How Arthur Hayes Thinks an AI Slowdown Could Become a Bitcoin Liquidity Catalyst

Arthur Hayes’s “Safety First” thesis is not a forecast that Bitcoin will rise simply because AI spending slows. It is a second-order macro scenario: a disappointing AI-demand cycle could impair infrastructure debt, spread stress through credit markets and eventually trigger a liquidity response from the U.S. government or financial system. Hayes argues that such liquidity would be supportive of Bitcoin. 142

The important distinction is between a real financing vulnerability and the much more conjectural path from that vulnerability to a rescue and then to higher BTC prices.

The thesis in one chain

Hayes’s argument runs through six steps:

1. AI labs slow frontier-model development.
2. That slowdown reveals weaker demand for expensive AI products and compute capacity.
3. Lower utilization weakens the expected cash flows behind data centers, chips and related infrastructure.
4. Debt tied to that buildout is repriced, causing downgrades, wider credit spreads and less willingness to fund new projects.
5. Private-credit and insurance-linked investors become a possible transmission channel for losses.
6. Washington responds by supporting compute demand or stabilizing the credit and insurance system, expanding liquidity in a way Hayes believes benefits Bitcoin. 142

Each step depends on the one before it. A slowdown in AI investment, therefore, would not automatically mean a Bitcoin rally; it could initially produce a conventional risk-off move instead.

Why AI infrastructure financing is the pressure point

The scale of projected AI investment makes revenue and utilization especially important. Apollo estimates that roughly $5 trillion could be spent on AI infrastructure through 2030 and argues that businesses and consumers would need to spend about $2 trillion a year on AI services to justify that investment. Apollo identifies whether AI can earn an adequate return on this spending as a defining question for debt and equity markets.

That spending includes data centers, power infrastructure, chips and networking. Apollo has also said the pool of capital required is large enough that investment-grade financing, across public and private markets, will be central to funding the buildout.

Hayes’s concern is that the infrastructure is being financed ahead of fully proven end-market cash flows. In his account, AI-lab demand for compute supports more than $1 trillion of investment-grade debt and hundreds of billions of dollars in lower-rated debt and loans. 142

The figure is part of Hayes’s framing, not an official tally of debt guaranteed to default if AI demand softens. But the underlying economic question is straightforward: if data-center capacity is built for usage that does not materialize at expected prices, projected revenues and asset values can fall at the same time.

Why “Safety First” matters in Hayes’s interpretation

Hayes interprets calls by AI companies to slow frontier development on safety grounds as an economic signal: in his view, demand for AI products at current prices may be insufficient to support continued compute-intensive expansion. 14

That is Hayes’s interpretation, not established evidence of the companies’ motives. A safety-based decision, if one occurs, does not by itself prove weak commercial demand. Still, the thesis asks investors to focus on the financial consequence if frontier-training and inference demand grow more slowly than infrastructure supply.

In that case, a reduction in compute demand could leave data-center and chip capacity underutilized. The concern would then shift from technological capability to debt-service capacity.

How an AI revenue shortfall could reach credit markets

A lower-than-expected AI revenue outlook could pressure the debt financing the buildout through several familiar channels:

• Asset values could decline as expected data-center utilization and cash flows are revised lower.
• Credit ratings could come under pressure if issuers appear less able to meet interest and principal obligations.
• Credit spreads could widen, raising refinancing costs and making new projects harder to fund.
• Capital spending could slow if large technology companies reassess returns on incremental infrastructure investment.

Hayes argues this process could affect investment-grade bonds as well as lower-quality loans, creating a broader financing retrenchment rather than an isolated technology-equity correction. 214

Apollo’s analysis supports the premise that the sector’s financing requirements are unusually large; it does not establish that a debt crisis is underway. Its central question is whether the AI economy can generate enough revenue to support the spending now being planned.

The private-credit and insurance channel: a real area of scrutiny, not a proven crisis

The most aggressive part of Hayes’s scenario is that AI-related credit losses could travel through private-credit funds and affiliated insurers or reinsurers. He argues that impaired AI infrastructure debt could expose vulnerabilities in those structures and force public intervention. 142

Regulators are paying closer attention to insurers’ growing links with private credit. Reuters reported that insurers and other large institutions were preparing to increase private-credit exposure even as investors worried about illiquidity and regulators examined the connection between private credit and insurance balance sheets.

That scrutiny matters because private assets can be difficult to value and sell quickly in a stressed market. But it is not evidence that AI-infrastructure losses are currently causing insurer failures. Reporting on Hayes’s claim likewise noted that U.S. insurance insolvencies tied to AI-related debt had not been established. 1112

In other words, the insurance leg of the thesis is best understood as a stress scenario: plausible enough to investigate, but not an observed systemic event.

Hayes’s two possible government responses

Hayes imagines two ways policymakers could respond if the stress became severe:

1. Government as a buyer of last resort for compute

The first route would be direct public support for computing capacity. In this version, the government would buy or otherwise underwrite compute demand, helping data-center operators maintain revenue and helping their debt remain serviceable. 214

2. Support for insurers or the wider credit system

The second route would be a financial-stability response if losses in private credit and insurance threatened a wider credit event. Hayes argues that a backstop for insurers, reinsurers or other creditors would also expand dollar liquidity. 214

His Bitcoin conclusion rests on a macro view: incremental liquidity and currency debasement concerns can increase investor demand for scarce, risk-sensitive assets, including BTC. That relationship is a market thesis, not a mechanical rule. Liquidity injections can coincide with stronger Bitcoin demand, but price outcomes also depend on risk appetite, rates, leverage and regulation.

Why the near-term timing case is weak

Several facts complicate a claim that an AI-credit rescue is already underway.

First, monetary policy remains restrictive. The Federal Reserve lifted its target range to 3.75%–4.00% in September, which is difficult to reconcile with an active rescue-driven easing cycle. 19

Second, financing for AI has continued. SoftBank was reported to be discussing a potential $10 billion to $20 billion bond offering to refinance borrowing associated with its OpenAI investment; discussions were ongoing and terms could change. Continued access to financing does not eliminate refinancing risk, but it does not fit a narrative in which capital has abruptly stopped flowing to AI either.

Third, the decisive test is longer term: whether AI demand, pricing and utilization are sufficient to support the infrastructure being financed today. Apollo’s $2 trillion annual AI-services estimate highlights the size of that revenue hurdle.

Bitcoin near $85,000 is not evidence of the Hayes scenario

Bitcoin’s move toward $85,000 is better explained by nearer-term market forces than by a hypothetical AI-credit backstop that has not happened.

Market reporting attributed the advance to factors including spot ETF demand and short liquidations. Bitfinex said a 5.9% rally on September 18 was supported by $433 million in U.S. spot ETF inflows, before BTC extended its move toward $85,000. Other reporting also described Bitcoin recovering despite the Fed’s rate increase and U.S. crypto-policy uncertainty. 1921

Those catalysts can move Bitcoin well before any structural AI-financing story is tested. They should not be confused with evidence that the U.S. government is preparing to purchase compute capacity or rescue insurers over AI debt.

What would validate—or weaken—the thesis

For the thesis to become more than a thought experiment, investors would need to see several developments together:

• sustained evidence that AI compute demand or pricing is falling short of buildout assumptions;
• lower data-center utilization or weaker cash-flow guidance;
• widening spreads, downgrades or refinancing difficulty among AI-infrastructure borrowers;
• visible losses spreading into private-credit portfolios or insurers; and
• actual policy measures aimed at supporting compute demand or stabilizing exposed financial institutions.

The opposite evidence would weaken the case: rising usage, durable AI-service revenue, successful refinancings and stable credit performance would indicate that the buildout is being absorbed rather than becoming a debt problem.

Bottom line

Hayes identifies a credible vulnerability: a massive infrastructure buildout financed before its eventual revenue base is fully known. Apollo’s estimate of roughly $5 trillion in AI infrastructure spending through 2030 makes the utilization and monetization question consequential for both debt and equity markets.

But the full “AI slowdown to Bitcoin boom” sequence contains several unproven links. A demand disappointment could first bring lower technology capital expenditure, tighter credit and risk-off trading—not immediate money creation or a Bitcoin rally. The thesis is most useful as a framework for watching AI revenue, infrastructure utilization and credit conditions, rather than as an explanation for Bitcoin’s latest price move.