# Modal's AI Infrastructure Strategy: Agents, Inference, and Elastic Compute

**Podcast:** Latent Space: The AI Engineer Podcast
**Published:** 2026-07-09

## Transcript

We're here with Akshat of Moro, CTO of Moro, together with Vibu.
Congrats on your 3C.
Thank you.
Your party yesterday was amazing.
Yeah.
From all the photos and all the swag.
We had a bunch of art installations, which is kind of fun seeing our products on pedestals next to Rodan.
Very nice, very nice.
When you started...
It was not the GPU inference company.
I mean, maybe it was in your mind.
Take us back to the origin story.
I actually first met Eric, who's CEO through an investor.
And back then, Eric was already thinking about building a new kind of runtime.
And he got there thinking through why are workflow orchestration products so hard to use?
It's because you have to run them on Kubernetes.
Kubernetes is hard to manage.
It's not built for burstiness and custom images and has a terrible developer experience.
And I'll interject for listeners who are new.
We interviewed Eric two years ago and there's a bit more of the story there from Spotify and all those things.
And I actually came across Eric through a data console because he did that talk on the serverless container stack that you guys did, which was like...
That was my first like, okay, I need to take models very seriously moment.
But it's still very unclear.
Like, do I actually need all this for just my data pipelines?
Yeah, I mean, initially what we were thinking about was if we build a better runtime, it's a very useful permittive in itself.
There's a lot of things that get solved by serverless functions.
Like you can do ETL stuff, you can do job queues, you can do...
all this like bursty processing, which it turns out every company had needs more.
But then we also were thinking about this as like, this is a primitive that we can build a whole collection of products on, which are very verticalized.
So perhaps data engineering would have been the first one, but we were thinking about inference.
Back then it was more classical inference, like computer vision stuff and running XGBoost and whatnot.
But we added GPUs to the product a year before ChatGPT came out.
Nice.
We just didn't think it would be that big of a deal.
Yeah, just like add A100.
Was there any like early key problem that really sparked off why you built it?
Yeah, primarily it's just none of the tooling that was out there was built for one, a really great developer experience.
And also there's a general trend of a lot of the workloads that we were seeing were very, this is, I wish there's a better word for it, but compute heavy.
Like they need one like, need a lot more resources, so you need to burst up and down a lot, versus Kubernetes designed for slow scaling and more for web server use cases.
And also, there's just a lot more specialization in what kinds of environments these workloads run in.
Sometimes they need accelerators, sometimes they need different kinds of images.
And this is just a consistent thing that we saw across a lot of companies.
That would be the next job.
Yeah, it'd be nice.
I don't know how much this factored into the early story, but I wrote a post when I was at Temporal about infrastructure, software-defined infrastructure or something like that.
The self-revisioning.
Self-revisioning, yeah.
I can't even remember my own post.
And then you put me on the landing page.
Yeah, we really like the term.
And so we stole it.
Because you had the insight that everything can just be in decorators next, co-located with the code, right?
Was that a big part of the original story?
It was just like a DX layer.
That was really important because we really didn't want people to spend so much time writing YAML.
And it seemed like you could really condense the surface area of what you're doing, put in code so you can actually operate on it, just so you can operate on other code and build stuff that's more expressive and dynamic.
And so, yeah, that was always a very important part.
Then the pushback is, this is a DSL.
It's your closed source.
I am locked in to Modo.
Yeah, we never really got pushback for that because the nice thing about Modo is you can bring whatever code you have.
And sure, the DSL is at the sort of configuration there for what hardware you're using, how you're scaling things up, but you still own the code.
And that's been an important part of our story, even as we do inference now.
How much of it do you think still stays the same today?
Like if you were to build something today, devX is obviously very important, but I feel like, you know, a lot of this has kind of been changed with just...
hook it up to an agent, have Cloud Code, have Codex, implement the tool.
There's very agent-native primitives that are kind of different than if I'm doing this myself, right?
We've actually changed our SDK team to think about agent experience instead of developer experience.
And we think that the same benefits that apply for DX also actually apply for AX, which is why would you have an agent read through hundreds of Kubernetes files and write YAML that's not even typed basically make a couple of changes in a decorator and it gets this sort of self-provisioning runtime of being able to see its changes live in action.
Yeah, it just seems from the customers we talk to, they actually find the model is way faster used for agents versus operating on a different substrate.
Yeah, because again, you collocate the infrastructure requirements to the code that runs it.
The negative thesis now is that nobody's looking at their code anymore, so there's no point.
Yeah, I mean, people aren't looking at code.
One thing we actually still see is really important is observability.
How good is your dashboard?
And of course, we've pushed a lot of it to the CLI so that agents can do their own investigation, but you still need humans to go interpret what's going on and make judgment calls and whatnot.
And that's, I feel like, maybe more important now than looking at the code itself.
Yes, because you can try to treat the code as a black box and then use, sort of see the observable action that comes out of it and then just prompt a change.
Yeah.
So I think actually it takes a bit of restraint to not specialize, to say, I want to ship a new primitive and then just be general purpose.
People ask you, what are you for?
You're like, I don't know, we can do this, we can do that.
I'd be curious to say, you know, like, okay, if we were to ask you, like, what is modal for even at a high level?
There's a lot you guys do, sandboxes, GPUs, everything.
How do you answer?
Modal is a cloud platform that's built for where we've built the primitives from scratch for AI applications.
And right now it basically covers inference, training, batch crossing, and sandbox workloads.
But we're building a lot more.
I noticed you didn't say web server.
So there is still a role for like the always on large scale Kubernetes type of things.
Yeah, absolutely.
We're not trying to compete with the renders of the world because, yeah, we think the differentiator for us is our other workloads that need specialized compute, need to scale up and down a lot.
Yeah, they're shaped differently.
I think you're building a lot of it alongside the startups, right?
They're innovating quite a bit.
Even in your latest blog post, even in the Series C, the customers that you mentioned here, the Cognitions, Decagons, Ramps, and whatnot, they're innovating with you, right?
And that's not something AWS is doing directly with them.
Yeah, absolutely.
I think, I mean, this is, again, classic.
We're a small team.
We can move really fast.
Our engineers are working with our customers and figuring it out.
week at Cognition, I walked in, there was someone wearing a moto shirt.
I was like, what are you doing here?
They're like, yeah, I just, I am embedded inside of cog.
Yeah, I think that was Peyton.
We sent him over because the latency of communication was too high otherwise.
Yeah, you know, distributed node, you have to place one and co-locate.
So I had a direct personal experience, right?
So I worked on Small Developer three years ago.
It was inspired by Cloud One.
I think you onboarded me at some point, like just before.
And I was like, oh, like I need some bursty compute.
Like I was just going to try using modal.
And it was a pretty pleasant experience.
Apparently I showed up in the board meeting.
I like the analytics.
Yeah, you blew up on Hacker News and we got a big traffic spike.
I actually, I think the way you use small developer was modal functions for running stuff, which was...
Like that was a good use case.
Yeah.
To me, that was proto-cognition.
Right.
If only I had like stuck to it.
Like that was like, if you just like draw the tech tree out, you're like, yeah, like probably this will happen.
Yeah.
Like he was so close.
I just didn't realize.
But the funny story there is at the same time, we were talking to a bunch of customers who...
needed something like sandboxing.
This is like 2023.
So we built a new API right after that.
Yeah.
Like we built sandboxes in May of 2023 before anyone was even knew this was going to be a thing.
And the first example we published was we took small developer and put in a loop so the agent can iterate on itself.
Loops are hard these days.
Loops in, when was this, 2023?
Yeah.
A small chick.
It's like mid-2023.
I mean, obviously for listeners, the problem was the models are not built for any of this, right?
They're not post-trained to understand looping and self-correction and tool calling was there, but also not that great.
Yeah.
I don't remember if you used tool calling in this one, but yeah, the models would just diverge after 10 iterations and not produce anything meaningful.
Yeah.
But like then, so I mean, okay, like now talking to myself three years ago, the answer is collect all the failures, build benchmark, and then...
collect all the examples, build the Aural environment, sell it for like $10 billion to Meta, and then also train a model and then sell that for $60 billion to Elon.
And this is a money machine.
It's actually about that hard.
I mean, it's hard to have that kind of inherent conviction that the stuff will get that much better.
In retrospect, it's so fucking obvious.
Fair enough.
Like what else were we doing back then?
I don't know.
Anyway.
Yeah, so that was the start of your sandboxing journey, right?
I feel like it didn't blow up, blow up until like last year.
Yeah.
So there was like a couple of years of quietness.
Exactly, yeah.
Very underrated product value.
My experience with Modal, Charles, before he had joined Modal, met this guy at a hackathon and he really insisted.
We wanted to run some small model, you know, not hosted anywhere.
And he's like, ah, there's this cool company, Modal.
They'll like spin up a GPU sandbox so you can throw it on there.
They'll take a hugging face link.
And, you know, like there's so much value just right there, right?
Like instant hosting, spin it up, spin it down.
It'll stay cold, but, you know, we run the demo a few days later, it'll come back up.
And like all this stuff in retrospect, like it's still what we needed like today.
Yeah, I mean, it's still needed today.
Obviously, workload shapes have changed a lot as we run stuff for people with really massive production scale.
And there, it's not about scaling from zero to one, but it's how do we scale really elastically from like 1,000 to 1,500 GPUs very quickly in a given region.
It's the same shape of the problem.
Okay, so you look at, say, Cursor Composer, right?
They'll do RL on a model every couple hours.
You guys have a whole version of RL, Inference, Jim, and whatnot.
When you look at workloads like that, you're basically doing train runs where you need to scale up, scale down every hour, thousands of GPUs, right?
That's the example for we do need it, right?
Actually, I'll take a step back and maybe talk about how people use Model today.
Because our biggest use case actually is elastic inference.
And the thing we first found product market fit with was inference for custom models.
So we kind of stayed away from the LM space and we were serving companies like Suno for audio, Runway for video, robotics, comp bio companies that train their own model elsewhere.
But modal is the best black box that...
for deployment, scaling to however many GPS you need as your traffic pattern changes.
And we saw all of them actually have a very unpredictable traffic pattern.
It's like diurnal.
Some days the company will do a launch and they'll need way more.
And it's not just one model that they deploy.
All these companies deploy lots of different models in different regions.
And so the autoscaling problem becomes even harder because then you have to scale within a certain region.
And those cycles sort of are offset.
So different times you can scale up in different regions.
So that's like our sort of...
And that in and of itself is a huge category.
There's a bunch of inference providers.
which, you know, provide this fireworks, does this as a service, together, whatnot, base 10.
That's kind of carved into its own niche for language models, at least right now.
Yeah, I mean, the thing that we have actually specialized in is the autoscaling aspect.
Because we found that it's not universally true that everyone else can autoscale.
And we've gone deeper into it on the tech side, but we've incorporated GPU snapshotting to the product so that we can actually...
take the GPU state, like your Torch compiler model, snapshot it, and the next call starts way faster.
And so going back to your question, that's why you need a lot of burstiness or inference.
But then people also do a lot of on-demand training.
Like for RL stuff, your rollouts are bursty, as you said.
People also do a lot of batch jobs.
So we'll see a lot of companies, before they have a training run, they'll need thousands of GPUs to run encoding or something like that.
And I think those things are much more bursty than I agree that agents are not that bursty.
Sandboxes are, except when you're doing RL.
RL is insanely bursty.
Yeah, like when you're doing rollouts, you sometimes need 100,000 sandboxes in your seconds.
Yeah.
I'm curious if you've seen early sparks of continual learning.
There's some people like our friends Ngram recently announced this.
They're trying to do training.
That also seems like a different workload, right?
If you're doing training 24-7 per se, there's a very weird dynamic of how you're using GPUs between people and whatnot.
But it seems like something you guys would work for.
As you said, we're fortunate to work with a number of customers of the frontier.
Ngram's one of our customers.
taking the primitives we have and trying to use them in very interesting ways like continual learning.
It's possible as the stuff gets better, some of that will be part of our offering as well if more people need it.
But we're just waiting to see how all this shakes out.
Is there a primitive that you added after sandboxing that was the next step in the story?
I guess we've been going much deeper into LLM inference because we've realized that Some of the advantages we have with autoscaling, again, especially in different regions and whatnot, are not present elsewhere.
And the place where we had a gap was we weren't working on the model there itself.
We were a black box.
And we realized that we actually can get to frontier-level model performance by having great people who work on...
all this, and we've actually been open sourcing a lot of our work in terms of recently we shared our work on DFLASH, which is a block-based speculator, and we've open sourced all of it, so you can get, by using open source DFLASH, you can get the same performance as you would with one of the proprietary providers.
And the next thing we're thinking about here...
I thought this was actually an interesting blog post as well, right?
I think in here you make a claim that, you know, not a claim, just that...
How effective speculative decoding really just gets you?
Anything you want to point out from this around, you know, what people should know?
Yeah, absolutely.
I mean, the high-level summary is, would it help to describe what speculative decoding is?
Yes.
So we've covered like Ego and Hydra and all those things, but it was like two years ago.
I think it doesn't hurt, right?
The speculative decoding is you have a smaller model called a draft model, Predict.
tokens ahead of the bigger model.
And then you have the bigger model verify all of this.
All the tokens are predicted.
And the reason it's faster is if you're predicting one token at once, you're kind of bound by memory bandwidth.
But if you can batch the verification of the draft model, then you're much more efficient using compute and it's faster.
And as long as your draft model is producing a lot of tokens that can get accepted, which is called the accept length, you can get...
speed up that's multiple times of the original model speed.
And that's what we highlight here.
It's like people talk a lot about we made these kernels faster and whatnot, but improving kernel will only give you like a few percentage points of improvement and increasing its set length literally is a multiplicative decrease.
Like 2 to 4x.
Yeah, exactly.
Without much head-on performance.
Yeah.
I think it may be, I mean, you are running a second model, right?
So maybe it's more expensive in the compute, but not much.
So there's no drop in quality performance because you're always, you're never accepting a token that's a big model.
It's strictly better or it's the same.
Exactly, right?
Yeah.
And so we've been working a bunch on DFLASH, which is a block-based speculator.
So it's, instead of predicting one token at a time, it's predicting a block.
And we've been open sourcing and worked with it.
The next thing for us here is for helping people train speculators and custom models.
It's something that traditionally is very FDE driven, support deployed engineer driven, like you work with customers and we can do that.
And our vision for, this is why we launched Auto Endpoints, is we want to make frontier level performance available to everyone.
And so we mentioned this announcement, we kind of teased it.
The next thing we're launching is, Basically, as you run an auto endpoint, we shadow traffic.
Do you want to explain what auto endpoints are?
Yeah.
So this is, I guess, going back to your modal is you touch the code, but sometimes people actually don't want to touch the code and they want to get started with...
an endpoint that works and has all the great performance and scalability that modal has.
So we've made that easier with basically a way to create an endpoint from our UI, from the CLI that has all of our optimizations that we talked about, like the deflash stuff already baked in, and this full transparency.
So we give you the code, you can go run it yourself, and if you want, you can eject out into the full modal experience, which we see as people get sophisticated, they do want to...
tweak the models, they want to fine-teat stuff.
You can still do all of that.
It's not a black box.
And yeah, the next thing, as we tease later in the post, is how do we give you value even beyond this in terms of having your draft models evolve as your data distribution evolves, again, without having to talk to a person.
And yeah.
I guess just to kind of understand it directly, I mean, you know, obviously you...
you have the GPUs, you have an endpoint that's compatible, you serve open model.
If someone was to do this themselves, what's the delta that you guys provide?
So you do a lot of open source great work on effective inference.
How does it compare to, say, I take the same model, GLM 5.2 FBA, take off the shelf inference engine, VLM, SGLang, you know, get compute of similar capacity, similar cost.
What's the kind of delta that...
plugging into something like this offers outside of the benefit of scaling?
It's kind of interesting because we've taken the approach of open sourcing our contributions and upstreaming them.
We work closely with the SGLang team.
We actually want the improvements that our team comes up with to be there in open source for others to use, even outside of modal.
The benefit to us is we have a team that has significant expertise in terms of if you do have something that is not there, our team can help you get that performance first.
The other thing is, with these endpoints, we are way more elastic, as you said, than anyone else.
And you have true scaling to zero.
You have true burstiness.
And in practice, that matters a lot more to people than just finding the GPU and running model code.
Yeah, and I will say it's actually not that.
straightforward to just, like what I said is easier said than done, right?
It's, I think still for the average person, still hard to just gut check using different, there's quite a bit of combinations you can make there.
The trade-offs aren't really known at face value.
Yeah, I mean, it's not just that.
I think it's that running production grade inference is a hard and fair problem, even if you subtract out the autoscaling.
controlling things like tail latency and making sure every request is delivered at least once and whatnot.
There's a lot of innovation that you can do here.
I think it's very interesting that you're starting to encroach on, as you become a full cloud, you're starting to encroach on other people's turf.
What will you not do?
Well, we want to follow our users and make sure they get a platform that has everything that works.
well together.
So right now we're kind of focused on the model life cycle and the agent life cycle.
So both like going from data prep to training to inference and then also if I want to deploy a background agent, let's say sandbox, persistent storage, a whole bunch of other stuff.
We talked to Cole who did Open Inspect.
Yeah.
And obviously Real Inspect also is on model.
Yeah.
So Ramp Inspect was a great example of a background agent that was really successful because they were able to use some of the primitives like snapshotting and fast scaling to just have something that feels really reactive and works well.
Yeah.
That's the new CTO of Ramp right there.
Yeah, Ravel.
It was really, really fun.
Yeah, I mean, you know, I think all very bullish.
Like, you know, one of my reflections was also...
I did not originally, because obviously when I met you guys, you weren't that much in the GPU game, and now you're all about inference.
And one of the points that I hinged on for Jensen's keynote at GTC this year was what we're calling the inference inflection, right?
That, let's say, in AI workloads or machine learning workloads, it used to be like, let's call it 8 to 1 GPU to CPU, and now it's more like 1 to 1.
which is like an interesting, like because of how much agents basically are blocked or call out to CPU heavy stuff, the actual like limiting factor like swings back and forth from GPU to CPU a lot more than it used to be all GPU and then occasional CPU.
Yeah.
GPU, CPU.
And now it's like just constantly and you just have to co-locate everything.
Yeah.
And that's one of the things that actually, again, we see is something appealing about modal, which is we've built this capacity pool that spans 17 cloud providers.
So we're very good at running on various kinds of cloud capacity across the world.
You don't have your own data centers?
We don't have our own data centers.
We just run across a lot of neoclouds and modal providers.
Yeah, you're running the math and you're like, what's the cutover point where you're like?
Yeah, it's a good question.
I mean, part of it is we see our differentiator in the software layer.
And being capitalized and focusing on the software helps us move really fast.
So far it's worked out well because there are so many other people building data centers that we're able to work effectively with them and again, focus on what makes us special.
17 gets you into the local providers sometimes.
Which was the most interesting one?
There are actually a lot more neoclows than you expect.
And they all have various degrees of...
various levels of reliability.
And that's why it's something we've invested a lot of time in is actually building our own reliability layer on top.
So if the GPU falls off the bus or something happens, user workloads are not affected.
And that actually lets us use a lot more capacity than you as a user would be able to.
a useful thing to have because now everyone knows what layer you are and you sort of optimize for being the super cloud of all clouds.
Yeah, that's the idea.
And so I guess when you mentioned co-location, that's another interesting thing where one thing we've seen is people come to us when they want very specifically located CPUs or GPUs like they want.
Or to pin it in like EU?
Exactly.
Or EU, EUS, Australia.
Data locality thing or performance or what?
It's either data locality or latency.
Yeah.
Like you want your, they're running sandboxes in modal.
They want them to be right next to it.
Yeah, it's easy.
That's, you know, that is important in all those things.
And so like you've kind of accidentally, I don't know if it's accident, but like you've built the perfect primitive for agents to express themselves.
And then, you know, like it's almost very funny how every extra development just involves more file system.
just involves more CPU.
Yeah.
Just like the things that you already have.
I don't know much about if there's any networking usages that are interesting, but you've also done some good work on networking.
Yeah, I mean, that's exactly right.
We're sort of just taking compute stories and networking and building stuff on that layer for, again, a bit of stuff people need.
We see a few interesting networking things coming up.
One is people actually want networked sandboxes.
So we have...
For like a Docker cluster type thing.
Sorry, Docker Swarm.
What is it called?
Compose.
Compose type thing.
Yeah.
So actually, if you want Docker Compose, our sandboxes now support this thing called sidecarves.
So you can, a sandbox is actually a pod of containers and you can run multiple containers in the sandbox.
Also useful because going back to networking, people want a lot of control over...
outbound networking from a sandbox.
They might want to run a man-in-the-middle proxy for maybe logging stuff for RL or controlling how egress can happen to a domain, injecting credentials.
And yeah, so we've kind of had to build a lot of that stuff ourselves.
But then also sometimes people actually want sandboxes spanning multiple nodes to talk to each other, which is an emerging thing we're seeing.
We have support for that for a different reason.
And yeah, we'll see if that becomes stable.
Like just an open socket.
This is directly like MTLS.
We do support that, which is you can expose a tunnel inside a sandbox.
And then you can either expose the public internet or it can be, you can add like a HTTP odd layer above it.
But we have this thing called i6PN, which we haven't talked about, which is this like overlay network using IPv6 addresses.
So if...
modal containers within the same workspace, when this is enabled, can actually address each other using this private IP v6 address, and no one else can.
This is like sort of private networking for containers.
We actually built it because we needed it as primitive for our distributed training product.
So we have this other feature, which is you can add a decorator to a function, and you get a cluster of GPUs, and they have already been networking.
So you can run a speed training job that's truly serverless.
And we did the overlay network for that.
But then we've seen that people are using it for other reasons and I'm kind of intrigued to see what would people do with it.
Build primitives and let people figure it out, right?
Yeah, exactly.
They're like, they read the docs page.
Let me use that for something that you never intended to.
This was literally not even in our docs page.
somehow found it and they're using it.
I mean, the way you portrayed it with like RDMA versus TCP, like very well laid out, but just the transfer speed change at scale for RL, like, yeah, you have it, you have it built in.
I'm sure someone found it, it's found it to be a lot more efficient before you actually made a thing out of it, right?
Yeah, and not to split hairs, I guess the overlay network actually is the TCP overlay network.
The reason we have that is you need that to do the key exchange for RDMA before you set up the RDMA network on top of that.
But then people found the TCP part.
Can I tell you, this is like a big aha moment for me because, so I, you know, I review 2,200 submissions for the World's Fair.
And then I got this from John Osterholt.
I don't know if you know John Osterholtz.
The name sounds familiar.
He's a well-known professor, published a lot of interesting software design books.
And this is the talk he chose to submit.
It's on RDMA and inference.
And I'm like, you wouldn't think that this guy, who is a kind of operating systems guy, would care about RDMA.
I mean, it makes sense to me because it's just a cloud, right?
Like the way you move around your KV cache and how efficiently you can do it, how efficiently you move your weights from your training GPUs to your inference GPUs and RL is there's a lot of degrees of freedom and it is basically a systems problem of moving memory around and scheduling.
This shows you how primitive my understanding of networking stuff is.
Is this like the domain of wire guard?
as well?
Not quite.
So, adjacent?
So, explain everything.
Sure, sure.
How do we move memory around GPUs?
Sorry, yeah, that is memory.
Sorry, I was talking more, and maybe I was talking like five minutes back, about the private IPv6 addressing that you've set up.
It's basically a VPN.
Yeah, it's sort of like a VPN.
And yeah, WireGuard is, yeah, you're right.
It is...
Yeah, you already moved on to any topics.
A similar, in the same space, WireGuard is encrypted.
And this is...
You don't need encryption.
Yeah, it's not encrypted.
There's no difference.
This is TCP, and we have eBPF programs that will reject or allow the TCP connection based on whether you're allowed to do it.
It used to involve a full sidecar, but now you have eBPF in the Linux kernel.
Yeah.
I don't know if this is a natural follow on to the topic of like my skepticism on distributed training is that, well, people spend a lot of money on like cables to hook up GPUs.
And even that is not like fast enough.
And that's the bottleneck.
Is your networking fast enough?
Yeah.
So I guess you're talking about sort of fully distributed training, like a dialoka or something, which is like cross center.
Yes, that's the extreme.
Yeah, you're kind of in the middle, and then other people would have the Melanox cables up in their actual data center.
When you run multi-endort training on modal, RDMA, I think Melanox is, or InfiniBand is, also use RDMA.
But basically it's a way to bypass the TCP networking stack and transfer stuff much faster between one node to the other.
And we have, I think, like...
three terabit per second internal networking, which is the standard that's needed.
Okay, so I misunderstood what part of the stack you were.
50 gigs over TCP.
Yeah, yeah, yeah.
Okay.
Yeah, I mean, very impressive work.
So effectively, you're extending sort of like the modal philosophy to the trading cluster.
Yeah, and we're not going for obviously like large-scale pre-training runs.
The thing that we've built multi-handle training for is We see a lot of smaller scale post-training.
Like people are post-training like medium-sized QN models so they can get higher quality on inference.
This is the perfect fit for something like that.
Yeah.
That is my impression of how a lot of these labs explore branches in post-training and then eventually merge whatever they find in.
Yeah.
The other use case we've seen for multinational training is even if you have a big cluster, your researchers are still doing small runs.
And having elasticity there matters a lot more.
This is actually the current limiting factor for auto research, which is you basically need to give your model some GPUs.
We have a blog post on auto research, and model turns out to be a pretty good substrate for that.
So my impression is auto research means many things, like anything that Andre coins.
Right now, still science fair, right?
Not actually, like, I don't know how many people are actually doing this.
I thought the same thing.
Yeah, you would know.
We, like, our internal both training and inference teams actually use this sort of, the general shape of this quite a bit.
Like, we have this one internal repo called auto-inference, which essentially, we've automated our own FDE efforts using this harness, which is the agent will just spin up.
a sweep of different things.
It'll even run like a NVIDIA inside profiler and it'll like tweak configs and it'll arrive the right thing.
It'll change your GPUs from H200 to B200 and actually works really well.
Nice.
By the way, I enjoy that your FDE is so technical that you have to do these things.
It's very different from FDE from other people.
Yeah, our FDE team is essentially they're like applied inference researchers or applied training researchers.
Someone told me they have to be able to build, but they also have to be able to sell.
Do they have to sell or are they like, they're good?
It's just like post-sale type of thing.
It does.
Being able to talk to a customer and engage effectively with that matters a lot.
They all want the same thing.
But it's not really a sort of sales thing.
We pair them with, we have solution to architects as well that are more on the pre-sales side.
Okay, let's spend a bit more time on auto-research.
This is a big focus for me for this year.
Where does this go?
Have people explored enough?
There's all these beautiful charts of improve, improve, and then sort of level off a bit, and then you find the next thing.
Is this basically sort of one abstraction up from normal training?
Is that how we think about it?
Or do you think about it differently?
Like model-level training versus basically AI-driven hyperparameter search.
Some people call it neural architecture search or whatever, right?
Yeah, I mean, so the stuff I've seen people do with it is nowhere on the architecture level.
It's pretty much tweaking parameters, but it's basically a hyper-proprietor sweep that's guided by some sort of model intuition.
So it's like much more efficient than whatever other sweeper you would have.
Yeah, I mean, you know, it's just a question of where you want to spend your computer, you know?
Right.
Because you can just throw infinite amounts of money on this and somehow you'll...
bang on Shakespeare, you know.
Yeah, infinite monkey.
I mean, so like very good for modal.
And I think it's also very important that agents can spin up other agents.
You can spin up their own infrastructure.
Like very, very good for you.
How good are LLMs at generating modal code?
Like, you know, the benefit of existing pre-LLMs is that you are in the data.
Yeah, they're actually surprisingly good.
I think like pre-Cloud 4, they were not.
And then now they're able to one-shot.
stuff out of the box.
We're playing around with releasing a modal bench for the harder things that the LMs cannot do yet.
What's an example of that?
I think the things that sometimes agents struggle with without right guidance and a skill is how to use the rest of our observability.
Something is failing, how do you look at the logs and then...
update the right thing.
It's sort of reasoning about that.
But they're able to one-shot like...
Yeah.
You can just add a skill to it.
Yeah.
So we have a model skill now that which is kind of actually why we built this model bench.
It's to find things like that so we can address them in our unit skill.
No, no.
I mean, it's good.
Are you facing any shortages?
You know, like we talk a lot about GPU shortages but also CPU, also memory.
Yeah.
We have had a lot of growth, which means that we've had to be much better about proactive capacity planning.
Which, by the way, it's like an MBA's dream job, just planning this stuff.
I think last time you and I talked about this.
Yeah, I mean, we have a really competent team of people that we call, the role is called compute strategy.
Anyone listening here wants to work on that.
Compute strategy?
Yeah, yeah.
I mean, the normies call it FP&A or something.
Well, it's not FP&A.
There's a lot of interesting financial questions of like, what is the blend between one year and three reservations?
How do we forecast our own capacity?
How do we basically, especially since our capacity is very...
fungible across different GPU types and different regions.
Like you basically have to model a lot of it.
And you also have to have an opinion on how the supply chain is going to evolve.
And then you have to like take bets based on that.
Tokonomics.
This is like probably not a real point, but I was trying to think about like what other industries, I was trying to think about like, you know, we cannot be first to like these kinds of problems.
And what other industries have had this?
And I was like airlines with fuel.
And they have to hedge their fuel.
And I think for a long time, Southwest, because they made a hero fuel bet, they were super low cost compared to everyone else.
Yeah, I hadn't thought about that.
We're in a fun time too, you know?
Yeah, a lot of the compute business in general for us is also about being very good about capacity management.
That is how you have great economics.
But also over time, it's how you can unlock more value for customers.
One of the things we're building now is a way for customers to get, if they don't care about latency, get much cheaper pricing and they'll get results back in the next 24 hours or something.
Like a batch tier, essentially.
And those are levers we have because we control the whole stack and scheduling and whatnot to give people a sufficient...
Yeah, I feel like they're not as popular.
The Frontier Labs have all those APIs.
They're not as popular as they should be.
The demand that we see for something like that is actually not for LLMs.
Although sometimes people want to run evals and do synthetic data prep and there it makes sense.
But it's from a lot of non-LM companies, like people who are doing computational bio, they haven't run really big bad jobs and they don't care about when they get it back.
Yeah, and they have a reasonable...
It's also a cousin to the stopping problem of like, well, this...
finish in time?
Yeah, you can bound it.
You can give people SLAs on it.
I think what's interesting is the next phase of Moto.
What do people expect from you now that you're established and you're a well-known computer player among all these leading companies?
You had it in France Launch Week and we talked a little bit about the launches.
What else?
What else should people know?
We are building permittance that make our users' lives much easier.
So I think, for example, LM inference, thousands more companies are going to post-training their own models and deploy open-source models for inference.
So we're thinking a lot about what is the best product shape for that, and that involves everything from our training gym to then endpoints that get frontier-level performance.
Again, I haven't talked to anyone.
It looks somewhat different on other verticals.
We're also seeing a lot of real-time audio-video stuff in there, which is why we're working on things like regional routing with fallbacks.
So you can get sort of GPUs that are as close to users as possible.
So you get low latency for video streaming and whatnot.
And then on the agent side, we're still working very closely with our customers because stuff is changing so fast in terms of what they need.
And I think beyond sandboxes and persistent file systems, there's a lot of other things people need from this agent stack as they build production agents.
So yeah, we're thinking about those other things that fit in there.
I want to ask what the other things are.
Yeah, I probably should share right now.
I think, okay, so I do think a lot about the principal components of cloud and you do talk about compute storage networking.
Because so far for me, it's fine.
So far for, I mean, the first couple generations of cloud, it's fine.
What's different, qualitatively different about agents that you need some new permission level?
Like a lot of people, you know, obviously, okay, and I'll just kind of spew tokens at you until it like hopefully sparks something.
Yeah.
Like the new level now is whatever cloud code does, which is...
dangerous these permissions or like allow this by command or like whatever, right?
And sometimes they're like, okay, we have like this adaptive thinking mode where like just trust me, bro, I want to make the calls for you.
Is that it?
You know, like basically like sort of LLM mediated permissions.
And now you're looping it with a goal and flooding it, bro.
Yeah, I mean, I'm skeptical of LLM media permissions for stuff that is at the sandbox level because you do want hard boundaries.
Otherwise, obviously someone can exfiltrate stuff.
Maybe that's old school thinking.
Maybe we're the dinosaurs.
Maybe the AI OS or the LLM OS is really the kernel as a goddamn LLM.
It makes you feel uncomfortable.
But that's what trusting the LLM is.
Imagine a spherical cow perfect LLM.
Let it.
Maybe.
I want to test the boundaries.
Obviously, I don't believe that.
but I want to see where I'm wrong.
Because that's the non-consensus.
Yeah, I mean, I think you always need hard guardrails when you want, and you can pair those with softer guardrails, right?
And the ask can be all unmediated.
I'll end with a couple of your commentary on the ecosystem outside of Modo.
Managed agents, everyone has one.
Gemini, OpenAI, Cloud.
Very useful for you.
But also like...
it is their way of starting to edge into your space.
Yeah.
What's going on?
Yeah, I mean, we're very excited to partner with Anthropic and some of the other foundational labs, who we're also working with.
The way we see it is the managed agent thing is a great place to start if you're starting out building an agent.
But then when you get to building something more production-grade, like you're...
company that's like Ramp, that's building their own.
Ramp also runs their accounting agent on us, so their external facing agent.
You need a lot more control over your compute primitive on things like what sort of, how do you persist different files that the agent has access to and how do you snapshot and restore?
How do you control the networking?
Maybe you want GPUs.
When you get to that point, you kind of want a specialized sandbox provider that gives you those things.
And that's the role that we are trying to play.
We don't really have an opinion on the harness, whether it runs in, it's a Cloud Managed Agent, and you hook it up to a little sandbox, we run the harness in a little sandbox.
We'll see where people converge with that.
Yeah.
Do you have any opinions on the meta harnesses?
It's just another layer on top of these things.
You mean like the OpenPi?
OpenPi is one.
I think Vercel had one, which I can't remember the name of right now.
Fred Shot had one.
And then to me, most recently was Databricks that had Omnigin.
All these are sort of meta-hardists, like kind of pseudo-agent cloud type things.
I personally have not played around with them.
I mean, everything is bullish modal as long as it consumes more infra.
That's why we're focusing on the infra there.
It's somewhere where our relative competences and also...
It's a hard problem to solve.
Yeah.
I mean, I will say like just generally reflecting on, I don't know if there's other topics on model, but like just generally reflecting as an infra person, not as intense as you, but in that field, this has like been the most exciting time in infra.
Like it was boring actually for a while and you couldn't really get people excited about data infrastructure.
Like Eric would get on data console.
Everyone just watched the video.
And like, it's like, look at how many sandboxes I can spin up and no one gave a crap.
Yeah.
And like now everyone gives a crap.
That's true.
It is a very exciting time.
And I think a lot of that's driven by just the amount of scale all of this stuff needs.
I think a lot of your initiatives, a lot of your like product directions make sense in retrospect, which is like the best kind.
but I wouldn't necessarily have thought about it myself.
We need the predictions.
I mean, I think there's a lot that you just don't even see, right?
Like you have the batch, you have the voice, you have the multimodal, but what else, you know?
What else is coming up for us?
Where do you see things going?
Yeah, I mean, in general, it's clear that there's a, obviously there's a huge shift happening.
I think one thing that's not, As obvious to people, because LLM inference gets talked about so much, we work with a lot of companies that are doing things like drug discovery and computational bio, like the China discoveries world.
Big things are probably going to happen there.
We work with a lot of robotics companies that are actually putting robots in active deployments and getting good results out of them.
Is there an air gap model?
Is there a version that is on-prem, air gapped, whatever?
No, we...
We use it cloud only.
Yeah.
Okay.
But yeah, I mean, so what you're saying is like, because you're focused on primitives and there are good primitives, you find use cases and all these kinds of things.
Actually probably diversifies you a little bit away from LMs all the time.
Yeah, absolutely.
Our goal isn't to only serve the LMs in Sparkhead.
Just on the website, the audio, the...
Yeah, we've had both on the pod.
Yeah, I mean, there's a lot here.
There's Qtai TTS, customizing, Oh, chatterbox.
You know, there was a customizing whisper.
This screen reminds me of a fallen competitor, which Replicate.
What's your postmortem on what happened?
This is one thing we've kind of stayed away from is providing an API for models.
Because I think providing model APIs is...
Some of it ends up serving like a really hobbyist market, which is much less sticky.
And we've always wanted to build for companies that are building sort of products and need sort of more flexibility that's not just an API.
Which you can build an API for a model, and this is clearly what it is.
But you're saying you can wrap it into a more fully functioning backend that you run.
Yeah, so actually all of our examples, it's not that spin up this model, here's an API token, use it.
They're actually all code.
And so the point is that, this is an example.
Starter code.
Yeah, but you can tweak it however you want.
And if you're like a company building a product, like a computational bio or whatnot, yeah.
I guess I'm trying to tease out for listeners, when does it stop becoming, oh, you're just an API call and you're just a wrapper on an API to becoming what you call a product, right?
Like what is that layer?
Like what, like, you know, obviously more lines of code, but like beyond that, what is the substance that people add?
that qualifies it to be something more?
I think there's a little bit of like a selection effect of like a lot of companies who do want to get deeper into that level are probably building something that's more differentiated.
And I think an example is like we, with LM inference, originally we worked with companies that were building their own post-training frameworks or they were Ramp actually early in the day was training their own tokenizer and swapping out the tokenizer in Lama and whatnot.
I'm not saying that was successful in that case.
A better example is like, let's say Suno, because Suno does not use modal for training.
But they use modal for all their inference, and that's because they have a completely custom modal architecture, and that means that they have to be at the code level.
tweak things that are not, you know, it's an API.
It's interesting as well, like we had Ethan most recently on the XAI Grok team make a prediction that actually like the next tier in video gen is not a better video model.
It's a better model or agent that orchestrates video models.
Language model backbone that can use tools and write.
Like, yes, I can make my six second video or my 10 second video from Grok, but actually I want my six minute video and I'm not going there through normal video gen.
Yeah, that's interesting.
I actually, so we have GPU sandboxes and recently have seen a few companies doing sort of agents that do video manipulation or.
Yeah, give it FFMPEG and just do whatever.
Like that, that's not enough.
You need to give it Adobe.
Yeah, I hadn't put it together with like.
this would actually be a video production thing.
In my mind, these things were going more towards editing.
Yeah.
I think about this a lot.
Obviously.
Luma.
Luma Agent is a version of this for video production, but you know, it's a one-off.
I was going to get your quick takes on some other stuff that happens in recent news and see if you have anything interesting.
Gitpod?
Very, like, somewhat, like, you know.
different market.
They're in the CI, CD market, but actually technically very impressive.
I don't know if you've taken a real look at them.
Yeah.
People on our team have talked to the Gitpod team and they're technically very strong.
I actually am...
We're very bullish and modal on the CI market as well because there's more agents, coding agents.
They're going to run a lot more CI and the prunitives there can be much better.
I think there's a lot of wasted CI.
Yeah.
So is it just like, let's filter?
Like, what is the highest order bit here in improving CI for agents?
Well, there's a lot of wasted time in CI on like preparing your artifacts and like, you know, getting you to the, basically preparing your dependencies and whatnot.
And obviously like build systems help with that.
But like, if you have promoters that are like memory snapshot in the store, can you just run CI more efficiently?
Oh, okay.
Okay.
Okay.
Interesting.
Yes.
I mean, another form of like, you know, on-demand compute.
Yeah, exactly.
Yeah.
It needs the same, again, platform.
Yeah.
So, for those who don't know, Gitpod rebranded to Ona.
It was like, there was this whole thing.
I actually, I like sort of semi-sounded the alarm and cognition.
I was like, you should take these guys seriously because they're infrared very good.
Yeah.
And, but, you know, then they joined OpenAI and presumably we'll see.
codex cloud from the owner team, which I think would be very, very strong.
To me, teams like that, that can set up the networking and the secure boundaries for your agents to have their own cloud each effectively is what you're doing, kind of.
And I'm just trying to draw the analogy or the differences.
If you have studied them, what is the philosophical difference?
My sense is maybe they didn't go off to the right.
market at the right time because we, I guess I also got lucky with like, it's in to give cases really taking off and needing like more of like a sandbox shape thing than like a mind setting is, yeah.
Sandboxes work, like CRCD is sandbox.
Yeah, yeah.
It's just like build time sandboxes versus runtime sandboxes and actually it turned out runtime was better.
Right, and the difference there is runtime sandboxes have a different configuration surface of like, how you configure images, how you attach storage.
Yeah, it's fascinating.
Other people, Astro, also OpenAI, also Python tooling ecosystem people.
Are you still bullish building on top of Python?
Also recently, Modular also got bought by Qualcomm.
Just any of your takes there?
Yeah, I mean, we had Python as our first SDK language because that was the language that people did data and ML in.
I actually now have Go and TypeScript SDKs as well.
And our runtime is completely language-knife.
It isn't in Rust, but it's not tied to Python by any means.
We haven't seen anything with inference and training stuff.
People are still very Python.
And the interesting thing with the agent stuff is people use their TypeScript SDK a lot more because they're not actually doing anything that needs ML.
I don't think we'll have to go beyond that super soon because Python and TypeScript are just still dominant.
The last two languages in the world.
Yeah.
That's it.
Well, English and prompting.
English and prompting.
I occasionally talk to people who try to build new languages.
They're like, even, what's it say, Brett Taylor, who's chairman of OpenAI, I was like, we need a new language for LLM.
So no one has come across one.
And I keep looking.
Python and TypeScript, you have a lot of data plus, but then also they are very imperfect as languages themselves.
Then my close is, I think Modal used to be a big bet on developer experience and you've pivoted the team to agent experience.
Is it like the way now, can entire companies and unicorns, multi-unicorns be built on just having better agent experience?
Do you need something else?
It's a big part of our identity.
It's not just you know, like the very tactical, how does an agent use a CLI?
But it's also how easy is it to spin something up?
Like, what is your iteration time when you want to spin up a new service and you want to get something going in prod?
In practice, that matters a lot to people.
And I think it will continue to matter.
Like, people are building stuff even faster.
give them ways to do it quickly and not have overhead then I think the debate for me has been do you do anything differently that is like very fundamentally different for developer experience versus agent experience you seem to be on the side of they're like this they're like cosine similarity on like 0.9 or whatever yeah I mean pretty much it's the main shift for us has been as I said like we built this benchmark, modal bench, to see where agents are lacking and actually literally add surface areas to a product if they're reaching for something.
Maybe this should just be a CLI.
They hallucinate their own features.
Yeah, and sometimes it makes sense.
If they're reaching for this thing, it's product feedback.
Give it to them.
And then, yeah, actually moving, we used to only have logs and metrics in our UI, just moving all those things to CLI as well.
So they're accessible in that form.
Simple as that.
Cool.
Thank you so much.
Yeah, this is great.
Yeah, it's a great update.
And I can see why you guys have succeeded so much.
It is really focused, but also really good execution.
Thanks.
I mean, we have a long way to go.
All right.
Thank you.