Multiplier Partners โ Executive Briefing
What has bolstered AI infrastructure valuations in the public and private markets โ and what makes those valuations vulnerable to repricing
Prepared by Matt Teeple, Managing Partner & Founder, Multiplier Partners ย โขย July 2026
01 ยท Executive Summary
AI infrastructure is the largest private capital-formation event in history, and its valuations reflect that: Nvidia alone reached a nearly $5.3 trillion market capitalization in April 2026, OpenAI was valued at $852 billion in its March 2026 round, and Anthropic surpassed it in May 2026 at $965 billion. Hyperscaler capex is on pace for roughly $725 billion in 2026, up from $150 billion just three years earlier. Every one of those numbers is real, and each is defensible on its own terms โ enterprise adoption is genuine, token volumes are climbing, and the compute build is filling in years-long order books.
The purpose of this briefing is not to argue that AI is fake or that the technology will not matter. It is to separate what is genuinely driving these valuations from what is manufacturing the appearance of demand โ and to name, specifically, the mechanisms that make today's marks vulnerable to a fast, disorderly repricing rather than a slow, orderly one. The short version: the equity story (adoption, revenue growth, market share) is mostly sound. The credit story quietly built underneath it โ take-or-pay contracts, GPU-collateralized debt, vendor financing, and a backlog concentrated in two cash-burning borrowers โ is not, and it is the credit story that determines how this unwinds if growth merely decelerates rather than reverses.
02 ยท What Has Bolstered the Valuations
Public markets: the compute supercycle
Nvidia captures the clearest version of the story. Data-center revenue of $75.2 billion in a single quarter (Q1 FY2027, reported May 2026), an estimated 85โ90% share of the AI training chip market, and fiscal 2026 revenue of $213 billion (+63% year over year) have supported a market cap that briefly reached the highest valuation of any publicly traded company in history. The four major U.S. hyperscalers โ Alphabet, Amazon, Meta, and Microsoft โ have guided to a combined $700 billion-plus of 2026 capex, and each has been rewarded by the market for doing so: capital spending has been read as confidence, and confidence has been read as a buy signal.
Private markets: a valuation race with no historical precedent
Private AI valuations have compounded even faster than public ones. OpenAI's valuation moved from roughly $86 billion in early 2024 to $852 billion by March 2026 across a sequence of rounds; Anthropic closed a $65 billion Series H in May 2026 at a $965 billion valuation โ nearly tripling from $380 billion just three months earlier โ to become the most valuable private AI company in the world. Both have filed for IPOs. Venture capital has concentrated around this story to a degree without precedent: U.S. venture deployed $412.7 billion in H1 2026, with AI absorbing 86% of it ($355.9 billion), and OpenAI and Anthropic alone captured 43% of all global startup funding in the period. Fund commitments earmarked for AI reached roughly $62.4 billion through the first five months of 2026 โ among the largest first-half totals on record โ led by Thrive Capital, Andreessen Horowitz, and Founders Fund.
The narrative reinforcement engine
Beneath both markets sits a self-referential support structure. Nvidia has invested in OpenAI and in CoreWeave; CoreWeave uses that capital to buy Nvidia chips and sells capacity back to Microsoft and Oracle; Oracle's Stargate buildout for OpenAI is itself Nvidia-supplied; Microsoft, Nvidia, SoftBank, and Amazon are the repeat investors behind OpenAI's own valuation marks. Across the four largest cloud platforms, contracted backlog (remaining performance obligations, or RPO) has swollen to roughly $2.1 trillion โ read by Wall Street as locked-in demand and cited as the primary justification for continued capex. Each round of investment, each new contract, and each capex guidance increase currently reinforces the others, which is precisely what has made the valuations look self-evidently justified.
03 ยท The Structural Fragility Beneath the Story
The same features that have supported valuations on the way up are the ones that make them vulnerable on the way down. Four are worth naming specifically.
The capex-to-depreciation gap
The four U.S. hyperscalers purchased $433.9 billion of property and equipment in the trailing four quarters through March 2026, against roughly $149 billion of reported depreciation over the same span. Depreciation schedules assume 5โ6 years of useful life for servers and GPUs; several analysts, including Michael Burry, have argued the real economic life is closer to 2โ3 years given the pace of chip generational turnover โ which would understate true depreciation by an estimated $176 billion across 2026โ2028 and flatter reported earnings in the meantime. Separately, Moody's has flagged roughly $662 billion in hyperscaler data-center lease commitments that are signed but not yet commenced โ an off-balance-sheet obligation larger than the combined on-balance-sheet debt of the same companies.
Capex is crossing 100% of operating cash flow
Aggregate hyperscaler capex as a share of operating cash flow has climbed from roughly 30% in 2022 to about 60% in 2025, and consensus spending puts it at approximately 100% in 2026. Above that line, every marginal dollar of new capacity is funded from debt or equity rather than internal cash โ meaning the "self-funded, fortress balance sheet" framing that has anchored hyperscaler valuations is no longer accurate on a forward basis. Morgan Stanley estimates investment-grade hyperscaler leverage has roughly doubled over the past year to about 1.8 turns of gross debt, now higher than the entire energy sector.
Concentrated, low-quality backlog
Of the $2.1 trillion in aggregate RPO across the four large platforms, roughly half โ about $1.05 trillion โ is owed by two companies, OpenAI and Anthropic, neither of which is yet profitable. Oracle's backlog is about 54% OpenAI-derived (roughly $300 billion from OpenAI alone); Amazon's is about 51% concentrated in the two labs; Microsoft's roughly 49%; Google's about 43%. An RPO is a forward payment promise, not a liquid asset โ its value is only as good as the counterparty behind it, and a large share of this backlog is owed by pre-profit borrowers whose ability to pay depends on continuing to raise capital at ever-higher valuations.
Leveraged neoclouds and vendor-financed demand
CoreWeave, the largest independent GPU cloud provider, carries roughly $24.9 billion in total debt against $3.34 billion of shareholders' equity, faces a $4.2 billion principal repayment in 2026, and paid $536 million in interest in a single quarter โ 25.8% of revenue and 46.3% of adjusted EBITDA. Nvidia is contractually obligated to buy CoreWeave's unsold capacity through April 2032 under a facility worth at least $6.3 billion, making Nvidia the buyer of last resort if the business it helped fund cannot fill its own racks. Nvidia has since extended similar backstop and revenue-share arrangements to smaller neoclouds (reported cases include Firmus and Sharon AI), effectively financing the purchase of its own chips โ and in May 2026, CoreWeave closed a $3.1 billion syndicated GPU-backed facility (DDTL 5.0) exposed to OpenAI and Cohere, the first such paper structured to trade in the secondary credit market.
What the risk managers are saying
Larry Fink (BlackRock) has said the AI investment race will produce bankruptcies. Norway's $2.1 trillion sovereign wealth fund warned in March 2026 that an AI bubble could erase 35% of its value. A Bank for International Settlements study concluded the circular financing in this cycle has "a closed valuation loop with no external reference point" โ a structural feature with no precedent in prior technology booms.
04 ยท Two Historical Parallels, One Mechanism
Most public debate frames this as "is AI a bubble like dot-com?" โ a question that produces dot-com answers: some leaders survive, some laggards wash out, multiples compress and recover. That framing understates the risk, because the financing architecture underneath this cycle more closely resembles two other episodes.
1. Telecom vendor financing, 1998โ2001
In the late-1990s telecom buildout, equipment vendors โ Lucent, Nortel, Cisco โ financed their own customers' purchases of network gear to keep order books full. When end-user demand failed to materialize at the assumed pace, both the vendors and the carriers they financed collapsed together. The parallel today is direct: Nvidia is now financially backstopping smaller neocloud providers (reported arrangements with firms including Firmus and Sharon AI) in exchange for a share of their revenue, and is contractually committed to repurchase unsold capacity from CoreWeave. SoftBank has launched a new U.S. neocloud venture; Together AI raised $800 million; hundreds of neocloud entrants are now chasing the same pool of demand. Vendor financing does not create demand โ it defers the moment demand is tested, and concentrates the eventual loss in the vendor that extended the credit.
2. The 2008 credit cycle, re-read as a loan book
A widely circulated July 2026 analysis ("The Second Derivative," Groundbreaker) makes a more specific argument worth summarizing here because it reframes the mechanics precisely: a take-or-pay compute contract is, in economic substance, a loan. The data center is the collateral, the lab's contracted payments are the debt service, and the RPO booked by the hyperscaler is the lease receivable. Capex reported as confidence is better read as loan origination volume, and its credit quality is exactly the credit quality of the tenant behind it. The analysis's central claim is that markets are watching the level of AI spending and its growth rate, but not its second derivative โ whether growth is accelerating or decelerating. By its estimate, aggregate hyperscaler capex growth ran +51%, +81%, +77%, and +52% across 2023โ2026, meaning the acceleration itself peaked in 2025 and has already turned negative โ while headline capex levels continue to set records. Historically, credit-financed buildouts (real estate is the clean analogue) break not when demand collapses but when its growth decelerates against fixed supply already under construction.
The credit divergence that matters most: OpenAI vs. Anthropic
The same analysis draws a sharp distinction between the two largest borrowers behind the $2.1 trillion backlog โ a distinction obscured when both are described simply as "frontier labs."
| OpenAI | Anthropic | |
|---|---|---|
| Valuation (2026) | $852B (Mar. 2026 round) | $965B (May 2026 Series H) |
| Revenue mix | ~60% consumer | ~80% enterprise, contracted |
| Reported burn | ~57% of revenue through 2027 | Converging to ~9% of revenue by 2027 |
| Financing backstop | None โ Microsoft ended revenue share, exclusivity, and its AGI license clause on Apr. 27, 2026; SoftBank separately seeking a personal-guarantee margin loan against its stake | Google guarantees lease shortfalls; Broadcom guarantees residual chip value on $30B of a $35โ36B Apollo/Blackstone TPU financing facility |
| Effective credit position | Unguaranteed โ dependent on continuously raising capital at rising valuations | Synthetically investment-grade โ backed by two investment-grade guarantors |
This does not make Anthropic risk-free โ it is still a pre-profit company scaling into an uncertain demand curve. But it is a materially different credit than OpenAI, whose implicit Microsoft backstop was formally removed in April 2026 and whose next funding step-up (from $852 billion toward a reported IPO target above $1 trillion) is, by the same analysis, the smallest in its round-over-round sequence โ which may explain the delay in its IPO filing.
05 ยท Why This Matters for Security and Agentic AI Leaders
None of this is an argument to slow enterprise AI adoption โ the underlying technology and the productivity case for agentic workloads are real, and the risk described here sits in the financing layer, not the capability layer. But it is a direct argument for how enterprises architect their dependency on that financing layer. An enterprise that wires agentic workflows tightly to a single model provider, a single cloud, or a single orchestration path inherits the correlated risk sitting underneath that provider's balance sheet โ whether or not the enterprise ever sees that balance sheet.
If the repricing described in Section 4 plays out even partially โ contract renegotiation, capacity pullbacks, neocloud consolidation, or a disorderly OpenAI refinancing โ the disruption will not stay confined to Wall Street. It will surface as changed pricing, changed terms, changed model availability, and changed continuity guarantees for every enterprise workload built on top of the affected counterparty. This is precisely the scenario that identity- and orchestration-first architecture is designed to absorb: an enterprise governed by a portable identity and policy layer across agents, models, and compute providers โ rather than one hard-coded to a single vendor's roadmap โ can re-route, renegotiate, or substitute without rebuilding its security and governance posture from scratch.
This is the same discipline this firm has advised early-stage infrastructure and identity companies to build toward for three decades: durable architecture that does not depend on the continuity of any single sponsor. For security leaders evaluating agentic AI pilots today, the practical takeaway is to treat model- and compute-provider diversification, and identity continuity across agents, as a risk-management requirement โ not merely a governance nicety โ given how concentrated the counterparty risk has become one layer beneath the infrastructure they are building on.
06 ยท Bottom Line
The bull case and the bear case are not actually arguing about the same variable. Bulls point to the level of demand and its growth rate; the more serious bear case points to the second derivative โ whether that growth is still accelerating โ and to the fact that a large share of the financing underneath this buildout requires acceleration, not just growth, to avoid a credit-cycle unwind. AI does not have to fail for that unwind to happen. Growth only has to decelerate below what the contracts, the leverage, and the next funding round already assume. The tells to watch: hyperscaler capex growth rates quarter over quarter, whether OpenAI's delayed IPO clears anywhere near its target valuation, and whether any hyperscaler is rewarded by the market โ rather than punished โ for slowing its own spending.