Goldman has to convince the public that the demand for AI is going hyperbolic in the next 5 years. The entire private AI complex is priced on the answer. That's why, next week, Goldman Sachs has to do the most difficult and most important thing in capital markets: convince the world that the boom is real.
On 12th June, SpaceX is going to be the largest IPO in history. SpaceX is going to be valued at about $1.75 trillion, with $75 billion in new capital at a price of $135 per share. What most people agree is that the SpaceX valuation has very little to do with rockets. It is almost entirely on the AI business of the company (the old xAI, which was merged in early this year), which is a business that lost $6.4 billion in 2025. To get investors the $1.75 trillion, Goldman is saying that the $3.2 billion AI business is going to grow to $322 billion by 2030. That is a 100x times growth in 5 years. That is the single most important projection of the entire deal, and is the most important projection of the AI boom, and a considerable part of the entire market.
This IPO is like no other, and here’s why. Demand most of the time exceeds supply during the AI era. For the past three years, the demand is what most analysts believe to be real, and the investment will eventually follow. The majority of the time, those beliefs are sustained and valued in the private market. The AI players, large and small, have mainly been transacting privately and most of the time, inter-company. The company going public, for the first time, allows retail and institutional investors to price it in the open market. If demand, for whatever the potential market value is, for the IPO does not exceed the supply, then all private market AI valuations will have lost the market justification they have been hinging on. The biggest risk lies in this being the first price private investors have to contend with. The rest of the AI infrastructure will be negatively impacted and realign to this price. The market will have a better sense, based on the price, of the demand for AI services after June 12. June 12 is the day we will learn that.
To understand why that belief is so unstable, look beyond the IPO to what is driving the bulk of revenue at chipmakers and purchases by AI labs: vendor financing. The mechanic is simple and ancient. A supplier gives a customer the money to buy the supplier's product and books the sale as revenue. Lucent did this during the telecom boom and, unsurprisingly, it went bad. Between 1999 and 2000, Lucent would finance $8.1 billion of vendor financing, or 24% of its sales of $33 billion. That ruined Lucent, and relative to what is happening today, it is pocken money. Nvidia's commitment of vendor financing to OpenAI, announced in September 2025, is $100 billion, and on its own, dwarfs Lucent. Considering all the vendor financing, the web of GPU-backed debt, and the direct financing, Nvidia's vendor financing commitment to OpenAI is $110 billion, making the Lucent case look mild, and Nvidia's case is three times worse.
The whole deck of cards rests on OpenAI, Nvidia's largest customer, which in March 2026 raised an astonishing $122 billion, at an $852 billion valuation, over 30x its annualized revenue. OpenAI is projected to have an approximately $14 billion loss this year, and estimates are that they will have to spend $50 billion on compute in 2026 and will make about $13 billion. That’s several times all revenue, just spending on chips. If you remove that block, you will quickly see how much of Nvidia’s “demand” was demand, and how much was just the same money circling around. You don’t even have to guess how this will get the nerves on the inside. OpenAI, in February, quietly reduced its stated ambitions for this infrastructure from $1.4 trillion to approximately $600 billion, and began the process of renting data centers instead of building them, the Stargate project. Its Chief Financial officer stated that she did not have confidence that revenue would meet the obligations and suggested delaying the IPO. When the CFO goes from stating a firm position to ambiguity in writing, you know that is not background noise, that is the tell.
Then there is the crack that is almost universally unobserved: Chips don't last nearly as long as the loans that are taken against the chips. Nvidia makes a substantially better chip every year. From Hopper in 2022 to Blackwell in 2025 and Vera Rubin later this year and Rubin Ultra in 2027: All of the new Chips are better and Faster for a lower cost. Yet the companies that purchase depreciation over 5-6 years. Microsoft and Google managed to increase the estimated lifespan of their chips from 3 years in 2020 to 6 by 2023. Meta managed 5.5 years, which single handedly saved them $2.9 billion in depreciation for 2025. The only interesting one was Amazon, who increased the lifespan to 6 years in 2024 and then decreased to 5 in 2025. The lifespan of the chips, under normal use, is probably 2-3 years. So, after 2 years the chip is collateral for a 6 year loan. However, the collateral rots faster than the loan. Michael Burry estimated $176 billion in unaccounted depreciation for 2026-2028, and claimed that Oracle profits for 2028 would increase by 27% and Meta by 21%. You can argue with his prediction.
The depreciation schedules themselves cannot be objected to because they are publicly available and audited.
The really unsettling aspect of the risk presented by the Nvidia xAI deal is that It cannot be found on the expected balance sheets. Modern data centers are often constructed within special purpose vehicles with 10-30% equity and 70-90% debt. These purpose-built data centers are leased back to the operators, and the debt is kept off the operators' balance sheets. In this case, that small percentage of equity means that equity is annihilated in the case of a decline in the value of the chips long before anyone feels it in the debt. In the Nvidia, xAI deal, $3.5B of the financing was provided by Apollo, and was securitized and sold to Athene, which is Apollo’s insurance company. Athene, which is in the business of providing annuities to American retirees, has over $217B in assets that have been moved to a captive insurance subsidiary in Bermuda, which is outside the normal US regulations. Of those assets, which are very close to 35%, $103B are classified as Level 3, meaning that they are valued using internal models rather than by an observable market price. So, several layers removed from the silicon, a retiree who invested in a fixed annuity is the ultimate holder of fast-depreciating, leased-out GPU risk. If this case is reminiscent of the 2008 financial crisis, that is not a coincidence, as the two cases are very similar, albeit with different underlying assets.
Athene disputes the characterization and states that it would hold the same reserves in Bermuda as it would in the United States.
This is ultimately going to unravel in one of two ways, both of which capture the same essence, which is demand. Either financing runs dry, or customers disappear. Back in 2000, the funding for telecoms went from billions to zero in a single quarter, and when it happened, ~60% of vendor loan books were uncollected because the customers went bankrupt and the equipment became worthless. The other type of tremor we have already seen. At the end of January, Microsoft reported that 45% of their $625 billion cloud contract backlog, which is a single customer that will deliver $281 billion in revenue, is in OpenAI, and the stock declined badly, wiping hundreds of billions in a single day. Oracle, which is the only other hyperscaler that has that sized backlog, is in the same situation with its $300 billion OpenAI contract, and its credit is already on watch due to the Stargate exposure.
To be honest, the Lucent gives us the closest proxy to who this could play out. Lucent’s highest revenue was reported at $37.92 billion in 1999 before declining 69% to $11.80 billion in 2002, and they never got back on their feet after that. The same situation may be happening with AI. The question that everyone has been skirting around is the same situation Lucent dealt with. Is the AI demand real and sustainable?
June 12th means that Goldman employees will be at the forefront of having to really prove that AI is currently real. Something that is expected to grow a hundredfold after a $6.4 billion loss, and if that is achieved, everyone’s businesses will continue the status quo. If that event fails, a benchmark reset will occur, and everyone will have to recalculate.
We are just about to answer the question of whether this time as a different asset, the same structure, and the same situation as last time, will it really be the same?
