Dhaval Joshi's AI Bubble Warning: The Rolling Sequence Thesis Explained

DDR3 RAM — an aging, largely obsolete memory standard — is up roughly 600% in under a year. Silver has nearly tripled. And BCA Research's former chief strategist Dhaval Joshi says neither has much to do with AI models themselves. That's the short answer. The longer answer is more interesting.

Joshi's argument, laid out in a Fortune interview published August 16, 2026, is that the question everyone keeps asking — "is AI a bubble?" — is the wrong one. The better question, he says, is which specific AI-adjacent bubble is popping today, because capital is rotating through a "rolling sequence" of them rather than inflating one bubble that eventually bursts all at once.

Dhaval Joshi's AI Market Bubbles Thesis, Explained

Dhaval Joshi argues AI-adjacent markets aren't experiencing one giant bubble but a rolling sequence of smaller ones — capital sprints into a sector, inflates it fast, then rotates out just as quickly into the next trade. His test for what counts as a bubble is blunt: "If you can make a fortune in weeks or months, then lose it all just as quickly, that constitutes a bubble."

By that measure, 2026 has already produced a chain of examples that have little to do with AI models themselves and everything to do with AI-adjacent supply chains: legacy memory chips, industrial metals, software-as-a-service stocks, and semiconductor names have each seen the same run-up-then-rotation pattern, with capital moving from one to the next rather than sitting still in any single trade.

DDR3 memory and the $50,000 Corolla

The starkest example in Joshi's framework is DDR3 — an aging, largely obsolete RAM standard that surged roughly 600% in under a year as AI data-center buildouts strained memory supply chains and buyers scrambled for any available capacity, old or new. Paul Burchard, president of Artificial Genius, likened paying today's DDR3 prices to "paying $50,000 for a beaten-up 2007 Toyota Corolla" — old technology repriced purely on scarcity, with no improvement in what it actually does.

The mechanism is real, not manufactured: Samsung, SK Hynix, and Micron have collectively shifted the overwhelming majority of their combined production toward high-bandwidth memory built specifically for AI servers, leaving a shrinking pool of output for general-purpose and legacy RAM. DRAM prices broadly were up roughly 90% quarter-over-quarter in Q1 2026 and around 172% year-over-year, according to industry pricing trackers — DDR3's move is the most extreme slice of a supply squeeze affecting the whole memory market, not an isolated speculative event.

The rolling sequence, ranked from most to least bubble-like

Joshi's framework implies not every AI-adjacent trade is equally fragile. Ranking the current examples by how disconnected the price move is from any underlying earnings or demand story puts commodity-driven spikes at the top and revenue-backed AI companies at the bottom.

Six AI-adjacent trades, side by side

Laid out with the earnings question attached, the pattern behind Joshi's ranking gets clearer — the trades with no revenue story are the ones that moved fastest.

The earnings-quality question sitting underneath

Joshi's sharper point is about earnings quality at the hyperscalers actually building AI infrastructure. Google went free-cash-flow negative for the first time in its history in 2026, and BCA Research strategist Peter Berezin's modeling projects Microsoft, Alphabet, Amazon, Meta, and Oracle will collectively see capital expenditure overtake free cash flow by 2027 — aggregate hyperscaler capex is estimated to approach $1.2 trillion that year, per BofA estimates cited in industry coverage, up from roughly $860 billion in 2026.

The four largest spenders aren't moving in lockstep, though. Microsoft is projected to remain the free-cash-flow outlier, with FCF estimated to rise from $19.6 billion in its June 2026 quarter to roughly $46.2 billion in 2027 on the back of Azure's margin profile, according to analyst estimates reported by Benzinga. Alphabet and Meta, by contrast, are modeled at roughly negative $18.4 billion and negative $21.7 billion in free cash flow respectively for 2027, with Amazon around negative $6.8 billion — meaning three of the four biggest AI spenders are expected to still be burning cash on a free-cash-flow basis two years from now.

Joshi frames the market's real underlying question as: "How is the E high?" — shorthand for whether the reported earnings backing today's AI-infrastructure valuations are as durable as the price-to-earnings multiples assume. That question lands in the same week Anthropic reported preliminary Q2 2026 revenue above $11.5 billion — a genuine data point on the other side of the ledger, showing at least one AI-native company with hyperscaler-adjacent growth backed by actual usage revenue rather than capex narrative alone.

What the headline misses

A 600% DDR3 price move sounds like the cleanest bubble evidence available, but it's also the easiest one to overread. Legacy memory prices are a supply-side story as much as a speculative one — fabs deliberately shifted capacity toward high-bandwidth memory, and DDR3 is a tiny, largely irrelevant slice of total memory output getting squeezed in the process. That's a real distortion worth flagging, but it says less about whether "AI is a bubble" broadly than it does about a niche supply-chain mismatch.

The capex-versus-free-cash-flow numbers deserve the same caution. Announced and projected capex isn't the same as capital already spent, and multiyear infrastructure buildouts — like railroads or fiber networks before them — routinely overbuild in the short run while still leaving durable value once demand catches up. Negative free cash flow at Alphabet or Meta in 2027 reflects a deliberate reinvestment choice their boards signed off on, not a company running out of money. The risk Joshi is actually flagging is narrower and more specific: that the earnings supporting today's multiples might prove less durable than assumed, not that the entire AI capex cycle collapses on a fixed timeline.

Who else is watching the same signal

Joshi isn't a lone voice. Arm co-founder Hermann Hauser made a similar case, telling CNBC that the AI revolution is real but so is the bubble risk sitting on top of it. JPMorgan CEO Jamie Dimon, Goldman Sachs CEO David Solomon, and Amazon founder Jeff Bezos have all separately flagged some version of AI-valuation excess in 2026, even while their own institutions keep underwriting and financing the buildout — a tension that's become its own recurring theme in 2026 markets commentary. OpenAI CEO Sam Altman has made similar comments about parts of the AI trade being overextended, an unusual admission from an executive whose own company's valuation depends on investors believing otherwise.

You can track the companies and valuations at the center of this debate on our AI Valuations Dashboard, and see how the capex side of the equation compares across the largest spenders on our Big Tech Earnings tracker.

Is AI a bubble in 2026? What to watch instead

The practical takeaway from Joshi's rolling-sequence framing is that treating an entire AI portfolio as one undifferentiated bet is the wrong instinct. A commodity-driven spike like DDR3's 600% run has no earnings underneath it and is the most exposed to a fast unwind; a revenue-growing frontier lab reporting real quarterly numbers, like Anthropic, is a structurally different risk — even if both get lumped into the same "AI bubble" headline.

The concrete checkpoint worth watching is whether Microsoft, Alphabet, Amazon, Meta, and Oracle's actual 2027 capex-versus-free-cash-flow numbers land where Berezin's model projects. If they do, Joshi's earnings-quality warning was right. If the hyperscalers find a way to keep the E genuinely high — through AI-driven revenue that actually shows up in the numbers rather than just capex announcements — the rolling sequence narrows to the commodity plays it started with.