Transcription
We are witnessing the most sophisticated act of financial engineering of the 21st century. Billions of dollars move in circles, valuations rise, debts vanish from balance sheets, and the bubble keeps inflating. 2008 had mortgages. 2025 has GPUs. It looks like a Ponzi scheme. It trades like a bubble. And yet, this is the system that allows more people than ever to become entrepreneurs. By the end of this video, you will understand exactly how financial engineering built the modern AI economy and see for yourself whether the bubble is about to burst. Let's dive in.
About 2 weeks ago, Bloomberg released an article about the web of circular deals shaping the AI infrastructure market. Highly recommend giving it a read. It sparked a lot of attention across all tech community and media. But when you start following the money, something strange starts to appear. The same names keep repeating. Investing in each other, buying from each other, and reporting explosive growth in the process of while they're buying from each other.
But when you zoom out, all of those lines, every connection, every dollar lead back to one company, Nvidia. Nvidia will invest up to $100 billion in OpenAI. Why? Because then OpenAI uses that money to buy Nvidia's chips. It's like if you're running a chocolate shop and you give your friend $100 and be like, "Can you come to my shop and spend this money at my store?"
OpenAI inks a $300 billion cloud deal with Oracle. Oracle desperately wants to compete with Amazon and Microsoft in cloud services. Landing OpenAI as a customer gives them instant credibility in AI. Oracle gets massive revenue and AI market presence and in turn openai gets cloud infrastructure without upfront costs.
NVIDIA buys $6.3 billion of cloud services from Coreeave. Nvidia already owns 7% of Coreeave. By buying 6.3 billion in services from a company they partially own, they're essentially paying themselves. OpenAI to pay Coreeave as much as $22.4 billion. OpenAI needs massive computing power, but it doesn't want to build its own data centers. Coreeave provides Nvidia powered cloud infrastructure to OpenAI.
OpenAI agrees to deploy billions of dollars worth of AMD chips. AMD is desperate to compete with Nvidia in AI chips. They're literally giving OpenAI 10% of their company to secure them as a customer. AMD gets a major AI customer, definitely the most well-known one on the planet, and OpenAI gets chips and equity in AMD. They become part owners of their supplier.
Nvidia invests $5 billion in Intel and plans to co-develop chips. Intel was Nvidia's enemy for 30 plus years, but now Nvidia needs them because they need their x86 architecture to expand beyond just the GPUs. And lastly, the United States takes a 15% cut of Nvidia's and AMD chip sales to China. And so instead of completely banning chip sales to China, which would hurt the US, the government takes a 15% cut of all China revenue.
So think about it. You've got billions flowing between the same handful of companies. And the deeper you go, the stranger it gets. Every deal loops back into another deal. The money is flowing in circles and each transaction is feeding the next. On paper, everybody's growing. In reality, they're passing the same money between the same players.
And here is the thing. This circular economy doesn't just create fake demand, it hides the risk. And the tool that makes it possible is called the SPV. The special purpose vehicle. A special purpose vehicle or an SPV is a subsidiary company created by a parent company for a specific limited financial objective. Its primary purpose is to protect the parent company from the liabilities of a particular project or transaction.
In the AI market, it is estimated that there is approximately $24 billion in debt. And of the $24 billion in debt, 49 is completely hidden from balance sheets. The special purpose vehicles are being heavily used in the AI infrastructure market because they serve as financial chameleons because they make the corporate debt disappear from the balance sheets of the companies that actually control and benefit from these structures.
One of the most striking examples is the XAI's $20 billion financial structure. Here's how it works. Elon Musk's AI startup called XAI is raising $20 billion to build out the next generation AI supercomputer code named Colossus 2. Main investor, you guessed it, Nvidia. The funding round combines $7.5 billion in equity and $12.5 billion in debt, but instead of direct investment, all of the money is being funneled through an SPV. Who do you think is the largest beneficiary of this deal? Nvidia. Why?
Let's break it down step by step. If you optimize for deep work and practical spending, Flexaspot gets it right. Flexaspot's whole bet is simple. Make ergonomic work setups durable and affordable. The company's building furniture and tools that help people work healthier and longer. They ship a full system of ergonomic office furniture, standing desks, converters, chairs, even cable and power accessories, so you can upgrade a home office or a small team with minimal setup. Flexis Spot leans into clean, minimalist design and sustainable materials. The desk I'm using is the E7 Pro Premium. It's wide enough to fit several monitors. The sittost stand transition is smooth and quiet, and the frame feels planted even at full height. No wobble while typing or whether you're on camera. Controls are straightforward. A small keypad with presets so you can hop between your preferred heights on autopilot. There is a built-in cable tray and a magnetic cover which means power bricks and loose cords actually disappear. You can pick the top to match your office space. They come in a variety of sizes and finishes while keeping a minimal and modern look that blends into any office. The standing desk and specifically the electric parts, frame, motor, controller, and switch, and electronics all come with a 15-year warranty. This desk really becomes a long-term asset in your office. Flexispot operates in North America, European Union, and some countries in the Asia-Pacific region. If you're looking for a quality workspace and want a stable, reliable standing desk, the E7 Pro Premium is the best value for your money. It's built for real work, not just a nice photo. Check the link below to learn more. Huge thanks to Flexispot for sponsoring this part of the video. We absolutely love this desk and we'll be using it in our future filming setups.
Step one, XAI creates a shell company or an SPV. XAI does not raise money directly. They use their SPV as a middleman that only exists to receive money from investors.
Step two, the Shell Company raises $20 billion. Like I said, 7.5 billion from investors, 12.5 billion borrowed money, meaning debt. Nvidia invests $2 billion and the rest comes from smaller investors. Of all of the invested money, Nvidia owns 2 billion. The rest comes from smaller investors.
Step number three, the Shell company buys GPUs. The SPV uses all of the $20 billion to buy Nvidia's chips. Important. It is not the XAI that owns the chips. It's the SPV.
Step number four. XAI rents the GPUs from its own SPV. They sign a 5-year lease to rent their own GPUs from the SPV and pay monthly or quarterly payments. And the Shell company essentially becomes the GPU rental business.
Step number five, everybody wins. XAI gets access to 20 billion worth of GPUs without debt on their books. Nvidia gets $20 billion in guaranteed GPU sales plus the $2 billion that they invested. That's equity. Other investors get steady rental income secured by physical GPUs and debt holders get interest payments backed by valuable hardware. And the reason why this works for all parties involved is because XAI can claim we're not raising capital, which is true because the SPV raised it, not XAI. The 12.2 billion debt does not appear on XAI's balance sheet. XAI only shows rental expenses, not the massive debt.
And now let's come back to the circular web that Bloomberg reported. Nvidia invests $2 billion. SPV uses it to buy Nvidia's chips. Nvidia gets sales revenue and also don't forget about the risk. If XAI fails, the SPV investors are stuck with depreciating GPUs. But XAI comes out clean. The debt is secured by hardware that can very much become obsolete fairly quickly. And we know this because GPU's rental rates have already dropped 75% in some markets.
Let me put this in perspective and give you an analogy that we have seen before. And the closest I can think of is actually the synthetic CDOS's from 2008. The current financial state of the AI market has created a 2008 synthetic CDOS, but instead of mortgages, we got GPUs. Just think about it for a second. Let's recall what caused the 2008 collapse. In simple terms, John bought a house that cost $200,000 and he took a mortgage from bank A. Bank A now has J's promise to pay the $200,000 plus interest. Bank A takes J's mortgage and a ton of other mortgages, packages them into a mortgage bundle one and calls it an MBS. Then bank A sells shares of this bundle to investor B for $200 million. Now investor B owns a piece of the mortgages that were in that bundle. Investment bank C takes mortgage bundle one and nine other mortgage bundles and creates a super bundle. Let's call it super bundle number one. And that super bundle contains 10,000 mortgages. And then they sell it to pension fund D for $2 billion. And now that pension fund D owns a ton of mortgages including John's. And this is where we move to the most dangerous part. The artificial cos hedge fund E creates bets on the super bundle number one without owning it and this becomes a superficial CDO. They start selling insurance on super bundle number one to five different investors and the result is that John's $200,000 house is now riding on $1.2 million in bets.
And now map this to the 2025 AI bubble. AI infrastructure companies create GPUs. They then offer finances and investments to customers to buy their GPUs. Customers create SPVS backed by the GPUs and then their SPVS issue more debt to buy more GPUs. What this essentially means is that multiple financial instruments are betting on the same GPU demand and the entire risk is placed on a piece of hardware that can very much depreciate. The whole industry is betting on the fact that we need more GPUs. That was the US loop.
But there is a second even more complex loop, the geographic one. Let's come back to the graph and see how far it spreads. Where are all these companies based? US, US, US, mostly US, France, but that's Mistrol. Makes sense. And the Netherlands, Nebus. What about Nebus? Nebus is an AIcentric cloud platform ready for intense workloads and full stack infrastructure for AI. Hold up. Wouldn't they be Nvidia's competitor? Yes. But remember the good old rule. If you can't beat them, join them. Nebus is Nvidia's European representative. And if you study them closely, you will see that there is a sophisticated legal workaround that takes advantage of the regulatory geography. When it comes to AI, especially given the current geopolitical situation, location is everything. And in many ways, the US itself has created a three- tiered system for AI chip exports that creates massive advantage for companies in the right locations.
Tier one, unlimited access. All US allies, including the European Union, UK, Canada, Australia, Israel, and Japan. Restrictions, none, can buy H100s, H200s, and the future Blackwell GBUs. The company in question, Nibbius, is headquartered in Amsterdam, which means tier one access.
Tier 2 countries, limited access. India, Brazil, UAE, Saudi Arabia, and most other countries that are not heavily sanctioned. Restrictions, listen closely. Maximum of 50,000 H100 GBUs. Put a pin on this. I'll come back to it. This is important.
And lastly, tier three countries that are severely restricted and sanctioned. And you guessed it, China, Russia, North Korea, and Iran. Restrictions only downgraded chips. H20s instead of H100s, H800s instead of H200s. Nvidia, as an American company, must comply with American laws. But they're not going to miss out on this entire chunk. So, what did they do? They partnered with Nebas.
And here is how Nebus, I don't want to say the word exploit because they're not really exploiting. They're using the system that lets them do this. So this is how Nebius and Nvidia's schema works. Nebus operates from Amsterdam, tier one location, unrestricted access to the world's most advanced AI chips. Nebus buys Nvidia's chips. This creates a legal way for Nebius to circumvent export controls and sell AI power to tier three markets, which let me remind tier three markets contain two of the largest countries on the planet. Chinese companies or any other tier three country cannot buy H100s directly, but they can use Nebia's cloud services for tier 2 countries.
And let's come back to that pin. Remember when I said that they get up to 50,000 GPUs? You may ask, is 50,000 a lot? It is enormous. 50,000 H100s is more AI computing power than most countries will ever need. The next logical question would be, so why does Nebus model work? If 50,000 units is more than a country will ever need, why buy from Nebus? Because countries hit the GPU limit, not because 50k is not enough, but because building and operating AI infra requires massive capital, expertise, and ongoing operational costs and capacity that most nations cannot or will not have.
Then Nvidia invests $700 million in Nebus in December 2024. No, it is not the largest investment, but through this partnership, Nvidia provides priority access to latest high performance GPUs and Nebus starts distributing AI compute. And coming back to the Bloomberg's diagram, this deal fits right into the circular flow. Nvidia invests, Nebius buys Nvidia's chips, Nvidia profits twice. On top of it, they create market access to restricted customers through European or Middle Eastern cloud services. They generate revenue from both chip sales and equity appreciation. They bypass export controls without violating a single law. I mean, say what you will, but this is genius.
Again, to put this in perspective and give you a more familiar analogy, imagine the US government said Americans can buy unlimited Ferrari sport cars. Europeans can also buy unlimited Ferraris. Middle Eastern countries can only buy 50 Ferraris, but Chinese can only buy Toyota Camry. And then Nebia steps in and Nebas sets up a European Ferrari rental company. They get the Ferrari to invest in their rental company. They buy Ferraris with no restrictions because they are a European company and then they start renting them to restricted customers who cannot buy directly from Ferrari. They charge premium prices because they're the only game in town and they use the rental contracts as collateral to borrow money for more Ferraris. And there you go. Should have probably included this into our billion-dollar ideas video.
The interesting thing is that Nebius, unlike many others, does not have an SPV. Nebus is actually a formidable independent competitor that Nvidia chose to invest in rather than compete against and there is proof that it works. Microsoft's 17.4 billion agreement with Nebius demonstrates the power of this arbitrage. The reason why Microsoft uses Nebus instead of building their own data centers is because Nebus guarantees faster deployment, lower capital investments, regulatory compliance, and risk transfer. And the best part is that Microsoft isn't stuck with depreciating assets.
And the circular benefit comes into play again. Microsoft pays Nebia 17.4 billion. Nebus buys Nvidia's GPUs. Nvidia profits from both chip sales and equity stake in Nebas. This is why Nvidia is in the very middle of this whole web. Nvidia benefits from its own restrictions. The three- tiered system that the US has created stimulates a regulatory arbitrage and in this arbitrage, the location matters more than the technology itself. Expert controls that appear to hurt Nvidia in fact increase demand for Nvidia's products because they create artificial scarcity and that drives premium pricing. The Nebus or the European representative model is actually spreading. G42 the UEE based where I actually applied for a job a couple of years ago serves the Middle Eastern market and there are many more NeoCloud companies that now position themselves as regional gateways. And this financially engineered system where you have a combination of expert controls, SPV financing, and circular investments create a self-reinforcing system where geography gets leveraged into massive financial returns. The bubble is indeed forming and it's growing day by day.
Which leads me to my final question. Will it burst? After studying lots of data related to this topic, it does seem like the bubble conditions are more extreme than in any previous technology cycle. And the pinnacle of all of this is that the GPU demand is largely artificial because AI has not proven to be profitable. Productivity does not equal profitability. 20% of US adults use Chad GBT, but enterprise AI ROI averages only 6%. 80% of companies report no significant bottomline impact from AI. 95% of Genaii pilots fail to reach production. Developers are not becoming extinct. Yes, the entry- level market is very difficult right now, but it's due to two factors. post-pandemic overhiring correction and because seniors have indeed become more productive. But GitHub copilot numbers confirm that the increase in productivity is very small. A ton of examples of companies with unprofitable business models, no moes, risks to be wiped out the moment the foundational company releases a native product. There has never been so much money spent on technology with unproven profits. This creates a massive paradox and the paradox is that even though it looks like a bubble. Even though parts of it trade like a Ponzi scheme, the AI infraarket will definitely deflate and balance itself out. There has never been a better time to build a technology product. These are mad times we live in. Absolutely mad.
As always, I do not mean to criticize anybody. My goal is to deeply research and truly understand what's happening. And I really hope that this video gave you a little bit more clarity in terms of what happens in the financial engineering behind the AI market. Would love to hear your feedback. Thanks for listening. Until next time. Bye.