By Roger Boyd
The whole thing reeks of the “dot com” bubble, but on an even larger scale. Added to by such things as “private credit” that are a creation of the massive deregulation of the financial sector in the 1990s, and the repeated bailing out of the financial sector from its self-created destruction, for example in 2008 and 2019/2020.
One of the greatest tools of the bubble has been the ability of private companies to openly misrepresent their revenues and costs, and for even public companies to manipulate depreciation schedules to misrepresent their capital investment related expenses and to engage in circular deals that misrepresent future revenues. With the whole thing dependent upon the ability of a very small number of companies to raise colossal amounts of capital from lenders, venture capital and private equity. These three pools are now approaching exhaustion and the loss-making companies now must reach out to the wider financial markets to tap pension fund, mutual fund and the wider public’s investment pools. This is why we are seeing the rush of SpaceX, OpenAI and Anthropic toward initial public offerings.
The problem with IPOs is that they require actual audited reports on revenue and expenses, which give the lie to the previous misrepresentations. The need to show a somewhat believable path to profitability, even if that is still a lie, is also requiring these companies to raise prices from the previously deeply subsidized levels; leading to customer sticker shock and reductions in usage. Many corporate CEOs are suddenly waking up to the colossal cost of their company’s usage of AI and the lack of any identifiable benefits.
The token cost increases, together with the reality of lack of ROI on AI token spending, directly contradict the growth forecasts of the SpaceX, OpenAI and Anthropic; undermining their claim for a path to profitability.
The future profitability of the hyper-scalers, those building the vast number of massive data centres to supply the computing power required for AI service providers, such as Microsoft, Alphabet (Google), Amazon, Meta and Oracle which are jointly investing US$570 billion in AI data centres, is utterly dependent upon AI service providers becoming profitable; otherwise they will have no paying customers for much of these massive investments. To hide the expense costs of these investments, these companies are engaging in two accounting tricks:
Not recognizing depreciation on the assets (GPUs, technical infrastructure, buildings etc.) until they enter economic usage, possibly delaying depreciation by years.
Stretching out depreciation on such things as GPUs to up to six years, when it would be three years at most if it reflected the real world depreciation. This stretched out depreciation schedule may also be after years of no depreciation due to not being recognized as being in economic usage. Such assets may already be next to worthless before they even enter depreciation in a company’s P&L reporting.
Even if the data centres are built and successfully start operation, a lack of paying customers will result in massive capital write offs.
The need for such colossal amounts of capital investment has also led to increasing levels of financial “innovation”, such as with Coreweave which buys much of its capacity from new specialist data centre companies such as Core Scientific which is building data centres explicitly for Coreweave; raising capital on the basis of its deals with Coreweave. With Corweave raising capital on the basis of its contracts with AI service providers and other hyper-scalers. With many of the AI hyper-scalers and chip providers such as Nvidia investing, and engaging in usage contracts with the AI service providers and even hyper-scalers such as Coreweave. OpenAI has revealed that Microsoft is responsible for both a significant part of its funding and provides a significant amount of its computing resources. This is a new version of the “vendor financing” so infamously practised by Cisco during the dot com bubble. Much of this “financing” does not even include real revenues, but rather barter deals that involve little or no exchange of cash. Such as an AI provider giving a financial stake in the company to a hyper-scaler in return or an agreed amount of computing power, counted as revenue on the hyper-scaler financial reports. Or even access to AI tokens being exchanged for computing resources, being counted as revenue on both the hyper-scaler and the AI providers financial reporting. A significant amount of the industries revenues are so created, involving no actual paying customers.
The AI sector did try to make a play for tapping huge amounts of government funding through “public private partnerships” in 2025, but that was quickly shut down.
As long as all the financial plates keep spinning the AI bubble continues, but many of the plates are looking decidedly wobbly. The SpaceX IPO, a very obvious pump and dump scheme to keep the company alive for now and to enrich insiders (insiders are allowed to sell stock only 2 months after the IPO rather than the usual 180 days), has now had to reduce its IPO valuation from the utterly insane US$2 trillion to the still insane US$1.8 trillion. Quite a good overview of the strings being pulled, increasing changes to IPO rules designed to protect public investors, to bail in retail investors through index funds while the insiders bail out.
The company is a hodgepodge of businesses:
xAI which had a US$63.6 billion operating loss on revenue of US$32 billion in 2025. Repeating his previous 2016 SolarCity acquisition by Tesla, designed to rescue his family’s investment in a failing venture, xAI was purchased by SpaceX in February 2026 at a ridiculous valuation of US$250 billion. The same was done with the failing X (aka Twitter) with xAI buying it for a ridiculous US33 billion in March 2025, plus the taking on of US$12 billion in debt obligations. Both all-stock acquisitions that were financial engineering designed to hide the failures of X and xAI, while also bailing out Musk from related financial obligations.
The space rocket company called Space, which in 2025 lost US$6.6 billion on revenue of US$40.9 billion. The revenue growth is greatly limited by the global launch market. This is in non way a rapid profit growth company.
A space-based internet service provider, Starlink, that in 2025 made US$44.2 billion on revenue of US$113.9 billion. The profitable business that is being used to keep the loss making businesses, especially xAI, alive.
The failure of xAI as an AI services provider is underlined by the failure of Grok, which has direct access to the whole X community of users, together with the leasing of much of xAI’s computing capacity to Anthropic. As with the AI bubble in general, Musk has to constantly work to find new spinning plates, in his case through having the wobbly plates taken over by the still spinning plates together with claims of future earnings that repeatedly turn out to be fantasy-based. Also as with the AI bubble in general, once the mystique is burst and reality intrudes, the financial fallout could be both fast and vast. Interestingly, the conditions of the IPO remove the ability of any shareholders to sue the company as they are forced into compulsory arbitration instead.
OpenAI is planning for an IPO in September with a valuation of about US$1 trillion, Anthropic has filed for its IPO so it will be before September, with a valuation of US$965 billion. Anthropic’s recent US$65 billion private capital raise, and OpenAI’s US$122 billion in March are seen as their last before the planned IPOs. xAI raised US$20 billion in January 2026, and SpaceX took on a US$20 billion bridge loan in March 2026. The incredible cash burn rate of these companies is shown by their need to go to the public markets so quickly after raising so much money. The IPOs also facilitate attempted insider exits before the plates stop spinning. SpaceX plans to raise up top US$75 billion in its June 12th IPO. If that IPO fails, all of the plates will rapidly start wobbling. Even if its is successful, the newly public SpaceX’s financial reporting that will somewhat reflect reality may cause severe difficulties for the OpenAI and Anthropic planned IPOs.
To get ahead of this, there have been big money raises prior to the SpaceX IPO:
xAI raised US$20 billion in February 2026, in a private offering
OpenAI raised US$122 billion at the end of March in a private offering
Anthropic raised US$65 billion at the end of May in a private offering
Alphabet (Google) public/private offering to raise US$80 billion at start of June
Doubled capital spending vs. 2025, to US$190 billion in 2026
With such massive amounts of money being raised (approx. US$275 billion plus insider share sales) in the financial markets by Alphabet, SpaceX, OpenAI and Anthropic over the next five months, that will enrich both insiders and Wall Street financiers, it is imperative to keep the overall US stock market levitated. This is very much understood by US Secretary of the Treasury Bessent and the new Federal Reserve Governor Warsh, as well as the US President. A self-feeding financial collapse would trigger a large recession, which would then explode a US government deficit already running at 5.9% of GDP and planned to grow more with President Trump’s proposed gargantuan increase in war spending. The government fiscal plates also need to be kept spinning, otherwise the US$ plate may start wobbling disastrously.
The colossal investment in data centres that may never be used, mirroring the “dark fibre” of the dot com bust, is what is driving the insatiable appetite for the products of the IT manufacturing giants such as Nvidia, AMD, Intel, Dell, Hewlett Packard, Supermicro, Arista, Cisco (yes them again!), Samsung, SK Hynix, Micron, Sandisk and a legion of other IT suppliers. Not only will their revenues collapse with the bursting of the AI bubble, with massive amounts of unused products overhanging their markets, but their share price-to-earnings ratios inflated by assumptions of rapid growth will also collapse; on the scale of the Cisco collapse in the dot com bust. Along with the hyper-scalers, they represent a huge chunk of the US stock and corporate bond markets. As the dominant supplier of high-end computer memory, South Korea is very highly exposed to any AI bubble burst. Will the South Korean leadership reach out more to China as they oversee a broken economy? About 40% of South Korea’s exports come from the IT sector, with the Chinese competitors already forcing deep price drops over the lower-technology memory and other IT markets. While the second biggest exporting sector, cars (10%), sees falling exports and is being increasingly challenged in its home market by China-based brands such as Tesla and BYD.
The Iranian leadership understand the above, why they are comfortable “running out the clock” against the US war of aggression. That is why the US administration will agree to a 60 day interim armistice, most probably with hidden terms to hide the scale of the surrender. This armistice will then be extended to get past at least the AI IPOs, the mid-terms and the renegotiation of the trade truce with China. But even then, there will be significant economic damage done by the fall in oil supples and the probable slow recovery of passage through the Strait of Hormuz, and more than US$275 billion will have been sucked out of the US financial markets by the IPOs and probable significant insider selling. The plates will need to be kept spinning by negative real interest rates, facilitated by US statistical manipulation and Warsh’s use of an even less representative inflation statistic, and obfuscated monetary expansion.
And the three AI amigos will be forced to report somewhere near truthful financial results every quarter. The plates will become increasingly wobbly, and the US administration will not dare risk a resumption of the Iranian War. Or any resumption of trade hostilities with China, which the Chinese leadership sully understands. Expect a very strong stance from the Chinese Party-state in its trade negotiations with the US, and increasing assertiveness about any US interference in Taiwan.
And in addition, there is the threat of the Chinese AI competition based upon cheaper to run models, Chinese AI chips and infrastructure, and much, much lower per token costs and even open source. Designed to facilitate rapid mass adoption rather than huge profit moats. More and more Western and other companies are utilizing Chinese Ai offerings, taking over the “countryside” surrounding the data centres of the three AI amigos; a classic game of Go with the Chinese winning.
From Steve Hsu on Twitter
We can expect increasing US actions based upon “national security” to wall off the US against the Chinese AI competition, just as with Chinese electric vehicles, but they will not stop the world outside the US walls adopting the much cheaper Chinese AI services. And the Chinese economy from massively benefitting from the much cheaper and more accessible domestic AI services.
As I have said before, a financial bubble has the ability to last much longer than thought possible and long enough to destroy the capital of short sellers. But they do burst, and the AI bubble will burst. And with it, the already decaying image of the US as a technological leader will be shattered, and quite probably the stability of the US financial system, the ability of the US state to fund its foreign wars and bases, and the value of the US$. The effects will be geopolitical as well as financial and economic, just as the 2008 GFC augured in the end of the second unipolar period.
