The Second Act: Where the AI Correction Moved Underground

The equity sell-off got the attention. The debt layer is where the risk lives now. The first act of the AI correction played out in public, on screens, over days. Between June 1 and June 22, IT-services stocks collapsed first — Accenture's worst single-day decline on record after cutting revenue guidance to 3-4%, Cognizant down 10%, EPAM down 9% — and then the megacaps followed, with Alphabet losing $256B in market cap and Amazon, Meta, and Microsoft shedding a combined $248B ^[1]. SpaceX shares slid 10% for a third straight session while the company was actively selling debt through a notes offering ^[1]. That act was fast, visible, and is largely done. What replaced it is slower and harder to see. The second act is in the financing layer — the debt and private credit that funds the physical AI build — and it is still unfolding. Morgan Stanley reported that hyperscalers doubled their gross leverage ratio from 0.9x to 1.8x in just two quarters, while UBS tracked $800B in additional U.S. credit over the past year flowing to AI projects ^[2]. Nvidia raised $25B in its largest-ever bond sale on June 15 — seven tranches maturing out to 2056, with demand of roughly $85B against a $20B target ^[3]. The IMF and Bank of England warned that this debt-funded buildout could create systemic risks resembling the dot-com era if demand falters ^[3]. Meanwhile, cloud giants have pushed total future data center lease commitments past $850B — 20-year obligations — with Meta at $182.9B and Microsoft at $196.6B ^[4]. The spending trajectory runs from $650B in 2026 toward a projected $1T in 2027 and potentially $3-4T annually by 2030 ^[4][5]. This is where the second act starts to look like the first: the market is splitting in two. The safe tier — hyperscalers with contracted backlogs and accelerating revenue — can probably carry the debt. Google Cloud grew 63% to $20B quarterly with a $460B contracted backlog ^[6][7]. AWS grew 28% to $37.6B, with Amazon's CEO defending $200B in 2026 capex against real financials: Q1 sales up 17%, operating income up 30%, AWS AI run rate exceeding $15B ^[8]. Nvidia reported Q1 FY2027 revenue of $81.6B, up 85% year-over-year, with Q2 guidance near $91B and an $80B share repurchase authorization ^[9]. CoreWeave posted $2.1B in Q1 revenue, up 112%, against a $99.4B backlog including a $21B agreement with Meta ^[10]. Jefferies identifies a 12 GW global data center capacity deficit — only 8.9 GW operational against 21.1 GW in demand — arguing that supply, not demand, is the binding constraint ^[11]. These are contracted multi-year revenue commitments, not speculative TAM. The debt taken against them is collateralized by real cash-flow obligations. The speculative tier is a different story. OpenAI filed for IPO with a $20.9B operational loss in 2025 — up from $8.78B the year before — on $13.07B in revenue, with total spending of $34B ^[12]. The company is projected to spend $600B before seeing a return, potentially not until 2030 ^[13]. The IPO is a financing event to cover the gap between spending and internally generated cash, not a victory lap. Anthropic raised $65B at a $965B valuation despite expecting its first operating profit only in Q2 2026 ^[14]. The three mega-IPOs — SpaceX, OpenAI, Anthropic — could add roughly $4T in market value to U.S. markets, a massive supply of equity that is itself a financing mechanism for the physical build ^[13]. The financing chain feeding this speculative tier is already stalling. SoftBank sought a $10B margin loan against its OpenAI stake in May; lenders hesitated to price a private entity during the market correction, and the loan was cut 40% to $6B ^[15]. SoftBank faces a $40B bridge financing obligation due March 2027, and S&P lowered its credit outlook to negative in March ^[15]. The mechanism for converting speculative AI stakes into deployable capital is seizing up. The $2T global private credit market — where AI companies accounted for over a third of deals in 2025 — is showing the same stress. The Financial Stability Board warned on May 6 that opaque valuations, lack of standardized data, and deepening interconnections between asset managers, insurers, and banks create systemic vulnerabilities, and that default rates, though still moderate, are rising ^[16]. Apollo's MidCap Financial default rate climbed from 3.9% in December 2025 to 5.3% in Q1 2026, posting a $61M net loss ^[17]. Apollo's private BDC received redemption requests totaling 11% of assets; the firm capped payouts at 5%. BlackRock, Blackstone, and Blue Owl all reported NAV declines and markdowns ^[17]. HSBC took a $400M loss from a collapse in the sector ^[16]. Marc Rowan attributed the stress to concentration:

If you can't, as a first lien credit manager, meet 5% redemptions per quarter, I'll say it frankly: You're an idiot. — Marc Rowan

. Central banks reinforced the warning. The Bank of Canada's Gravelle noted that

Our overall view is that the Canadian financial system remains well positioned to weather shocks. Over the past year, it has faced repeated tests. While there have been some strains, those episodes did not lead to broad-based financial stress. — Toni Gravelle

, and Ireland's Makhlouf cited AI threats and high market valuations as specific risks ^[18]. Ray Dalio framed the problem precisely: not equity valuations, but liquidity.

All great technology changes produce bubbles. — Ray Dalio

. He said bubble indicators are nearing levels seen in 1929 and 2000 ^[19]. The bridge between the two tiers is Meta — the first safe-tier company showing signs that financing pressure is biting capex behavior. Meta halted share buybacks, borrowed $55B including a $27B bond sale for its Hyperion data center, and is considering an equity sale of tens of billions to fund AI infrastructure ^[20]. The stock fell 6% when the equity sale was reported and is down 21% from its August peak. What makes Meta the tell is not the borrowing — Alphabet chose equity over debt for up to $84.75B in what could be the largest equity capital markets deal in history, a different route to the same destination ^[21]. What makes Meta the tell is the explicit hedge. CFO Susan Li told investors:

If we end up not needing as much as we anticipate, we can choose to bring it online more slowly or reduce our spending in future years. — Susan Li

Meta is also considering selling excess compute capacity — a sign of potential oversupply ^[20]. When the safest of the safe-tier companies starts talking about slowing the build and offloading compute, the financing layer has reached capex decisions. The mechanism that could carry the credit event from the speculative tier into the physical infrastructure — and from equity prices into the real economy — is the demand-side crack. Enterprises are pulling back on AI consumption as token costs exceed returns. Uber exhausted its entire 2026 AI budget by April ^[22]. Microsoft canceled most internal Claude Code licenses after monthly costs per engineer reached $2,000 ^[22]. One client received a $500M monthly bill for unrestricted Anthropic Claude access ^[22]. Amazon shut down its internal AI leaderboard to curb unnecessary usage ^[22]. Meta's own CTO, Andrew Bosworth, told staff:

Nobody should be using AI tools just for the sake of using them. — Andrew Bosworth

That is a company spending $145B on AI infrastructure telling its own employees to stop using AI gratuitously ^[23][20]. Tom Essaye of Sevens Report Research drew the dot-com parallel directly:

While people connected to the internet, their connection wasn't nearly as profitable as quickly as everyone assumed. — Tom Essaye

. He identified a chain reaction: if enterprises cancel data center projects over poor ROI, it triggers massive order cancellations for Nvidia, Micron, Broadcom, and SanDisk ^[24]. Oracle shares declined roughly 25% from June 1 as investors questioned whether AI hardware spending would translate into timely corporate profits ^[24].

The propagation pathway runs through collateral. Hyperscalers have $850B+ in lease commitments backing the debt they are taking on ^[4][2]. If the demand-side crack erodes the reliability of those commitments, the credit event reaches the physical build even in the safe tier — but only if enterprises actually pull back at scale, which the 12 GW capacity deficit and contracted backlogs argue against ^[11][6].

The counter-evidence is real and should not be waved away. The picks-and-shovels layer is delivering accelerating revenue, not just burning cash: Astera Labs grew 93%, Credo 201%, HPE 40% with 2028 targets reached two years early, Micron nearly tripled year-over-year ^[10][25]. Palantir hit $1.63B, its fastest growth since IPO ^[10]. Alphabet's capex could rise to $300B in 2027 ^[26]. The supply constraint — engineering labor, cooling, power infrastructure, copper — may be the real binding limit, not demand ^[11]. Capital is still flowing to the picks-and-shovels suppliers even amid debt warnings: Brown Capital Management and Carillon Tower Advisers both added Credo positions in Q1 2026, citing 30-50% projected revenue growth ^[27]. Alibaba's Cloud Intelligence Group grew revenue 38% while its CEO declared that

We aim to maintain growth that is faster than the market average in order to gain larger market share and firmly cement our absolute market leadership position... those are the primary objectives, and margin is still secondary. — Eddie Wu

— prioritizing share over profit ^[28]. That is a company deliberately widening its spending-revenue gap to chase position. Physical projects are also stalling for reasons beyond financing. Sentinel Data Centers halted a $550M Maine project after the governor stripped tax exemptions ^[29]. A Microsoft-G42 $1B Kenya project stalled because the Kenyan government could not meet Microsoft's demand for guaranteed annual cloud-capacity purchases; President Ruto said the 1-GW facility would require

We would need to switch off half the country for the data center to be powered. — William Ruto

^[30]. Microsoft demanding pre-committed demand before building is itself a sign that hyperscalers are aware of demand risk — they want sovereign guarantees before they pour concrete. The open question is not whether the physical build continues. It does, funded by hyperscaler cash flow and contracted backlogs. The question is whether it continues at the scale the spending-revenue gap implies — or whether Meta's hedge becomes the template and the build decelerates from the financing side before the demand side proves it out. The FSB's June 10 warning about agentic AI as a financial stability risk, with 52% of the financial sector having adopted AI agents ^[31], is a separate dimension of systemic exposure, but it shares the same root: the financial system's deepening entanglement with the AI build. The first act was a price correction. The second act is a credit event working its way through a chain that runs from private equity stakes, through private credit funds, through bond markets, through data center lease commitments, to the physical infrastructure itself. The safe tier can probably hold. The speculative tier is already straining. The hinge is whether enterprise demand holds the line — and the earliest evidence, from Uber's exhausted budget to Meta's CTO telling his own people to ease off, suggests it may not.