NVIDIA's $10B Anthropic bet changes what you actually own

Jensen Huang Just Rewrote the Rules of What a Chip Company Can Be

There's a moment on every great earnings call where the CEO stops reporting and starts evangelizing. Jensen Huang hit that moment about four minutes in and never came back down.

NVIDIA's Q4 numbers are almost beside the point now. Data center revenue at $62 billion, up 75% year-over-year. A Q1 guide implying something close to $300 billion annualized. These are numbers that would define any other company's decade. For Jensen, they're table stakes for the argument he actually wants to make.

The argument goes like this: compute is revenue. Tokens are money. Performance per watt is dollars per watt. Once you accept that framing (and every hyperscaler CFO is being forced to accept it right now), the capital expenditure debate collapses. Spending $10 billion on NVIDIA chips isn't a cost. It's a factory floor. You don't debate whether to build the factory when the factory prints cash.

That's not spin. It's structurally true. And it's the reason NVIDIA's customer base doesn't negotiate the way normal customers do.

But here's what most people missed in the transcript. There were two disclosures buried deep that matter more than the headline numbers. First: the $10 billion investment in Anthropic. NVIDIA is no longer just the picks-and-shovels supplier to the AI gold rush. They're now financially tied to the frontier model layer. If Anthropic's models win, NVIDIA wins twice. Once on the infrastructure side, once as an equity holder. That's a different company than the one most investors think they own.

Second: the Groq deal. Jensen compared it to Mellanox. That comparison wasn't casual. Mellanox was a $125 million acquisition that became a $31 billion annual revenue business inside five years. If Jensen is reaching for that analogy, he's telling you exactly what he thinks Groq becomes.

CFO Colette Kress was, as always, the steady hand. She flagged gaming supply constraints plainly, acknowledged real uncertainty around the Rubin ramp timeline, and didn't let Jensen's energy pull her into overclaiming. That's not a problem. That's a healthy org.

The bear case is real, to be fair. Rubin timing is genuinely uncertain. Custom silicon from Google, Amazon, and Microsoft is getting better every year, and at some point those internal chips eat into NVIDIA's share of incremental spend. The $300 billion run rate also means comps get brutal fast. 75% growth doesn't compound forever, and anyone buying NVIDIA today is paying for a lot of future perfection.

But the honest read is this: NVIDIA is operating at a scale and velocity that has no historical precedent in semiconductors. The nearest comparison might be Cisco in 1999, which is exactly the kind of thing that keeps bulls up at night. The difference is that Cisco was selling to enterprises that were going to stop building. NVIDIA is selling to hyperscalers that are constitutionally incapable of stopping.

The chips-equal-revenue argument only breaks if AI ROI breaks. And right now, nobody betting serious money thinks that's next.

✦

Quick Takes

Three more stories worth knowing about

Tesla's CFO and CEO Are Reading From Different Scripts

Elon Musk ran Tesla's Q1 call like a man planning a decade out: $25 billion-plus in CapEx, two Optimus factories, Terafab with SpaceX, and language that was expansive in a way that felt genuine rather than rehearsed. CFO Vaibhav Taneja, to his credit, didn't follow him there. He flagged negative free cash flow through the rest of 2026, tariff exposure on energy, and one-time items propping up the auto margin beat. Tesla has 1.3 million paid FSD subscribers with falling churn, and the company has officially reframed itself as a software business that uses cars as distribution. The Street is underweighting both the Musk upside and the Taneja caution. That's a strange place to be.

Google Cloud Is No Longer Chasing Anyone

Alphabet's Q1 was the kind of quarter that forces a rethink. Cloud backlog nearly doubling to $462 billion in a single quarter, with 800% year-over-year growth in AI solutions revenue, is not a rounding error. The real story is the TPU external sales announcement. Google is now selling AI hardware at enterprise scale, which means that $462 billion backlog is attached to a larger total addressable market than any current model assumes. CFO Anat Ashkenazi said 2027 CapEx will significantly increase above an already-raised $180 to 190 billion range based on 'tangible demand signals.' That's a CFO telling you the returns are real. Take it at face value.

Microsoft Turned Its Most Expensive Partnership Into a Cost Advantage

The headline numbers from Microsoft's Q3 were clean: Azure at 40% growth, AI ARR at $37 billion growing 123% year-over-year, 20 million Microsoft 365 Copilot paid seats. Impressive. But the structural story matters more. The renegotiated OpenAI deal gives Microsoft royalty-free model IP through 2032 while OpenAI becomes a paying Azure customer. Microsoft turned a celebrated and expensive partnership into a long-duration cost advantage. Amy Hood's comment about adding a consumption meter on top of per-seat licensing across GitHub, Dynamics, and eventually M365 is the unlock that turns the $600 billion-plus RPO book into an open-ended revenue engine. This is a platform transition, and it's moving faster than anyone modeled.

✦

Overlooked Stock

SKYT

SkyWater Technology, Inc.

While every investor on earth is trying to own NVIDIA at 35 times revenue, almost nobody is looking at the domestic foundry enabling the next layer of the AI hardware stack. SkyWater Technology is a pure-play US semiconductor foundry, the only one offering full commercial services on American soil, and it sits at the intersection of three forces that are not going away: the CHIPS Act, the defense sector's desperate need for domestic rad-hard chips, and the AI buildout's demand for specialized analog and mixed-signal silicon that TSMC and Samsung weren't built to prioritize. The customer base spans aerospace, defense, biohealth, and advanced computation. Revenue is growing. The government contract pipeline is real. And the market cap remains small enough that institutional money hasn't fully arrived. This is a speculative position. Execution risk is genuine and the foundry business is capital-hungry, but the setup for a company doing what Washington has been begging someone to do for a decade is as clean as it gets.

The Contrarian Take

Amazon's Chip Business Is Worth More Than Ford

Andy Jassy dropped a number almost casually on the Q1 call: Trainium is worth $50 billion as a standalone business by Amazon's own internal assessment. Almost no analyst has updated their models to reflect this. Amazon is now, by their own reckoning, a top-three data center chip company (behind NVIDIA and Intel but ahead of almost everyone else), and the market is pricing it as a retail and cloud business with a side hobby in silicon. That mispricing matters. AWS at 28% growth on a $150 billion run rate is already remarkable. But if Trainium compounds even a fraction of what Jensen's roadmap suggests is possible in custom inference silicon, the chip business alone closes in on the entire current market cap of Ford. The risk is real. Custom silicon is genuinely hard, and AMD and Broadcom are not standing still. But Jassy was specific, not vague, and specific executives with specific numbers usually have something to show you.