Arthur Hayes: Nợ của AI, bom nổ của bảo hiểm, nhiên liệu của Bitcoin

Original author: Arthur Hayes

Original translation: Saoirse, Foresight News

Have you heard? Those AI industry bigwigs suddenly grew a conscience and started worrying about the survival of humanity — because the near silicon-god product they have been building is about to be born. In fact, it has been close to silicon-god status for a while now, but with the final step seemingly imminent, they suddenly began to reflect on the path of artificial general intelligence (AGI) development.

(Note: Silicon-God, the super AGI that Silicon Valley says is about to be born; the author uses this term with sarcasm, questioning whether this grand narrative is merely a fundraising and lobbying tactic.)

That is the narrative they tell the public. But being naturally suspicious, I checked the calendar: the end of the third quarter is almost here, and Anthropic is still not public. I cannot help but wonder what their financial statements actually look like. Has the high-growth annualized revenue curve repeatedly mentioned in press releases already slowed? They claim that after excluding all operating costs, the company is profitable. I cannot wait to study the S-1 prospectus they are about to file and figure out exactly how much it costs to provide each token to users. Also, how many customers actually generate profit for them, and is this profitable customer base expanding or shrinking? Unfortunately, I cannot get answers to these questions, and the reason is — safety first.

Readers can tell from my tone that I believe Anthropic, OpenAI, and SpaceX claiming to slow AGI research on the grounds of “safety first” is not out of concern for the well-being of ordinary humanity, but rather stems from harsh economic reality: the market is unwilling to buy the AI products they sell in sufficient quantity at current prices. More specifically, demand for AI is strong, but what people want is Chinese pricing — only one percent of the U.S. price.

When this “Chinese pricing” hits self-important American AI practitioners, their first reaction is to shout: “But those Chinese products are low quality.” And when the quality of Chinese models quickly catches up, they then complain: “China only built its own products by distilling our models.” The market does not care why Chinese models are cheap; the market only wants the cheapest intelligence services. So these AI practitioners turn to lamenting: “We care about human safety, so we are pausing research.” How noble that sounds... But this exaggerated “injury acting” worthy of a World Cup match later comes with a demand: “But we still need to beat China, so the government should step in, introduce regulations, and continue funding this AGI race.”

The reason these three leading U.S. AI labs using “safety first” as an excuse to slow AGI research is so crucial to financial markets and global fiat liquidity is that these labs’ demand for compute supports more than $1 trillion in investment-grade debt, plus hundreds of billions in lower-rated loans.

This shows the purchase commitments OpenAI and Anthropic have made to the four major U.S. cloud providers as a share of each provider’s unfulfilled revenue orders, revealing the massive long-term orders the two major AI model companies bring to cloud providers.

Together, these AI labs generate no profit at all. Therefore, they need profitable tech companies such as Nvidia, Broadcom, Google, and Microsoft as backers to provide off-balance-sheet guarantees for debt tied to data center leases and chip purchases. Subsequent purchases of chips and hardware depend on AI labs continuing to train the most cutting-edge model, the one that is “almost, almost, almost about to become a silicon-god,” while also processing inference requests for customers. But if “safety first” becomes the new core principle, spending on training new models will not fall to zero, but it will certainly decline from current highs; companies will focus on improving the efficiency of converting electricity into intelligence, which means customers will spend less on compute. In essence, safety first means destroying compute demand.

If AI capital expenditure were financed by operating cash flow, there would be nothing to worry about. But the problem is that regardless of whether AI labs continue buying compute, these trillions of dollars in debt still exist. Default will not happen immediately, but once AI labs stop consuming compute at the scale previously expected, the price of this debt will fall. The real core question: who bought this debt, and did they buy it with leverage? The answer is obviously that these speculators used leverage to buy this low-quality debt. So who is the ultimate bag holder?

Millions of American insurance policyholders are in fact indirectly betting on the AI story. And “safety first” will hurt them without any buffer. I did not fully understand this scam at first; it was Nick Nameth on Substack who explained it very thoroughly, and next I will write it in simple language for my crypto readers. The conclusion is: if AI-related debt is repriced at fair market value, a large part of the U.S. insurance industry is in fact already insolvent. This leads to the core proposition for investing in a global economy dominated by fractional reserve banking: the U.S. government has two choices — either act as the buyer of last resort for compute in the name of national security, or print money to bail out deeply loss-making insurance companies.

Whichever path is chosen, we Bitcoin holders and crypto investors are the winners. If the government ignores market signals and insists on spending money to develop this commercially unprofitable “silicon-god,” it will need to print money to fund such unproductive spending, which will inevitably fuel more financial speculation and push up the Bitcoin price. If the government chooses to bail out the insurance industry, it will print money to absorb bad AI debt, expand the money supply, and thereby push up the Bitcoin price.

The rest of this article will break down this mechanism.

In the Name of China

As long as it is labeled national security, the U.S. government can find a justification for almost anything. Think about how much destruction was caused after 9/11 when the United States stoked public fear and launched the global war on terror. This time, the imagined enemy is the Chinese, who are supposedly buried in mathematics, stealing top American technology, and selling products back to the United States at one percent of the price. To defeat China, it is necessary to promote state socialism within the capitalist system.

AI bigwigs successfully persuaded Trump and his staff to ignore two realities: the market has already proven that the AI business is unprofitable, and bipartisan voters oppose building large numbers of new data centers while also demanding compensation for data theft. Since China can provide affordable AI products, the U.S. response is to spend even more money to build that “almost, almost, almost, almost about to be born silicon-god.” (I will keep adding “almost,” because the only things missing before AGI are faith, data theft, and taxpayer funding.)

War, the economy, robots, and everything else ultimately depend on AGI. Therefore, the rest of the world must use AGI according to America’s wishes. According to this narrative, the United States has the world’s most inclusive and fair culture, and this silicon-god absolutely must not be controlled by countries outside Judeo-Christian civilization, such as China. (Eye roll, followed by a giant eye roll.) I have my preferred place to live, and others have theirs. Even if I believe the moral culture I identify with is superior to that of other countries, I am unwilling to spend all my wealth and even my life forcing these values on the whole world. You can believe American or Western culture is superior, but do not hand trillions of dollars of taxpayer money to Elon, Sam, and Dario. (The heads of the three leading U.S. frontier AI companies: xAI, OpenAI, and Anthropic.)

The alleged superiority of the American system is that, most of the time, hundreds of millions of informed citizens determine the prices of goods and services through free markets, with the government staying out and letting market signals decide what to produce and how much. But now, because of national security, and based on an assumption — that pouring in enormous amounts of money and feeding data to predict the next token can produce AGI — the market’s signal is deemed wrong. So the U.S. government must increase spending to build the next generation of frontier models and use cultural superiority to suppress China. This is hubris. Remember what happened to Icarus when he flew too close to the sun?

All right, enough grandstanding. Let’s talk about the bailout plan.

“Safety first” means the compute demand of the three major U.S. AI labs declines. At that point, the government can step in and sign offtake agreements to guarantee stable profits for the AI labs, similar to the contracts the United States gives to some defense and mining companies. The government will use this compute to advance AGI research. Finally, the government can lease its own models back to the AI labs, which will sell inference services to customers in the United States and allied countries at extremely high prices.

This plan would allow the government to control frontier models for the Western world to use as it wishes. The private AI lab model has a problem: these labs are global companies, and sometimes they sell model access to anyone in the world for profit, which conflicts with the government’s national security goals. If Chinese labs have partly relied on distilling American frontier models to make technological progress, direct government control of R&D could set China back months or even years in the AGI race. Even if this idea holds, it will ultimately fail, just as the attempt to block advanced chip manufacturing equipment failed to stop China from producing cutting-edge chips. Information inherently wants to flow freely. In the internet age, information blockade is simply impossible. Even before the internet existed, after the United States successfully developed the atomic bomb, it could not stop the Soviet Union from obtaining related intelligence. Those who think AGI can be an exception do not understand human ingenuity and adaptability when interests are at play at the national level.

Funding this silicon-god requires issuing more debt. This plan is easy to sell because the key monetary policymakers — Treasury Secretary Bessent and Fed Chair Warsh — both believe AI can boost productivity. They are convinced that by fully embracing AI, the United States can grow its way out of its enormous debt burden, and this judgment is not entirely wrong. In June 2026, U.S. nominal year-on-year GDP growth was 6.6%, while the effective federal funds rate was about 3.6%. Bessent continues to issue short-term Treasury bills; the government can earn a 3% return on this debt issuance, but savers bear the loss. If the fiscal deficit is kept within 3% (a very big if), the debt-to-GDP ratio will fall. Economic growth is mainly driven by AI data center construction, and what supports all of this is the compute demand of AI labs. So from a financing perspective, if short-term Treasury issuance can still earn 3%, it is financially feasible for the government to act as the buyer of last resort for compute.

There is no free lunch. The U.S. government runs chronic deficits and can only spend by borrowing. If Warsh cooperates, this is easy. But so far, the Fed under his leadership and the Treasury under Bessent are not moving in lockstep.

For the first time since July 2023, U.S. monetary policy has tightened: at last week’s meeting, the Fed unanimously approved a 0.25% increase in the policy rate. The total amount of money created by the Fed is no longer growing, and as of August 14, the RMP short-term Treasury purchase program has stopped.

If the government pushes this plan but the Fed does not lower the cost of funds or expand its balance sheet, large-scale debt issuance will push interest rates higher. Rising interest on mortgages, credit cards, and auto loans will only provoke voter anger at AI-related policies. If Trump and Bessent cannot secure the support of at least seven FOMC members, the feasibility of this plan will be greatly reduced.

The above describes Fed policy from a traditional perspective, which easily makes people bearish on the market. But do not forget there is another powerful money-printing machine: commercial banks. Both Warsh and Bessent argue that the banking sector should take on the task of money creation. Since RMP purchases stopped on August 14, banks have created hundreds of billions of dollars in new money by expanding total assets, backed by looser liquidity regulatory constraints.

In addition to balance sheet expansion, after this 0.25% rate hike, banks’ excess reserves held at the Fed can earn an extra $7.5 billion in interest per year. This money will be used for new loans and financial market speculation. Therefore, one cannot look only at the Fed raising rates and stopping balance sheet expansion; the impact of the commercial banking system must also be considered. Combined, the overall effect is still stimulative. That is, if the government is willing, the liquidity environment is sufficient to support additional borrowing for investment in AI compute construction.

But if the government does not take over compute procurement, the debt will be impaired. Those holding such debt with leverage will fall into crisis. Next, let’s go deeper into the captive insurance scam.

Captive Insurance

I had never studied the insurance industry before. Large private equity institutions using captive insurance company assets to raise funds for investments did not initially seem like a scam to me. But after digging deeper into how this mechanism works, I found the ultimate victims.

Every credit bubble has a group of ultimate bag holders. Usually it is ordinary retail money managed by highly credentialed fiduciaries. The fiduciaries invest other people’s money and earn double returns: collecting management fees while also selling their own assets to retail investors and marking up the price of their own holdings. This time, the victims are policyholders who bought American life insurance and annuity products. To understand this scam, we first need to understand the survival tactics private equity bigwigs devised after the golden age of private equity ended.

After the 2008 global financial crisis, the private sector deleveraged, and the Fed cut rates to near zero. The classic private equity playbook: find mature companies with stable cash flow and almost no debt, load them with leverage, then pay themselves dividends, and finally leave the broken company in the private market. When public market sentiment runs hot, relist the bad company and complete the cycle. In a low-rate era, this logic worked perfectly. Ordinary people, stuck with negative-equity mortgages and struggling with monthly payments, had no ability to increase consumption. Private equity bigwigs did not want to expand production or provide better products; they only wanted to maintain existing cash flow, cut costs, and extract cash dividends for their own investors.

U.S. total credit as a share of GDP. Orange is private market credit, blue is government-related credit. After the 2008 financial crisis, private credit continued to fall while government credit continued to climb, reflecting a shift in the debt structure from the private sector to the government sector.

Assets under management (AUM) of private equity and venture capital (PE&VC). Even through economic recessions, it kept expanding continuously after 2000, surpassing $15 trillion by 2025. The gray shaded areas mark NBER-designated recessions.

In the post-pandemic era, the cost of capital rose, and the law of diminishing marginal returns hit private equity fund returns hard. Because financing was cheap, winning a deal required paying a higher valuation to acquire cash-flow assets. Private equity fund returns then declined. To complete the next round of fundraising, private equity bigwigs began looking for long-term capital pools that did not care about short-term redemptions — and insurance companies entered the picture. Insurance companies sell life insurance and annuity policies; policyholders pay premiums, and the insurance company invests that money to earn returns, then pays out on policies decades later. This is exactly the perpetual capital pool private equity dreamed of, one that can be invested in private equity funds full of overvalued private companies and high-yield private credit.

So private equity bigwigs acquired insurance companies, appointed themselves as investment managers, and packaged bad assets to sell to unsuspecting policyholders. This is captive insurance.

The most outrageous part of this arrangement is how captive insurance companies satisfy legally required capital buffers. Asset prices fluctuate, and regulators require insurance companies to set aside capital buffers to ensure policy payouts. This gave rise to the reinsurance industry, which assumes the claims risk of primary insurers. Normally, the primary insurer and reinsurer are independent institutions, and reinsurance prices risk at fair market value. But this rule does not work for the private equity insurance scam, because the core of the scam is to find a bag holder while private equity itself does not have to put in much of its own money. So private equity-controlled insurance companies set up affiliated captive reinsurance entities. The parent company only needs to contribute very little of its own capital to obtain reinsurance protection.

These are all regulated entities and are required to make regular public disclosures. If policyholders knew the insurance company was operating such a system, would they still buy policies? To hide the scam, the U.S. capital system sided with private equity. Regulations in some states, such as Vermont, conflict with national prudential regulatory standards. Primary insurers and affiliated reinsurers can privately set up reinsurance risk assets, capital buffer sizes can be set arbitrarily, and after state regulators approve them, the relevant materials are sealed and kept confidential.

There are even more details. Before continuing to dismantle this bold scam, let me use a crypto example as an analogy. Do you all remember Terra Luna? Luna collapsed because USDT holders sold the stablecoin and broke the dollar peg. What would have happened if Do Kwon had the resources of those nearly 70-year-old New York