Inside AI's $518 Billion Problem

For months, AI bulls have dismissed the rise in credit default swaps (CDS) across AI companies because broad high-yield spreads remained relatively calm.

Fair point.

But there’s a fundamental problem with that view.

These spreads measure the entire market, and the debt/credit boom is happening inside a much smaller group of companies.

That small group is now becoming risky enough to move the much broader credit basket.

So I started pulling on that thread, here’s what I found:

Eventually, I landed on one number:

$518 billion.

That’s the amount of cloud, compute, and infrastructure obligations tied to Anthropic (before leverage) relying on $65 billion of annual revenue.

With token prices down 60% this year alone, which now command a 150% volume increase just to keep their old revenue targets, this gap is widening by the day.

Let’s not even add OpenAI’s commitment pile, this one’s bad enough.

And it’s not even the worst part…

NVIDIA is now reportedly looking to insurance companies to help absorb the losses if GPU-backed loans collapse.

If GPU prices decline by 50% (they already have) then hundreds of billions of loans, and bets riding behind them, all go into default.

The question is, who blows up first?

Let’s get into it,

Inside AI’s $518 billion problem.

So far this year, Anthropic has indeed posted an amazing revenue growth run-rate going from $9 billion in January to $65 billion in August.

Compared to the annualized obligations of ~$74 billion a year, this seems to be under control as long as revenue continues to grow past that annual mark.

However,

There is one major issue these models are not accounting for, and perhaps the reason why most AI stocks are now trading at declining forward P/E valuations.

Remember, declining multiples signal markets are losing confidence in the future EPS growth/quality of the underlying, and one of the reasons is found in the following dynamic:

Let’s break these down further and connect them to what’s actually happening.

As Amazon, Google, Meta, and other neoclouds continue to expand on their cloud/compute demand…

Further obligations land on Anthropic’s and OpenAI’s desk.

At the same time,

Token prices are falling because of model competition, as new labs release the latest model with ever cheaper costs to the end user.

Take Chinese open source models for example, asking 10% of the U.S. model costs and delivering just as much output.

This brings up the issue of customer routing (switching to different models) and price wars amongst the several models now online.

As an AI power user, I have switched between Anthropic, OpenAI, and DeepSeek effortlessly by transferring chat history and meta data as if I never stopped using any of them.

I bet hundreds (if not thousands) of other users are doing the same.

On that topic, I have taken the company’s base and bullish guidance against what’s actually happening today in a price war + customer routing scenario.

The effects quickly bring the forecasted revenue from ~$350 billion in 2030 down to ~$80 billion.

I still believe $80 billion is too conservative, so let’s assume it’s somewhere around $150-$200 billion for now.

The point is that it is still well below the base guidance every single AI company is relying on.

If Anthropic cannot turn this goal, nobody gets paid, and the loans we talked about all go under.

Let’s go over to the margin side now…

On a base case, Anthropic will reach a steady ~20% EBIT margin by 2030, which is nowhere near what they have been touting lately at 80%.

When you add in the compute, GPU, and model training costs, that reported 80% quickly goes below 10%.

Now if we do continue down this path of price wars + customer routing, Anthropic will be projected to have negative ~70% EBITDA margins.

What’s important is that the AI race doesn’t break until the loans go bad.

So I would be a fool to tell you AVOID everything having exposure to AI.

I also would be a fool to tell you to buy and not hedge your book!

Which is exactly where a long/short equity strategy finds the middle ground between those two approaches.

Click here to see our winning trade in Data Center Cooling Stocks.

Here are the scenarios for depreciation in GPUs and other chip assets held by most of these companies.

From chipmakers, hyperscalers, and neoclouds, to some banks even…

The current depreciation rate is of 20% a year, which is significantly slow compared to how quickly some of these chips have already depreciated as new models come out.

A more realistic base case can be set at 35%/year.

But,

The reality is that these chips are already depreciating at 50%/year!

Which means the loan goes “underwater” during the first year, and during the second year if we take the 35% base case.

Take the NVIDIA H100 rental prices for example:

Given that most of these loans were originated in the fourth quarter of 2025 and the first quarter of 2026…

That would mean we are 1-4 quarters away from seeing reports of all these GPU-backed loans be underwater.

Effectively a collapse in the system.

Now here’s the biggest takeaway from today:

This is now a hot potato game

And the bag holder is about to shift.

At the base case of 35% depreciation per year, the insurer only takes on ~10% of the loan’s risk.

Effectively no need to transfer the risk into an insurance company.

But,

Given the most likely reality is for 50% per year, suddenly risk cannot be tolerated at the sponsor level.

That’s why NVIDIA has just coerced some insurance companies to create policies against GPU loan defaults.

Crazy right?

Anyways, at 50% depreciation, the insurer suddenly is responsible for over 70% of the loan’s losses.

We can see the waterfall of GPU loan exposure begin to shift to different places.

The way I see it, APO and KKR are the most exposed so far.

But,

As defaults begin to hit with the 50% depreciation scenario, my guess is that a lot of that portfolio will be sold for pennies on the dollar to a reinsurer like AIG, and then down the list we go.

All until it is too much for the system to take in, and the government has to step in and bail out the gamblers.

So no, this time is not different.

Don’t think so?

The only way this doesn’t happen is if token volumes can increase by over 150% at Anthropic alone (don’t forget Muse, OpenAI, Gemini, Copilot.)

And,

For these 8% interest loans to be within covenant before 2028, meaning - in Anthropic’s case - over $550 in revenue by then.

You’re asking for a ~10x jump in revenue over two years, where the commodity being sold is down 60% in unit price.

In other words…

I am investing in a company whose products are down 60% due to competition and price wars, but I’m telling you the company will be okay because they’ll 10x their revenue in two years to offset a current 10x Debt/Sales ratio.

Possible? You tell me.

Since we know that the risk waterfall is quickly - and surely - shifting from APO to KKR and now toward AIG as the next reinsurer…

Let’s take a look at what the whales are doing with their money in the options market:

Seems like a bearish spread is going on right now.

This bet is a one directional view that AIG stock will decline (quickly) over the next quarter.

The ATM calls provide more than enough premium and liquidity to buy up twice as many long puts in this case, which decreases the risk and still allows the holders to sell out of the stock at the current market price or keep the stock if it doesn’t fall further.

There is a major fight happening at this $75 level as well, seeing that both call and put walls sit at that strike at the options dealer’s desk.

Very rarely do we get such a tight wall dissonance, meaning MASSIVE tail risk for either dealer to take on here.

My guess is they are willing to assume this much tail risk because of the economics described today, a very low probability that these loans make it out okay in the next 1-4 quarters.

Meanwhile, here is the latest from the AI retail voice Jordi Visser.

Notice how he has dismissed ALL of the risks I pointed out today, citing that we are no longer operating in human time but rather in some distortion of the earthly laws.

In other words, yields and interest don’t matter because AI is going to accelerate time and push us into another continuum.

Sure, let me tell my bank I won’t be paying interest on my loans because we’re no longer in human time… I’m sure that will work!

Until next time,

OFFSIDE RESEARCH

Against the Tape, Ahead of the Curve.