Nvidia just posted $96.2 billion in revenue in a single quarter, up 106% from a year ago. Jensen Huang stood on stage and said, "Compute is revenue." Congratulations. Generally, that is an extraordinary amount of money. Can I have some? The situation around that money, however, is complicated in ways that most coverage either ignores or oversimplifies.

I've already covered Nvidia's circular financing structure in a previous video, so I won't relitigate all of that today. What I want to walk you through instead is something quieter. Wall Street has been building entirely new financial instruments around AI compute. I promise the next video will have plenty of quips about Jensen Huang's jacket collection, but today we need to talk about what's being built behind the scenes. And I'm going to need you to stay with me on this one because what I'm about to show you is either the most sensible piece of financial engineering in a decade or the most elaborate attempt to keep a bubble inflated since someone invented the credit default swap. Possibly both.

There was a time in my life when I didn't know what a derivative was. What a blissful, innocent time that was. I had my PhD. I was doing perfectly fine. Well, as fine as a newly minted PhD can be. And then one day I made the catastrophic mistake of asking somebody to explain how futures contracts work. And the more I learned, the more I kept thinking, what? People bet on this. You can place a wager on the future price of something you never intend to actually buy. And that's considered a functioning financial market rather than say a casino with better lighting. Except of course in a real casino. The house eventually has to let somebody win a hand. In derivative markets, Wall Street just invents a new deck of cards every time the current deck stops paying out.

Now, you can make the argument that futures and options serve a legitimate purpose. Of course, they let businesses hedge risk, lock in prices, and plan ahead. A farmer can sell next season's harvest at today's price, and sleep at night. That price is reasonable, but of course we live on this planet with these specific people. And so the question becomes what happens when you take that machinery and point it at artificial intelligence?

Because that is what's happening right now. Both the Chicago Mercantile Exchange, CME, and the Intercontinental Exchange, ICE, the two largest derivatives exchanges in the world, are launching compute futures. Contracts that let you trade the future price of renting AI chips. The CME is rolling out contracts tied to the hourly rental price of Nvidia's H100 GPUs, while ICE is launching a broader index of AI compute capacity.

To understand why this is happening, you have to look at the problem AI companies are currently facing. Right now, buying and selling compute is basically like buying a used car in the 1970s. You negotiate a private contract with a cloud provider, AWS, Azure, Google Cloud, or one of the newer neoclouds like CoreWeave or Lambda Labs. The pricing is opaque, the terms are bespoke, and if you need thousands of GPUs next quarter for a training run, you have no way to lock in that price without signing a massive multi-year lease that could bankrupt you if your next model doesn't work out. It's like trying to run a restaurant where meat suppliers charge wildly different amounts for the same cut of meat and there is no public market where you can see what the going rate actually is.

You would have no way of knowing whether you were getting a fair deal. You couldn't plan your menu prices 3 months out. You're essentially flying blind. A futures market fixes that. It creates a public tradable reference price. Everyone can see what GPU time is expected to cost next month, next quarter, next year. Companies can lock in their compute cost today instead of hoping prices don't spike before their next training run. Lenders can look at the forward curve to assess whether a data center's revenue projections are realistic before extending a loan.

For the beautiful non-nerds in the audience, a forward curve is just a chart of what the market expects a commodity to cost at each point in the future. If GPU rental prices are expected to drop, a lender can see that before agreeing to finance a data center whose business plan assumes they won't. The CME's head of energy products said it directly. I'm quoting here. "Just as oil fueled the 20th century's economy and evolved from spot trading into a global derivatives market, our futures contracts will now turn compute into a standardized tradable commodity."

So that quote is interesting. They are not comparing compute to software or to cloud services or to any other technology product. They're comparing it to oil, a physical resource that powered the global economy for a century. And they're building the financial infrastructure to treat it accordingly.

Now, there is a version of this story that is entirely positive. Transparency and price discovery are good for any market and hedging is of course good for any business planning more than a quarter ahead. BCG estimates that a reliable forward curve of compute prices could reduce borrowing costs across AI infrastructure investment by about $116 billion through 2030. That is real money saved by real companies making real planning decisions. If you're building a data center and you can hedge your GPU cost the way a mining company hedges metal prices, that's genuinely useful financial engineering, of course.

But here is where my personal history with derivatives makes me twitch a little bit. Every major commodity bubble in modern history was amplified by the financial instruments built around it. The 2008 financial crisis wasn't caused by people buying houses. It was caused by the derivatives, the collateralized debt obligations, the credit default swaps that let Wall Street make leverage bets on housing prices disconnected from whether anyone was actually living in the houses. The futures market for oil is used for hedging, yes, but it's also used for speculation by traders who will never take delivery of a single barrel.

The AI giveth and the AI taketh away. When you create a liquid futures market for something, you don't just enable prudent risk management. You also enable people to pour money into bets on price movements, which can drive prices in directions that have nothing to do with actual supply and demand.

And compute futures have one additional feature that oil futures don't. Both the CME and ICE contracts are tied specifically to Nvidia chip rental prices. Not AMD, not Intel, not Google TPU, Nvidia. In the oil market, a barrel from Chevron is chemically identical to a barrel from ExxonMobil. The commodity is fungible. No single producer controls the benchmark. But Nvidia holds an estimated 80% market share in AI GPUs. An hour on an H100 is specifically Nvidia's product. If compute truly becomes the new crude oil, Jensen Huang isn't just the CEO of a chip company. He is the commodity baron of a type that has never existed in the history of energy markets because the commodity itself is his. At this rate, by 2028, Jensen won't just sell the hardware. He'll probably charge you a monthly subscription fee for the ambient heat radiated by his leather jackets.

Which brings us to the question underneath all of this. Is the demand real? But before I answer that, and trust me, the numbers are worth the wait, let me tell you about today's video sponsor. Morph from modelcode.ai is built for one of the least glamorous problems in software: the enormous legacy codebase that everyone knows needs modernizing, but nobody particularly wants to touch. Mainstream LLMs can be useful for rewriting a function or explaining a strange piece of code. But modernizing an entire system is a very different problem. Morph from modelcode.ai is the world's leading generative AI platform built specifically for legacy code modernization. It analyzes and maps the codebase, creates a modernization plan, executes and tests the work, and lets you approve every step before anything reaches production. So whether you're moving from .NET 4.5 to .NET 10, web forms to React, or breaking apart an ancient monolith, it is designed to handle the wider modernization process rather than just individual snippets. Modelcode.ai customers report reducing projects that might once have taken years down to months, weeks, or even less. You can sign up and analyze your codebase completely free with no credit card required. Morph gives you a realtime cost estimate before any work begins and you only pay if you decide to proceed. Visit modelcode.ai/m and use code house of l. You'll get 60,000 free credits every month for your first year, unlocking nearly $10,000 of real world value completely free just for using my link.

Right, the demand, because futures help or hurt depending entirely on what's underneath them. If the underlying demand for compute is genuine, broad-based, and durable the way global demand for oil is genuine, broad-based, and durable, then futures are exactly the right infrastructure to build. Great. They'll make the market more efficient, more transparent, and more resilient. But if the demand is concentrated, fragile, and particularly manufactured, then futures are a derivatives layer on top of a bubble. And that's how you turn a correction into an actual crisis.

So, let's look at what's underneath here. Nvidia's $96 billion quarter is real money, but 70% of their accounts receivable come from five customers. 60% of total revenue came from a single customer they won't name. If you ran a restaurant and 70% of your tabs were from five regulars, you wouldn't call that a thriving business. You'd call it maybe a private dining club with a very expensive lease.

I've covered the circular financing structure in detail before, so I'm going to keep this part brief. Nvidia invests in companies. Those companies use the investment to raise debt. They spend that debt buying Nvidia GPUs. Nvidia books the revenue. Just look at Sharon AI as an example. That's an Australian cloud company. Their quarterly revenue in Q2 2026 was $1.9 million. Nvidia signed a $4.9 billion deal with them. Sharon AI's own SEC filing states that the company has limited experience in delivering, implementing, and managing such contracts at scale. A company making less than $2 million a quarter just signed a deal worth nearly $5 billion with its largest investor. I mean, at their current revenue, it would take Sharon AI roughly 650 years to earn the value of that contract. For context, 650 years ago was 1376. Nobody had discovered the Americas yet. To believe this deal pays for itself in current numbers, you'd need Jesus to personally return, take a seat on the board, and just start closing enterprise sales. And even then, he would probably spend the first two quarters just trying to figure out how to put water into wine down as recurrent SaaS revenue.

Nvidia has also extended payment terms from 90 days to up to a year for certain investment grade customers. That sounds reassuring until you realize that investment grade in this context does not necessarily mean Microsoft. It could include CoreWeave which has had to raise billions in fresh debt every single year just to stay operational.

Meanwhile, Elon Musk's xAI built a 500,000 GPU cluster that was running at only 11% utilization as of April this year. That is the computing equivalent of building the world's largest power station and then leaving 9/10 of it sitting in the dark. Or to reframe it in terms Elon might appreciate in case he's watching this video: it's a bit like buying a fleet of Cybertrucks, parking 89% of them in a field, and claiming you've revolutionized logistics. All of this, the circular financing, the futures contracts, the extended payment terms, all of it is a bet, a very large, very expensive bet that the demand for AI compute will grow fast enough and lasts long enough to justify the debt.

And demand only grows if the technology is actually useful. So, before I go any further, I want to address something I see in every single comment section on every single video that I've made so far. I know a lot of people in the comments are adamant that generative AI is useless, and I understand the frustration driving that, but the research doesn't support this blanket dismissal.

A peer-reviewed Harvard and BCG study of 758 consultants found that those using AI completed tasks 25% faster with 40% higher quality output, though the gains depended heavily on the human knowing when and how to use the tool. A separate study of over 5,000 customer support agents found a 15% average productivity increase with the biggest gains going to less experienced workers. The picture is genuinely mixed and I want to do a proper deep dive on this one in a separate video because it deserves one and I don't want to derail this one. But the claim that nobody benefits from generative AI is simply inaccurate. These companies are not attracting hundreds of billions in investment because the technology does nothing. The question is not whether AI is useful. It's whether it's useful enough to justify what's being spent on it.

But despite generative AI being useful, for demand to grow, something else has to change: the way people feel about AI. And right now that sentiment is moving in the negative direction. 7 in 10 Americans oppose building a data center near them. At least 48 data center projects representing $156 billion in investment have been blocked or stalled by local resistance. Sam Altman himself admitted recently clearly people hate data centers right now at least. That's the CEO of OpenAI conceding that the public does not want the infrastructure his business model requires.

Meanwhile, the number of Amazon workers relying on food stamps has nearly tripled since 2020, while the company announced $200 billion in AI spending for 2026. I'm not going to editorialize any of that. I'm just going to leave it there. We all know what it means.

On top of all of this, Sony and Warner just sued Anthropic, alleging one of the largest and most blatant ongoing thefts of intellectual property in history, a year after Anthropic already paid $1.5 billion in the largest copyright settlement in US history. So, to summarize the demand situation, the public hates your buildings, your workers cannot afford groceries, and the music industry thinks you're a pirate. Other than that, the vibes are immaculate.

Here's where I land, and I want to be careful because I really don't think that it's simple. Debt is not inherently a bad thing. Debt buys time. Time for cost to decrease, time for demand to grow, time for the technology to find its footing, time for further research and innovation, beautiful things. Every major infrastructure buildout in human history was financed by debt. Railways, electricity grids, telecoms networks. The question with debt is always the same: is there a viable path to the revenue that services it?

Right now, both paths are being a bit obstructed. Costs are not decreasing. GPU prices are up. Debt is getting more expensive. Copyright liabilities are accumulating. And demand isn't scaling the way the financial structure requires. Public hostility is blocking infrastructure. The killer app still doesn't exist. And token spend is beginning to plateau.

The compute futures market launching in October will give us something we've never had before: a real-time market price signal of where this is heading. If the forward curve goes into contango, meaning future prices are higher than current prices, the market believes demand is growing. If it goes into backwardation, meaning future prices are lower, the market believes demand has peaked. That's the most honest signal available because it's people putting actual money behind their predictions rather than posting on X.

My prediction, and I'm being completely honest, predictions are hard, and I could definitely be wrong, is that this doesn't end in a dramatic crash, but rather it ends in a slow, boring deflation. The technology is definitely real. The productivity gains are often measurable and definitely real, but the financial structure around it is priced for a revolution that hasn't happened yet and may not happen on the timeline the debt requires.

The most useful historical parallel here is the dot-com era. The internet is of course a real thing. Amazon is around. Google as well. Email is a real thing. But the NASDAQ back then still lost 78% of its value and took 15 years to recover because the market had priced in a future that was directionally correct but wildly premature. Most of the companies riding the wave didn't survive. The ones that did became the most valuable businesses on Earth. The technology was not the problem, but the financial structure around it definitely was.

I think big tech AI companies follow the same pattern. The useful applications will likely survive. The infrastructure likely rationalizes. Several of the Neocloud companies and possibly one or two AI labs either fold or get acquired at a fraction of their peak valuations. Nvidia will likely remain a large profitable company. They make genuinely excellent chips, but at a significantly lower market cap because the market will eventually price in realistic demand rather than $30 trillion fantasies. Or of course, people could just become pro-AI as well and everything's going to be okay. Who knows? Only Oracle, actually.

The responsible position here isn't hoping that generative AI fails. Hoping a technology fails because its corporate stewards are reckless is a bit like hoping medicine fails because pharmaceutical companies overcharge. The responsible position is demanding that the financial structure be stress tested before it gets bigger. That the people building this show their working, that someone counts the chairs before the music stops.

But if you're wondering how the AI industry ended up with a public that hates it this much, I made a separate video about how the word slop broke the internet's brain and what it actually means for anyone using AI responsibly. Thank you so much for watching. I'll see you in the next one.