tensor processing unit

supplying these active copper cables.

And this is going to really matter as Google scales up its TPU program, its custom chips, especially with a partner named MediaTek across the next year.

So what you're going to see for Semtech is across 2026, the tail end of it, and 2027, massive growth rates from this active copper cable opportunity.

And then after that, in 2028 2029 2030 you're going to see a massive boom in opportunities from these networking speeds increasing so this is a company most people don't know what semtech is this is this is one of the most boring backwater technologies has existed but you know what also used to be a boring backwater technology memory you know what every single investor is chasing today the returns from memory, and the unique technology position it employed that allowed it to receive these returns.

maintaining an open platform effectively.

Google highlighted a mix of predefined and custom CPU shapes, NVIDIA graphics processors, its own TPUs, Axion CPUs, and you can run it on Kubernetes and dynamic workload schedulers.

And, you know, they've got tools in there for managing costs.

So...

say? So if you want to use Gemini in Australia and in Australia only, it's still 2.5%.

We don't have any of the 3.X models in Australia.

Is that because there's no Ironwood TPUs here? Probably.

Can't get TPUs into the country? US has blocked the export of them.

Um, it also, I think, helps counterbalance a lot of the, um, pull that Anthropic and OpenAI, uh, have right now, and all you have to do is see OpenAI launch something like, uh, was it Jalapeno, their own chipset, uh, to see-

And Anthropic's thinking about it, and obviously their TPUs-

I imagine Anthropic is thinking-

And everything.

They've just been on like a wild terror with hardware.

The M series of CPUs, which I guess also has this neural engine thing and the like TPU thing.

I mean, they've just been on terror.

It's crazy.

And my last question is like, what kind of like hardware platforms do you support today? And what is actually a roadmap? Is it just like for Nvidia based platforms or do you have support for AMD and like maybe any other like ASIC that you are exploring?

So, so we have engagements with the, with all this, all the GPU vendors and the TPU guys.

And as part of the platform, yes, of course we have plans to validate it across all the hardwares, hardware platforms out there.

Yeah, but it should be seamless.

platform itself

differs significantly, but if this can seamlessly work across the NVIDIA's existing GPU platforms itself, that's kind of a proof point that this is pretty much hardware agnostic.

I will say one more thing, which is that you sort of look at the process of what inference is doing on a TPU versus what inference is doing on a GPU.

We have some beautiful opportunities with TPU, I would say, and I'll leave it at that for now.

A lot of talk about Broadcom.

It's also re-accelerating its AI semiconductor revenue right now, thanks to some of those closed model AI labs, especially Google and the TPU, as well as others like OpenAI.

So lots of discussion there about NVIDIA being able to hold on to its market share leadership.

But if you compare the companies side by side, yes, Broadcom has lots of software revenue in the mix as well that dilutes the growth.

If a request or some information comes into a data center, you have these special logic chips.

Sometimes actually it's an application specific integrated circuit, custom silicon, like a TPU, decides where to route that request or that information within the data center.

It may forward that request to a GPU cluster Or it could send that request on to a storage array, which is where a company or a data center operator is housing its customers data for long term use.

Right.

I don't think it's serving any inference traffic yet, but you've got that.

You've got TPUs.

You've got companies like Etched making the silicon.

There are companies that are actually going to the far end of the extreme and burning models onto the chip.