YT YouTube2Text

YouTube2Text → Transcripts → Google Cloud CEO: Anthropic, TPUs, Mythos, NVIDIA and more

Google Cloud CEO: Anthropic, TPUs, Mythos, NVIDIA and more — Transcript

by Matthew Berman · 9,109 words · 1,508 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00We're not just a distributor for other
0:01people's IP. We own our own IP.
0:03>> But how do you change the hearts and the
0:05minds of the broader US demographic with
0:08regards to artificial intelligence?
0:12>> Mythos, I think it's rumored to be the
0:14first 10 trillion parameter model.
0:17>> It's better to have your own chips and
0:18demand than not having your own chips.
0:21>> Where's the next big bottleneck?
0:22>> The next big bottleneck will be largely
0:25around
0:25>> Is there some line in the sand, some
0:27benchmark you would determine Gemini is
0:29no longer safe to release publicly?
0:31>> We have more demand than we can possibly
0:33meet from all the other AI labs.
0:36>> Thomas, what keeps you up at night?
0:42All right, Thomas, thank you for joining
0:43me today. We're at the Google Cloud
0:45campus and I really appreciate your time
0:47today.
0:47>> Thanks for having me.
0:48>> I'm super excited to talk to you. I have
0:50a bunch of questions for you.
0:51>> Sounds good.
0:52>> The first question I've been thinking
0:53about so much lately is TPU capacity.
0:56When you look at the other frontier
0:58labs, like Anthropic and OpenAI, all
1:00they talk about is being uh compute
1:03constrained.
1:04>> Mhm.
1:04>> But then you see Google over here, who,
1:07you know, you have the full stack, you
1:08have your own chips,
1:10but you're not only serving your own
1:11inference, you're training, you're
1:13selling inference, uh
1:15you are also allowing some of your
1:17competitors to build on top of your own
1:19chips, you're also selling your own
1:21chips. How do you have so much capacity
1:24or how do you think about it, whereas
1:26the other frontier labs can't seem to
1:28get enough?
1:29>> If you think about what percentage of
1:31the world are we monetizing, and in some
1:33places we monetize the tokens as the
1:37token and the chip,
1:39in other places we monetize the token,
1:41somebody else's model, but using our
1:43chip underneath it. And part of the
1:45reason is we go back many, many years
1:48and we do long-term planning.
1:51And so when we saw this AI moment
1:53coming, we looked at a number of
1:56different factors to ensure we were not
1:58physically constrained. We diversified
2:00how many energy sources we had.
2:03We
2:04locked in
2:05real estate so we could build data
2:07centers. We changed how we manufacture
2:10data centers. We don't build them in
2:12construction. We shifted a lot more to
2:14manufacturing because manufacturing you
2:16can always do
2:17faster than you can do with
2:19construction. We reduced cycle time to
2:21deploy machines. All of that was stuff
2:24we've done and uh that helps us
2:26capacity-wise. And then on the silicon
2:29side,
2:30we've always
2:31worked with Nvidia as a partner, but
2:33we've also wanted to build our own
2:34silicon and we've done it for
2:36for I think it's like 11th year now or
2:3812th year.
2:39Eight generation TPU will be announced
2:41at our event.
2:42>> Yeah, we're going to talk about that.
2:43>> And so that's something we've got real
2:45art of doing over and over and over and
2:48delivering that advantage over and over
2:50and it's something we've
2:52delivered and the interesting thing is
2:54we see demand now from not just from the
2:56AI labs,
2:58but from other segments. You'll see
3:00Citadel, for example, in capital markets
3:03talking about how they're using our
3:05TPUs. You'll see Department of Energy
3:08and high-performance computing customers
3:10talk about it. So we're also seeing TPUs
3:12becoming more general purpose
3:14infrastructure, not just for AI
3:16algorithms.
3:17>> And so when you're looking at monetizing
3:19TPUs across all of the different avenues
3:21that you have to allocate your compute,
3:24how how does it compare? Can you maybe
3:27if you want to share specific numbers,
3:28great, but if you're looking at just
3:29comparing and contrasting selling a TPU
3:32versus
3:33allowing Anthropic or OpenAI to serve
3:36their inference through your
3:38infrastructure versus your own Gemini
3:40models, how do those different avenues
3:42compare?
3:43>> We balance investments across all these,
3:46you know, and we make great margins no
3:48matter which way we're selling it
3:49because we own our own IP. We're not
3:51just a distributor for other people's
3:53IP. I think that's helped us and you've
3:55seen us improve both top line and
3:57operating margin. We've also moved TPUs.
4:00Now, as for example, when you look at
4:02capital markets, one of the things we
4:05find that's very interesting is that
4:07algorithmic trading was done using
4:10numerical computation, which was largely
4:12done on traditional compute. That's been
4:15constrained by the Moore's law. The
4:16incremental improvement you see
4:18generation to generation is getting
4:20slower.
4:21And so many of the top firms have seen
4:25the huge improvements they can get by
4:26shifting to inference.
4:28Instead of doing computation using
4:31numerical techniques, if you shift to
4:32inference, you can ride the improvements
4:35you're seeing in inference time. And as
4:36they've come in, they want our machines
4:40in their venues, where it's closer to
4:43where the exchanges are, for example.
4:45>> Right.
4:45>> So, we've started taking TPUs and making
4:47available in other people's data center
4:49for some of our key customers and that's
4:51a slightly different uh business model.
4:54And we have, you know, in in macro, I
4:56would say diversification improves
5:00product because you see requirements
5:03from many places.
5:05Diversification in monetization also
5:07helps us grow. I mean,
5:10uh when we deal with supply chain
5:11vendors, for example, because we are
5:13using these chips not just for our own
5:15needs, but also offering them in market,
5:16they say, "Well, Google's demand is a
5:18sum total of a much larger pool and as a
5:21result, we get favorable terms."
5:23>> I want to stick on this point for a
5:25moment longer. If
5:27compute demand is infinite, I mean, even
5:30just for the R&D side, why not why not
5:32just hoard the compute? Why not just
5:34keep it? And to put a finer point on it,
5:37if AGI is is really the goal that all of
5:40the AI labs are heading to towards and
5:42whoever hits it kind of first and is
5:45able to scale it out wins, it seems like
5:49keeping the capacity, keeping it for
5:51yourself, keeping it for your own models
5:53is actually quite beneficial. What am I
5:54missing?
5:55>> You have to make money to fund all of
5:57this.
5:57>> Well, Google makes a ton of money.
5:59>> But you have to keep generating cash
6:01flow to do it.
6:02>> Mhm.
6:02>> And this is one more lever for us to
6:04generate sufficient cash flow. And the
6:06amount we allocate to other people is
6:08balanced always against our own needs
6:10and our own capital requirements.
6:12And you know,
6:14no matter which lab you're in,
6:16venture capital cannot fund you
6:18indefinitely.
6:19>> Yeah.
6:19>> And as compute costs grow, if you're
6:21running a loss leader business where
6:23you're losing money,
6:25uh and you're not making enough money
6:27from inference and other techniques to
6:29cover the cost of training, as that gap
6:31was gets wider,
6:33the number of sources you can go to get
6:35smaller.
6:35>> I've been talking about how Google is in
6:37such a unique position. They have the
6:39cash cows.
6:40They have the chips. They have the
6:42models.
6:43>> That's right.
6:44>> Do does does your Gemini team ever come
6:46to you and say like we don't have
6:47enough? Or I I I know I'm really stuck
6:49on this point. It's just so wild for me
6:50to hear where where these other
6:52companies are just not able to keep up.
6:54>> There's always demand for these things
6:57and it'll
6:58I think for the next 10 years there will
7:00always be more demand than supply.
7:02And that's a good place to be in if you
7:04have your own chip. If you don't, you're
7:06reselling other people's stuff. And in a
7:08capacity constrained environment, your
7:11unit economics get more expensive.
7:14And in our case, because we control the
7:16chip,
7:18>> [clears throat]
7:18>> the unit economics remain attractive.
7:20So, that is going to be an advantage for
7:22us because we own the silicon.
7:25>> So, if you look at the whole pie of your
7:27TPUs, of your compute infrastructure,
7:29can you talk a little bit about what the
7:31split is between training, inference,
7:33selling TPUs, serving inference for
7:36other labs?
7:38>> Broad brush, I think we don't talk about
7:40the details, so I'm not going to go
7:41through every element.
7:42>> Sure.
7:43>> But broad brush, if you look in macro,
7:45cloud is about half of Alphabet's
7:49capital, and it's growing because it's
7:50growing much faster, as you know. So,
7:52that's like a split, and then on our
7:55side, a significant part of our growth
7:57is coming from Gemini and our models.
7:59And so, you can use that as a rough
8:01approximation.
8:02>> Okay.
8:03You you mentioned data centers and
8:05building out data centers. Can you
8:07explain what the difference is between
8:09construction and manufacturing when
8:10you're talking about data centers?
8:12>> Yeah, it's just uh what is the unit at
8:14which you're deploying capacity.
8:17So, you could take, for example,
8:20uh a rack of machines,
8:22assemble it in a data center. You could
8:24take an entire row of machines and
8:26deploy it in a data center. The higher
8:28the grain at which you can deploy,
8:31the more you can pre-construct it,
8:33pre-test it in a central location, which
8:36means you're much faster in deployment.
8:38>> When you're
8:39planning deployment of a new data
8:42center, I I you're probably more aware
8:44than anybody there's a pretty negative
8:46sentiment about data centers in the US
8:49specifically. I think it's 20%
8:51favorability.
8:52H- h- how do you think about that?
8:55And and h- how do you think the broader
8:57AI industry can start to change the
8:59opinion, change the sentiment around
9:02artificial intelligence, and
9:03specifically deploying data centers,
9:05which gives the US a strategic
9:07advantage. It's I'm I tend to be very
9:09optimistic about AI in general. So, how
9:11how do you think about that?
9:12>> What people are really concerned about
9:14on data centers is a couple of things.
9:16Number one,
9:18you know, will energy costs in my state
9:20or my county go up?
9:22>> Right.
9:22>> Second, will there be sufficient
9:26employment in the local community in
9:28which the data center operates?
9:31And so, there's a couple of things that
9:32we're doing.
9:34First one is we're investing in
9:36behind-the-meter technology where we're
9:39not taking energy off the grid and we
9:41cross connect to the grid if the state
9:43wants it so that in case there's short
9:45supply on the grid our energy can fuel
9:48the grid. We're investing in alternate
9:50forms of energy because we think that
9:53the traditional mode of generate and
9:54distribute is not necessarily the only
9:56way that energy supply will come into
9:58the market. And so one of the things
10:01we're we're looking at is can you
10:04reduce the unit cost of energy with new
10:07forms of energy delivery
10:09created by the demand for AI but then
10:11can serve the broader market. Third, we
10:14take a lot we pay a lot of attention to
10:17making sure the unit of energy we're
10:19consuming what's called PUE that we have
10:22the best in the industry meaning
10:25if you need a 100 megawatts of compute
10:28how little incremental megawatts do you
10:31need from the energy source so you're
10:32not wasting energy? And we're by far the
10:35most efficient of that in the world.
10:37And there's a thousand things that go
10:39into that on thermodynamic exchange, how
10:42we do heating, all of that. Lastly,
10:45we're investing in the communities in
10:47which we're in.
10:48And to avoid the communities
10:51feeling like Google's deploying in one
10:53giant location, we distribute it in many
10:56places so that no individual state feels
10:59like we're becoming a big burden on
11:00their resources. And we've had a great
11:03track record. I travel to many many of
11:06our data centers and when you go to the
11:08local economy and see the kids in the
11:11school systems and you see
11:14the employees who operate our data
11:16centers who are super important to us,
11:18how much economic development we bring
11:20to those rural communities where you
11:23know, we think that's part of our
11:24responsibility.
11:25>> That's fantastic.
11:27What about the broader sentiment in not
11:30the local community cuz when you go in,
11:32you create jobs, you're investing,
11:34you're not using the electricity and
11:36driving up the prices directly. Um
11:40that's all wonderful, but how do you
11:41actually change the the hearts and the
11:44minds of the broader US demographic with
11:48regards to artificial intelligence?
11:49>> That's going to be a process. You know,
11:51and I think it's
11:53finding places where you can apply the
11:57technology in a way that's good for
11:59society rather than just uh you know,
12:03causing issues where people are worried
12:04about job displacement. I'll give you
12:06just a few examples of things.
12:09>> You'll see in our you know, keynote the
12:11company called Signal, we do not talk
12:13about it.
12:14They're a health insurer based in
12:16Germany. They're the largest health
12:17insurer in Germany. They today deploy
12:20a lot of agents built on Gemini uh
12:23Enterprise
12:24to to help their teams do work.
12:28Now, really interesting thing was
12:30there was a lot of anxiety when we
12:32started the work with them that it would
12:34mean job displacement. They have not let
12:36go anybody.
12:38And in fact, what they found is the
12:40accuracy with which and the speed with
12:43which they can answer
12:45questions from people about
12:48uh am I eligible for this treatment or
12:51not? It's cut down from in some cases 23
12:55minutes to research and answer now to
12:58less than a few seconds.
13:00And so, that's improved efficiency, it's
13:03improved quality of their customer care,
13:05and they've not let go a single job.
13:08Uh we work for example with the American
13:11Society for Clinical Oncology. They're
13:13the largest 51,000 members, every
13:16oncologist in the United States.
13:19They wanted an application where AI was
13:22helping a doctor sitting down to deal
13:24with a patient to understand standard of
13:27care guidelines, which is this person's
13:29come in
13:31they have breast cancer, what's the
13:33guideline? It turns out they're also
13:34diabetic. I can't prescribe chemo if
13:37they're diabetic of this kind. There's a
13:39lot. These rules are incredibly
13:41complicated. In many cases, they
13:43overlap.
13:44And they wanted some help on providing
13:47answers. The answers have to be 100%
13:50correct. You can have hallucination. And
13:53we've helped them. And it helps doctors
13:55take care of patients. And the feedback
13:57from their membership has been
13:59incredibly rewarding to see. So, there
14:02are lots of examples. And we always say
14:04the most important thing, Citigroup, for
14:06example,
14:07is building a wealth advisor. So, today,
14:10if you think of the average citizen,
14:12they get
14:13uh if you're a high-net-worth person,
14:16you can go to a private bank and you
14:17have wealth management professional
14:19advise you.
14:20If you're an average person who doesn't
14:21have those financial resources, you may
14:24not get high-quality advice. Citigroup
14:27is building a wealth advisor. They're
14:28going to be showing it at the event,
14:31which allows you to use our reasoning
14:35and task management capabilities in
14:37Gemini to advise people and also to help
14:40them take investment actions if they
14:42want. So, these are examples of
14:45things that society's going to find
14:47beneficial. It takes time to get the
14:50balance between the AI's going to cause
14:53massive job displacement to also hear
14:56from this side of things. And that's
14:58part of the journey we're on as a as a
15:00society.
15:01>> I I do agree. I think job displacement
15:03in particular is is something the
15:07general population in the US is
15:09extremely worried about. Let me ask you
15:11directly for your organization, Google
15:13Cloud,
15:14now that you're seeing your engineers
15:16and and other parts of your organization
15:18be more productive because of artificial
15:19intelligence, more automation, are you
15:22hiring? Are you letting go folks? Are
15:24you stable? Where where are you on that
15:27front?
15:27>> We're adding people for products and
15:29sales. We're hiring a lot of people in
15:31our go-to-market organization. We're
15:33hiring a lot of forward deployed
15:34engineers. And in places where we're
15:37building new product, we're adding
15:38capability. Like an example I will tell
15:40you
15:41here's an example of what people don't
15:43see. A long time ago we said as models
15:47became more sophisticated
15:50in understanding code
15:53and second as models learn how to use
15:55computers to do tasks
15:57there's lots of things they can do
15:58amazingly well.
16:00But one of the issues with understanding
16:02code is they can also find
16:03vulnerabilities in code.
16:04>> Mhm.
16:05>> And so enormous anxiety about
16:09cybersecurity vulnerabilities from some
16:11of the new models.
16:12>> We're going to talk about that.
16:13>> made a decision a long time ago to do
16:16three things. Number one
16:19helping improve Gemini as a way to
16:21detect issues in code and we've a lot of
16:23customers using it.
16:25Second, helping build a model that can
16:28repair code.
16:30Because if you're finding
16:31vulnerabilities very quickly, people may
16:32not be able to keep up. So can a model
16:34assist you in fixing it?
16:36And we have new capability coming for
16:38that. When we acquired this company Wiz,
16:42you'll see us showing new capability
16:44with Wiz and it's really about
16:46continuous detection.
16:48>> Mhm.
16:48>> You know, people call it continuous red
16:50teaming. So we're going to show three
16:52different types of agents. An agent that
16:55continually attacks you to make sure
16:57your vulnerabilities are being fixed and
16:59you don't get caught off guard which you
17:01couldn't do before.
17:03An agent that prioritizes
17:05the issues that are being discovered so
17:07that you can get okay, these are the
17:08ones I really need to fix.
17:10And then a third one that helps you fix
17:12>> I'm glad to hear you're you're still
17:14hiring.
17:15>> Yeah. more
17:16>> and hiring. I
17:18I There are companies out there. I think
17:20Block is the the big one. You know, Jack
17:22Dorsey put out this blog post. Block
17:24laid off half of their organization,
17:26blamed AI
17:28or pointed at AI as the reason. What do
17:31you think the difference is between like
17:33how Google sees this productivity
17:34increase and is still increasing
17:36employment versus Block who said, "No,
17:39we're actually going to transform the
17:40company. We need We need half as many
17:42people and we're going to do things
17:43better." What Where's the discrepancy
17:45there?
17:47>> Every company has demand for its
17:48products and services and each CEO makes
17:50their own decisions. We're seeing plenty
17:53of demand and so we're investing.
17:55>> Let's talk about Nvidia for a moment. Um
17:58Jensen just did a podcast with Tarkash
18:00and he talked about how
18:02uh Nvidia and their architecture is the
18:05cheapest on a per token basis overall
18:08total cost of ownership. That's because
18:11of CUDA and NVLink networking tooling
18:13delivering better tokenomics.
18:15Can Do you Do you agree with that
18:16assessment? Do you think Google is the
18:19best overall total cost of ownership?
18:22Uh and if not, how does Google catch up?
18:25>> We have a lot of customers who say we
18:27are the best total cost of ownership.
18:29>> Yeah, I guess that was the answer,
18:30right?
18:30>> I mean, the reality is if you're an AI
18:33lab, you choose the best platform.
18:35It's not just our own teams that use it.
18:39We have more demand than we can possibly
18:41meet from all the other AI labs.
18:44And so, I would just tell you that
18:47uh they would not be asking for TPU if
18:49you were much more expensive.
18:51>> Is a big factor of what makes TPU
18:53special the speed? I've noticed the
18:55Gemini family of models is very fast and
18:58as a speed maxi myself, I very much
19:01appreciate the speed. And typically when
19:04you're looking at ASICs, they're
19:06specialized, they tend to be a lot
19:08faster than the generalized GPU. Like is
19:11that a selling point for a lot of AI
19:14labs or for your own customers or are
19:16they still saying quality all day,
19:18quality, quality?
19:19>> It's a combination, I would say, three
19:21core elements because I think
19:24it's not the chip, it's the system. A
19:27TPU system, for example, uh 8T has 9,600
19:31chips. Uh 8I is, I think, 1,152.
19:36All on a single optical torus network.
19:39So, there's incredibly high bandwidth,
19:41super predictable latency across all the
19:45the the chips in a pod. And that gives
19:48you, for example, when you look at the
19:49speed with which we're able to take
19:51stuff out of
19:52the memory for processing and to put
19:55stuff back in memory, it's
19:56extraordinarily efficient.
19:58Uh just to give an example, 8T, the
20:01training chip, can fit 2 petabytes of
20:03memory in a single system.
20:052 petabytes is like 100 times
20:08the size of all the Library of Congress
20:10digitized.
20:11>> Yeah.
20:12>> Uh and because it's the super low
20:15latency network, your throughput from
20:18memory into
20:20the chips themselves are extremely fast.
20:23Third, as as if you look above that
20:25layer,
20:27uh from a programming stack point of
20:29view,
20:30uh you have a lot of tools that Google
20:33has built and given to the industry
20:35that's used for compiler optimization.
20:38For example, JAX, we've done great work
20:40with PyTorch, you know, XLA, Pathways.
20:44These are all technology that Google's
20:46built. And so, do you you put all of
20:48that together.
20:49And even if you look at inference via
20:51LLM, there's a number of technologies
20:53that are very super optimized. It's that
20:55whole stack that makes that TPU system
20:58so efficient and so powerful. And you
21:00see that measured through what we call
21:03goodput. Goodput is how much effective
21:07throughput are you seeing? We also made
21:09some decisions a long time ago, like 3-4
21:12years ago, for instance, we saw that
21:15energy, to your point earlier on energy,
21:18was going to be short supply. So, we
21:20focused on optimizing
21:22the dollars per watt or tokens per watt,
21:26and I think that's another element that
21:28you see a lot of people wanting.
21:29>> So, you you've talked a little bit about
21:31planning and the TPU remind me you said
21:3311 years?
21:34>> Yes.
21:35>> 11 years ago. It's kind of wild to see
21:38that a decision made so long ago, long
21:41in in the tech world, has bared so much
21:44fruit in the last few years. Um how how
21:47much
21:49change, how much variance in your
21:50planning happens based on what you're
21:52seeing in the market today?
21:54Uh do the decisions that you made years
21:56and years ago apply and they're just
21:59steadfast, or are you having to change
22:01things constantly?
22:03>> I would say the the the history we have
22:07across the different layers of the stack
22:10has compounded over time. When we did
22:13TensorFlow, we realized you needed a
22:15large-scale distributed programming
22:17model for training, and we built JAX,
22:19for example. That was something that was
22:22compounded on our history of learning
22:24from what people were trying with
22:26TensorFlow and needing a new distributed
22:29you know,
22:30training model, right? So, some of these
22:33accumulate over time because we learn
22:35from what we're doing from prior, and
22:38we're we're making new improvements.
22:40We're also incredibly attuned to the
22:42market, listening to customers, making
22:44decisions like
22:46people asked us, "Why did you build 8i,
22:48you know, the inference chip?"
22:50It's because it's we've seen that as you
22:54eventually, no matter how rich you are,
22:57you cannot fund training without making
22:59money on inference. And so, you have to
23:01at least cover the cost of your training
23:03from a break-even point of view over
23:05time. You can't just always depend on
23:07venture capitalist to fund you.
23:10Um and so, we said there's going to be a
23:12big demand for inference.
23:14We knew what the factors we needed to
23:17optimize for inference. And you know,
23:19frankly, the demand for the inference
23:21the 8i has been way, way more than we
23:24expected.
23:25>> So, let's talk about the eighth-gen
23:26chip. So, this is the first time where
23:28you have split out two two different
23:30chips,
23:32family, but two different chips, one for
23:34inference, one for pre-training. Um
23:36first,
23:37just confirm Ironwood
23:40was
23:41uh
23:42more built for inference.
23:44>> Ironwood was a mixed chip that's used
23:46for training and inference.
23:48>> Yeah, I think there was
23:49>> people
23:50run to for example, inference, there's a
23:54lot of diurnality to it. During for like
23:57chat,
23:58during the daytime, people wake up and
24:00they ask a bunch of questions. At night,
24:02even some people still sleep. And so, at
24:05that time, a lot of people were using a
24:07spot for inference. Like post-training,
24:09a lot of people were doing on spot
24:11instances at night.
24:13So, it's a it's a general-purpose chip.
24:16Uh with eight,
24:18T is mostly for training. Some people
24:21are considering using it for inference.
24:24And I is
24:26primarily for inference. Although,
24:28people with smaller models also use it
24:30for training.
24:31>> Based on the fact that you decided to
24:34split the chips, what does that say
24:35about where the workloads are heading?
24:37What do you see today? And then what
24:39what do you think we're going to see
24:40over the next, let's say, 5 years? Where
24:43are the major workloads going to be?
24:45>> So, we You see that in the work we're
24:48doing with Gemini as much as in the
24:50silicon. So, if you look at Gemini,
24:54we've seen sort of three phases, if you
24:57will, with the models. The first phase
25:00was where
25:02people were asking the model a set of
25:04questions,
25:06and it was answering,
25:08and you may iterate on it in a
25:10multi-turn,
25:12but it was primarily uh kind of a a
25:14search chatbot-like experience.
25:17So, our Gemini Enterprise product does
25:19provide the ability to do search and
25:21answer questions. It also added deep
25:24research to do deep analysis.
25:26>> Mhm.
25:27>> Then the second phase came along where
25:30people used to use diffusion models
25:32primarily to create content,
25:35uh like images, audio, video,
25:38and then with 25 nano banana,
25:42we added
25:43uh media in was always true, but media
25:46out became part of the main model.
25:48And so, we saw people from creative, for
25:51example, WPP,
25:53uh a variety of CPG firms using Gemini
25:57Enterprise, our enterprise AI platform,
26:00to create content.
26:02And uh and there's all kinds of content
26:04creation now going on with it.
26:06>> Mhm.
26:07>> And then, the models became
26:10really good at dealing with the
26:12abstractions of the world. And when I
26:14say abstraction of the world, if you go
26:15to a company, the model has to be hooked
26:18into a variety of different systems. It
26:20has to talk to your CRM system to ask
26:22and answer questions about customers.
26:25It may have to look at your supply chain
26:27and planning systems. And as the models
26:30became really good at dealing with
26:32those, and the ultimate abstraction is
26:34abstracting the rest of the world as a
26:36computer.
26:38Because if you can talk to a computer,
26:40the computer can talk to everything.
26:41Because all these forms of software are
26:44just abstractions for how a computer can
26:46talk to it.
26:47>> Do you do you think that is the ultimate
26:49abstraction, a model being able to
26:50control computers, computer use, browser
26:52use.
26:53>> But understanding the information that
26:55comes from those systems as well. It's
26:57not just I can talk to a computer, but I
26:59need to be able to respond to the
27:01information that computer gives me. Do
27:03you see what I mean? And so that's then
27:05led to this notion of agent. An agent is
27:09uh you know, a module, let's call it
27:11that, that you can delegate tasks to.
27:14The agent describes itself with a set of
27:16skills and it knows how to operate a set
27:18of tools, and then it can operate those
27:21including a computer and do tasks in
27:23your behalf.
27:24Now, for us, that allows people whether
27:27it's Xfinity using us for their you
27:31know, to schedule and manage all their
27:33customer care,
27:35uh Walmart using us for a variety of
27:37things in their organization from
27:39planning to scheduling, Bosch in
27:41manufacturing using us. Merck has talked
27:44about how they're using us for research
27:47and uh you know, uh patient from the
27:50drug discovery out to delivering uh to
27:53patients the whole cycle being
27:55automated.
27:57And that's the next phase of evolution.
27:59And so part of it is we're sort of
28:02co-designing as the skills of the model
28:04advance, we're able to broaden the set
28:07of things that can be done.
28:08>> Tie it back to how that informs the
28:10decision to split up the two chips
28:12between inference and
28:14training.
28:14>> So, if you look back when you asked
28:17search a quest the first phase,
28:20there were a lot more input tokens than
28:22output tokens. When you asked the model
28:24a set of questions, cuz you would ask it
28:26a very complicated long described
28:28question and it would say this is the
28:29answer.
28:31Then, when you came to
28:34uh content generation,
28:37you would give it a simple prompt,
28:38create a video that shows my dog wearing
28:41a Superman cape and driving a car. And
28:44then it would take a while to generate
28:46the output tokens. So, that generated a
28:49very different mix of tokens of the type
28:51of tokens. Multimodality
28:53was one big thing, and then the volume
28:56of output tokens grew.
28:59Then you come along to agents.
29:02It informed chip design in three or four
29:03different ways. It informed us on how
29:06long you need to maintain stuff in
29:07memory.
29:09So, for example, how what kind of KV
29:11cache would you want? Because now you're
29:14delegating something that could run for
29:166, 7, 12 hours. You don't want to be
29:19shuffling things in and out as tokens
29:21because they get expensive. So, that's
29:23one example of something.
29:25Second example.
29:27You want this talk this system to
29:30operate a computer.
29:32By the way, that computer is a
29:33traditional classical compute machine.
29:37>> [snorts]
29:37>> So, when people asked us, "How did that
29:38inform your chip effort?"
29:40We not only work with Intel. We all have
29:42our own arm chips.
29:44And so, we built that because we saw a
29:46general purpose compute usage coming
29:48from these tools. When you run for
29:51inference an agent that does many, many
29:53different steps,
29:55there are things about how you want to
29:58hold and pin objects in memory in the
30:01model so that the model runs things
30:03super efficiently because that can
30:06really optimize the cost of inference.
30:08There are many things we've done
30:09internally on how the chip can hold
30:12things in memory.
30:14Uh and then because people wanted even a
30:16practical example,
30:18people want inference in many locations
30:20cuz they want to manage latency, unlike
30:22training where you can put it in a few
30:23big locations.
30:25So, uh practical example is 8i can be
30:29run in non-water-cooled mode
30:32so that you can put it in many more
30:34locations cuz air cooling is still the
30:37primary thing in most data centers. So,
30:38there's a lot of thought that goes into
30:40these decisions. I'm just giving you
30:42three simple examples to illustrate.
30:44>> Yeah, I I think that the aging piece is
30:46really interesting because it really
30:47changes the way that those tokens are
30:50actually used in practice.
30:52Um obviously Nvidia talks a lot about
30:54extreme co-design. Google seems like
30:57they have extreme co-design on every
31:00layer.
31:01>> Yes.
31:01>> Um
31:02First talk about like with agentic usage
31:06especially if you're doing a lot of read
31:08and writes to a hard drive or you you
31:09know there there's like a lot of pieces
31:11that you need to optimize for. What's
31:12the latest thing that you've optimized
31:14for in the TPU stack and then where do
31:16you think the next big bottleneck is
31:18based on agentic usage growth?
31:20>> So we look at the whole system all the
31:22time.
31:24Uh couple examples.
31:26We're announcing two new storage
31:28solutions next week.
31:30Uh one is our managed Lustre solution.
31:34We've improved it to run 10 terabytes
31:36per second throughput.
31:38It is really designed for large-scale
31:40training.
31:40>> Mhm.
31:40>> So you can cross-connect it to a giant
31:43cluster.
31:45And because you have large data sets you
31:48can read them from
31:50the the large-scale Lustre cluster now
31:53into a large training fleet for super
31:56efficient scale. So that's one.
31:59Second thing we introduced is a
32:02uh a new
32:04uh ultra-low latency inference storage
32:08system
32:09uh a rapid storage.
32:11Uh and the idea is you can centrally
32:14keep the information you want to
32:16inference on
32:18in our cloud storage. But you can mount
32:21it close to where think of it as a
32:23forward proxy like thing wherever your
32:26inference chips are running.
32:28And so from your inference uh processor
32:32down to the storage system rapid storage
32:35to fetch for inference
32:37It's incredibly fast. Uh it's 15
32:40terabits per second.
32:41So you get ultra low latency. You want
32:44to optimize all this stuff on a common
32:46network backbone. So they're introducing
32:48a new form of networking called Virgel,
32:51which gives you ultra low connectivity
32:53speed across a giant cluster.
32:57Uh so there are many many other parts to
32:59the stack we're also co-designing uh
33:01because of agents coming in.
33:03And the idea is to give people
33:07the most efficient cost structure to run
33:09agents with the best performance and
33:10quality.
33:11>> Where's the next big bottleneck?
33:13>> The next big bottleneck will be largely
33:15around when consumers use virtual
33:18machines. You know, they Let's say I'm a
33:20consumer at home.
33:22I build an agent.
33:24And the agent is going to schedule
33:26travel for me, just hypothetically, if
33:28you're going on vacation.
33:30And you ask it to do a bunch of tasks
33:33like
33:34go look up eight travel sites, which are
33:36exposed as tools. You know, this common
33:39thing now people call MCPs or APIs.
33:43Let's go find all the travel sites.
33:46Let's say it's booking a trip to Europe
33:49or to Southeast Asia. Run that,
33:51calculate for me the total cost, and
33:53tell me my budget. Consumers cannot
33:55afford to have VMs running forever.
33:58It's extremely expensive, as you know.
34:00So people want to activate deactivate
34:03VMs whenever a task gets done.
34:07And because they these
34:09these tools need local storage,
34:12uh they these virtual machines can be
34:15over subscribed, but you can also have
34:17local disk in from which you read and
34:19write super efficiently. And so that's
34:22going to be a bottleneck because it's
34:23going to directly affect how widespread
34:26you can make this technology available.
34:28Because companies can pay for stuff.
34:30Obviously, the cheaper and more
34:32efficient they can use more stuff.
34:35But if you want to bring this to
34:37consumers,
34:39you know, for them it gets expensive
34:41very quickly. And if you want to reach
34:43everybody, you're going to have to
34:45engineer the cost structure of these
34:46things. And having again that ability to
34:49go across the layers
34:51from the agent down to Gemini, down to
34:55the storage system and the compute
34:56systems, that allows us to co-design.
34:59>> Thank you for sharing that.
35:01I want to talk about Anthropic a little
35:03bit.
35:04Anthropic is one of Google's customers.
35:06>> Yes.
35:06>> They are a unique company in a lot of
35:08ways.
35:10Claude is
35:11one of Google's biggest rivals at the
35:13same time. Yet, you are essentially
35:15their backbone for a lot of the
35:16training, a lot of the inference.
35:18How do you think about that decision?
35:21And I know we touched on it earlier, but
35:22I want want to go into more detail. How
35:24do you think about powering Anthropic's
35:27models? And then they are also competing
35:30with Google. Is that the AWS playbook
35:32where it's power everybody and we're
35:35just not going to play favorites,
35:37or is it something different?
35:39>> Google's a platform company, you know,
35:41so
35:42when you're a platform company,
35:43different parts of business compete with
35:45different players in the market. Some
35:47parts of business may supply them and
35:50some part of the business may compete
35:52with them.
35:53And so we're determined to be best in
35:55class in the models.
35:57And we're very proud of what we've done,
35:59not just with Gemini the model, but also
36:02the whole tool chain that we're bringing
36:03around Gemini with our enterprise
36:06portfolio of tools.
36:08At the same time, you know, there are
36:09customers who want, for example, our
36:12TPUs, and so Anthropic is an example of
36:15them.
36:16And it's just part of being a platform
36:18company. It's the same way that people
36:20ask us, how well do you optimize
36:23your model with Apple? Apple, for
36:25example, has signed a contract with us
36:27for the model, as you know.
36:28And so, people go, "Isn't that competing
36:30with your Android platform and
36:32ecosystem?" Yes,
36:34but that's part of being a platform
36:35company.
36:36>> Yeah.
36:37I think I'm I'm kind of stuck on the
36:39Anthropic piece because
36:41I mean, they are competing on the
36:43enterprise level where Apple is not.
36:45Yeah, I'm I'm just thinking you're you
36:47know, you're you're powering them, and
36:48then at a certain point we may get and
36:50although there's plenty of TPU capacity
36:53to go around right now, as you said, but
36:55at a certain point there might have to
36:57be a difficult decision. How do you make
36:58that decision of
37:00well, can we give the capacity to an
37:02Anthropic, or do we keep it for Gemini,
37:04do we keep it for our own research? How
37:06do you make that decision?
37:08>> We have an executive team with Sundar,
37:10and we discuss these, and as any mature
37:13company, we make those decisions.
37:14There's
37:15difficult calls every day.
37:17For instance, we have demand not just
37:20from Anthropic. So, what percentage for
37:24even if you said there's X amount for
37:26Gemini, and there's Y amount for the
37:27rest of the world, what amount do you
37:29give Anthropic versus the hundreds of
37:32other labs and other customers ask us
37:34for it? That's all complicated decisions
37:36that anybody has to make. What I'll tell
37:38you is this, it's better to have your
37:40own chips and demand than not having
37:42your own chips.
37:44>> Yeah, well said. Mythos, you you kind of
37:47hinted at a little bit. I think it's
37:49rumored to be the first 10 trillion
37:51parameter model. Is Google
37:54playing in the 10 trillion parameter
37:56model space yet? Are you Are you close
37:58to it? Where are you in that in that
38:00life cycle?
38:01>> You'll see new stuff from us on Gemini
38:05uh with announcements coming both at
38:07next and soon after. I think on the the
38:10capability of the model, we're very
38:12proud of where Gemini is.
38:14I mean, it's been state-of-the-art for a
38:15long time. We have a new version of
38:17Gemini coming very, very soon.
38:19And from all the benchmarks we've seen,
38:21we've been very confident on that as
38:23well.
38:23>> So, hypothetically, if you think about a
38:2610 trillion parameter model based on
38:28what you oversee on the TPU side, is
38:31that even a feasible size to serve in
38:33the current state of the the the world?
38:37>> We've had a capability to do
38:39disaggregated serving,
38:41uh which allows us to scale very large
38:43dense models super well, and that's been
38:45in place for a long time. And so, we're
38:48able to we would not design a model that
38:50we couldn't serve.
38:51And so, we're very confident if TPUs can
38:53serve
38:55the largest models in the world,
38:57and most importantly, our serving stack
39:01we use for disaggregated serving, by
39:04definition, is the most efficient on TPU
39:07uh in of all the model providers in the
39:08industry. So, we're very confident we
39:10can serve the largest models,
39:12particularly the largest Gemini models.
39:14>> Does this mean that we're not seeing any
39:15slowdown on the scaling pre-training
39:17side? You're you're not feeling it at
39:20all, cuz there was for a while in the
39:21industry people talking about
39:22pre-training is slowing down. Now, let's
39:24focus on RL, let's focus on the thinking
39:28time.
39:29You're not seeing that at all.
39:30>> We're not seeing that from the point of
39:32view of chip design or system design or
39:35lack of capacity or any of that.
39:38>> And then what about the underlying data?
39:39Are you seeing more
39:41uh
39:42an an effective use of synthetic data?
39:45>> We are seeing I mean, I'll give you two
39:47or three examples of things we are
39:49seeing.
39:50Historically, a lot of the data that was
39:52fed into models
39:54was unstructured data, like text, audio,
39:57video files, etc. Those continue to
40:00grow.
40:01But, the reality of those is
40:04like there are many elements in an
40:06enterprise context which make them
40:08actually really simple to deal with.
40:10When you ask a question to an agent and
40:12you have the agent respond and tell you,
40:15"Tell me the citation of where did you
40:17derive this answer from?
40:19It's easy if it's in a document because
40:21you can just show a link to that
40:22document.
40:23Now, just imagine you ask the model a
40:25question, tell me how much inventory
40:28we'll need to meet demand for this
40:30product.
40:32That is going to translate to a query
40:35on a system like an SAP system or some
40:38kind of
40:38supply chain system.
40:40That's dynamically going against set of
40:42tables.
40:44So, first being accurate on decom-
40:46decomposing that query into which table
40:49is it getting it from and showing the
40:51response like where is the citation?
40:53Like, how did you get this How do I know
40:56the answer that you gave me is correct?
40:58Is a much more complicated problem.
41:00And so, because of
41:02the work we do in enterprise, we're able
41:04to feed Gemini a lot more cycles into
41:09our uh you know, uh
41:11trajectory optimization harness with
41:14structured data. Complex things like
41:17complex fields. Have you ever seen a You
41:19know, when you talk about computer use
41:21in a browser use? If you ever see an
41:23enterprise application with a thousand
41:25fields, drop-down lists, etc.
41:29There's no consumer app that would ever
41:31have that complexity.
41:32>> Yeah.
41:33>> So, being in this space also allows us
41:35to teach
41:36our Gemini system some of those things
41:39and put it into the harness.
41:41>> I I I let's let's continue on harnesses
41:44and agentic coding in general. Um
41:47I I've been doing a lot of coding
41:48myself. There was kind of a viral tweet
41:50that went around about somebody who had
41:52a friend at Google who basically said
41:55Google isn't on the frontier of agentic
41:58coding internally.
42:00What's your take on that? How How has
42:02Google adopted agentic coding? And
42:05especially again, I have to bring up
42:06Anthropic. The rate at which they're
42:08shipping is incredible. How is Google
42:10adopting the frontier of agentic coding
42:13today?
42:14>> We have a lot of engineers using Jetski,
42:17which is our internal coding harness.
42:20And that feedback has been going
42:22directly to DeepMind in the
42:24you know, in that reinforcement loop,
42:26and it's improving quality of Gemini for
42:28coding every day.
42:29And we have a lot of people in my
42:31organization using it.
42:32>> One thing I've noticed, I'm more
42:34productive than ever.
42:36>> Mhm.
42:36>> I'm shipping so fast. I'm having so much
42:39fun doing it. I'm not reviewing every
42:41line of code.
42:42>> Mhm.
42:42>> Actually, I'm reviewing very few lines
42:44of code.
42:44>> Mhm.
42:45>> But Google can't do that. I have little
42:46toy projects. Google, you have
42:48high-stakes
42:50uh projects and services and products
42:52that you're serving.
42:53How how can you both
42:56be on the frontier of agentic coding and
42:59be produ- producing so many lines of
43:01code, but also making sure that you're
43:04maintaining that quality. Also, make
43:05sure that you are actually reviewing
43:07every single line of code that gets
43:09deployed.
43:09>> So, when we talk about software
43:12engineering productivity,
43:14we look at it slightly differently than
43:16is reported externally.
43:18So, if you work in a in a company that
43:20builds products like Google does,
43:23the reality is that there are two or
43:26three examples of things that you find
43:28that are really important.
43:30Like a senior engineer writes much more
43:32compact code than a junior engineer.
43:35So, we don't count how many lines of
43:37code as a measure
43:38because that's generally, you know, a
43:42a weaker engineer writes a lot more code
43:44to do the same task that a senior
43:45engineer does.
43:46>> It's kind of a cliché, right? Over the
43:48years, that's don't don't count lines of
43:49code. But I think now more than ever,
43:52it's just the shipping speed overall.
43:54>> so it's how much functions do we add?
43:56That's important.
43:57>> Yeah.
43:57>> The second thing, we've always had a
44:00tradition at Google that
44:02when you go to check in code, you need
44:04peer review. Typically, the peer review
44:05is done by senior managers, right? So,
44:08and they become the bottleneck. So,
44:10we've introduced and people are using
44:11Gemini and we, for example, recently in
44:14Cloud introduced it to scan for security
44:17vulnerabilities in code.
44:19So, it's not just that the tool is being
44:21used to generate code, we're also using
44:24it to inspect code.
44:25And that helps us get when the senior
44:29engineers come in for the review, a
44:31bunch of free work has been done.
44:33The third one is, for the long-term, in
44:36any real software company, the bulk of
44:39the time of the engineers where they
44:40find doing less productive work is
44:43debugging issues.
44:44So, we built a version of Gemini and one
44:48of the things we're going to show next
44:49week is,
44:50you know, what's the most complex
44:52computer in the world?
44:54The most complex computer in the world
44:56is a cloud.
44:58It makes a PC look like a toy.
45:00And so, we've taken all of our cloud and
45:03exposed it as tools to the model.
45:06And so, now we're using Gemini to
45:09troubleshoot incidents happening.
45:12And so, that's also helped us improve
45:14the speed with which people can function
45:17and in turn improve the quality of the
45:18model itself. So, there's a number of
45:20dimensions that through which we look at
45:22the issue.
45:23>> But as productivity increases and you're
45:26shipping more features more quickly, I
45:28know lines of code is not the
45:29measurement, but it's certainly an
45:32an output of this increased velocity.
45:35>> Yes.
45:36>> There comes a a point where you just
45:38cannot review every single line of code.
45:41And then, I think kind of if you think
45:43about abstracting and going beyond that,
45:46there's a point at which humans are
45:48understanding the actual code less and
45:50less over time, especially as you
45:51mentioned, if you're using AI to review
45:53the code to debug. So, if you're having
45:56AI create code, AI review code,
45:59are we losing the core understanding of
46:02of code and the functionality being
46:04deployed.
46:05>> That's a risk that we have to manage as
46:07an industry. People talk about I'm going
46:09to give you a prompt and the prompt's
46:11going to generate a block code.
46:13>> Mhm.
46:13>> And [clears throat]
46:14you don't need to understand the code
46:16because you understand the prompt. In
46:18reality, for a complex system,
46:21the prompt will not explain
46:23all the potential behavior of the
46:25system.
46:26>> Possibility.
46:26>> Right?
46:27>> Yeah.
46:27>> And so, for example, how do you deal
46:30with exceptions? Uh
46:32and so, that's uh something that I think
46:35every time you find there's one area,
46:38like if some time ago, people said
46:41you won't need all these software
46:43engineers.
46:44And then
46:46along comes the model and finds a lot of
46:49security vulnerabilities.
46:51And just when we need a ton of software
46:53engineers to work with models. Like
46:56we're introducing a version of our model
46:58that can actually fix bugs. A fix
47:00security vulnerability, specifically.
47:02>> Yeah.
47:02>> But you still need a human to use the
47:05tool and and focus on it. Sometimes the
47:08industry over-rotates and so, you say
47:11you don't need anybody just when you
47:12need it.
47:13>> Yeah.
47:13>> And so, we we take a much longer term
47:15view of things.
47:17And so, we're constantly looking at for
47:19instance,
47:21do you need a supervisor model
47:23to look at code in a different way to
47:25actually review the code? And that's why
47:27when I said we still do peer review of
47:30the code. And we're helping our senior
47:33engineers use the tool to do the
47:35reviews.
47:37And then the question comes, will the
47:38tool be self-aware enough? If it
47:41generated the code, will it find an
47:43issue with code that it generated?
47:46Because it's not self-aware of certain
47:47patterns. That's something we're looking
47:49at approaches to solve. And so,
47:52our goal has always been to make sure we
47:54have the best model is to apply it at
47:57scale. And in my team alone we have
47:59thousands of people using it every
48:00single day. I mean, if you walk just
48:02over there to the campus, you can see
48:04people like six different windows open,
48:06one in which they're coding, one in
48:08which they're compiling, one in which
48:10they're you know, deploying and testing
48:13and another one where they've got a
48:15background job running to run code
48:17review. I mean, the the whole there's a
48:19lot of people using the JetSki tool
48:22harness and
48:23it's part of just evolving how work is
48:26getting done.
48:26>> You touched on cybersecurity. Let's
48:28finish on that.
48:30Uh
48:31Anthropic decided the Mythos model was
48:34too advanced in the cybersecurity
48:37uh capabilities to release publicly, at
48:39least not yet. For Google, how do you
48:41think about that? What was your
48:42reaction? And then also, is there some
48:44line in the sand, some benchmark that
48:46you think or you would determine Gemini
48:48is no longer safe to release publicly?
48:51>> We're working through that on what would
48:53that line be.
48:55But our you know, our issue has been So,
48:58if Mythos finds a set of issues,
49:02what percentage of those issues could be
49:03found with an open-source model?
49:05And the reason I mentioned open-source
49:07model is, no matter how much you defend
49:10and you can say, "Well, I'll make sure
49:11closed-source models don't fall in
49:12adversary's hands." Open-source models
49:15for sure are going to fall into
49:16adversary's hands.
49:17>> And they're just getting better.
49:18>> And they're getting better. So, sooner
49:19or later, some part of this, it may not
49:22be all of the patterns, but some part of
49:24it can be detected.
49:26So, what should you do in response? And
49:29we are unique because
49:31we're a hyperscaler, we're a model
49:33provider, and we also have a
49:35cybersecurity organization.
49:38Both our Mandiant team and Wiz. So,
49:40we've done three practical things.
49:44If people are going to find issues using
49:47a model, you need to have a model help
49:49fix issues because they're going to find
49:51them way faster than humans can fix. So,
49:54you need a model to help fix and so
49:55we're looking at something there.
49:57Second,
49:59if they're going to find issues with
50:01models,
50:02they're going to cause use the model and
50:04computer use to launch a large-scale
50:07attack.
50:08And so to defend that,
50:10using like I'll red team my system once
50:13a month is not going to be sufficient.
50:15So, introducing agents that can do
50:18continuous red teaming
50:21and agents that can actually help fix.
50:25Like, for example, it's one thing to fix
50:27the code. It's the second thing to find
50:29all the places the old code was running
50:31and to remove it and then deploy the new
50:34code that's been patched and updated,
50:36right? So, that's the second piece.
50:39And the third piece is there's so much
50:41code out there. What do I start with?
50:43>> Yeah.
50:44>> So, again, that's another thing where we
50:46built tools to help people identify and
50:49prioritize what.
50:51>> Is it Is this an argument for or against
50:53open-source software? Not models, but
50:55software. If you're open source, your
50:58your your your all your code is out
51:00there. It is ripe for models to go look
51:02at it, find uh vulnerabilities, and
51:04exploit them.
51:06Close source,
51:07you don't have that problem. But on the
51:09other hand, open source is going to get
51:10hardened much more quickly. What is your
51:12take on that? Is that an argument for or
51:14against?
51:15>> No, we we as Google use a ton of open
51:18source and we contribute a ton of open
51:19source. We're going to help the
51:21open-source community using our tools to
51:23actually go fix these things. I'm just
51:25pointing out the reality of where things
51:27are is that adversaries are going to use
51:30the model and the first place they're
51:32going to try and scan is popular
51:34open-source libraries.
51:35>> Yeah. Yeah.
51:36>> Because that gives them the maximum
51:38surface area to try and attack.
51:40And so it's those are all elements where
51:42we think it's important to go and
51:44address and fix and we're in process
51:47with the rest of the industry.
51:49>> Thomas, last question for you. What
51:50keeps you up at night?
51:52>> We, you know, we're balancing so many
51:55things. Making sure we have to
51:58one part of your discussion, do we have
52:00right term long term plans for capital
52:03infrastructure, for data centers,
52:05networks, enough of those lovely TPUs to
52:08go around.
52:09Second, are we constantly pushing the
52:12domain problems, the important problems?
52:15Three years ago when we said we should
52:18solve the problem of as AI gets better,
52:21cyber is going to be definitely an area
52:23that's affected.
52:25And when we made the offer to buy Waze,
52:27people asked like, why would you guys be
52:29doing that? When we look at our Gemini
52:32enterprise platform, just to give you an
52:34example, between January and now,
52:37our token count has jumped from 10
52:40billion a minute
52:42to 16 billion a minute.
52:45And the number of enterprise users of
52:47Gemini enterprise has jumped by 40%
52:50sequentially. So, we're always looking
52:53at how we're solving the right problems
52:56for for customers and users,
52:58and that's always the focus for us.
53:02And as long as we keep pushing
53:04aggressively and solving those problems
53:07and staying ahead of the market when the
53:09the technology is evolving so quickly
53:11that when something happens, you got to
53:14have solutions before that occurs to the
53:16most part, and
53:18our teams have done an amazing job and
53:19we're super proud of what they've done
53:22and
53:22looking forward to the event.
53:25>> Thomas, thank you so much. Really
53:27appreciate it.

About this transcript

This page contains the full transcript of Google Cloud CEO: Anthropic, TPUs, Mythos, NVIDIA and more by Matthew Berman, generated from the public captions YouTube serves with the video. The transcript has 9,109 words across 1,508 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.