Yesterday I watched Nvidia drop almost 5% in a single session and, for the first time in months, the comment sections weren’t screaming “buy the dip” with quite the same religious fervor.
Something felt different this time.
The reason? Google quietly announced that its brand-new reasoning model – the one that’s suddenly topping every benchmark leaderboard – was trained from scratch on its in-house tensor processing units, designed together with Broadcom. Not a single Nvidia GPU was invited to the party.
And then, just to twist the knife, a report surfaced that Meta is seriously considering those same Google TPUs for some of its 2027 infrastructure. Cue panic in the options pits.
Why This Actually Matters More Than the Last Ten “Nvidia Killer” Headlines
Look, we’ve seen this movie before. Every six months someone claims they’ve built a better, cheaper, faster chip and Nvidia shareholders get the jitters for about 48 hours until the next quarter blows the doors off again.
But this time the plot feels different. Google isn’t some scrappy startup with a whitepaper and a dream. It’s the company that already has the third-largest cloud on the planet, a decade of real-world TPU deployments, and now the best publicly available reasoning model on the market.
In other words, the proof isn’t in a press release – it’s live right now for anyone to test.
The Two Questions Every AI Investor Is Asking Right Now
- Does this actually threaten Nvidia’s moat?
- And if custom silicon really is the future, who wins besides Google itself?
Let’s take them one at a time, without the hype and without the fear-of-missing-out nonsense.
First: Nvidia’s Moat Is Still Ridiculously Wide
Everyone loves to talk about hardware specs, but the part people keep forgetting is software. Nvidia didn’t become the default AI accelerator because its chips are magically better at math. They won because literally every researcher on earth already knows CUDA.
Rewriting millions of lines of code to move to a new stack isn’t a weekend project. It’s a multi-year, multi-hundred-million-dollar bet that your new hardware will stay better long enough to justify the pain.
Google knows this. That’s why even though Gemini 3 itself runs on TPUs, Google Cloud’s public AI offerings still run mostly on… you guessed it… Nvidia GPUs. Because that’s what paying customers demand today.
Flexibility still beats pure speed for 95% of the market.
Think of it like this: custom TPUs are like building your own race car when you own the racetrack (Google Search, YouTube, Maps, etc.). You can tune every millimeter for your exact workload. Fantastic.
But if your business is renting out race cars to other teams, most of them still want the familiar Porsche (Nvidia) they already know how to drive.
Where Custom Silicon Actually Wins – And It’s Growing Fast
That said, the scale at which these hyperscalers operate is getting absurd. When you’re spending $50 billion a year on capex, saving even 20-30% on the chip bill is real money. Real enough to justify building your own weird-looking race car if you’re never going to rent it out anyway.
And that’s exactly what Google, Amazon, Microsoft, Meta, Apple, and now even Tesla are all doing. They’re all designing custom silicon for their biggest internal workloads.
In my view, the winner here isn’t the company that kills Nvidia. The winner is the company that gets paid to design everyone else’s custom chips.
Enter Broadcom.
They co-designed Google’s TPU. They’re building Meta’s next-gen MTIA chip. They’re working with ByteDance, and rumors say they’re in deep with at least two other hyperscalers we can all guess.
Broadcom isn’t trying to replace Nvidia across the entire market. They’re happily taking the most profitable slice – the custom designs for the companies with basically infinite budgets – and letting Nvidia keep the broader, lower-margin rental market.
What This Means for OpenAI (And Every Other Nvidia-Dependent Lab)
Here’s the part that actually keeps me up at night.
Google just showed the world that you can build a frontier model without renting Nvidia GPUs by the tens of thousands. That’s a problem if your entire valuation is based on the assumption that you need to keep spending $50 billion+ per year on exactly those GPUs just to stay in the race.
I’m not saying the next model from the big independent labs suddenly becomes impossible. But the margin of error just got a lot thinner. If you’re burning cash at the current rate and you’re not clearly ahead anymore, investors start asking very uncomfortable questions.
The Sovereign AI Angle Nobody Wants to Talk About
One more thing that barely gets mentioned: countries building their own national AI stacks.
Do you think Saudi Arabia, the UAE, Singapore, or France want to be permanently dependent on an American company’s proprietary hardware stack for their most strategic compute? Of course not.
They’re already buying Nvidia today because there’s no real alternative at scale. But the moment a credible, flexible, non-locked-in option appears – even if it’s 20% slower – politics will beat pure performance every time.
Google making TPUs available outside its own cloud (still unclear if/when that happens) would instantly become a national-security conversation in half the world’s capitals.
So Where Does This Leave Investors?
Here’s my personal take after turning this over for the last 48 hours:
- Nvidia remains a core holding. The installed base advantage and software moat are still massive, and demand is nowhere near peaking.
- Broadcom just became even more attractive. They’re the “picks and shovels” play for the custom silicon gold rush without taking on Nvidia’s execution risk.
- Google itself looks wildly undervalued trading at the same forward multiple it did when people thought it was falling behind in AI twelve months ago.
- The pure-play training labs that don’t control their own silicon are suddenly on a much shorter leash.
The AI infrastructure build-out is still in the early innings. Custom silicon will take a bigger and bigger share of the highest-volume workloads. But the idea that we’re heading toward a world where Nvidia becomes irrelevant feels like wishful thinking from people who’ve been trying to call the top for three years straight.
If anything, Gemini 3 and its TPU roots are a reminder that the pie is growing faster than even the bulls imagined – and there’s room for multiple winners with very different strategies.
The only question left is which ones you want to own for the next decade.
(Full disclosure: my personal portfolio is long Nvidia, Broadcom, and Google. That probably tells you where I come down on the ultimate outcome.)
