TL;DR
Arthur Hayes argues in his Sept. 22 essay “Safety First” that public pauses framed as AI “safety” could actually signal weaker demand for very expensive training and inference compute. If compute demand collapses, Hayes warns, that could stress the financing that underpins data centers, chips and private-credit deals, potentially prompting fiscal or monetary backstops that expand dollar liquidity and, in his view, favor scarce assets like Bitcoin. The chain is coherent but speculative. Big numbers and active financing make the risk plausible. Regulators and markets have not yet shown systemic distress. Treat this as a low-probability, high-impact stress case worth monitoring.
Hayes’s argument, boiled down
Hayes writes:
“Safety First is by definition compute demand destruction.”
He lays out a clear causal chain:
- Frontier model pauses, publicly framed as safety, reflect weaker demand for extremely costly compute.
- Lower compute demand undermines cash flows backing data centers, GPUs and related infrastructure.
- Those cash-flow hits show up in credit markets, especially private credit, long-duration bonds and insurer exposures.
- To limit systemic damage, governments or quasi-public actors might intervene (Hayes calls this Washington becoming a “compute buyer of last resort”).
- Policy backstops or large fiscal purchases would expand dollar liquidity; Hayes expects that to tilt capital toward scarce stores of value such as Bitcoin.
How big is the AI financing story right now?
The scale matters because the larger and more levered the buildout, the easier it is for demand shifts to create credit stress.
- Apollo, in a Sept. 21 note, reported that AI-related issuance accounted for nearly 40% of longer-duration investment-grade corporate bond supply at that time and estimated the AI ecosystem could support more than $2 trillion of additional investment-grade debt. Apollo also projected roughly $5 trillion of AI infrastructure spending through 2030 and argued businesses and consumers would need to spend about $2 trillion per year on AI services to justify that infrastructure level.
- In August, Nvidia said a group including Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR was working on independent AI-compute financing platforms intended to mobilize more than $500 billion of third-party capital over time, a planned capacity figure, not capital already deployed.
- SoftBank has been reported to be marketing more than $11 billion of high-yield bonds to help finance its OpenAI investment.
- Apollo’s Sept. 21 note also cites consensus forecasts that operating cash flow at the five hyperscalers (Alphabet, Amazon, Meta, Microsoft and Oracle) could rise from roughly $600 billion to $2 trillion by 2030, underscoring the expectation that hyperscalers will be the primary demand engines for compute.
Evidence that supports Hayes, and where the chain weakens
There are solid reasons Hayes’s scenario is at least plausible:
- Massive planned capital commitments and new financing vehicles create concentrated exposures across the financing stack.
- Private-credit funds, long-dated bond issuance and structured off-take or offtake financing commonly back data-center and GPU investments. Those instruments are less liquid and less transparent than public equities.
- Sector watchers have flagged areas of attention: the Regulators and rating agencies have highlighted private credit as a supervisory focus and counted 139 private-equity-owned U.S. insurers by June 2025; Moody’s (as reported by the Wall Street Journal) estimated direct U.S. insurer exposure to data centers could be up to $20 billion.
But several critical links are fragile or speculative:
- Regulators and rating agencies are monitoring exposures, but as of the dates cited they have not reported an AI-driven insolvency or systemic failure. Attention is not the same as confirmed losses.
- Some large figures are estimates or planned capacity rather than committed, deployed capital. The >$500 billion tied to Nvidia-partner platforms is planned capacity; SoftBank’s >$11 billion bond marketing was in progress rather than closed capital.
- Nick Nemeth has estimated $1.54 trillion of affiliated reinsurance credits across the U.S. life and annuity industry, but that number is his estimate, not a regulator finding, and “affiliated reinsurance credits” can reflect internal accounting and statutory treatment rather than immediate cash shortfalls.
- On the policy side the mechanics are uncertain. The Federal Reserve raised its target range by 25 basis points to 3.75% – 4.00% on Sept. 16, and Fed staff emphasized that scheduled reserve-management purchases are not equivalent to broad monetary stimulus. That tightening stance runs counter to the immediate liquidity expansion Hayes’s thesis requires.
Demand signals are mixed
Public statements about “safety” and continued spending create ambiguity. OpenAI and Anthropic have publicly discussed pacing and safety. Reuters reported on Sept. 19 that Anthropic was nonetheless considering another model release. Industry forecasts and announced financing programs show ongoing capital momentum. Firms are debating pace even as large financing plans proceed.
What would a “compute buyer of last resort” look like, and how realistic is it?
Hayes’s image of a government stepping in is evocative. Mechanically, plausible interventions include:
- Direct offtake contracts or public purchases of cloud capacity (guaranteed future demand to keep projects cash-flow positive).
- Underwriting or guarantee programs for lenders and insurers that have financed data centers and GPUs.
- Targeted fiscal support for critical semiconductor or data-center infrastructure (grants, tax incentives, purchase commitments).
Analogues exist, governments have guaranteed debt, purchased assets or underwritten industries during crises (TARP and various guarantee programs being historical examples), but buying compute at scale is different politically and technically. No U.S. agency has announced a compute-buying program tied to preventing AI-finance stress, and modest programs wouldn’t produce the broad liquidity surge Hayes envisions. A very large, fast policy swing would be required to change macro flows significantly.
Does this path lead to higher Bitcoin?
Hayes argues that broader dollar liquidity and investors seeking scarce stores of value would lift Bitcoin. That channel is plausible historically, accommodative policy has often pushed capital into perceived stores of value, but it depends on a rare combination of events: a deep, durable drop in compute demand; concentrated losses across private-credit and insurer balance sheets; and a large, rapid policy response that meaningfully eases financial conditions. With the Fed in a tightening posture at the dates cited, an immediate liquidity surge is not the default scenario.
Practical steps for finance, treasury and risk teams
Treat Hayes’s thesis as a structured stress test. Concrete, actionable moves:
- Map exposures. Build a ledger of offtake contracts, cloud commitments, data-center leases, GPU purchase or lease obligations, vendor financing terms and private-credit counterparties. Capture contract terms: expiries, termination clauses, step-downs and change-of-control triggers.
- Stress-test cash flows. Model scenarios such as a 30-50% reduction in training hours, a 25-40% decline in inference margins, or 12-24 months of muted adoption. Re-run EBITDA, free cash flow and covenant headroom under those scenarios.
- Ask insurers and lenders specific questions. Request counterparty breakdowns of private-credit exposure, affiliated reinsurance relationships, collateral valuation practices and loss-recognition timelines.
- Establish a monitoring dashboard. Weekly or monthly KPIs to watch: GPU utilization rates, hyperscaler capex guidance, data-center vacancy rates, private-credit spread movements and covenant amendments, insurer loss ratios and regulatory statements from the Fed/Treasury/NAIC.
- Re-price optionality. Where possible, convert fixed commitments into more flexible arrangements (shorter lease terms, capacity on demand, renegotiable offtake pricing) and preserve liquidity to weather adverse scenarios.
How to judge probability and impact
Think of Hayes’s thesis as low-probability, high-impact. The necessary conditions are multiple and independent: a deep, persistent drop in compute demand; concentrated, realized losses across the financing stack; and a swift, large policy response that expands liquidity materially. Each link is possible, but none is guaranteed. The practical conclusion: don’t bet the company on the shock not happening. Model it and prepare contingency playbooks.
Key questions executives are asking
- What exactly is Hayes arguing?
He says that pauses framed as AI safety could mask weaker demand for expensive compute; that would stress financings backing data centers and chips; and that large fiscal or monetary backstops could follow, expanding liquidity in a way that may benefit scarce assets such as Bitcoin.
- Is AI financing really that large?
Apollo’s Sept. 21 research projects roughly $5 trillion of AI infrastructure spending through 2030 and suggests the ecosystem could support more than $2 trillion of additional investment-grade debt. Planned private financing platforms tied to AI compute target over $500 billion of capacity, but that is planned, not the same as deployed capital.
- Are insurers or private-credit funds already at risk?
Regulators and rating agencies are monitoring exposures: NAIC has highlighted private credit for supervision; Moody’s (reported by the Wall Street Journal) estimated up to $20 billion of direct U.S. insurer exposure to data centers; and researcher Nick Nemeth has offered an estimate of $1.54 trillion in affiliated reinsurance credits. Those signals warrant attention but do not, as of the dates cited, equal confirmed systemic insolvency.
- Would the government likely act as a “compute buyer of last resort”?
No U.S. authority has announced such a program. Governments can and do intervene in crises, but a compute-buying program would be politically complex and technically novel. The scale required to materially expand broad dollar liquidity would be large.
- If Hayes is right, will Bitcoin rise?
Expanded liquidity has historically supported scarce or uncorrelated assets, so it’s a plausible channel. But it hinges on sizeable, coordinated policy easing tied to the stress event, an outcome that is neither certain nor immediate, especially given the Fed’s tightening stance at the dates cited.
Hayes provides a useful mental model more than a forecast. The main takeaway for executives: the AI buildout is huge and financed in ways that create concentrated tail risks. Running the Hayes scenario as a formal stress case, mapping exposures and tightening contingency plans will be time well spent, whether or not Bitcoin ends up the unintended beneficiary.