The $300B machine nobody knows how to value

There's a moment on every great earnings call when the company stops reporting results and starts describing a different kind of future. Jensen Huang hit that moment about twelve minutes in, and he never really came back.

NVIDIA's Q4 numbers are almost beside the point at this stage. Data center revenue at $62 billion, up 75% year-over-year. A Q1 guide implying nearly $300 billion in annualized revenue. Those figures are staggering, but they're also backward-looking. What matters is the argument Jensen is making — and whether it holds.

Here's the argument in plain English: compute is money. Every token generated is revenue for someone. Every watt of performance delivered is a dollar of value created. Once every hyperscaler CFO internalizes that framing — and they are internalizing it, fast — the debate about AI CapEx stops being a debate. It becomes circular. You spend more because the return is measurable. The return grows because you spent more.

That's not spin. It's structurally true, and it's why AWS, Azure, and Google Cloud are all guiding CapEx higher into 2027 despite every macro headwind you can name.

But two disclosures from this call deserve more attention than they're getting.

First, the $10 billion Anthropic investment. NVIDIA is no longer just the infrastructure layer for AI — it's now financially tied to the frontier model layer too. That's a different kind of moat. It means NVIDIA wins whether the compute goes up or the models get smarter. Both roads lead home.

Second, the Groq acquisition. Jensen compared it to the Mellanox acquisition in 2019. That comparison shouldn't be brushed off. Mellanox was a $7 billion deal that became a $31 billion per year networking empire within five years. If Groq does something similar for inference infrastructure, we're talking about another compounding engine bolted onto an already-compounding machine.

Colette Kress was measured throughout — flagging gaming supply constraints, acknowledging uncertainty around the Rubin ramp timeline. That's not a warning sign. That's an adult keeping the books honest while the founder bets the decade.

Now for the honest counterargument. NVIDIA trades at a premium that assumes everything goes right. Rubin needs to ramp cleanly. Groq needs to become Mellanox. Anthropic needs to matter at the model layer. Custom silicon from Google, Amazon, and Microsoft keeps losing. Regulatory scrutiny stays manageable. That's a long list of dependencies, and any one of them can disappoint.

The bear case isn't crazy. It's just losing, for now.

What NVIDIA has built is a de facto standard — the CUDA stack, the NVLink fabric, the software ecosystem — and standards are extraordinarily hard to dislodge once enterprise software is written on top of them. The switching costs are real and growing.

At this revenue velocity, with this capital return profile and this strategic optionality, the question isn't whether NVIDIA is expensive. It's whether the conventional tools for valuing semiconductor companies are even the right tools anymore.