Record AI Chip Capex Bets Collide With Historic Semiconductor Market Selloff

The story of the week is not Broadcom losing $444B of market cap in two sessions. It is that Alphabet raised $84.75B in equity the same month, the largest US corporate equity raise on record, to fund the spending public markets just punished Broadcom for forecasting. Private capital is pricing AI infrastructure as a generational utility build. Public capital is pricing it as a peaking cycle. Both cannot be right, and the resolution will determine who controls the asset base when the music slows.

The capex-to-revenue arithmetic is now worse than 2001

Allianz Research puts the current divergence between AI capex growth and AI revenue growth at 46%. At the peak of the 2001 telecom build it was 32% [1]. Sequoia's David Cahn calculated the gap between hyperscaler infrastructure spend and AI sector revenues at roughly $600B annually in 2025, and the arithmetic has worsened since: combined 2026 capex from Amazon, Microsoft, Alphabet, Meta and Oracle is now guided at $700B–$900B, a 36% step up, with Amazon alone at $200B [2]. Evercore and Bank of America see hyperscaler capex above $1T in 2027 [3].

What matters is not the gap itself; it is that the buyers know about the gap and are accelerating anyway. Capex intensity at the big five is now running between 45% and 57% of revenue, ratios you see at regulated utilities, not at software businesses [4]. In 2025 the same group raised $108B of new debt, and CreditSights has flagged that aggregate capex now exceeds cash flow after buybacks and dividends [5]. Morgan Stanley and JPMorgan estimate the sector may need to issue $1.5T of new debt to fund what is on the drawing board [6]. Hyperscaler CFOs are not stupid. They have decided that the cost of underbuilding is worse than the cost of overbuilding, because no single player will cede position, and the capex cycle is therefore self-reinforcing irrespective of near-term ROI [7]. That is a game-theoretic argument, not a financial one. It works until a balance sheet breaks.

Where the marginal dollar is actually coming from

The most important development of the past month is not the equity raise; it is where the marginal dollar is now coming from. Meta and Blue Owl closed roughly $27B for a single Louisiana data center campus, the largest private credit transaction ever recorded [8]. Oracle sold $18B of bonds in a single day to fund its Stargate commitments, part of a programme priced near $500B with SoftBank and OpenAI [9]. Nvidia has put roughly $110B into direct financing or chip-backed lending to its own customers, keeping the exposure off hyperscaler books while still booking the revenue [10]. OpenAI closed $122B in March at an $852B valuation; Anthropic raised $65B in May at $965B, having raised $30B three months earlier [11].

Read these together and a pattern emerges that the consensus capex narrative misses. The visible hyperscaler balance sheets are being preserved with some care. The real debt is being pushed into SPVs, private credit funds, vendor financing books and pre-IPO equity. No aggregate figure exists for the total encumbered debt load behind the AI build, and that absence is itself the point. When CFOs at Microsoft or Alphabet are asked about capital discipline, they can truthfully say their balance sheets remain investment-grade. The debt is real; it is just somewhere else. This mirrors the financial engineering that preceded the 2001 telecom bust, where vendor financing from Lucent and Nortel masked deterioration in customer credit until both ends collapsed simultaneously. Nortel's optical networking revenues grew 133% to $9.2B in 2000. The company filed for bankruptcy in 2009 [12].

Public markets doing the job private markets refuse to

Broadcom's $444B two-day loss, and the Nasdaq's worst session in 14 months that accompanied it, were not about Broadcom's guidance [13]. They were about public investors discovering they have no instrument to price the capex-to-revenue gap except the chip stocks themselves. Meta fell 9.25% in a single session after raising capex guidance, with Mark Zuckerberg justifying the spend by invoking "personal superintelligence to billions of people" [14]. That is the language of a founder who has stopped trying to persuade public shareholders and is instead daring them to sell.

The Asian repricing is more violent and more revealing. South Korean equities fell 12% and Taiwanese equities fell 6% in three sessions from record highs [15]. TSMC is 41.5% of the TAIEX. Samsung and SK Hynix together are 55% of the KOSPI. The three together account for almost a third of MSCI's Asia Pacific ex-Japan Index, exceeding the Baidu/Alibaba/Tencent concentration in MSCI China at its October 2020 peak of 37.14% [16]. May saw a record $27.9B outflow from Korean equities even as Nomura tracked an unprecedented $20.4B YTD inflow from US-domiciled funds into Korea and Taiwan [17]. Forced selling and forced buying are happening simultaneously in the same names. That is what a liquidity event looks like before it has a name.

The MIT Project NANDA study, if its 95% figure for enterprise GenAI pilots producing zero measurable P&L impact is even directionally right, suggests the demand thesis underwriting the build has not yet shown up in customer income statements [18]. TS Lombard's Dario Perkins frames most current AI revenue as "capex recycling": hyperscalers buying from each other, model labs buying compute from hyperscalers who took equity in the labs, Nvidia financing customers to buy Nvidia chips [19]. The revenue is real. Whether it represents end-user monetisation is a separate question, and the one public markets are starting to ask.

The bull case, stated honestly

The strongest version of the bull case is structural, not a matter of earnings multiples. TSMC's leading-edge capacity is sold out. C.C. Wei said on June 4 that chip supply will not meet AI demand for "a very long time," and TSMC is spending $165B on Arizona alone, with land for roughly a decade of expansion [20]. Nvidia's data centre revenue ran at $62.3B in Q4, up 75% year on year [21]. Anthropic's annualised revenue went from roughly $10B at end-2025 to $47B by May 2026 [22]. None of this looks like 2001 telecom, when dark fibre sat unlit for a decade.

The bull case is correct that current infrastructure is being used. It is wrong that utilisation guarantees the returns. The 2001 analogy that matters is not capacity glut; it is that Cisco, Nortel and Lucent built the right infrastructure for an internet economy that did emerge, and still destroyed equity value, because the value migrated to the application layer and the infrastructure itself commoditised. Per-token inference costs fell 280-fold between 2022 and 2024 [23]. OpenAI spent $2.22 for every dollar of revenue in Q1 and projects no profit until around 2030 [24]. The infrastructure can be fully utilised and the returns still terrible if pricing power deflates faster than the asset base depreciates. That is the scenario public markets are beginning to discount and private markets are still refusing to.

What to watch

1. Hyperscaler debt issuance through Q3 2026 versus the $1.5T Morgan Stanley/JPMorgan multi-year estimate. If gross issuance from Amazon, Microsoft, Alphabet, Meta and Oracle exceeds $250B in any single quarter, the off-balance-sheet financing channels are saturating and the debt is coming back onto the visible balance sheet. If issuance stays below $100B per quarter, the private credit and SPV channels are absorbing the load and the divergence can persist into 2027.

2. Q2 2026 enterprise AI revenue disclosure from Microsoft and Google Cloud, specifically the share attributable to other AI firms versus non-tech enterprises. If more than 40% of disclosed AI revenue traces to other model labs, hyperscalers or Nvidia-financed customers, Perkins's capex-recycling thesis is confirmed and the revenue quality argument collapses.

3. The TSMC, Samsung and SK Hynix concentration ratio in MSCI's Asia Pacific ex-Japan Index by end-Q3. If it crosses the 37.14% Baidu/Alibaba/Tencent peak in MSCI China, accompanied by another month of $20B-plus forced outflows from Korean equities, the index-level forced-selling feedback loop becomes the operative risk, independent of fundamentals at any individual chipmaker.