Google Plans New ‘Frozen’ AI Chip, Preventing Scams in the AI Era, 160 Software Startups For Sale

**Akash Pasricha** (0:13)
Welcome, everyone, to The Information's TITV. My name is Akash Pasricha. It is Monday, July 20th. Today on the show, Google is developing a new server chip that would allow its Gemini AI models to run more efficiently. We'll talk to The Information's Google reporter who broke that story. We'll also look at how AI is changing the cybersecurity sector, for better or for worse. We'll talk to the CEO and co-founder of Savi Security. Plus, last week, The Information published a list of more than 160 enterprise software start-ups that could be acquired in the near future. We'll speak with the reporter who put together that list. It's going to be a great show, so let's get right on into it.
Google is working on a new server chip called the Frozen chip complimenting its TPU offering. I want to bring on Erin Woo, our opening-eyed Google reporter who broke that story with our Asia correspondent, Chana Liu.
Erin, welcome to the show. It's great to have you back.

**Erin Woo** (1:08)
Hey, thanks for having me.

**Akash Pasricha** (1:10)
Tell me about this new chip that Google is working on.

**Erin Woo** (1:13)
Yeah. Google is working on a new chip. It's informally dubbed Frozen V2 inside the company.
Essentially, the idea is that it could be way more efficient, even than the TPUs at the time of launch. They're projecting it could be six to 10 times more efficient than the TPUs at the time of launch, and that's because it works slightly differently by etching the model itself into the chip. Fewer decisions have to be made at the time that the model is running.

**Akash Pasricha** (1:39)
Etching it into the chip, so why is it called Frozen?

**Erin Woo** (1:44)
The name Frozen has to do with the idea that the model is frozen into the chip to some degree.
Essentially, the way that a regular chip works, it's generalizable, which means that it can do a lot of different things, but that means that there are a lot of different decision points that the chip has to run. This chip makes things more efficient by combining some, sorry, by pre-setting some of that, essentially, and so it doesn't have to make all those decision points.

**Akash Pasricha** (2:11)
So how would it be different than from the TPU?

**Erin Woo** (2:14)
Yeah. So the TPU is more of the generalizable kind of chip versus these frozen chips, which would specifically run this model architecture.

**Akash Pasricha** (2:23)
I see. So if this frozen chip, I guess it's only meant to run Gemini models, sounds like. So is it only meant to be used internally by Google then? Is that the idea?

**Erin Woo** (2:38)
I mean, so a lot of Google Cloud customers also run Gemini models, and so this is something that could theoretically also be run by external customers. But this is still something that's pretty early, so it's not expected to launch until 2028 at the earliest. So I think a lot of these strategy questions might still be in the process of being worked out.

**Akash Pasricha** (3:01)
Right.
I mean, the reason I'm asking this question is because we, of course, know that the TPU was mostly used internally, but now they're at the point where they're really trying to outfit customer data centers, if I recall correctly, with these TPUs. So your point is well taken that it's not just people at Google running Gemini, it's all sorts of customers. So that could very much be the strategy. How does this compare with other inference folks' chips?
We've had the CEO of Salmonova on the show, OpenAI is working on their own chip. How would this compare to that?

**Erin Woo** (3:36)
Yeah. So a lot of people are trying to do the same thing, which is essentially driving down the cost of inference because everyone's in this big compute crunch. So Google's chip works a little bit differently because of this Frozen design. It's actually somewhat similar to what this Canadian chip startup TALIS is doing. So they're also trying to specifically etch a model into the chip so that it is more efficient.

**Akash Pasricha** (4:04)
Has Google been compute constrained? Is this a way to get at that problem?

**Erin Woo** (4:09)
Yeah. So they talk about that constantly on earnings calls, the idea that their Cloud division is compute constrained. Google reports earnings this week. I would expect them to also talk about that this week.

**Akash Pasricha** (4:20)
And tell me, TSMC capacity, that I imagine, is also top of mind here for this new chip?

**Erin Woo** (4:28)
Yeah. So Google is going to have to find fabrication space for this chip. It's not totally clear to me that it's going to be like a one-to-one exchange in the sense that like, in order to make these chips that like seals from TSMC space for the TPUs, like that's definitely possible. But it's also been suggested to me that this is enough of a different fabrication process that might use different space. So it might not be as easy or some game for them.