Google releasing a new family of open source models today that it says can run on a single chip,

a GPU or Google's custom AI chip called the TPU.

Dear TBOA has more in today's Tech Check D. Interesting development here.

Hey Leslie, good morning. Yeah, this is really Google's answer to Deep Seeks breakthrough

and flexing its custom chip advantage on the way.

Now, this chart the company put out with the announcement, it shows that Gemma 3 can be run

on a single Nvidia H-100 versus the equivalent of 34 of them needed to run deepseek's R1

reasoning model, 16 for meta's Lama 3.

Now, this really underscores Google's efficiency edge, making

advanced models more accessible and cost effective for enterprises, an edge that is actually

pretty well known to those that follow Google closely, but a fact that is often overlooked by

the market. Now, the other piece of this announcement, it highlights Google's custom chip

advantage and a potential growing threat to Nvidia's dominance as the AI race shifts to an inference versus a training phase.

Hyperscalers, we know, they are developing their own in-house AI chips to bring down costs and reduce their reliance on Nvidia.

It's notable in this demo release that Google says its TPUs, its custom in-house chips,

can achieve what GPUs can.

And the worst kept secret in tech is that Anthropic widely uses Google Cloud, and thus their

TPUs for the majority of their inference.

Amazon 2 has been making inroads for its tranium chips through its own investment in Anthropic.

This week, Meta began testing its first in-house AI chip.

And, of course, OpenAI has big ambitions here.

It is pouring resources into developing its own as well.

...