AI Bubble Timeline 2024–2026: Data Center Debt, Chip Volatility and the $500 Billion Question

On 20 June 2024, Sequoia Capital partner David Cahn published “AI’s $600B Question,” a follow-up to his earlier “$200B Question.” His method: take Nvidia’s run-rate revenue forecast and roughly double it, since GPUs represent only about half of a data center’s total cost of ownership (the rest is power, buildings, cooling and backup generation). Even under generous assumptions — each of Google, Microsoft, Apple and Meta generating $10 billion a year in new AI revenue, plus $5 billion each from a wider set of companies — Cahn still found a revenue gap of roughly $500 billion against the implied ~$600 billion needed to justify the infrastructure being built. ⚪

This was an analytical framework built on stated assumptions, not a proven industry break-even requirement. Sequoia did not “prove” AI needed exactly $600 billion to succeed — it estimated that the pace of GPU and data-center spending implied roughly that much annual revenue under its model.

Has the $600B warning “come true”? That framing is too strong. AI revenue has grown substantially since 2024 — Nvidia’s own data-center revenue alone was $89 billion in a single quarter of 2026. But the infrastructure buildout has grown too, with 2026 hyperscaler capex consensus near $527 billion. The honest question in 2026 is not whether Cahn was “right,” but whether revenue growth has kept pace with a buildout that has also gotten much bigger. That remains genuinely debated. ⚪

On 5 August 2026, Goldman Sachs Research estimated that AI-related borrowers had raised roughly $489 billion in 2026 — nearly $500 billion — already exceeding the bank’s estimate for all of 2025. 🟡 Crucially, only about 40% of that issuance came directly from the hyperscalers (Microsoft, Alphabet, Amazon, Meta, Oracle); the remaining roughly $412 billion was raised by the wider AI ecosystem — software vendors, semiconductor companies and data-center financiers. Goldman expects direct hyperscaler debt issuance to rise from about $250 billion in 2026 to around $400 billion in 2027, when debt could fund roughly 35% of that year’s capex. 🟠

Do not confuse this with “hyperscalers borrowed $500 billion” or “$500 billion of bad debt exists.” Goldman’s figure describes issuance — new debt raised — across a broad ecosystem, not one company’s balance sheet, and issuance says nothing on its own about credit quality or repayment risk.

A separate, narrower measure from BNP Paribas data reported by Reuters found that Alphabet, Amazon, Meta, Microsoft and Oracle alone issued about $220 billion of bonds in 2026 through 10 August — specifically to finance AI infrastructure. 🟢 The gap between $220 billion and $489–500 billion is not a contradiction; it reflects different definitions (five hyperscaler issuers only, vs. the full AI-linked ecosystem including private credit and data-center developers).

Why cash-rich companies still borrow

A company can hold enormous cash reserves and still issue debt. Reasons include preserving liquidity, matching financing duration to a data center’s decades-long life, tax and capital-structure efficiency, funding several projects simultaneously, and spreading risk. Debt issuance alone does not indicate financial distress.

US corporate bond markets felt the effect directly. Bloomberg reported that August 2026 set a record for high-grade bond sales for that month — supply topped $145 billion, beating 2020’s prior August record of $136 billion — making it the third straight monthly record of 2026, with AI infrastructure financing a significant contributor. 🟢 Not every dollar of that issuance is “AI debt” — broad corporate bond markets carry many other borrowers — but AI-linked names were disproportionately responsible for the record pace.

September 1: AI Borrowing Meets a Global Bond Sell-Off

As this article publishes, global bond yields are rising at the same time technology companies continue issuing large amounts of debt. Higher yields raise financing costs, which makes data centers more expensive to build and raises the revenue bar AI services must clear to justify that cost. Government-debt concerns, inflation and monetary policy are broader forces behind the global bond move — AI corporate issuance is one contributor to elevated supply, not the sole cause of a worldwide sell-off. ⚪

Nvidia’s role is evolving beyond simply selling chips. Recent reporting (including Morgan Stanley research) has examined Nvidia’s contingent commitments, customer financing arrangements and infrastructure-linked investments — sometimes described informally as “balance sheet as a service.” That is an analyst/media description, not an official Nvidia term. The underlying concept: Nvidia benefits when customers build more GPU infrastructure, so it has a growing incentive to help finance, support or invest in that expansion. Estimates of Nvidia’s potential future financing exposure under particular scenarios are analyst estimates, not measured Nvidia debt. 🟡

A software startup’s costs are mostly code, people and a cloud bill — easy to scale down if demand disappoints. Hyperscale AI infrastructure is not.

Samsung Electronics and SK Hynix sit near the center of the AI supply chain, together representing more than half of the KOSPI’s value. In May 2026, South Korea’s exchange debuted single-stock leveraged ETFs tracking the two chipmakers. Retail investors piled in: by early July, SK Hynix, Samsung and leveraged products tracking them accounted for over 70% of all trading value on Korea’s roughly $4.3 trillion market. 🟢

Assets in leveraged ETFs tied to Korean chip stocks peaked near $52.5 billion on 22 June 2026. Within about a month, that figure fell to roughly $19 billion — a decline of over $33 billion. Leveraged ETFs tracking SK Hynix alone lost about $17 billion in market value from their peak; leveraged KOSPI 200 products lost about $10.5 billion; leveraged Samsung ETFs lost over $5 billion. One SK Hynix-linked leveraged fund fell roughly 70% from its peak. 🟢 Citi estimated total retail losses across these leveraged ETFs at approximately $38.7 billion — about 92% of holders were retail investors. 🟡

Do not confuse this with the KOSPI itself falling 40%. Samsung Electronics and SK Hynix shares fell as much as roughly 24% and 29% respectively from their peaks before rebounding — large, but not a “halving” of the underlying stocks. The far larger percentage losses (up to 70%) occurred in leveraged products, which amplify daily moves and are structurally different from owning the shares outright.

Why a 2× leveraged ETF made the swings worse

The volatility ran both ways. On 31 July 2026, the KOSPI surged a record 17.91% in a single session to close at 6,595.45 — the largest daily point and percentage gain in the index’s history — as Samsung and SK Hynix each jumped nearly 30%, their own biggest single-day gains on record, before the index fell nearly 5% again the following Monday. 🟢 One investor quoted by Reuters described the episode as “a leverage event, not an earnings event” — the sell-off was amplified by leverage, forced selling and concentrated retail positioning, not a collapse in underlying chip demand.

The paradox: stock volatility, strong business

Despite the market turmoil, South Korean semiconductor demand stayed strong. On 28 August 2026, SK Hynix’s CEO said the company saw no clear signs of a memory downturn and expects the current shortage to persist through 2030, while breaking ground on a $4 billion Indiana packaging facility targeting HBM4E production in 2029. 🟢 SK Hynix held about 58% of the global HBM market by revenue in Q1 2026, according to Counterpoint Research, with Samsung and Micron each holding roughly 21%. 🟢

On 1 September 2026, SoftBank-backed SB Energy Corp. — founded 2019, focused on pairing power generation with AI data centers — filed for a US IPO, reportedly seeking a valuation above $50 billion. 🟢 For H1 2026, the company reported revenue of $138.7 million (up 66.4% year over year) against a net loss of $3.21 billion (versus an $215.5 million loss on $83.3 million revenue a year earlier). SB Energy holds 8.8 gigawatts of data-center capacity contracted or under construction, and has granted OpenAI warrants valuing SB Energy at $5.5 billion to help secure it as a long-term tenant.

SB Energy’s project backlog and long-term commitments are not guaranteed revenue, and stock warrants granted to a tenant like OpenAI are not cash revenue — they are potential future equity value, contingent on outcomes. SB Energy is also significantly dependent on a small number of large customers, a real financial-concentration risk distinct from any judgment about the technology itself.

Nvidia’s Q2 FY2026 results (reported 26 August 2026) showed record revenue of $96.2 billion (+106% year over year) and record data-center revenue of $89.0 billion (+117% year over year), with Q3 guidance of $108 billion. For the first time, Nvidia issued a long-term forecast: CFO Colette Kress projected roughly 70% revenue growth in fiscal 2028, calling the figure “supply-constrained” — well above the ~44% analysts had expected. The outlook assumes no China data-center sales. 🟢 This is not evidence Nvidia demand has collapsed; the bubble question is whether Nvidia’s customers ultimately earn adequate returns on the infrastructure they are buying, not whether Nvidia currently has buyers.

OpenAI’s own monetization is also expanding: its ChatGPT advertising business crossed a $1 billion annualized revenue run rate by 31 August 2026, with the company targeting $2.5 billion in ad revenue this year and tracking toward over $40 billion in total annualized revenue. 🟢 Separately, internal documents reported by the Wall Street Journal suggest OpenAI’s own forecasts show operating losses reaching roughly $74 billion in 2028 — about three-quarters of that year’s projected revenue — driven mainly by compute costs. 🟡 This is a media report of an internal projection, not an audited financial result; it should be read as one data point in a company still scaling both revenue and spending simultaneously, not as a verdict on OpenAI’s viability.

A better historical analogy than “dot-com” may be the telecom fiber boom: the internet was real, demand eventually became enormous, but too much fiber was built too quickly and many telecom firms failed even as the underlying technology succeeded. Railways changed the world and many railway investors still lost money. Technology can win while some investors lose — those are not contradictory outcomes, and AI could plausibly follow either the “technology succeeds, some investors don’t” pattern or a more benign, well-financed path. Which one plays out is genuinely undetermined today. ⚪

1. AI capex growth slows sharply — a real deceleration from the ~$527B 2026 pace, not just a smaller increase.
2. GPU or HBM orders fall — the clearest sign hyperscalers see less need for capacity.
3. Data-center vacancies rise — built capacity sitting unused would be a direct overbuild signal.
4. AI revenue growth disappoints against the capex already committed.
5. Credit spreads widen or financing dries up for AI-linked borrowers, especially outside the hyperscalers.
6. Model efficiency improves dramatically — cheaper inference could strand expensive infrastructure, though it can also expand demand (the Jevons paradox), so its net effect is ambiguous.
7. Power and grid constraints bind harder, preventing built data centers from actually operating at capacity.
8. Interest rates rise further, raising financing costs and lowering the present value of distant future AI cash flows.

There is no single, universally accepted number for “AI debt” — estimates differ because they measure different things. Goldman Sachs’ ~$489 billion figure counts AI-related borrowers across the whole ecosystem, including data-center developers and non-hyperscaler tech firms. Reuters/BNP Paribas’ ~$220 billion figure counts only bonds from five named hyperscalers through 10 August. Neither figure is “wrong” — always check what a given AI-debt number actually includes before comparing it to another one.