The AI boom has created a surge in earnings that has been the justification for moves higher in SPX and NDX for 2026. Among the hyperscalers, the majority of the earnings growth is coming from non-operating one-time gains from the revaluation of private equity AI holdings. The private equity valuation increases for Anthropic and OpenAI are the main reason for the increased earnings. These type of earnings are not sustainable. Two companies whose main source of revenues are from LLMs which are quickly becoming commoditized, thanks to China. Two money losing companies that have huge upcoming debt payments when you get a wave of AI data center completions starting next year.
For the semiconductors, the earnings are real and come from operations, but depend on continued spend from Anthropic and OpenAI which fuels hyperscaler and neocloud capex demand. Hyperscalers/neoclouds are going increasingly into debt to fund their AI data center buildout so the Street has been flooded with AI data center paper. A look at ORCL 5 year CDS (around 200 bps) shows you the consequence of going all in on AI data centers. Before ORCL's spending spree, their 5 year CDS was around 35 bps.
The earnings growth story is dependent on this AI capex boom continuing. It all flows back to Anthropic and OpenAI continuing to spend beyond their means through equity and debt offerings, or getting enough profits to pay for it without financing. But a recent Substack post from Groundbreaker shows the similarities between AI financing and Subprime mortgages in the mid 2000s.
Instead of a mortgage rate reset wall, you have a compute commencement wall from AI data center completions. When the data centers are completed, that's when OpenAI and Anthropic are obligated to start paying for the compute they ordered 2 to 3 years before. The bulk of the completions happen in 2027 and 2028. That is when they really need to start paying for their compute with profits or go back to the Street for more financing. Eventually, the Street will balk if they remain unprofitable. If the AI labs can't get more financing or become profitable enough to fund themselves, they will default on their debt obligations to the hyperscalers, and could force NVDA to backstop them, leading to big losses.
Already seeing reduced demand for AI data center debt. From a recent FT article:
So far, with the rapid revenue growth from selling their services below cost, the AI labs have been able to get all the financing they need. This is the honeymoon period where the hyperscalers book huge order backlogs with each signing with AI labs, raising their revenue outlook. But the hard part comes later, when the AI labs' payment obligations really ramp up with data center completions in 2027 and 2028. The AI labs need to profitably sell their compute at increasing scale or borrow more/sell more equity. If the AI labs aren't sufficiently profitable by next year and the financing window closes, then defaults start happening, hyperscalers have to use non AI cashflow to pay for their AI capex debt, and the shit hits the fan.
Few are bracing for this. It all depends on demand for AI compute at high prices that can make Anthropic and OpenAI profitable enough to meet their growing debt obligations. In order to improve their poor operating margins, the AI labs will have to use less compute for training, meaning their lead over Chinese AI labs will shrink even more, or completely disappear, or they will have to raise prices, reducing demand for compute. In either case, there will be less demand for compute at unsubsidized pricing. Less compute demand is bad news for semiconductors, the most important sector in the stock market. Only if the demand for AI compute continues to grow even at unsubsidized pricing will you see the AI boom continue. Otherwise, AI investors will be heading full speed towards the compute commencement wall in 2027/2028 without airbags.
If the AI labs remain unprofitable and keep burning cash, as I expect in 2027, then they will need more financing rounds to pay when the compute commencement wall hits. I have a feeling that those financing rounds will be much tougher to get than the ones this year. I doubt there will be much appetite to fund money losers when they have so many future debt obligations coming down the pike. 2028 will be even worse for AI data center completions and debt payments. If the AI labs can't get the financing, then they will burn through cash and have to default on their loans.
Those loan payments are going towards hyperscalers and neoclouds. Then the hyperscalers will have to write down huge losses on their books, along with the steady depreciation of their AI assets. They will have to repay AI infra debt from non-AI data center cash flow. This will impair their ability to do stock buybacks and be a huge hit to their earnings. The worst of the storm will hit the semiconductors, as their order books will dry up, along with their profit margins which will plummet. Then people will understand why semiconductors are considered cyclicals.
You will need at least 150% revenue growth in 2027 for openAI and Anthropic to be able to meet their debt obligations without additional financing. From a demand perspective, I am highly skeptical that AI compute at unsubsidized pricing grows 150%+ vs. 2026 levels. If AI was so great, why aren't we seeing call centers all replaced with AI automated chat bots? If you ever called a customer service phone number and gotten an AI voice bot trying to solve your problem, you know how annoying and unhelpful it is. You just end up having to waste time going through various chatbot menus instead of speaking to a call center worker directly. You have to wonder why there haven't been more job cuts at the tech companies which are the heaviest users of AI agents. If these AI agents were so productive, why not do more drastic job cuts? Many are ignoring the limitations of AI use cases as they project huge growth rates for years on end.
Instead of scaling laws leading to exponentially increased capabilities, its been more like a logarithmic increase in capabilities that show less progress for each new model vs previous ones. ChatGPT is a clear example of this. Their first models hallucinated a lot. The next iteration vastly improved these shortcomings, but with each new model, the low hanging fruit having already been picked, you get less incremental improvements. You are already seeing pricing pressure at the AI labs, as demand for the latest expensive models has been subpar. There is very limited demand for frontier models when previous models get most of the AI capable jobs done for a fraction of the cost.
BofA clients have been aggressively buying tech for the past 4 weeks.
BofA clients bought heavily last week.
Back to the markets. You are seeing a lot of discussions about bond yields these days, and the kneejerk reaction would be to say that we are nearing a top in bond yields because people are worried about them. If investors were really worried about bond yields, the MOVE index or VIX wouldn't be so low. You have been getting very little volatility in both stocks and bonds, and thus investors are complacent. There hasn't been a 1%+ down day in SPX in over a month. Investors tend to extrapolate recent history into the future. Hedging demand for SPX is way down from earlier in the year. Investors are increasingly unhedged going into 2 big catalysts: Anthropic IPO and midterm elections. Gamma fueled selloffs may seem scary, but they don't last for long. The scarier selloffs come from real money selling, not dealer delta hedging. With customers light on put protection, they are more apt to panic sell in a correction.
SPX avg stock call skew is high, avg stock put skew is low. Investors are leaning heavily towards calls.
DBMF, the biggest trend following ETF, substantially increased their long SPX and MSCI index futures positions to a total of 63% of AUM. They now have one of the biggest net long equity positions for the history of the ETF.
The imbalances in the market are huge. So many are leaning towards the long side here, aggressively long, just when the risks are about to be the greatest. Maybe I am off my rocker, but I think the next 2000 points in SPX is down, not up. It is a long term view looking out into 2028, so I try not to let those views affect my shorter term trades.
That being said, the probabilities on the short side are lower than trading on the long side, but the payoffs are more explosive and come in waves, all at once. Shorting the index is a low probability/large payoff structure. I see a big wave coming before the midterm elections. It could start next week, it could start in late September. If we are still above current levels by the last week of September, then my thesis is incorrect and I will exit my position. In the meantime, I will hold a full position short, with initial target of SPX 7400, and depending on price action and how the crowd reacts, may look for a move down to 7000. Looking to hold the trade at least till the last week of September (may close out some earlier to put out again on a short term bounce), and will look to close it all out at the latest in the last week of October.
Posts
