“Safety First” — Arthur Hayes’ New Essay. A Brief Recap

In the Name of China

The state can justify almost any action on national security grounds. In the past, that argument was the global war on terror after the September 11 attacks. Now the main threat is China, which is rapidly developing its own technologies and offering AI products at prices far below those of US companies.

Competition with China is no longer just a normal market fight. To preserve its technological edge, the US effectively has to combine a capitalist system with elements of state funding.

Representatives of the AI industry have managed to convince the Trump administration of the need for such a policy, despite two signals:

• the private market has not yet confirmed the profitability of large-scale AI investment
• some voters oppose further data center construction and the use of their data.

If China can offer a cheaper product, the US responds with even higher spending to build AGI. The logic is simple — because AGI could potentially underpin military technology, the economy, robotics, and other sectors, the US wants to retain control over the most advanced models and set the terms of access to them.

Chinese Kimi K3 Brought Moonshot AI to $1B in Revenue, but Anthropic Claimed Claude Was Used Covertly

13.09.2026

Read now

This approach seems contradictory to me. The American economic model is largely built on the idea that the free market determines the price of goods and the volume of their production. However, in the case of AI, the authorities are effectively ready to ignore these signals for the sake of national security and keep funding new models.

Buyer of last resort

The “security first” principle implies a decline in demand for compute from the three largest US AI labs. To offset the reduction in purchases, the government could sign guaranteed contracts similar to those used in defense and some commodities industries.

Guaranteed procurement would ensure profits for AI companies, and the government could use the acquired compute for AGI research and development. Over time, the state could theoretically transfer the models it creates to private labs, which would sell access to them to US and allied users.

This approach solves two problems at once:

• the state supports demand for compute
• and, at the same time, gains direct control over the most advanced models.

The problem with private AI labs from a national security perspective is that they are global companies. Their drive to sell services to any customer can conflict with the authorities’ interests.

If Chinese labs are indeed accelerating the development of their models through distillation of American solutions, direct state control would allow the US to preserve its technological edge for several months or years.

However, I doubt such restrictions will be effective. A similar strategy has already been applied to advanced semiconductor technologies, but it has not been able to completely stop the development of China’s industry. In the internet era, stopping the spread of information is becoming increasingly difficult.

How to fund the AI race?

Even if controls do not stop China, the very attempt to preserve an edge will be costly, and those costs will have to be financed somehow. Since the private market no longer provides the same level of demand, further development and infrastructure buildout will require new borrowing.

Officials can frame such borrowing as an investment in a future economic upswing. U.S. Treasury Secretary Scott Bessent and Fed Chair Kevin Warsh view AI as a source of productivity gains. In their logic, if the technology accelerates the economy enough, GDP can grow faster than the debt burden.

In June 2026, U.S. nominal GDP rose 6.6% year over year, while the effective federal funds rate was around 3.6%. As long as the economy’s return exceeds the cost of short-term government funding, officials can view additional debt as economically justified.

But the math only works under one condition — if the budget deficit stays at 3% of GDP or lower. Only then can the debt-to-GDP ratio theoretically decline even with new borrowing.

I attribute a significant share of the current economic upswing to data center construction and AI labs’ demand for compute. Under these conditions, it is in the government’s interest to keep data centers running at high utilization, as long as the cost of short-term financing remains below nominal economic growth.

However, this whole logic only holds with sufficiently low rates — the U.S. is running a budget deficit and has to borrow to cover new spending, so the Treasury’s course runs into Fed policy — if the central bank tightens conditions, servicing additional debt becomes more expensive.

The Fed Raised Interest Rates for the First Time Since 2023. Bitcoin Held Near $76,000

17.09.2026

Read now

At its latest meeting, however, the Fed unanimously raised the rate by 0.25%. This was the first hike since July 2023. At the same time, the amount of money created directly by the central bank is not growing, and Treasury bill purchases under RMP stopped on August 14.

If the government starts aggressively ramping up borrowing without rate cuts or additional money supply expansion from the Fed, Treasury yields could rise. That would increase the cost of mortgages, auto loans, and consumer credit, and could potentially fuel public backlash against AI spending.

To scale this strategy without a sharp rise in yields, the administration needs a more dovish stance from the majority of FOMC members. However, beyond the Fed, there is another money-creation mechanism — commercial banks.

Both Bessent and Warsh have suggested that the banking sector could take on most of the credit expansion. After RMP purchases stopped on August 14, banks partially offset the effect by increasing total assets by more than $100 billion. This coincided with a loosening of some liquidity requirements.

In addition, after a 0.25% rate hike, banks will receive roughly $7.5 billion more per year in interest on excess reserves held at the Fed. Those funds can be used to expand lending and activity in financial markets.

The rate hike and the end of Fed balance sheet growth cannot be viewed in isolation from what commercial banks are doing. Taken together, the current setup remains stimulative, and the financial system is still capable of absorbing the additional debt needed to build out AI infrastructure.

But if the state does not become a buyer of compute, some AI-related debt could be impaired. Investors who bought these instruments using leverage would be especially vulnerable.

Affiliated Insurance

I’ve never studied the insurance sector in depth, so for a long time I didn’t pay much attention to how large private equity funds use the assets of affiliated insurance companies to finance their own deals. Once I understood the mechanics, I realized who ultimately ends up taking the risk.

In this case, the risk holders are buyers of life insurance and annuities in the United States. To understand the mechanism, you first need to unpack why the golden age of private equity funds ended, and how the industry tried to preserve its former returns.

The classic PE model looks like this:

• a fund buys a company with stable cash flow
• increases its leverage
• extracts part of the capital via dividends
• then tries to sell the business or take it public again

After the 2008 financial crisis, this strategy worked especially well — the private sector was deleveraging, and the Fed cut interest rates to near zero. Funds could borrow cheaply and acquire companies with resilient cash flow and low debt loads. Instead of scaling the business aggressively, they cut costs, ramped up borrowing, and funneled the freed-up capital back to investors.

Over time, returns from this model began to decline — cheap money pushed up valuations for attractive assets, and after the pandemic, the cost of raising capital started to rise. The industry needed a more stable source of capital that would not demand a quick payback.

The solution came from the insurance sector. Unlike a typical fund, an insurance company receives capital for a very long time and does not face constant pressure from investors to return their money.

An insurer sells life insurance policies and annuities, collects premiums, and invests them until it is time to pay out — sometimes decades later. For private equity, this is close to ideal long-term capital that can be deployed into private companies and lending.

Some funds have started acquiring such companies while also becoming their asset managers. Client money was then routed into investment products tied to the insurer’s owner. This is the structure I call affiliated insurance.

Reinsurance and hidden risk

However, buying such a company alone is not enough. Assets can fall in value, so regulators require a certain capital reserve to be maintained. Some of these risks can be transferred through reinsurance.

Normally, the insurer and the reinsurer are independent, so the risk is priced on market terms. Here, the insurance company creates its own affiliated reinsurer. This structure can take on a significant volume of liabilities with relatively little capital from the parent company.

Formally, these entities remain regulated. However, in some jurisdictions, the rules allow such arrangements with limited disclosure and relatively low capital requirements. The mechanics are easier to show with a hypothetical example.

Suppose Luna acquires an insurance company, Alameda Insurance, with a multi-billion-dollar asset portfolio. Alameda Insurance cannot directly buy a high-risk token, but it can buy investment-grade corporate debt.

Luna issues such high-yield bonds, and Alameda Insurance buys them with client premiums. Then an affiliated reinsurance company, Three Daggers, is created, which supposedly assumes the risk on these assets while having only a small amount of its own capital.

As long as Luna’s bond rating remains high, the structure looks stable. But if Luna stops servicing its debt, rating agencies downgrade it to speculative grade. After that, Alameda Insurance is required to increase capital against those assets.

Three Daggers is supposed to cover the difference, but it may simply not have the money to meet the new requirements — meaning the losses remain on Alameda Insurance’s balance sheet.

Who ultimately takes the hit? Policyholders. In most states, the guaranteed payout in the event of an insurer bankruptcy is capped at roughly $250,000-$300,000. If a contract promises a payout of several million dollars, the client can lose the difference.

There is another wrinkle — unlike the banking system, where the FDIC collects premiums from banks in advance, insurance-industry guarantee mechanisms in many cases are funded by the remaining participants only after a company fails.

AI debt on insurers’ balance sheets

Now let’s apply the same mechanics to traditional finance. The names of the assets and the players change, but the core risk remains the same — the insurer holds the debt, while the capital of an affiliated reinsurer may turn out to be far smaller than stated.

Instead of the debt of a hypothetical crypto project, you can plug in private credit extended to software developers facing competition from AI. The same goes for data center liabilities, whose value depends on demand from the largest AI labs for compute.

In the table below, the Affil Reins column refers to affiliated reinsurance, which is supposed to serve as a backstop for these kinds of investments by large PE firms, including Apollo, KKR, and Brookfield. The quality of the assets on the balance sheets of such structures is not reliably known, because they are typically registered in jurisdictions where financial disclosure is limited.

According to estimates by Nick Nemeth, the total volume of assets tied to such affiliated reinsurance structures reaches $1.54 trillion. But how much of that amount is backed by real capital is unknown.

As an example, he cites one of Brookfield’s reinsurance structures. An asset linked to it was valued at $1.48 billion. At the same time, the entity that was supposed to provide coverage told the regulator that it effectively bears no payout obligations.

High ratings allow this system to look resilient, but the private credit and AI infrastructure debt markets are already under pressure due to the massive volume of issued obligations. The tipping point could be a downgrade of the credit ratings of securitized data center obligations.

Insurance companies bought the highest-rated instruments because they offered higher yields than Treasuries with comparable duration. If the largest AI labs consume less compute than expected, data centers’ cash flow may be insufficient to service the debt.

Then rating agencies will start downgrading these instruments. After such a decision, the parent insurer would have to raise capital, and it may find itself unable to obtain the necessary funds from its affiliated reinsurer.

For investors betting on money supply growth, a potential multi-trillion-dollar hole in insurers’ balance sheets creates the conditions for government intervention.

I see a parallel here with the 2008 crisis. Back then, AIG accumulated massive risks tied to complex debt instruments and ultimately received government support. Funds from the TARP program helped cover the company’s losses, and a significant share of the payouts went to its counterparties, including Goldman Sachs.

I believe a similar scenario could play out around AI debt. The mechanics differ from the 2008 crisis, but the end problem is the same — if losses become systemic, authorities have a strong incentive to prevent full recognition of them.

However, this time, authorities will most likely try to avoid a major public insurer bankruptcy. During the 2008 crisis, Paulson and Bernanke initially tried to preserve market discipline and allowed Lehman Brothers to fail. The fallout made the financial sector’s problems obvious to the broader public.

Nothing New

So, “safety first” does not mean an immediate, sharp surge in money printing. First, this principle forces the Trump administration to choose which part of the system to backstop.

For us, holders of bitcoin and other crypto assets, there is almost no difference between these scenarios — both lead to money supply growth.

I expect the period of choppy sideways action in the crypto market after the local rally in late August will end soon. The amount of dollars in the financial system will keep rising, which should support bitcoin and select altcoins.

This policy benefits both AI and crypto projects that need low-cost computing power. I also expect that rising dollar liquidity will increase investor demand for crypto assets as instruments that are sensitive to monetary expansion.