Building Real-World Infrastructure for AI: Google, Cisco & a16z

Source material

Transcript

The good news is infrastructure is sexy again.

So that's cool.

This is like the combination of the build out of the internet the space race And the Manhattan Project all put into one where there's a geopolitical implication of it There's an economic implication with national security implication and then there's just a speed implication.

That's pretty profound I mean, I think it's easy to say I've seen nothing like this.

I'm fairly certain no one's seen anything like this the internet In the late 90s early 2000s was big and we felt like oh my gosh can't believe that built out the rate This makes it I mean 10x is an understatement.

It's 100x what the internet was The AI boom isn't just changing software.

It's transforming the physical infrastructure that runs it today You'll hear a conversation with a mean vadot from Google G2 Patel from Cisco and ragu ragu ron from a 16z on what it takes to build a real-world systems behind large-scale AI From chips and power to data centers and networking They discussed the scale of the current build out the new constraints will compute power and interconnect and how specialization in hardware and architecture Is reshaping both the industry and global geopolitics?

It's a grounded look at how infrastructure itself is being reinvented for the AI era and what comes next Let's get into it What better time and place to talk infrastructure So we were back in the green room and just as The first question was getting answered.

I got cut off so this could be an entire repeat for all I know Anyway, let's go right the first question is similar so both of you freshly welcome and thank you for being here And hope you'll have a great day and a half as well Both of you be in the industry for a while and both of you have lived through many infrastructure cycles, right?

So have you seen anything like this cycle from your vantage point not from an investor vantage point but from your internal Vantage point where you are responsible for building things and planning for things and so on Anyone of you very much talk you want to start a meeting?

I mean, I think it's easy to say I've seen nothing like this I'm fairly certain no one's seen anything like this the internet in the late 90s early 2000s was big and we felt like oh my gosh Can't believe the built out the rate this makes it.

I mean 10x is an understatement It's 100x what the internet was I think the upside is as big as the internet was same thing 10x and 100x Yeah, nothing like it.

Yeah, I'd agree I don't think there's any priors to this size the speed and scale I'd say the good news is infrastructure sexy again So that's kind of cool.

It was a long time or wasn't sexy.

The thing I would say that's really interesting is this is like the combination of the build out of the internet the space race and The Manhattan Project all put into one where there's a geopolitical implication of it.

There's an economic implication There's a national security implication and then there's a just a speed implication.

That's pretty profound So yeah, none of us have ever seen it at this size and scale on the other hand I think we are grossly underestimating like there's the most common question I asked right now is is there a bubble?

I think we're grossly underestimating the build out.

I think there's going to be much more needed than what we are putting the projections towards So that's the following question is where are we do you think of the capex spin cycle, but more importantly?

What are the signals that you guys use internally right in your thinking?

I mean you have to plan data centers, whatever four or five years in advance You have to buy nuclear reactors and whatnot so how do you think about the demand signals as well as your technology signals and G to the same thing for you, but from the point of view of enterprise and new clouds, etc I mean we're early in the cycle is what I would say certainly relative to the demand that we're saying our internal users are We've been building with TPUs for 10 years.

So we have now seven generations in production for internal and external use our Seven and eight year old TPUs have a hundred percent utilization That just shows what the demand is.

Everyone will of course prefer to be on the latest generation But whatever they can get so this tells me that the demand is tremendous, but also Who we're turning away and the use cases that we're turning away.

It's not like oh, yeah, that's kind of cool.

It's oh my gosh We're actually not going to invest in this and there's no option because that's where we are on the list Same with many of you in the room We're working with many even the room and many of yours are telling me directly and thank you.

We need more earlier Now the challenge here though is as you said We're limited by power.

We're limited by transforming land.

We're limited by permitting and we're limited by Backup delivery of lots of things in the supply chain.

So one worry I have is that the Supply isn't actually going to catch up to the demand as quickly as we'd all like I heard the previous session some of the discussions of the Trillions of dollars that we're gonna be spending which I think is accurate I'm not sure that we're gonna be able to cash all those checks You know was it literally you all have so many you can't spend it all as fast as you want I think that's going to extend for three four or five years Wow How do you deal with the depreciation cycles that are involved there?

The demand curve and the depreciation cycle curves we match up well fortunately by just in time but the nice thing is just in time for the hardware the depreciation cycle for the space power is more like Somewhere between 25 and 40 years.

So we have benefits there.

I Think if you think of on the networking side and you look at both enterprise and the hyperscalers as well as neoclouds I think the story is quite different.

So the enterprise is pretty nascent and it's built out of true infrastructure I just don't think that The data centers like if you assume that hundred percent of the data centers at some point in time will need to get re-react and You will need a very different level of power requirement Per act that's going to be there compared to what used to be there in the traditional data centers I just don't think that the enterprises are far enough along Maybe the few enterprises that are at super high scale might be there but I don't think the enterprises are far enough along hyperscalers and neoclouds is a completely different story and to a means point on this notion of Scarcity of power compute and network being the three big kind of constraints in this thing.

I would say right now that Because there's not enough power singularly in one location data centers are being built where the power is available rather than power being brought to where the data centers are and that's why you're seeing a lot of projects that are being built out all throughout the world and the other point though is the lion's share of the constraints that we're gonna have I think are gonna be sustainable for a long period of time and as you have data centers that are being built farther and farther apart One there's gonna be a huge demand for scale up Networking so that you can have a rack that gets more and more networking for scale up The second is you're gonna have a lot of demand for scale out But you have multiple racks and clusters that need to get connected together But we just launched a new piece of silicon as well as a new chip and a system for scale across networking where you might have two data centers that act as a logical data center that could be up to eight nine hundred kilometers apart and You will see that just because there's not going to be enough concentration of power in a single location So you'll just have to have different architectures that get built out Actually that brings us to the next topic that I want to discuss the future of systems and networking and so on and so forth so Google bought the first on at least large scale scalar commodity servers and production for the web revolution and Now Nvidia is bringing back the mainframe in a different form.

So what do you think happens next?

I mean this is this the new style of coherent cluster wide computing that we need and There's gonna be shared memory and all sorts of things or do you think the pattern changes again?

I don't think we're quite too back to mainframes in that it is still the case that people are running on Scale out architectures across these pools.

In other words, whether you have GPUs or TPUs, you're not necessarily saying hey That's my GPU supercomputer.

You're saying I've got 16,384 GPUs.

Yeah, and maybe I'm going to go grab some subset now I've got a uniform all tall connectivity in many cases, which is fantastic same with TPUs It's not like I say I have a 9,000 chip pod and I have to make my job fit on that Maybe I actually only need 256 Maybe I need a hundred thousand.

So I do think that actually this software scale out is still going to be there I'll note two things though one You're absolutely right that say about 25 years ago at Google and other places simultaneously There was really a transformation of computing infrastructure Like the notion that actually you would scale out on commodity PCs Essentially the same ones that you could buy off the shelf running a Linux stack and that's what you would do for disk That's what you would do for compute.

That's what you do for networking I mean you all take it for granted that this is sort of it was radical There are many people who thought that this was a terrible idea.

That wasn't gonna work I think the exciting thing about this moment right now is actually that we're going to be reinventing I'm not saying Google we are going to be reinventing computing and Five years from now, whatever the computing stack is from the hardware to the software Right is gonna be unrecognizable And by the way, there was this co-design because if you think about it I'll use Google examples because I know those best big table spanner GFS Borg Colossus They were hand-in-hand co-designed with the hardware cluster scale out architecture And if you want that done the scale out hardware if you didn't have the scale out software Yeah, same thing is gonna happen in this moment.

So I think actually the mainframe is gonna look very very different Yeah, I do think there'll be like this extreme demand for an integrated system because right now we are very fortunate at Cisco where we do everything from the physics to the semantics and you think about the silicon to the application and Other than power one of the constraints is how well integrated are these systems and do they actually work with the least amount of lossiness?

across the entire stack and so that level of tight integration is going to be Super important and what that means the industry will have to evolve into is We will have to work like one company even though we might actually be multiple companies that actually do these pieces And so when we work with hyperscalers like Google or others There is a deep design partnership that actually goes on for months and months together Ahead of time before we actually even do them deal and then once a deal is done Of course, there's a tremendous amount of pressure to make sure that they're moving pretty fast but I think the industry's muscle of Making sure that you operate in an open ecosystem and not be a walled garden is gonna get important at every layer of the stack Really great.

So let's talk about the This aggregate the stack a little bit One of the most interesting topic is processors, right?

clearly, there's an amazing render producing an amazing processor that has massive market share today, right and VC startups all the time doing all sorts of processor architectures.

You got an amazing processor inside your fortress What do you think happens next in process or land?

Yeah, we're huge fans of Nvidia.

We we sell a lot of Nvidia products and chips customers love them We're also huge fans of our TPUs.

I think the future is actually really exciting and actually Where it's not that I don't think that we've hit the point of okay, there's TPUs.

There's GPUs There's whatever traineeums or something else.

We're really seeing the golden age of specialization and that's my observation otherwise if you look at it a TPU I'll use that example again because I know it best for certain computation is Somewhere between 10 and 100 times more efficient per watt and it's this watt that really matters than a CPU It's hard to walk away from right 10 to 100 X and yet we know that there are other computations That if you built even more specialized systems for but not just a niche computation So we run a lot of at Google right for example, maybe for serving maybe phrase your intake workloads That would benefit from and even more specialized architecture So I think that actually at one bottleneck is how hard is it and how long does it take to turn around a specialized architecture?

Right now is forever.

Yeah, right for the best teams in the world Really from concept to in live in production Speed of light is two and a half years.

Yep.

I mean that's that's if you nail everything Right and there are a few teams that do but how do you predict the future two and a half years out for building specialized hardware?

So a I think we have to shrink that cycle But then be at some point when things slow down a little bit and they will I think we're gonna have to build more specialized architectures because the power Savings the cost savings the space savings are just too dramatic to ignore and this will actually have a really interesting implication on Geopolitical structures as well because if you think about what's happening in China China actually doesn't make two nanometer chips.

They make you know, seven nanometer chips and And so if you think about what?

But they have unlimited amount of power And they have unlimited amount of engineering resource And so what they can do is do the optimization on the engineering side keep the seven nanometer chips and make sure that they give people Unlimited amount of power we might have a different architectural design where you have to get extremely power efficient you don't have as many engineers as you might enjoy in China and You can actually go to two nanometer chips But and those might be power efficient in some ways, but they might have thermal lossiness in other ways Like there's a whole bunch of things that have to get factored in On the architecture that will get more specialized even by geo and by region and then depending on how the regulatory frameworks evolve You know how that that geo then expands like if China expands to different regions in the world You will have a very different architecture that they plays out than if America expands to different regions in the world So this is a very interesting kind of game theory exercise to go through on what happens in the next three years and in tech in general and No one knows right now Yeah, that's a beauty of the world that we live in.

Yeah, so we'll soon be measuring systems by engineers for token in addition to What's for talking?

All right, so let's turn to another topic which very much a year per kilowatt Networking right obviously you alluded to it Scale up scale out in your case.

He mentioned scale across So it seems to me that the networking is also going to get reinvented in a fairly significant way So what are the leading signs that you're seeing that and the signals that you're seeing in all the direction networking is going to take?

Yeah, networking is going to need a transformation for certain in other words The amount of bandwidth that's needed at scale within a building is just astounding I mean and and it's going up the network is becoming a primary bottleneck which is Scary so more bandwidth translates directly to more performance and then Given that the network winds up actually being a small power consumer That delivered utility you get per watt like it's a super linear benefit like spend a little bit here get way more there So I think that that side is absolutely there I'll put in a plug here in that in this for these workloads We actually know what the network communication patterns are a priority.

So I think this is a massive opportunity In other words, do you then need the full power of a packet switch?

When actually, you know what the rough circuits are going to be and I'm not saying you need to build a circuit switch But there is an optimization opportunity The other aspect of this here is these workloads are just incredibly rusty Yeah, and and work to the point where And we've written about this power utilities notice when we're doing network communication relative to computation at the scale of tens and hundreds of megawatts like massive demand for power Stop all of a sudden and do some network communication and then first back to computing So how do you build a network that needs to go at a hundred percent?

For a really short amount of time and then go idle Yeah, and then same actually for the scale across use case Which we're absolutely saying you don't run large scale free training across all your wide area data center sites 12 months out of the year So and then you're gonna this is the problem I think about a lot is let's say you build the latest greatest chips in these three data center sites How long are you gonna be there before you migrate to the latest latest chips and three other sites?

And then what do you do with the network that you left behind people are gonna run jobs on them Yeah, but you're not gonna need nearly the network capacity that you did for large-scale training free training anyway, so the shift of Needing massive networks for like five percent of the time It I don't know how to build a network like that.

That's if any of you do please please please let me know I mean if you don't know how to build is there's nobody that knows we're trying to figure it out It actually is a fascinating problem.

Yeah.

Yeah, I do think like if you think of if power is the constraint And if compute is the asset, I think network is going to be the force multiplier.

Mm-hmm because you know if a packet if If you have low latency and low performance and high energy and efficiency in the pack Every kilowatt of power you save moving the packet is a kilowatt of power you can give to the GPU.

Yeah Which is you know super important?

The the other thing is, you know when you think about Scale up versus scale out versus scale across you'll also need especially on inference versus training There are different things that get optimized like you might optimize for latency much more on training runs you might optimize much more for memory on inferencing that there's there's architectural and so I also feel like the way that networking will evolve is rather than it being a Training infrastructure that then gets applied to inferencing.

You might have inferencing native infrastructure that gets built over time and so there's There's good considerations to look at on like how all of the architectural components are are moving but in my mind like if I were to say strategically one of the biggest things that's happening in networking from our vantage point is If you're just a rapper around Broadcom Then you've got a monopoly that's going to be a very predatory one and so one of the big reasons where Cisco is Super relevant is you don't just have a Broadcom world with people just wrapping Broadcom I mean that their systems are on Broadcom But you will actually have a choice of silicon and that choice and diversity of silicon is going to be super important Especially for high volume, you know kind of consumption patterns So last question on the system since you brought that up and we'll move to use cases Inference both of you have mentioned when you talked about in the context of the processors you just started talking about the architecture Are you deploying today's specific architectures for?

inference, I mean Sorry Is it still shared workloads?

We are deploying Specialized architectures for inference and I think as much software as hardware, but the hardware is also Deployed in different configurations is the way I would say and then the other aspect of inference that is becoming really interesting is reinforcement learning Especially on the critical path of serving because latency just becomes absolutely critical And I think that so how you would build your system and how you would connect it up to one another and of course networking Plays a key role there becomes increasingly interesting but are there Singular choke points that if removed would accelerate the thousand-fold reduction In the cost of inference that we need or is this a natural curve that we are writing down?

So so we're massive.

I mean two things here one Again, maybe many of you are familiar with this pre-fill and decode on inference look very very different So actually ideally if you you would have different hardware actually the balance points are different So that's that's one opportunity that comes with downsides.

We can talk about that What I would say though is that maybe something people don't realize is that we're actually driving massive reductions in the cost of inference.

I mean 10 X's and 100 X's The problem or opportunity is the community the user base keeps demanding higher quality not better efficiency, so Just as soon as we deliver All the efficiency improvements we were looking for the next generation model comes out and it is that whatever intelligence per dollar is way better But you still pay more and it costs more Relative to the previous generation and then we repeat the cycle and it's almost like the longer The reasoning Yeah that you have the more impatient the market gets right so for example if you have a 20 minute reasoning cycle Like for example with deep research you could have autonomous execution of about 20 minutes.

That was interesting Now you have you know most of the coding tools that can go up to seven hours to 30 hours of You know duration of autonomous execution when that happens.

There's actually a greater demand for saying compress the time down And so you like it's it's kind of a self-fulfilling prophecy where you need to have more performance because of the fact that you've been able to go out and do things for a longer autonomous amount of time and So it's almost a never-ending loop where you you'll need to have more performance for inference.

Yeah in perpetuity Yeah, though intelligence intelligence per dollar is a business model metrics Metrics, so it is not just a processor capability.

No, it's end-to-end.

Absolutely.

Yeah, so okay So let's change topics and talk about actual usage right so both of you have massive organizations where are the key wins that you're getting today with with applying all the AI that's available to you and Then we'll talk about what your customers are doing, but I'm actually curious about what are you doing internally?

Within the teams.

Yeah, so I mean coding is the obvious one and that's actually picking up increasing traction and increasing capability We just actually in the last couple of days Published a paper that showed how we applied AI techniques to do instructions that migration so in other words, we actually had a fairly massive migration from x86 to arm making our Entire code base and at Google.

It's a very very large code base Sort of instruction set agnostic and including to you know future risk five or whatever else might come along Tens and thousands hundreds of thousands of individual Entire code base because we Want to need all of our code base to be crazy ass project Yeah, so so we it was and the the motivation though for this actually was a few years ago.

We had this amazing legacy system called Bigtable and then a new amazing system called Spanner and we decided to tell the company hey Everyone needs to move from Bigtable to spanner and better Bigtable was amazing for its time, but Spanner was better The estimate from doing that migration for Google was seven staff millennia How much seven staff millennia that we had a new unit that we had to actually To see what and and it was it wasn't like made up people being lazy It's like this is this is what it's enduring that they came up with that though.

And you know what we decided long live Bigtable I Decided what it just wasn't worth it.

Yeah, honestly like the opportunity cost was too high So did and we have these sorts of migrations tensor flow to Jack's We actually I mean began somewhat private but not not too secret.

We've affected this internally with a I assist went integer factors faster now there are other tasks which the tools probably aren't quite yet up to the Whatever standard for but the the area under the curve is getting bigger and bigger and bigger So we're seeing probably like three or four Really good use cases and then we're seeing some use cases which are not working yet.

And so What is working?

Code migrations are working relatively well so far we use largely a combination of codex Claude and Cursor some windsurf and so code migrations tends to work pretty well Debugging oddly enough has actually been very very productive with with these tools and especially with CL eyes The Where we've not done as good a job and then front end zero to one projects Tend to do extremely well like the engineers are super productive when you go to code that's older And especially for the down in the infrastructure stack Much harder to go out and get that to happen But the challenge that we have to orient our engineers on this is actually much more of a cultural reset problem Than it is a just a technical problem, which is if someone uses something and says this isn't working, right?

You can't put it back on the shelf saying this doesn't work for another six or nine months You have to come back to it within four weeks and see if it works again because the speed at which these tools are kind of advancing is so fast that you almost have to kind of get like so I was with a 150 of our distinguished engineers today and What I had to urge them to do is Assume that these tools are gonna get infinitely better within six months Yeah, and make sure that you get your mental model to where that tool is gonna be in six months And what are you gonna do to be best in class in six months rather than assessing it for where it is today and then Putting it aside for six months assuming that that's not gonna work for the next six months I think that's a big strategic error.

So I we've got 25,000 engineers.

I'm Hoping that we can get at least Do a 3x productivity within a very short amount of time within the next year And we'll be able to see what if if that happens The second the couple of the big areas that we are starting to see some good responses isn't sales Preparation going into an account call really good legal contract reviews actually much better than what we had thought and then the last one is not super high inference volume, but product marketing I Think the first chat GPT Take on competitive is always better than what my any product marketing person comes up by themselves So we should never start from my slate to start from chat GPT and then go from there Okay, you could be talking about the topic for a long time, but they showed me the two-minute warning So I want to focus on one last question here.

So you got a lot of founders here, right building amazing companies So what is the most interesting development?

They should look forward to In the next calendar year, let's call it power the next 12 months a from your company and be from the industry if you are Look at a crystal ball And I think to build on the point these these models are getting more spectacular By the by the month and then they'll be from whatever companies you like a bunch of really exciting including ours Oh, I forgot to say you're not allowed to say models will get better Yeah, everybody knows the models are gonna get but I mean they're getting Scary good is the part that I would say But I think that then the agents that get built on top of them and the frameworks for making that happen are also getting scary Good, so the ability to have things Go quite right for quite long over the coming 12 months is gonna be transformative I'm anything do you want to leak any aspect of your roadmap?

Next one not so not right now.

Yeah, okay Do you do I'd say the the big?

Shift and what I would urge startups to do is don't build thin wrappers around models that are other people's models I think the the combination of a model working very closely with the product and The model getting better as there's feedback in the product is gonna be super important So you are gonna need foundation models, but if you just have a thin wrapper I think the durability of your business will be very very short-lived So that would be something that I would I would urge you on and I think that intelligent routing layer of some sort that Says I'm gonna use my models for these things I'm gonna probably use foundation models for other things and dynamically keep optimizing will be I think cursor does that pretty well But that'll be a good way that the software development lifecycle will evolve What you should expect from Cisco is look truth be told for the longest time so people thought Cisco is a legacy company Like that there were has been I think there's a level of momentum in the business as a spring in the step in the employee base, so you should expect like I said from the physics to the semantics in every layer from silicon to the application a fair amount of innovation in Silicon and networking and security and observability and the data platform as well as applications, you know from us and we're excited to work with The startup ecosystem and so if you if you ever feel like you want to work with us make sure that you reach out to us What are you gonna say something?

I mean one aspect that I want to highlight about the models is Where we were with let's say text models two and a half three years ago They were fun like hey write me a haiku about Martine did a great job now.

They're amazing I think that what's gonna happen in the next 12 months is the same thing is gonna be happening with Input not but of images and video to these models and to the extent that even for images imagine them as productivity and educational tools not just okay, here's Martine as Superman on a by-law school to write but Using it for productivity gains and learning I think is gonna be really really transformative Awesome, so I'm not really in this session.

Thanks for a great conversation.

I mean thanks for detail Thanks for listening to this episode of the a 16z podcast If you like this episode be sure to like comment subscribe Leave us a rating we're of you and share it with your friends and family For more episodes go to YouTube Apple podcast and Spotify Follow us on x a 16z and subscribe to our sub stack at a 16z dot sub stack calm Thanks again for listening and I'll see you in the next episode as A reminder the content here is for informational purposes only should not be taken as legal business tax or investment advice Or be used to evaluate any investment or security and is not directed at any investors or potential investors in any a 16z fund Please note that a 16z and its affiliates may also maintain investments in the companies discussed in this podcast for more details Including a link to our investments, please see a 16z.com forward slash disclosures You