A sorting, not a crash
Why a financial mania that physically cannot overbuild is still not good news
On Tuesday last week, the Kospi – South Korea’s benchmark stock index – tripped its circuit breakers twice in a single session. SK Hynix and Samsung Electronics each lost more than 12 per cent, dragged down by an overnight AI and semiconductor sell-off on Wall Street and investor doubts about whether AI capex can keep growing at this pace.
By Wednesday, both stocks had mostly recovered: a Samsung buyback rumour and SK Hynix’s decision to raise its planned US ADR sale to $29 billion, up from an earlier $10 billion estimate, triggered a short squeeze that pushed both stocks back up within hours. The panic and the correction arrived a day apart.
That sequence is a useful test case, but not because it settles whether the AI bubble is bursting. Instead, it shows how badly that question is usually framed. A one-day scare in Seoul doesn’t tell you much about a financial mania, but the slower pattern sitting underneath it might: who actually captures the value this boom is supposed to create, and how concentrated that capture already is.
The overbuild that isn’t arriving
Every previous technology bubble, in Carlota Perez’s framework, follows the same arc: markets correctly sense a revolution, excitement produces a financing bubble, the bubble produces an overbuild, the overbuild produces a crash, and the resulting oversupply of cheap infrastructure produces a “golden age.” Railways, radio, the internet. Fibre got laid faster than demand could fill it, prices collapsed, and that cheap fibre is what the next twenty years of internet traffic ran on.
Gavin Baker, a hedge fund manager who has covered Nvidia for years, argues this cycle might not follow that arc, for a reason that has nothing to do with sentiment. His core claim: “hyperscalers would overbuild if they could, but they simply cannot” – because, in his phrase, “the world is fundamentally short both watts and wafers,” a bottleneck this newsletter flagged back in January, and Baker thinks the shortage could persist for years.
Writing ahead of Nvidia’s February earnings, he noted that rental prices for “nearly four-year-old H100s” had “gone vertical over the last two months” – the opposite of what you’d expect if compute were oversupplied – and that “even six-year-old A100s remain fully utilized per AWS.” None of the previous bubbles Perez studied ran into a comparable physical bottleneck. A classic crash needs an overbuild first, and an overbuild needs the ability to build past demand.
This is the detail the 2000 comparisons tend to skip. Owen Lamont points out that expected long-term S&P 500 earnings growth has hit 20.2 per cent, above the dot-com peak of 18.6 per cent in 2000, and that analysts have a forty-year track record of predicting roughly 13 per cent annual growth while the market delivers about 7. That gap between expectation and reality is real, and it is dot-com-shaped. But the mechanism that turned 2000’s gap into a crash needed spare capacity to dump on the market. This cycle is short of the one input that the mechanism requires.
Marking each other’s homework
That doesn’t make the boom healthy, only differently fragile, and Robin Wigglesworth has been tracking where the fragility actually sits: in the accounting, not the chip supply. Alphabet booked $37.7 billion of “other income” in the first quarter of 2026, more than half its net income for the period. Amazon’s equivalent figure was nearly $16 billion, up from $2.7 billion a year earlier, also close to half its quarterly profit. Wigglesworth’s assessment is blunt:
This is another sign of just how comically codependent the AI tech industry has become.
That income is the rising paper value of their own stakes in Anthropic, whose valuation has gone from $183 billion in September to $380 billion now. Alphabet and Amazon hold those stakes partly because Anthropic spends a large share of the cash it raises on cloud computing from Google and AWS: OpenAI and Anthropic between them now account for roughly half the cloud order books at Oracle, Alphabet, Amazon and Microsoft. The companies funding the boom and the companies whose earnings confirm the boom are, increasingly, the same companies.
What that circularity is propping up became visible in May, when Ed Zitron published OpenAI’s audited 2025 financials, independently verified by the Financial Times. OpenAI lost $38.5 billion on $13 billion of revenue, itself above the company’s own internal target.
Strip out the one-off non-cash charge from its conversion to a for-profit structure and the operating loss alone was still $20.9 billion against that same $13 billion in revenue, with research and development outrunning the entire top line. Microsoft collected $17.2 billion of it, largely for the compute OpenAI rents to train and run its own models. Just over $50 billion in assets sat on the balance sheet at year end, only about half of it actual cash, enough to buy time but not to chart a path to break-even.
OpenAI’s numbers are the extreme version of a sector-wide pattern, not an outlier. Exponential View’s bottom-up model of the AI economy, published last week, puts total AI sector revenue at $110 billion over the past year. Even so, the model finds that hyperscaler AI revenues only just cover the depreciation on the infrastructure built to earn them – break-even, with no margin for a bad quarter.
Where the rent actually goes
The circularity obscures a separate and more durable question: which layer of the stack keeps the money once the circular accounting is stripped out. Gennaro Cuofano’s reading of Apple’s WWDC 2026 announcements makes the case that the answer isn’t the model layer. Apple shipped an operating system in which the primary user of an app is increasingly an agent rather than a person: Spotlight as a knowledge graph the agent reads, App Intents as the surface it acts on.
His point is structural rather than Apple-specific:
The interesting question is not which layer is winning. It is which layer captures the rents over the next decade. And the structural pattern across previous platform shifts — desktop, web, mobile — is consistent: the layer that owns the interface to the human eventually captures the rents the layers below it generate.
Apple doesn’t need the best model, only ownership of the place the agent runs.
Satya Nadella made a version of the same argument about Microsoft, in a manifesto built around what he calls “token capital” and “human capital”: a frontier model without an ecosystem around it isn’t stable, because the model doesn’t know your company, your workflows, your compliance requirements. Saanya Ojha’s reading of that manifesto is the sharper version: “Of course Microsoft believes in ecosystems,” she writes – it has been an ecosystem company since before most AI founders were born. “The interesting part is that Satya is probably right.”
The frontier bet has started looking expensive and brittle, and if a model can be pulled from customers overnight by an export control, betting the firm on any single one of them is an architectural risk rather than a strategy. The value, in Ojha’s framing, “migrates from the model to the harness around the model” – permissions, memory, workflow integration, the connective tissue that turns a prototype into something a regulated enterprise can actually deploy.
JPMorgan’s numbers from this spring give that argument a shape you can measure: the ten largest S&P 500 companies now account for 41 per cent of the index’s market capitalisation, against 33 per cent of its earnings. That gap is usually read as a single bet on “Big Tech”, but the deployment-layer argument says it isn’t one bet – it’s several, bundled together.
Nvidia owns compute, not the interface, which is exactly the position Cuofano and Ojha say won’t hold its premium. Apple owns the interface and little else, which is exactly the position they say will. Microsoft and Alphabet straddle both layers, which is why they’re harder to call. If the deployment-layer thesis is right, the correction running through those ten companies won’t be even – it will sort them.
A pattern, not an anomaly
This concentration isn’t new to AI, which is the part worth sitting with. A Federal Reserve Bank of Chicago analysis found that the IT sector alone produced roughly 45 per cent of total US productivity growth over the past four decades, while accounting for only about 8 per cent of the economy’s output. Productivity gains from information technology have been landing in a narrow slice of the economy since long before anyone called it AI. The current boom looks less like a new phenomenon than an old one running at higher resolution.
The clearest snapshot of that unevenness right now isn’t a productivity statistic. It’s a spending one. Ramp’s data on corporate AI spend, tracked through its own card and expense platform, shows the top one per cent of firms spending $7,449 per employee per month on AI tools. The median firm spends $11.38. Both can describe themselves as “investing in AI” while being separated by a factor of six hundred.
What bursting would actually mean
If this cycle can’t overbuild the way fibre did, it probably won’t crash the way the dot-com bust did either, because there is no glut of cheap GPU-hours waiting to flood the market and collapse prices to zero. What it can still do is what that Tuesday in Seoul hinted at and then walked back within a day: a sudden repricing of which companies were ever going to capture the returns the market has priced into all of them.
The useful question isn’t whether the AI bubble bursts. Michael Burry, who shorted subprime in 2008 and has been betting against Nvidia since late 2025, is watching for that event; so is anyone watching US margin debt for a signal: it hit a FINRA record of $1.42 trillion in May, up 54 per cent on a year earlier. Owen Lamont’s warning is different in kind, not a crash to watch for, but a slow disappointment as earnings fail to grow as fast as expected over the next five years. Either way, leverage decides how violent a move is, not which companies were ever holding the value to begin with.
The Chicago Fed data and the Ramp numbers point to the question that actually matters, the one already running quietly in the background, without a circuit breaker to mark it: most companies buying into this boom were never going to be where the value lands. Call that a sorting rather than a crash. It’s underway already, while the headlines wait for Seoul to do it again, louder this time.
Photo by Erwin Bosman on Unsplash
Correction, 2 July 2026: Several passages have been revised after a source review. (1) Three phrases in the Baker section used language from Gavin Baker’s X thread “Some thoughts ahead of Nvidia tonight” (25 February 2026) more closely than intended; they are now quoted directly. A stale temporal reference – “over the past two months” reflected Baker’s February data, not the date of publication – has been corrected. (2) Two phrases in the Ojha section used language from “Satya’s Convenient Truth” (Saanya Ojha, 16 June 2026) more closely than intended; they are now quoted directly. (3) One phrase in the Azhar section closely echoed the wording of “The state of the AI economy” (Exponential View, 25 June 2026); it has been rewritten.