Nothing to see here, folks. Nvidia’s still light years ahead of the competition and Jensen Huang’s anyway “delighted” by Google’s great leap forward.
That was the message from the world’s most important company, which on Tuesday saw its shares trade down to a two-month low on concerns that Alphabet’s becoming the market’s go-to AI bet amid excitement around Gemini 3 and a burgeoning turf war in the chip space.
“Google is talking to cloud customers about letting them use Google AI chips in their data centers,” The Information reported, describing an “escalating rivalry” between Huang and Sundar Pichai.
Long story short, Google’s not satisfied to challenge OpenAI and Anthropic on the model front. They’re coming after Nvidia for chip supremacy too.
Meta’s reportedly considering Alphabet’s TPUs (on which Gemini 3 was trained) for its data centers in 2027, and could pay Google Cloud to rent the chip as early as next year, the same Information article said.
That’s ominous for Nvidia, to say nothing of AMD which was running a distant second already.
This isn’t a zero-sum game, but the market’s trading it that way for now. Nvidia’s down 13% in November and AMD 22% (albeit after a historic gain in October) while Alphabet’s up nearly 15%.
Maybe (probably, surely) Nvidia’s GPUs will remain the industry standard for a while longer but Alphabet with a superior model and a capable chip is going to be a formidable beast, particularly given that, as I put it Monday, Google basically is the internet.
If you ask me, it says a lot that Nvidia felt compelled to issue a statement about Google on Tuesday. “They’ve made great advances” it read, in the course of reminding investors that Nvidia is a Google supplier.
Nvidia “is a generation ahead of the industry,” the company went on, adding that Huang has “the only platform that runs every AI model” and his chips “offer greater performance, versatility and fungibility” than Google’s TPUs, “which are designed for specific AI frameworks or functions.”
It was short and colloquial as corporate statements go, but it nevertheless felt like a lot of explaining from a company that’s as confident in the perennial, categorical supremacy of its product offering as Huang claims to be.
Think about it this way: Mercedes can tweak and improve the Maybach variant of the S-Class all they want, but you’re never going to see Goodwood issue a statement explaining why a Rolls-Royce Phantom is a superior automobile. If you’re comparing two cars and one of those cars is a Phantom, it doesn’t matter what the other one is. A Maybach might as well be a Camry for the purposes of that comparison.
I understand there’s a difference in the two chips — i.e., that Nvidia’s GPUs are a different animal from Google’s TPUs — but the same general principle applies. If Huang’s platform was that much better — i.e., so much better that comparisons are asinine or even nonsensical — then Nvidia wouldn’t be saying anything at all, let alone in an off-the-cuff tweet, during the middle of a random Tuesday.
Want to see something remarkable? Have a look at this:
Now that’s a rally. Alphabet’s up eight months in a row, and the last three months’ gains are all 14% or better.
Google issued a statement of its own on Tuesday. “Google Cloud is experiencing accelerating demand for both our custom TPUs and Nvidia GPUs,” the company said, politely. “We are committed to supporting both, as we have for years.”
Bloomberg quoted Counterpoint’s Neil Shah. Google was the “sleeping [AI] giant,” he said. “Now it’s fully awake.”
Fee-fi-fo-fum.
If I both understand and recall correctly (doubtful?), GPUs were developed for the high precision calculations required for graphics, while LLMs use low precision calculations, so GPUs were not actually the optimal architecture for LLMs and maybe not for other forms of AI. GOOG’s TPU architecture was developed more specifically for AI, in which GOOG was dominant back when AI was called “machine learning”, i.e. pre-LLMs. GOOG claims TPU are much more efficient than GPU, and perhaps there is a fundamental architectural reason. Of course, all that may have changed by now. I also recall recently reading that a Chinese company developed an even more efficient AI chip using the low-precision aspect of LLMs. My understating of all this may be wrong, one of Dr H’s more technically knowledgeable readers will hopefully set us straight.
The AI game is a fast moving one with the rules, players and equipment changing rapidly. In one year what will be the ‘next’ chip that everyone needs. Impossible to say. What if Meta’s former AI chief Yann LeCun is right that AGI will need a new visual approach and not LLMs. And that’s only one of a hundred things that could change in a year.
I’m still wondering how AI is going to get me, the consumer, to spend more money. The savings I’ve read about all seem to be based on productivity savings for business. I have friends who use Comet to search out items online and make purchases, but it’s not for anything they couldn’t find themselves.
Gemini works pretty well, it’s right there in the Google Search window, and it is free to use (for now). Unless you have specific needs or uses, why fiddle around with anything else?
Software engineers want ai in their development environment. Lawyers need something in their environment.
Google NotebookLM
The thing is the Phantom is curtailed to a very small user base, while that of an S-Class is much wider. Why drive a Phantom when an S-Class is more than enough? Hell, an E-Class would do if you ask China.
I loved my E350. Wonderful car.
Hardware always loses in the end excepting two scenarios: 1) a luxury moat (e.g. Apple), 2) a software moat (also Apple). Where there’s margin, there’s competition, and margins erode like beaches on the Outer Banks. Never forget that IBM once held the title of Biggest Market Cap, planet Earth division. Then they outsourced their software to some nerds from Redmond. Presumably that saved them some money. Improved margins. Definitely worth it.
(I know Nvidia has software offerings. I just don’t know that it constitutes a “moat.” The vikings are coming. They promise to not be evil, nor to split infinitives.)
I was so proud of that last sentence, but just thought of a better wording; thus, amended for your edification:
They promise to neither be evil, nor split infinitives.
Much better. Syntax matters folks. When it’s good, you can get away with incomplete sentences.