NVIDIA Wrote OpenAI a $100 Billion Check and Called It Revenue

NVIDIA just committed $100 billion to OpenAI, who will use it to buy NVIDIA chips. They wrote a check and they are calling it revenue.

This convoluted financial engineering eerily mirrors the telecom era circa 2000, when Cisco became the world's most valuable company by lending money to startups that used it to lease and buy Cisco equipment. When the music stopped, Cisco wrote off $900 million in bad loans and their stock crashed 88%. Twenty-five years later, it still has not recovered.

Sharp leaders learn from history and Jensen Huang is exceptionally bright, so how did we reach this point where history might repeat itself?

LLM Economics Are Fundamentally Broken

It starts with LLM economics being fundamentally broken.

OpenAI burns $1.60 for every revenue dollar they bring in. They are structurally unprofitable, sitting between expensive compute and commoditized consumer-focused products.

They pay cloud margins, chip margins, and other exorbitantly high fees to third parties for everything they have productized thus far. They are essentially a pass-through entity for compute costs: a middleman in search of an actual business model.

Without AGI, which is winner-takes-all, the next best option for OpenAI is to go completely full-stack. It is only a matter of time before we see them publicly making plays to start taking things vertical.

The CUDA Lock-In Gambit

I believe NVIDIA's deal is precisely designed to prevent this: a trojan horse to lock OpenAI into CUDA dependency before they can go vertical.

At the same time, if Google/DeepMind achieves AGI first with a breakthrough in their own silicon through their TPU-centered efforts, NVIDIA's CUDA-based empire becomes a historical footnote overnight.

This is an existential threat to NVIDIA. Their $4 trillion valuation largely rests on CUDA remaining the indispensable layer for AI development.

When Your Biggest Customers Are Your Future Competitors

However, this is not NVIDIA's only threat.

With 40 to 50% of revenue flowing from just 3 customers, NVIDIA is essentially held hostage by hyperscalers (Microsoft, Amazon, and Google). All three have the resources to build custom silicon (Maia, Trainium, TPUs) and the incentive to eliminate their largest supplier. All three will start to migrate from NVIDIA to their own chips the second they can.

When your biggest customers are also your future competitors, vendor financing is not strategy. It is a countdown timer.

An Existential Bet on Architecture

In this context, the NVIDIA x OpenAI deal is NVIDIA's gambit to keep its CUDA ecosystem at the center of the AI universe: they have to ensure AGI happens on their architecture.

This deal is a bet that OpenAI will push the underlying complexity and hardware needs that power frontier AI so aggressively that nobody else's custom silicon can keep up.

Why This Matters

As we watch this play unfold, we can look forward to continued subsidized compute while titans burn the capital. Models will become cheaper and more powerful as companies race to win market share. Context windows will grow in size, making models more useful. And maybe, just maybe, we will get AGI.

But the structural fragility underneath this capital cycle is real. The question is not whether the technology works. The question is whether the financial engineering holding it all together can survive contact with economic reality.