Overview of Here's How The AI Bubble Bursts — With Paul Kedrosky
Alex Kantrowitz interviews investor and analyst Paul Kedrosky about whether the AI boom is actually a bubble and, if so, what an unwind could look like. Kedrosky argues that AI infrastructure spending has reached historically extreme levels, financed increasingly with external capital, while the economics under the hood are deteriorating: token prices are falling fast, GPU hardware depreciates quickly, and returns may not keep pace with the pace of spending. He believes AI is a genuinely transformative technology, but that does not mean current investment levels are economically justified.
The Core Thesis: AI Spending Is Historically Unprecedented
Kedrosky’s main argument is that today’s AI buildout is larger and faster than previous infrastructure booms.
How it compares to past buildouts
He compares AI capex to:
- railroads
- electrification
- interstate highways
- fiber optic buildouts
- World War II rearmament
His conclusion: by several measures — share of GDP, contribution to GDP growth, and non-residential fixed investment — AI spending is already larger than nearly all prior infrastructure booms, and it is happening in a much shorter time frame.
Why the pace matters
Unlike past buildouts that unfolded over decades, AI infrastructure is being built in a compressed window. That makes it harder for capital markets to slow down, reassess returns, and course-correct before overbuilding happens.
Why Kedrosky Thinks the Economics Break Down
He argues that the AI buildout is structurally different from a real estate or utility investment, even though investors often treat it that way.
The “data center as real estate” analogy is misleading
He says data centers are often financed like commercial real estate projects, but they behave very differently:
- GPUs and other hardware need frequent replacement
- data centers require ongoing capital expenditure, not just upfront build costs
- the core revenue unit — AI tokens — is rapidly deflating in price
Three key economic pressures
-
Continuing capex
These systems need ongoing reinvestment, unlike apartment buildings or most traditional real estate. -
Fast hardware churn
GPUs fail, get upgraded, or get replaced by newer inference-specific chips. -
Token price deflation
The thing being sold is getting dramatically cheaper over time, which squeezes margins.
Kedrosky says that combination makes the investment case far more fragile than a standard project-finance model suggests.
Why “More Demand” May Not Save the Model
A major counterargument is that cheaper tokens lead to more usage — a Jevons-paradox style dynamic. Kedrosky says that effect exists, but is likely far too small relative to the speed of price declines.
His rebuttal
- Token prices have been falling very quickly
- Model differences are converging
- Open-source and Chinese models are increasing price competition
- Harnesses and orchestration layers are becoming more important than the underlying models
His view: even if usage grows, it may not grow fast enough to offset the collapse in per-unit pricing.
The Role of Model Convergence and Harnesses
Kedrosky argues that the biggest gains are increasingly coming not from frontier model training itself, but from:
- harnesses like coding agents and orchestration layers
- post-training
- RLHF
- quantization
- product wrapping around models
Main implication
If the real value is shifting to wrappers and workflow tools, then pouring billions into frontier pre-training runs may become harder to justify.
He suggests that model quality is converging enough that:
- users can barely tell many models apart
- competition is increasingly about price and marketing
- switching costs are lower
- commoditization pressure is rising
Why AI Companies May Be Forced to Move Upmarket
Kedrosky argues that frontier labs may need to move into applications and vertical software to make the economics work.
Why this creates tension
If AI companies go after businesses like Palantir or Figma, they risk:
- alienating customers
- competing with the very firms that build on their models
- becoming generalized oligopolies instead of neutral infrastructure providers
He says this pattern is not new — Microsoft did something similar by moving up the stack and eating application markets — but it would still be painful and politically toxic.
Why the “AGI Call Option” Argument Doesn’t Convince Him
The strongest bullish argument, in his view, is that investors are effectively paying for a call option on AGI.
His response
- He thinks that is a rhetorical escape hatch, not a real financial framework
- You can’t justify unlimited spending just by invoking AGI
- Investors may use that story publicly, but privately they are comparing AI projects to other capital-intensive investments with similar expected returns
He says serious investors are usually much more cynical than the public narrative suggests.
What Could Burst the Bubble?
Kedrosky says there is no single trigger. The bubble could unwind through multiple paths.
Possible triggers
- higher interest rates or tighter capital markets
- public-market scrutiny after IPOs
- investors demanding better returns
- export controls and market balkanization
- re-rating of AI companies as utility-like businesses
- reduced enthusiasm once the true economics become clearer
His view is that the system is overdetermined: there are many ways it can stop, and only a few ways it can keep going.
What Happens When It Stops
Kedrosky warns the unwind would spread through the financial system much like the 2008 crisis.
Channels of contagion
He points to:
- high-yield debt
- investment-grade debt
- insurance companies
- asset managers like PIMCO
- S&P 500 index funds due to hyperscaler concentration
His concern is that AI-related debt and exposure are already embedded widely enough that many institutions are “in it whether they want to be or not.”
Why Investors Still Keep Funding It
According to Kedrosky, the money keeps flowing for a few reasons:
- Groupthink and momentum
- Check-size pressure among large funds and sovereign wealth funds
- Fear of missing out
- Technology, real estate, credit, and policy all reinforce one another
- Geopolitical competition with China
He argues that these forces make the boom self-reinforcing until sentiment or capital markets shift.
China: More Insulated, but Still Vulnerable
Kedrosky thinks China may weather an AI overbuild better than the U.S. because:
- its economy is less dependent on consumer spending
- government-directed investment can absorb losses
- overbuilding is already a familiar pattern in China
But he also says China is experiencing its own version of excess, with local officials racing to build data centers to impress the central government, similar to past overbuilds in batteries, solar, and real estate.
Kedrosky’s Bottom Line
- AI is real and transformative.
- Current infrastructure spending is historically extreme.
- Token prices are deflating rapidly.
- Hardware and infrastructure require continuous reinvestment.
- Public narratives about AGI often mask much more cynical return expectations.
- The bubble, if it exists, could unravel in multiple ways rather than one dramatic crash.
Final Takeaway
Kedrosky’s view is not that AI is fake or useless. It is that the technology may be extraordinary while the current investment structure is unsustainable. In his telling, the likely end state is not the disappearance of AI, but its normalization: AI becomes a utility-like layer in the economy, with far lower margins and returns than today’s exuberant spending implies.