforce for power infrastructure will fall short by more than 500,000 people by that year. The IEA (International Energy Agency), too, analyzed in detail the load AI places on power systems in its April 2025 report⁶. The simultaneous moves toward nuclear power and small modular reactors (SMRs), and toward extending the life of existing thermal plants, are all about filling this supply-demand gap.

Friction is emerging on the social side as well. As of July 2026, local residents' opposition movements have reportedly frozen data center construction projects worth about $130 billion. More than 230 groups have signed on in support of construction moratoriums (temporary halts). Heavy water consumption, pass-through to electricity bills, and impact on the landscape are the points of contention. Without all three, power, water, and community consent, no data center gets built, no matter how much capital there is. I've also organized AI's power problem itself in our article on AI and power consumption.

Where Is the Money Coming From?

Another thing I couldn't fully cover in the first version is the structural change in financing. Traditionally, hyperscalers funded capital expenditure from their ample operating cash flow. But the scale of investment has now exceeded their own funds, and they're entering a phase of relying on outside capital.

Morgan Stanley's outlook is emblematic. The firm forecasts $3 trillion in global data center investment from 2025 to 2028, with about half of it funded by private credit (lending by funds and the like, rather than bank loans). It also sees data-center-related debt topping $1 trillion by 2028. Data center debt issuance in 2025 already reached $182 billion. Meta's large project "Hyperion" was a $30 billion structure arranged by Morgan Stanley, one of the largest private-capital deals of its time. Financing collateralized by the GPUs themselves is also spreading; one cloud-specialist company raised $12.4 billion against GPU collateral.

This gave rise to what are called "neoclouds," a group of operators specialized in renting out GPUs. They set up an SPV (special purpose vehicle), borrow against GPUs as collateral, and lend compute to companies that need it. The mechanism can add supply quickly, but it carries the fragility of a chain reaction, if AI demand falls short of expectations, collateral values decline and debt cascades. Market forecasts through 2030 vary widely by institution: Citigroup estimates about $2.8 trillion, while McKinsey estimates about $7 trillion. That forecasts an order of magnitude apart can coexist is itself telling of this industry's uncertainty. In July 2026, Meta signaled a stance toward entering the cloud business, and Anthropic was reported to be considering procuring compute from Meta on the order of $10 billion⁷. The tug-of-war between capital and compute is still shifting violently.

How Japanese Executives Should Read This

So far this has been a global story, but I can't leave out the Japanese perspective. I've already noted that Stargate's core backer is the Japanese company SoftBank. At home, too, major overseas operators keep announcing large data center investments, and behind them the availability of grid (transmission) capacity and alignment with GX (green transformation) policy are becoming practical constraints. In Japan, there aren't many sites that can satisfy both a tight power situation and decarbonization demands at once.

Another thing not to overlook is economic security. Export controls on semiconductors to China affect the infrastructure plans themselves, and which country's chips you build your compute on becomes part of the management decision. The question of sovereignty, where you place your data, has grown heavier as well. The meaning of holding AI infrastructure domestically has to be measured not only by cost but from the standpoints of regulation and trust. It's in this context that TIMEWELL insists on domestic server operation for our enterprise AI, ZEROCK.

So should every company hold vast infrastructure like the hyperscalers? I don't think so. For many Japanese companies, the realistic path is to be on the "smart user" side rather than the "builder" side of infrastructure. The trend of compute becoming a capital-intensive equipment industry is exactly what I wrote about in The Heavy-Industrialization of AI and OpenAI's Next Ten Years. While power and capital concentrate among the handful of companies that own the equipment, the application layer that creates value on top of it is, if anything, broadly open. What's tested here isn't the size of the investment but the design of which parts of your operations you hand to AI and which you keep in human hands. TIMEWELL's AI consulting service, WARP, supports translating this "design of the using side" into management decisions.

Then and Now: What Changed in the Past Six Months

The very terms of the debate moved from January 2026, when I wrote the first version. Let me lay out the main changes.

Item	Then (January 2026, mid-2025 forecast)	Now (July 2026)
Central investment scale	$602B across five hyperscalers (forecast)	Roughly $650B combined across the major players (Bloomberg, February 2026)
Cumulative outlook	$1.15T for 2025–2027 (Goldman Sachs)	$3T for 2025–2028 (Morgan Stanley); a range of ~$2.8T–$7T by 2030 (Citi, McKinsey)
Largest project	Not mentioned	Stargate (initial $100B, up to $500B)
Focus of the article	Sheer size of investment and historical comparison	Recoverability (bubble debate, ROI gap)
NVIDIA market cap	$4T (July 2025)	Crossed $5T (October 2025), ~7.3% of the S&P 500
Power	Noted as "the biggest bottleneck"	A concrete picture: 122 GW by 2030, $720B grid investment, $130B frozen

Six months ago I was talking about "how big it is." Now the phase is asking "how long it holds." The absolute amount of investment has actually risen, yet the market's attention has moved to recovery and sustainability. That gap in temperature is, I feel, the key to reading AI infrastructure in 2026.

Summary

I've been chasing numbers, so let me finally translate them into the language of management.

• 2026 AI infrastructure spending is roughly $650 billion combined across the major players. In the stock market, concentration in NVIDIA and a handful of related companies keeps advancing.
• Stargate is one of the largest projects in history, with an initial $100 billion and up to $500 billion, and SoftBank is a core backer.
• The Bank of England and the IMF warned of a bubble; MIT reported about 95% of generative-AI pilots unprofitable, and NBER reported that about 90% of companies could not confirm productivity gains.
• Power (122 GW by 2030), financing (dependence on private credit and GPU collateral), and geopolitics (semiconductor controls and data sovereignty) are the three bottlenecks.
• The realistic answer for Japanese companies is to invest in "the design of the smart user side" rather than in building infrastructure.

Whether or not this is a bubble can honestly only be settled a few years from now. But there is one thing I can say clearly. The more extraordinary the capital flowing into infrastructure, the more the outcome hinges on "what you changed in your own business with that compute." Whether you land in the 95% with low returns, or in the remaining 5%, that split is decided not by the number of GPUs but by the quality of your operational design. First measure where you stand with the AI Literacy Check, and if you're unsure about investment priorities, let's sort it out together in a WARP consultation. Don't get swept up in the size of the numbers; work backward from your own path to recovery. That, I believe, is the right way to engage with AI infrastructure in 2026.

Related Articles

• The New Competitive Axis of AI's Heavy-Industrialization and Capital Intensity
• OpenAI's Next Ten Years and $1.4 Trillion in Infrastructure Investment
• 2026 Predictions for the AGI Era and Enterprise Data Strategy
• The Current State of AI and Power Consumption
• Thinking About Data Management and Infrastructure in the AI Era
• Three Strategic Options for AI-Agent Management

Footnotes

1. Bloomberg, "Major tech companies' 2026 AI data center spending estimated at roughly $650 billion" (reported February 6, 2026) ↩ ↩²

2. OpenAI, "Announcing The Stargate Project" (January 21, 2025) https://openai.com/index/announcing-the-stargate-project/ ↩

3. Bank of England, "Financial Stability Report – October 2025" (October 8, 2025) https://www.bankofengland.co.uk/financial-stability-report/2025/october-2025 ↩

4. IMF, "Managing Director Speeches" (MD Georgieva, October 2025 dot-com comparison remarks) https://www.imf.org/en/News/Speeches ↩

5. Goldman Sachs, "AI to drive 165% increase in data center power demand by 2030" (February 4, 2025) https://www.goldmansachs.com/insights/articles/ai-to-drive-165-increase-in-data-center-power-demand-by-2030 ↩

6. IEA, "Energy and AI" (April 2025) https://www.iea.org/reports/energy-and-ai ↩

7. DataCenterDynamics (reporting on Meta's stance toward entering the cloud business and Anthropic's consideration of compute procurement, July 17, 2026) https://www.datacenterdynamics.com/en/news/ ↩