Google working on initiative to challenge Nvidia's software advantage

Sarah (Reporter) 00:02.540

google now working on a new initiative that would challenge one of nvidia 's biggest advantages in AI it's proprietary software deirdre bosa has more in today's tech check segment morning deirdre hey

Deirdre Bosa (TechCheck Anchor) 00:13.220

good morning sarah so google 's chip challenge to nvidia 's moving from hardware to software the main knock against those custom AI chips tpus is that they're hard to adopt nvidia 's advantage is its ecosystem so that's its chips and its cuda software if google can crack the

Deirdre Bosa (TechCheck Anchor) 00:28.420

software and the adoption piece this becomes a very test for nvidia 's moat so google 's tpu 's they started as an internal tool built to run google 's own AI as efficiently as possible and optimized around google 's own software now that worked really well inside of google but

Deirdre Bosa (TechCheck Anchor) 00:45.910

as the company started selling these chips to outside customers that created friction tpus were harder to drop in because they required extra engineering work nvidia by contrast has cuda that's a software platform that is deeply embedded in how its AI is built and run making

Deirdre Bosa (TechCheck Anchor) 01:02.710

nvidia the easier default for pretty much anyone building now reuters first reported and a source familiar confirms to me that google is now trying to fix that with its own initiative torch tpu which is essentially trying to lower the software and engineering barriers that have

Deirdre Bosa (TechCheck Anchor) 01:17.030

held adoption back so this doesn't mean that nvidia is about to be displaced but it does change the stakes the AI race it's not just about peak performance anymore it's about margins and scale and that is where the real pressure is starting to build so there's google tpus but

Deirdre Bosa (TechCheck Anchor) 01:30.550

this is part of a bigger trend as well amazon is pushing its own custom chips meta is looking for ways to lower inference costs every hyperscaler they want more leverage in GPU pricing negotiations that has long been the case so even if nvidia keeps most of the market here the

Deirdre Bosa (TechCheck Anchor) 01:45.710

pricing environment around AI infrastructure we could start to see that shift and the big question for investors is what this chip neutrality means for the AI trade and that is just simply the idea that companies are less locked into one vendor and they can move workloads to

Deirdre Bosa (TechCheck Anchor) 02:00.950

where compute is cheapest or most available be that a tpu so in the short term guys it creates uncertainty but here's the thing in the longer term it may be what actually makes the build out more sustainable

Carl (Reporter) 02:14.530

although DI wonder what you make of the idea that we've seen sort of the logistical gating issues frictions on the build out the financing frictions on the build out and now the memory price frictions on the build out that's those are three big gating issues

Deirdre Bosa (TechCheck Anchor) 02:30.370

i mean this has been a supply and demand issue for the past few years right there's so much compute that needs to be built out and nvidia GPU 's have really been you know pretty much the only game in town so if you're looking at a tpu get the software piece right not just the

Deirdre Bosa (TechCheck Anchor) 02:45.150

hardware piece and by the way this could be true of google it could be true of amazon that's really been nvidia 's key mode is having that cuda software so if google gets close that would be a huge win because it would mean that those custom AI chips could work outside of google

Deirdre Bosa (TechCheck Anchor) 02:59.550

cloud which has really been the main knock that i hear from folks in the industry that could really alleviate some of those supply constraints of course carl you bring up a good point is the demand still going to be there and that is like you know the trillion dollar question in

Deirdre Bosa (TechCheck Anchor) 03:13.110

all of it

Sarah (Reporter) 03:14.190

multi trillion dollar now yeah thank you deirdre