AI - A Bubble or a Revolution?

Introduction - 2025 - Between a Capital Rally and Turbulence Around Nvidia

Much has been said recently about fears of a financial bubble in the AI sector, especially around AI infrastructure. I decided to dive into this highly important and fascinating topic and bring some clarity. The article below offers explanations that are accessible even to readers without a financial background.

Since early 2023, technology stocks have seen historic gains: the combined market cap of the Big-Tech group (Microsoft, Alphabet, Amazon, Apple, Meta, Nvidia) rose from 6.5 trillion dollars to roughly 19.8 trillion dollars - a jump of more than 200%. In parallel, the S&P 500 index rose about 85%, and the Nasdaq about 130%, with a significant portion of the gains driven by a very small number of AI-related stocks.

At the same time, capital investment in AI infrastructure reached unprecedented levels: the four major cloud giants alone (Amazon, Google-Alphabet, Meta, Microsoft) invested 282 billion dollars in the first nine months of 2025 - in long-term physical assets such as data centers, power infrastructure, and AI-chip technologies. This pace is expected to reach 375-400 billion dollars for the full year - a 52% increase over 2024 (247 billion dollars).

Recent weeks highlighted just how sensitive the market is to this topic:

• Before Nvidia’s quarterly earnings, investors were already concerned about an “AI bubble”, and the stock fell sharply after soaring hundreds of percent in three years.
• The results themselves were extremely strong and beat Wall Street expectations: 57B dollars in quarterly revenue, up 62% year-over-year. The stock jumped in after-hours trading - but showed high intraday volatility the next day.
• Days later, reports surfaced of negotiations between Google and Meta for a massive deal around Google’s TPU chips. The mere possibility that Meta might shift part of its workloads from Nvidia GPUs to Google TPUs sent Nvidia’s stock down and boosted Alphabet and Broadcom.

On one hand - enormous valuations and capital expenditures never seen before. On the other - extreme volatility, growing nervousness, and serious questions about long-term sustainability.

The big question: Are we in a financial AI bubble, or in the midst of a real Data-Center revolution that will remain with us for decades even if stock prices correct?

Part A: Why This Does Look Like a Bubble

1. Extreme Concentration of Market Value - Classic Late-Cycle Behavior

AI mega-caps now constitute about one-third of the S&P 500 and an even larger share of the Nasdaq-100. In practice, an investor who bought a “market index” actually bought concentrated exposure to a handful of AI-driven stocks rather than broad economic diversification.

A key metric signaling possible bubble conditions is the Shiller CAPE Ratio (Cyclically Adjusted Price-to-Earnings), which compares stock prices to average earnings over the past decade. During the dot-com peak it reached ~44. Today, at the end of 2025, it is around 40 – its second-highest reading in history. This is not a formal indictment, but it is a clear red flag.

2. Valuations, Debt, and Pricing Detached from Cash-Flow Reality

Nvidia- now the world’s most valuable company - briefly crossed a 4.5 trillion-dollar valuation this year, exceeding the GDP of most countries, largely based on expectations of exponential cloud-driven demand for GPU compute.

OpenAI is valued at ~500 billion dollars despite lacking proven long-term profitability. The company is expected to report cumulative losses of ~44 billion dollars by 2028. If you ask Sam Altman - money is secondary; he is trying to build a revolution. But a meaningful portion of the money comes from hyperscalers and flows back to them via cloud-service contracts - a circular flow in which the same dollar “moves between pockets” within the same ecosystem.

Traditional players like Oracle have taken on substantial leverage to fund aggressive data-center expansion. Oracle raised 18B dollars in debt in September 2025, and Moody’s warned of prolonged negative free cash flow.

Common denominator: valuations rely on the assumption that this growth spree continues at the same pace for at least a decade. Minor changes in interest rates, regulation, or competition could undermine this scenario quickly.

3. CAPEX Far Outpacing Profit Growth – “Money Leaves Faster Than It Enters”

AI infrastructure CAPEX is accelerating at breakneck speed. In the first nine months of 2025 alone, big-tech giants invested 282B dollars in servers, power, real estate, and chip capacity. Annual spending is projected to exceed 375B dollars – a dramatic revision from prior forecasts.

To justify this, companies must generate revenue and profits approaching that scale – not just user growth or hype.

Meanwhile, financial reports show free cash flow shrinking as CAPEX rises. Analysts forecast that CAPEX will reach 94% of operating cash flow (after dividends and buybacks) in 2025–2026, up from 76% in 2024.

Meta, for example, is expected to see free cash flow drop from 54B dollars (2024) to roughly 20B (2025) – a 63% decline.

4. Enterprise Adoption: Lots of POCs, Very Little Measurable Value

MIT and consulting-firm research shows that AI initiatives in most organizations are still at experimentation stages. According to MIT’s 2025 enterprise AI report: 95% of generative-AI projects fail to generate measurable ROI.

Another P&S Global survey found that 42% of companies abandoned most of their AI initiatives in 2025 – up sharply from 17% in 2024.

The meaning: plenty of marketing and hype, but limited real business impact so far.

5. Closed Money Loops and a “Hype-Based” Economy

When a chip manufacturer (Nvidia) invests in an AI startup (OpenAI), which commits to using its chips through a third-party cloud partner (you can guess who), and that partner also invests in the startup – a circular system emerges that inflates expectations and obscures basic questions:

Who is the real paying customer? For what? And for how long?

In such a system, weakness in one player (a failed funding round, a CAPEX slowdown) can trigger cascading cancellations across the chain.

Part B: Why It's Not Just a Bubble - The Infrastructure Is Real and Permanent

1. The Money Is Going into Concrete, Steel, and Chips - Not Just Screens

Unlike the dot.com bubble, where investments went into companies without physical assets, today’s AI boom is translating into massive physical build-outs:

• Hundreds of new data centers globally, each consuming hundreds of megawatts
• Substations, high-voltage lines, subsea-cable landing connectivity
• Advanced cooling systems
• Power-generation and energy-infrastructure companies benefiting from soaring demand

U.S. data-center power consumption is projected to rise 22% in 2025 to 61.8 GW, and double by 2030 to 134 GW.

2. Data Centers Resemble Long-Term Real Estate, Not Startups

Data-center operators sign 7–20-year contracts with hyperscalers and GPU-cloud providers. This behaves more like industrial real-estate leasing than a speculative software contract.

Facilities are pre-designed for high density, liquid cooling, modular power, and generational hardware refreshes.

3. Real-Economy Adoption Is Rising

Outside startups, AI adoption is growing in:

• Finance
• Energy
• Industry and manufacturing
• Logistics
• Healthcare

78% of organizations now use AI in at least one business function (up from 55% the prior year).

4. History Shows: Bubbles Leave Infrastructure Behind

Railroads, oil & gas, optical fiber, the early Internet – valuations crashed, but the infrastructure remained and powered decades of growth.

The same will likely happen here: even if AI-stock valuations decline 20–40%, data centers, networks, and power infrastructure will stay in place and evolve.

Sidebar: GPU-Only Future, or a Mixed World of GPU + TPU?

A key sign of market maturation is the emergence of real competition to Nvidia from specialized accelerators.

Google’s TPU ecosystem is now central:

• TPUs are ASIC accelerators purpose-built for AI workloads
• Google designs them; Broadcom manufactures them
• Google is adding MediaTek as a secondary partner to reduce costs and diversify

Reports from late 2025 indicate Meta is negotiating a multi-billion-dollar TPU deal, first via Google Cloud and later through on-prem deployment (beginning 2027).

This means:

• GPU dependency is no longer absolute
• Competition could improve pricing and energy efficiency
• Investors who bet exclusively on “the GPU story” may see value shift toward Google, Broadcom, and other accelerator vendors

Part C: What Happens If the Bubble Does Burst?

1. What a Burst Might Look Like - Even with Strong Underlying Demand

We must separate:

• Financial markets
• Real-world demand for compute and data-center capacity

A bubble usually bursts in prices first – not in real-world usage.

Possible scenario:

• AI-related stocks fall 30–60%
• Startup fundraising collapses; debt becomes expensive
• Public companies delay or pause CAPEX on borderline projects
• Small/medium GPU-cloud providers (heavily leveraged) collapse or get acquired

Yet at the same time:

• Chip shortages may continue in some regions
• Power scarcity remains
• Enterprise AI demand still grows

The market shifts from “grow at any cost” to “only profitable and efficient models survive.”

2. Implications for Data Centers and Chips

Data Centers:

• Early-stage projects with no signed customers or power contracts will be paused
• Facilities with long-term hyperscaler contracts will continue operating (possibly with pricing pressure)
• M&A activity will rise as strong players acquire distressed assets
• Regulators may slow approvals due to public pressure around land and electricity consumption

Chips:

• A temporary oversupply is possible if companies offload GPUs/TPUs or cancel orders
• Prices may fall, hurting margins but lowering compute costs for survivors

3. What Happens to AI Demand After a Burst?

• Real use cases do not disappear:
• Organizations showing real ROI will continue and expand AI initiatives
• “Nice-to-have” projects without clear value will be cut
• Market shifts from “everyone does everything” to “a few proven use cases dominate”

The post-2000 Internet is a good analogy: many companies collapsed, but usage skyrocketed afterward. The same is likely for AI.

Conclusion

From a capital-market perspective, there are clear signs of a bubble:

• Extreme valuation concentration
• Pricing based more on expectations than profits
• Daily volatility around every Nvidia or Google earnings report

But on the ground, a deep infrastructure layer is being built:

• Data centers, power systems, cooling, and networks at unprecedented scale
• Purpose-built chips like GPUs and TPUs with improving performance-per-watt
• Growing enterprise AI adoption across every major sector

Even if markets correct sharply, the world will not return to a pre-AI era - just as it did not return to a pre-Internet era after 2000. Those who remain standing will be the ones who connect real demand with stable infrastructure – not those riding the hype alone.