Big Tech is now officially held hostage by the AI bubble
So, the AI bubble drama is entering the home stretch.
By now, only a starry-eyed dreamer could fail to grasp what is happening with AI infrastructure investments.
What is happening? Let me tell you.
These are no longer investments — they are tribute. The “Magnificent 7 Stocks” are becoming hostages to sunk cost.
Amazon announces plans to spend $200B on capital expenditures in 2026. A year ago, the figure ($132B) already looked aggressive. Now it looks reckless.
Some might think this is a leap toward AI market dominance. But when you look at the free cash flow, the picture is turned on its head. This is a shift into territory where their math stops adding up.
Here are numbers anyone can verify. This matters, because I am certain that even inside Amazon there are clear-headed people. They know what lies ahead is not a stairway to heaven but a cliff edge. But there is nothing they can do about it.
So:
- $33B remained after capex in 2024.
- In 2025 — only $8B.
- With the announced $200B in spending, free cash flow turns negative.
This means one thing: more money is going out than coming in.
And that’s not all.
A company that spent years paying down debt is suddenly borrowing again. Data centers must be built, GPU must be purchased. Stopping is not an option.
Why not? Because losing the race is not an option.
Do you see what is happening? Fear of Missing Out (FOMO) is pushing them toward decisions that can no longer be rationally justified.
Let us ask ourselves: at what point does an investment stop being an investment?
When it can no longer be stopped. Then it becomes feeding an addiction. And yes, walking away starts to feel more dangerous than losing the return on investment.
The Lie We Choose to Believe
We are told this is a long-term bet. That it is the infrastructure of the future. That it is the “new electricity” (and all the rest of the marketing mythology).
But the plain truth is that data centers do not last for decades.
Two-thirds of this spending goes not to buildings but to rapidly obsolete hardware — GPU, server processors, HBMmemory, networking infrastructure, and cooling systems.
They do not become obsolete physically — they become obsolete economically, as computational efficiency requirements tighten and architectures shift.
Each new chip generation raises performance per megawatt. Older hardware begins losing ground on margins, compute density, and energy consumption.
Naturally, it gets replaced — and with it go the server nodes, network switches, cooling configurations, sometimes even the power distribution within the cluster.
This is anything but 30-year infrastructure. I would call it a pile of stranded assets in an accelerated replacement cycle.
And if a company starts spending $200B a year on infrastructure that will need replacing in a few years, it is signing up for a perpetual cost escalation where $200B today means $300B tomorrow, and so on.
These are tracks you cannot get off. Because, as I already noted, stopping means admitting failure.
The Real Numbers
Here is what we are witnessing:
Big Tech has started financing AI through a debt spiral that demands ever-escalating bets.
I do not want to follow the AI optimists. So let me present real CapEx / cash flow ratios that define the outlook for ROI.
Here are the numbers — and this is where the illusion starts to crack:
Amazon
(I have already mentioned the overall figures, but the details matter when comparing to other companies.)
2024:
- $116B operating cash flow
- $83B CapEx
- $33B free cash flow
2025:
- $132B CapEx
- free cash flow dropped to $8 B
2026 (planned):
- $200B CapEx
- growth of more than 50%
- an increase of ~$70B
Total: with $200B in capex free cash flow turns negative.
And yes, you noticed it too — its modus operandi is starting to look disturbingly like the doom loop that OpenAI fell into (and tried to drag everyone else along). I wrote about it here, but I strongly recommend bracing yourself before you start reading.
Amazon Debt
- Was $50B
- Borrowed an additional $15B in a single quarter
The Result
Amazon is turning from a cash-generating machine into a company that finances unfounded growth hopes through debt.
Here we see a doubling down on CapEx:
- 2025: $91B
- 2026 (planned): $175–185B
Meanwhile, net profit for 2025 was: $132B.
That means Google plans to spend on data centers more than the entire company earned in its last fiscal year!
Google Debt
- $21.6B — long-term debt at the end of the previous quarter
- $46.5B — long-term debt at the end of the following quarter (+$25B in a single quarter).
It would be extraordinary if this kind of trajectory did not worry Alphabet shareholders. So how did the company decide to reassure them?
Here is how: Google issued bonds with maturities of up to 100 years (I am not joking — a literal 100 years).
I will not go into this here. Perhaps I will dedicate a separate investigation to it. For now, I will simply note that the last company to do this was Motorola in 1997 — right before its collapse as a market leader.
Oracle
Has confidently fallen into the same spiral as OpenAI toward the end of 2025 (I have written about this before as well).
- The company’s debt exceeds $100B (once more, in words: one hundred billion dollars. That is more than the company’s total profit over 10 years).
- In late January and early February of this year, shares fell 50% from their peak and have been drifting around that level ever since.
The key takeaway: free cash flow is dying.
Investors have been used to valuing companies by:
- EBITDA — earnings before interest, taxes, and depreciation, where CapEx is not counted
or - FCF — the real cash remaining after capital expenditures
But now EBITDA looks fine while FCF is heading toward zero or negative!
The Impossible Math
Morgan Stanley forecasts roughly $400B in borrowing by cloud giants in a single year. There is no whiff of organic growth here for anyone. This looks far more like voluntary financial captivity.
The hard truth is that AI payback is not just uncertain — it is highly unlikely. And when companies say AI “will pay off,” nobody asks the most important question: where will the revenue come from?
A $20/month subscription will not pay back a multi-billion-dollar data center, since every AI query is effectively sold at a loss. Even $200 for a “professional” tier does not solve the problem if the infrastructure burns energy and capital faster than revenue grows.
To recoup hundreds of billions in CapEx, the average ticket price needs to be an order of magnitude higher.
For a mass-market user, the monthly payment would need to be at least $1,000!
Would you pay that?
No — you would be using free DeepSeek. Call it a hunch, but I suspect that curious Chinese AI-agents can afford to keep being that generous for a reason.
So the only hope is enterprise.
LLM would need to become so mission-critical that companies would agree to pay that much for workplace AI — for any of this to make sense for model providers.
In the enterprise segment, of course, we are not talking about run-of-the-mill AI agents like ChatGPT, but about dedicated clusters, API, custom models for business process integration, and so on.
But in my view, that scenario is unlikely. Here is why:
Even if you take the minimum enterprise subscription cost — say, $1,000 — you can effectively treat that as adding the same amount to an employee’s salary. That is a massive additional cost!
It would only make sense if AI could genuinely replace a significant share of employees — probably no less than 1/3.
But is that realistic? This is precisely where the whole scheme falls apart.
A Fundamentally Limited Technology
So-called AI is not capable of that. Not because the technology is “immature,” but because it is fundamentally limited. It is like trying to fly a commercial airliner to the Moon. What are the odds it gets there?
I will again (as I did in another article) refer to a large-scale study published in late October 2025 on arxiv.org — Remote Labor Index: Measuring AI Automation of Remote Work. Its results showed that while generative models deliver phenomenal performance on benchmarks, they fail completely in real-world projects. Not a single agent was able to complete more than 2.5% of tasks.
Models were found to stumble on basic things, unable to evaluate the quality of their own work, and making “rookie” mistakes.
Behind these results lies a pattern investors prefer not to discuss. LLM improvement slows as training costs rise.
This is called diminishing returns, and it is observed across all major model scaling studies. Moreover, such systems never become “error-free” — by definition they operate with probabilities. They can be impressive, but not absolutely reliable.
Business does not need “impressive probability” — it needs predictability. If a model is wrong 5% of the time, that is already a disaster for accounting, medicine, or legal decisions.
To replace a third of employees, AI would need to be near-perfect. But each additional fraction of a percent of accuracy costs more and more, and at some point spending growth outpaces quality growth. And then $200B a year starts looking not like an investment in the future, but like an attempt to buy the impossible.
Big Tech has started financing AI through a debt spiral that demands ever-escalating bets.
Conclusion
Based on everything above, the math that AI bubble participants are counting on will never work out.
The data centers and the debts will remain. The profits will not.
A staggering amount of resources has already been spent chasing the impossible — and more will follow. This is inevitable, because events unfold not spontaneously but according to a logic driven by the prospect of outsized returns.
The logic goes: as long as there is demand, windfall profits are possible — you just need to pour enough money into the technology that will deliver them. The possibility that the technology is fundamentally broken is simply not on the table. Because accepting that idea requires an entirely different mindset and different motivations.
Neither of those exist among those for whom the acquisition of ever-greater resources is the ultimate goal. These incentives and these mindsets rarely coexist — and that may be the most dangerous gap of all.