Alphabet, Meta Reportedly in Talks on Google AI Chips

Bloomberg

Watch Alphabet, Meta Reportedly in Talks on Google AI Chips - Bloomberg on Dailymotion

Transcript

00:00So these would be TPUs, apparently, Tom, as opposed to NVIDIA's GPUs.

00:04How are they different? Are they better for the likes of inferencing, for example?

00:09Yeah, so they were developed in-house by Alphabet, the parent company of Google,

00:13about 10 years ago, and they have since been adapted, and they can continue to be adapted

00:18because, of course, Google, as well as having its own chips, also has its own large language model,

00:23Gemini, of course, so there's this feedback loop between how the model performs,

00:26then the developers can take that to the engineers of the chip and adjust the chip

00:32to add in those performance changes that they may need.

00:36In terms of whether or not it's better or not, it kind of depends what your specific aims are

00:40when you're one of these AI researchers building one of these models,

00:44but certainly it seems to be meeting the needs, certainly not just it seems of meta,

00:48at least that's the expectation according to this reporting from the information.

00:52We know it is meeting the needs of Anthropic, and they think that's going to work

00:54because Anthropic is committed to a deal with Google as well about being supplied with close

00:59to a million of these TPUs.

01:01The key standout point, Voni, obviously, is that it seems that Alphabet, with these partnerships

01:06with Anthropic and potentially this partnership with Meta, one of the biggest spenders,

01:10hyperscalar spenders on AI chips, it seems that Alphabet is being better positioned with its TPUs

01:15as being potentially a bit of a competitive threat to NVIDIA,

01:19which essentially has dominated the market for these chips.

01:21Yeah, exactly.

01:22So what would switching look like in this arena?

01:24I mean, is there an incentive on price?

01:27Do we know anything about how relative pricing, relative margins would be?

01:32Well, they can work very well in terms of both training the models, but also inference,

01:36and it seems to be that is what the emphasis is for Meta in this potential deal,

01:40is they want to be able to use these chips for inference.

01:42So when we and consumers and enterprise are actually using these chatbots

01:46and these large language models, that is the process of inference,

01:49and TPUs work very well in that sphere as well.

01:51They do have to, and they will have to prove that they can meet both in terms of energy efficiency

01:55and in terms of cost as well, of course.

01:59And look, Jensen Huang and NVIDIA has been very keen to emphasize that with each new iteration

02:03of his own chips and the whole package that comes with it,

02:07so the CUDA software, but also the networking that comes with those NVIDIA chips,

02:10that you get those gains in efficiency, both on energy and on cost as well.

02:15And the latest, of course, generation of NVIDIA chips will be coming, the Rubin will be coming out in 2026.

02:22And he says there's already strong demand for that.

02:24So this is heating up competition, but in some ways, it's also a reminder,

02:27this is just as much Alphabet's world as it is NVIDIA's world,

02:31because not only are they performing and getting demand for their chips,

02:34they're also getting demand for Gemini 3, which, of course, dropped a few days ago

02:38and has had a very good response.

02:39And you've seen that impact in terms of the competitive challenge versus OpenAI,

02:43and you've seen the read-across to what is essentially a proxy now for OpenAI,

02:48which is SoftBank listed over in Japan, which has this stake, sizable stake in OpenAI,

02:53and its shares have been under pressure.

02:54So SoftBank's become something of a proxy for the privately held, of course, OpenAI

02:59and concerns that maybe OpenAI is facing more of a competitive threat from Gemini,

03:04and particularly Gemini 3.

Category

Transcript

Be the first to comment

Recommended