// learn · Market history

The AI Bubble: hype vs. what holds up

"Is AI a bubble?" is the wrong question. The useful one is narrower: which parts of the AI trade are paid for by real revenue today, which are priced on a future that may or may not arrive, and how do you tell them apart before the market does it for you?

Bubbles and real revolutions look identical early on

The internet was both a genuine revolution and a bubble that wiped out most of the companies riding it. Both things were true at once. Railways, electricity, the dot-com era, every real technology shift has drawn in far more capital than the winners could ever justify, then repriced hard. AI can be transformational and overvalued in the same breath. Treating "it's real" and "it's a bubble" as opposites is the first mistake.

Follow the cash, not the narrative

The cleanest filter is boring: who is actually being paid, and by whom?

Picks and shovels. Chipmakers and infrastructure providers are booking real, auditable revenue now. The risk isn't "is it real", it's whether current growth rates are already priced in for a decade.
Platforms. Large software and cloud firms are monetizing AI features into existing paying customers. Grounded, but the incremental revenue is often smaller than the valuation implies.
Pure story stocks. Companies whose AI revenue is a roadmap, not an income statement. This is where bubble dynamics concentrate, and where the biggest moves, up and down, happen.

The circular-financing tell. When the same dollars cycle between a chipmaker, its customers, and the funds backing both, demand can look stronger than end-user spending actually supports. Real end demand is what survives a downturn.

The red flags on the table right now

None of these prove a bubble on their own. Together, they're the reason serious people are asking the question (figures as of late 2025, via BrokerChooser and public filings):

Stretched valuations. NVIDIA has traded around a P/E of ~54, roughly double the market average, and Palantir's multiple has run into the hundreds, hard to square with its growth and margins.
Index concentration. A single chipmaker has reached ~8% of the S&P 500 and ~14% of the Nasdaq-100. When one name carries that much weight, the whole index rides its story.
Derivatives boom. The market went from ~100–200 new ETFs a year pre-2020 to 900+ in 2025, over half using derivatives, more than a third leveraged. Both record highs.
Margin debt. Investor borrowing against portfolios hit ~$1.2 trillion in late 2025, up ~45% year-on-year, much of it presumed to be in AI names.
Single-headline swings. Oracle added ~$250B in market value overnight on one AI-infrastructure announcement, then gave it back, a sign the market's conviction is thinner than the price implies.

The 2026 turn: the risk moved from valuation to physics and unit economics

Through 2025 the debate was about multiples. In 2026 three harder problems surfaced, and none of them needs a crash to bite.

The bottleneck is physical. Top-tier GPUs are effectively sold out across generations, B200 rental rates rose about 23% in a single month, and grid interconnection queues now run into years. Compute is capped by power and hardware, not ambition. Hyperscaler capex guidance jumped to roughly $720B for 2026, up from about $380B in 2025.
A token price war. The price per million tokens is falling 60 to 70% a year as custom silicon and rival models compete. One launch cut a frontier model's price 67% overnight. Great for buyers, brutal for anyone who borrowed to build inference capacity expecting fat margins.
The first margin call on returns. One study found only about 18 cents of every enterprise AI dollar reached production value, the rest lost to rework, bug-fixes and review friction. The spending is real, the payback is not there yet.
The capex-versus-revenue gap. The marquee case is roughly $1.4 trillion in datacenter commitments against about $13 billion of revenue, funded partly by circular vendor financing where a chipmaker invests in its own customers. Debt raised for datacenters could pass $1 trillion by 2028, much of it below investment grade.

Positioning is not proof. The most-watched bear, Michael Burry, calls a record-high market a warning sign. His book is not blanket-bearish, it is a split: short the AI and chip complex (the SOXX semiconductor ETF, Nvidia, Micron, Palantir, Tesla, Applied Materials, Caterpillar, Oracle), with the option contracts rolled out to 2027, while long a value and consumer basket away from the AI trade (Zoetis, Mercado Libre, Fiserv, Lululemon, Freddie Mac, Flutter, PayPal, Adobe). He rotates often, trimming his Oracle short and selling Microsoft and DraftKings, so treat any snapshot as dated. A famous investor's positions are a view, not evidence. Treat them as one more signal, never a verdict. Full book via the 13F tracker.

We track these as a live, sourced composite on the 2026 KPI dashboard: concentration, valuation, capex, the price war, realized ROI and the physical limits, each scored and weighted, with every number carrying its source and date.

Where the stress would show first

If the trade unwinds it won't hit everything equally. Read neutrally, the most exposed corners are the ones furthest from present cash flow:

Cash-burning pure-plays. Frontier labs and AI startups running large losses. Even AI leaders concede many labs will not survive, and funding is already concentrated in a few: OpenAI and Anthropic took about 43% of all H1 2026 startup funding.
Froth in exits. The IPO and acquisition market is the hottest since 2021, with more than 70% of venture money going to AI in Q2 2026. Hot exit windows tend to close fast.
Leverage stacked on top. Margin debt hit a record ~$1.5 trillion, up ~49% year-on-year, and leveraged-ETF assets roughly doubled in a year. Borrowed money amplifies the move in both directions.

Further reading, neutral sources: why some AI labs fail, the likely first casualties, the AI-led IPO boom.

Signs you're paying for hype

Valuation justified only by a total-addressable-market slide, not by margins or cash flow.
"This time is different" used to dismiss profitability entirely.
Revenue that depends on a handful of customers who are themselves unprofitable.
Price moving on announcements and partnerships rather than shipped, paid-for products.

If you think it's a bubble: how people short it, and why it's brutal

Deciding the AI trade is overvalued is the easy part. Profiting from that view is one of the hardest trades in markets, because being right and being right on time are two different things. As the old line goes, the market can stay irrational longer than you can stay solvent. These are the common instruments, and the catch built into each.

Short selling shares. Borrow the stock (the crowded AI names: NVDA, PLTR, MSFT, META, GOOG), sell it, buy it back lower. Downside is theoretically unlimited, you pay borrow fees, and if too many shorts pile in you can get caught in a squeeze. Needs a margin account with short access. Highest risk of the lot.
Short or inverse ETFs. Betting on the sector instead of one name spreads single-stock risk (one earnings call can move a stock 20–30%). You can short broad AI ETFs (QQQ, AIQ, ARKK) or simply buy an inverse ETF, no margin, no borrow: PSQ (inverse Nasdaq-100), SARK (inverse ARKK), QQQD (a Magnificent-7 short). The catch: fees are higher (PSQ ~0.95% vs QQQ ~0.20%) and daily rebalancing makes them drift over time, built for short holds, not a months-long bubble call.
NASDAQ-100 futures (NQ). The Nasdaq-100 is ~50% AI/big-tech by weight, so shorting NQ is a direct sector bet, and going short is as simple as going long. But margin is steep (~$30k for full NQ, ~$3k for Micro NQ) and the leverage is unforgiving, a ~6% index move against a full contract can wipe the margin.
Put options. Loss capped at the premium, the cleanest defined-risk way to position bearish, with built-in leverage. But time decay is the enemy: you must be right on direction and timing, spreads can be wide (favour high open-interest contracts), or the option expires worthless even if you're eventually right.
CFDs. Leveraged short exposure with a small account, but you trade against the broker (wider spreads), pay overnight financing, and it's banned or restricted in many places (notably U.S. retail). Roughly 3 in 4 CFD traders lose money over time.

Hedging is not the same as betting. Buying a little downside protection on a portfolio you already hold is risk management. Putting on a large, leveraged short to "call the top" is a directional bet with asymmetric pain, most people who try to time a bubble's peak lose money doing it. If you're a beginner and still want exposure to the downside, plain (non-leveraged) inverse ETFs are the least dangerous door. Know which one you're actually doing.

Fuller broker-by-broker breakdown: BrokerChooser, how to short the AI bubble.

How Peaky Radar treats the AI trade

We don't call tops or bottoms. The radar flags momentum, volume and new listings across the AI complex, including the speculative small caps, and labels them for what they are. A signal that a name is moving is not a claim that it should be moving. You get the scan and the context; the decision is yours.

Educational market information, not financial advice. Nothing here is a recommendation to buy or sell any asset. Some links on this site are affiliate links, always disclosed. Markets carry risk of loss, do your own research and consider a licensed advisor.

← Back to Peaky Finance