Semiwiki on LinkedIn Semiwiki on Facebook Semiwiki on Twitter Register / Log in 
Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/threads/meta-in-talks-to-spend-billions-on-googles-chips-the-information-reports.24092/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2030770
            [XFI] => 1060170
        )

    [wordpress] => /var/www/html
)

Meta in talks to spend billions on Google's chips, The Information reports

Daniel Nenni

Daniel Nenni

Admin
Staff member
1764087543419.png


Meta is reportedly in deep talks with Google to commit billions of dollars toward purchasing Google’s AI chips, specifically tensor processing units (TPUs), for use in its data centers starting in 2027. In parallel, Meta may begin renting TPUs from Google Cloud as early as next year, according to the report.

This would mark a major shift for Google, which has historically confined TPUs to its own infrastructure; by opening them up to external companies, Google could position itself as a serious competitor to Nvidia in the AI chip market. Google Cloud executives reportedly see the potential to capture up to 10% of Nvidia’s revenue through this strategy.

For Meta, this isn’t just about cost — the TPUs are being pitched as a more secure and potentially cheaper alternative to Nvidia’s GPUs, especially as demand surges and supply tightens. Yahoo While the deal remains unverified and all three parties have declined comment, the move could significantly reshape the landscape of AI computing infrastructure.

www.reuters.com

Meta in talks to spend billions on Google's chips, The Information reports

Meta Platforms is in talks with Google to spend billions of dollars on the Alphabet -owned company's chips for use in its data centers starting from 2027, The Information reported, a move that would cast Google as a serious rival to semiconductor giant Nvidia .
www.reuters.com www.reuters.com
 
I have heard this as well. Renting chips is an interesting business model. I can't wait to see the numbers behind this. Exciting times in the semiconductor industry, absolutely.
 
I think "renting chips" just means Meta gets a special deal on using TPUs in Google Cloud, which Google already supports.


IMO, there's no way Google offers TPUs to generally compete with Nvidia GPUs. Everything about TPU hardware, including the interconnect networks, are proprietary. Nonetheless, if Meta is thinking of using Google as a systems-level source for TPU-based nodes and racks, that could be very interesting for both companies. I strongly suspect Google systems would be substantially cheaper than Nvidia-based comparable systems, from any vendor, and also more power-efficient.
 
blueone said:
I think "renting chips" just means Meta gets a special deal on using TPUs in Google Cloud, which Google already supports.
IMO, there's no way Google offers TPUs to generally compete with Nvidia GPUs. Everything about TPU hardware, including the interconnect networks, are proprietary. Nonetheless, if Meta is thinking of using Google as a systems-level source for TPU-based nodes and racks, that could be very interesting for both companies. I strongly suspect Google systems would be substantially cheaper than Nvidia-based comparable systems, from any vendor, and also more power-efficient.
Click to expand...

Are you saying Reuters is wrong? :ROFLMAO: Google selling TPUs would be a really big deal. I wonder if Amazon and the other hyperscalers will follow suit? Or is this just a jab at Nvidia to get more competitive pricing?

Here is the non-paywall version:

 
Daniel Nenni said:
Are you saying Reuters is wrong? :ROFLMAO:
Click to expand...
Who me? :rolleyes:
Daniel Nenni said:
Google selling TPUs would be a really big deal. I wonder if Amazon and the other hyperscalers will follow suit? Or is this just a jab at Nvidia to get more competitive pricing?
Click to expand...
I think this falls into the NFW category. Take a look at the TPU Inter-Chip Interconnect (ICI):

cloud.google.com

Using Cloud TPU Multislice to scale AI workloads | Google Cloud Blog

Learn how new Cloud TPU Multislice functionality can enable 2x higher scale than alternate accelerators, with 2x higher performance/dollar and near-linear scaling out-of-the-box.
cloud.google.com cloud.google.com

1764094062912.png


And that Data-Center Network at the top of the diagram? That's Google's all-optical and proprietary Jupiter network:

cloud.google.com

The evolution of Google’s Jupiter data center network | Google Cloud Blog

Thanks to optical circuit switching (OCS) and wave division multiplexing (WDM) in the Jupiter data center network, Google enjoys a host of benefits.
cloud.google.com cloud.google.com

IMO, there's no way Google sells that network to anyone else, and I can't imagine what field support would look like. So, yes, I think Reuters is wrong. :)

Edit - it just occurred to me that the TPU sales model is in the cloud, or very likely looks a lot like the Cerebras model - no chip selling, just systems. I think Google's solution and the Cerebras solution are equally proprietary.
 
blueone said:
Who me? :rolleyes:

I think this falls into the NFW category. Take a look at the TPU Inter-Chip Interconnect (ICI):

cloud.google.com

Using Cloud TPU Multislice to scale AI workloads | Google Cloud Blog

Learn how new Cloud TPU Multislice functionality can enable 2x higher scale than alternate accelerators, with 2x higher performance/dollar and near-linear scaling out-of-the-box.
cloud.google.com cloud.google.com

View attachment 3878

And that Data-Center Network at the top of the diagram? That's Google's all-optical and proprietary Jupiter network:

cloud.google.com

The evolution of Google’s Jupiter data center network | Google Cloud Blog

Thanks to optical circuit switching (OCS) and wave division multiplexing (WDM) in the Jupiter data center network, Google enjoys a host of benefits.
cloud.google.com cloud.google.com

IMO, there's no way Google sells that network to anyone else, and I can't imagine what field support would look like. So, yes, I think Reuters is wrong. :)

Edit - it just occurred to me that the TPU sales model is in the cloud, or very likely looks a lot like the Cerebras model - no chip selling, just systems. I think Google's solution and the Cerebras solution are equally proprietary.
Click to expand...
agree, that's why the most likely model is to rent the existing infrastructure for external work load.

As for Google's approach, they are for now the only one deploying OCS in production. Rumors are all hyperscalers are exploring it. this could really give them key leverage in all the different scaling ranges (up/out/across) into the next couple generations.
 
You must log in or register to reply here.
Back
Top