AI's Circle of Horror

FUTURE
AI’s Circle of Horror

There’s a secret in the AI industry that is observable in plain sight, but no one is willing to discuss it openly.

It’s like Lord Voldemort in Harry Potter; people are afraid to even mention it. And that secret is a word that, among all possible words, is a seemingly harmless one:

But how can that word hold so much (dead)weight? Read on to get the answer.

In particular, we will review the following:

1. Macro estimates on the size of AI investments through the decade,

2. Financial deep dive into the agreements between OpenAI, NVIDIA, Oracle, and Coreweave, and the scary secret they share.

3. AI’s very troubled relationship with debt

4. AI’s even more troublesome relationship with, well, making money

5. And a rather doomish view coming from one of AI’s Godfathers (not Yann LeCun), arguing we are betting in the wrong direction.

The research for this week’s post will leave you very concerned. If you’re an investor in the S&P500, you’re probably not going to like this post. If you’re a public investor in NVIDIA, you’re definitely not going to like this post. And if you’re an investor in Oracle or CoreWeave, you’re going to hate it.

Let’s dive in.

The Frenzy Continues… For Now

Over the past few weeks, we have seen a significant escalation in the values and hype surrounding several AI mega-deals. But before we dive deep, a few numbers to ground ourselves.

How much are we talking about?

Now, to make sense of all this, to understand the magnitude, let me ground your perspective based on three questions:

1. What is the baseline figure that drives most large deals in AI today?

That number is the total capital cost of AI compute per gigawatt (GW) of power. That is, the value of a mega deal is based on how many GW you need and the cost per GW.

2. What is that number?

Jensen Huang recently stated that each GigaWatt of AI compute has a total capital cost of $50 to $60 billion, divided into ~60% in NVIDIA chips and gear, and the remaining 40% of other expenses (land, power, distribution, etc.).

3. And how much power is a gigawatt?

A data center requiring one gigawatt of power is equivalent to the power requirements of ~830,000 US homes, or ~2.5 million EU homes.

A single NVIDIA state-of-the-art server (each data center will have thousands of them) consumes the equivalent of a US home’s entire yearly electricity bill in three days.

And now, let’s get to it.

Numbers that don’t make sense anymore

Huge compute deals have been a thing for years now, but things have escalated recently.

The latest ‘mega deal’ is the NVIDIA – OpenAI announcement of a massive infrastructure pact: NVIDIA intends to invest up to $100 billion in OpenAI, tied to the deployment of 10 gigawatts of NVIDIA-powered AI systems.

The deal is structured incrementally, with NVIDIA making capital injections as each GW comes online (installments of ~$10 billion), and OpenAI committing to using NVIDIA’s GPUs and systems to power new data centers. 

Almost simultaneously, CoreWeave expanded its deal with OpenAI by $6.5 billion. This is the third such significant step in 2025 alone (after the $11.9 and $4 billion phases earlier in the year).

And these announcements come only weeks after Oracle and OpenAI’s significant $300 billion partnership over several years, and expected to reach 4.5 GW, starting in 2027.

But how much is OpenAI looking to build?

According to a Wall Street Journal article, they are openly discussing around 20 GW, which is about $1 trillion (using our above GPU maths). That same article mentions some executives are already envisioning up to 100 GW of demand, or up to $5 trillion in total investment, more than Japan’s or Germany’s entire GDP.

And OpenAI is far from being the entire picture, with:

• Meta openly eyeing about $600 billion investment in AI through 2028, including a $20 billion deal with Oracle,

• xAI, Elon’s AI company, potentially raising $20 billion in just a few weeks in two consecutive rounds

• Anthropic having raised in the tens of billions in total investment since inception

But how much AI CapEx in total are we talking about?

What’s the total AI CAPEX investment in the next five years?

McKinsey puts that number at close to $6.7 trillion through 2030, assuming a necessary addition of up to 125 GW of compute in those years.

However, we both know McKinsey isn’t an exceptionally reliable figure in these conversations. So, let’s cross-reference these numbers.

1. Taking McKinsey’s 125 GW value for granted and using Jensen’s back-of-the-envelope maths of ~$50 billion per GW, that is $6.2 trillion, so that side makes sense.

2. On the power side, SemiAnalysis predicted around 70 GW of AI data-center capacity by 2028 at a CAGR of 25%, giving ~110 GW by 2030.

3. EpochAI has predicted that the most extensive training runs in 2030 could draw as much as 4-16 GW of power, and that’s just training a single model, so total capacity must indeed be at leas¡st an order of magnitude above that.

Therefore, yes, most estimates align with a power demand of 100-130 GW by AI in 2030.

Which raises the question: Is the AI dream possible, or are we leaving on a dream that will soon turn into a nightmare?

To address this, we will examine the issue from two perspectives: financing and revenues. That is, we will clarify where the money is coming from (and will come from) and assess whether all this is just a disaster waiting to happen.

Follow the Money

Money can come from three places: equity, debt, or cash flows. But when you follow the money, a concerning and troubling pattern emerges.

And to answer where this money is coming from in this case, let’s examine the particular instances of three companies: OpenAI, CoreWeave, and NVIDIA, and how all three companies merge with Oracle. This may seem reductionist, but trust me, it explains what’s going on in this industry to uncanny detail.

OpenAI, The Culprit

No company illustrates the AI craze like OpenAI does.

To summarise: They have evolved from a non-profit with <$200 million in early donations into the most heavily financed AI company in history, raising tens of billions since 2019.

The key milestones are:

• 2019–2022: Microsoft’s initial $1 billion investment (mostly Azure credits) and later $13 billion stake tied OpenAI’s infrastructure to Azure, granting Microsoft profit-share rights until payback.

• 2023–2024: Massive rounds followed: $6.6 billion in late 2024 (convertible notes) plus a $4B credit facility from major banks, boosting liquidity above $10B.

• 2025: A record $40B equity round led by SoftBank (with Microsoft, NVIDIA, Thrive, etc.) at a $300B valuation. And as mentioned earlier, NVIDIA has recently pledged up to $100 billion in staged investments for compute buildouts; Oracle committed $300 billion in cloud capacity over five years; and Project Stargate announced $500 billion in AI data-center plans via consortium financing.

Overall, a potential total investment of between $500 billion and $1 trillion is expected.

And what about revenues? Isn’t the hottest private company in history printing money? Well, revenue growth is robust, with current run rates between $12 and $15 billion.

However, losses heavily outstrip revenues. In fact, OpenAI projects $115 billion in cumulative cash burn through 2029, implying tens of billions more financing needed before reaching breakeven around 2030, based on Q3-updated projections (as early as Q1, they projected 2029 would be the ‘profitability year’):

Thus, OpenAI will need substantial financing to sustain operations throughout the decade, with a projected total cash burn of approximately $115 billion through 2029, which, by the way, is $80 billion higher than the last projection a few months back.

And all this is assuming the huge rounds coming from Softbank and others (+$40 billion) to be counted?

Yes, but that still leaves out $80 billion to be financed, and that’s assuming they hit their revenue projections, which is saying something, as they project the massive value of $200 billion in revenue by the end of the decade, with revenue almost doubling most years:

If true, that would put them in the top 30 of companies with the largest revenues (in today’s values).

Bottom line, it’s safe to say numbers aren’t adding up. This explains the unique particularities of NVIDIA’s latest deal with OpenAI we’ll explain below (suggesting the party is beginning to end) and helps us understand NVIDIA’s moves, which, in my view, represent the most significant risk in capital markets today: NVIDIA’s infinite money glitch.

NVIDIA, ‘What Are You Doing?’

Financial engineering occurs when revenues fail to meet expectations. And NVIDIA is taking these games to unprecedented (and potentially dangerous) lengths, setting off a cycle of doom/infinite money glitch that could soon drag many notable AI companies along with it.

Become a paying subscriber of Full Premium package to get access to this post and other subscriber-only content.

Already a paying subscriber? Sign In.

A subscription gets you:

• • NO ADS
• • An additional insights email on Tuesdays
• • Gain access to TheWhiteBox's knowledge base to access four times more content than the free version on markets, cutting-edge research, company deep dives, AI engineering tips, & more