00:00AI inference startup Base10 has raised $1.5 billion across two tranches, one at an $11 billion valuation, the other
00:07at $13 billion.
00:09Base10 specializes in delivering software and computing capacity to companies tapping open source, typically lower cost AI models.
00:17The firm's co-founder and CEO, Tihin Srivavastra, joins along with one of its investors, Altimeter Capital Partner, Apoor of
00:25Agrawal,
00:25whose firm co-led the round, and actually it was really interesting, the list of participants that backed you.
00:32Let's start with the basic questions that your audience always has, which is, wow, $1.5 billion round is a
00:38lot of funds, but the stories seem to not change, right?
00:41Compute is a constraint, talent is a competitive marketplace. I'm assuming that's where those proceeds go.
00:48Yeah, absolutely. Look, with inference right now, what we are seeing is there's a lot of demand for inserting intelligence
00:55everywhere we possibly can.
00:57There's open source models are getting very good.
00:59Right.
00:59And post-training techniques to specialize those models for specific tasks are getting very good.
01:04And as the app layer comes online, you know, there's an enormous surge of demand.
01:08And for us, what that means is, look, we need to go procure a lot of compute to be able
01:14to fulfill that demand and also hire amazing infrastructure engineers and research engineers to be able to build the software
01:23layer on top of that.
01:23Just real quick, what's that like trying to acquire the compute?
01:28Do you have any leverage in that market? If everyone is supply constrained, how is the experience of being a
01:36company of base 10's age and size going to NVIDIA, as an example, saying we kind of need this?
01:42Yeah, look, we have a very strong relationship with NVIDIA and, you know, they've been very supportive of us over
01:46the years.
01:49Look, one thing that we have realized is that you cannot be reliant on one compute source.
01:53You have to, you know, diversify, not necessarily from a chip perspective, from a cloud perspective.
02:00So, you know, we acquire compute from 18 different clouds now.
02:03We sit in around 90 different clusters.
02:06And that's the flexibility we need to be able to acquire compute and fulfill the demand.
02:10I posted on X that you're both coming on the show and a poor kind of put this question out
02:15there that if you could only invest in one part of the five layer cake, as Jensen calls it, what
02:20would it be?
02:20But Altimus has been so busy, you know, with the frontier labs at one end and other parts of the
02:29five layer cake at the other end.
02:31But what was the thesis around base 10 and why you wanted to get in here to co-lead the
02:36round?
02:37Ed, it's good to be here.
02:39Some truths cannot be said enough times.
02:41And one of those is that inference is going to be one of the largest, if not the largest markets,
02:46not an AI in the world.
02:48And why is that the case?
02:50You know, the early innings of AI were Q&A.
02:53Tell me about my trip.
02:55We've come a long way since then.
02:56You know, it's multi-agent.
02:58They've gotten longer context.
02:59You go retrieve, analyze, synthesize, verify, and then you go again.
03:03And each request is kicking off hundreds, if not thousands, of inference requests.
03:08The second thing is, you know, I might correct something you said, Ed.
03:11It's not just cost.
03:12It's capability, control, and cost.
03:18Frontier models, the post-stream open source model comes very close to frontier model performance, sometimes even better for specific
03:24workflows.
03:26Control, as Satya Nadella said last weekend, you know, you've got to compound your unique advantages as an enterprise, your
03:32data, your knowledge, your know-how,
03:34in a way that you're not giving away your intelligence on rent.
03:38And that's what Basetime allows.
03:40And so to your question from this morning is, like, where does value accrue?
03:43Is it at the app layer?
03:44Is it at the model layer?
03:45You know, the way I like to frame it is, look, the model layer is a phenomenal business for very
03:49few people.
03:50It's a game of emperors.
03:52You've got to be at the frontier to accrue a lot of value.
03:54And it's a knife fight, as you know.
03:56I've started joking there's four seasons to the year.
03:59It's Anthropic, OpenAI, SpaceX, Google, with an evergreen of open source.
04:04And so that's...
04:05That part is so fascinating, right?
04:06If we could just go to open source, when this story started, for us, let's say, 2022, but for you
04:12guys, probably before that,
04:14it was this idea that it was only the model, the largest models with the hundreds of billions of parameters
04:21that mattered.
04:22And in the world of open source, the issue was that it just didn't work.
04:28People couldn't make any real use of them.
04:31I go back to what Apov just said.
04:32You know, you are making utility out of open source models, but probably also at a slightly smaller scale of
04:39model.
04:40Is that fair?
04:41Yes.
04:41Look, I think if you look at, you know, when you think about open source, it's kind of gone through
04:46a bit of a journey.
04:47Right.
04:47You know, there was Lama 3 maybe two and a half years ago, and then there was a bit of
04:52a winter.
04:53And then I think the DeepSeek moment last year really put it back on the map, and there was definitely
04:57the capability gap shrunk.
04:59I think this is happening again just over the last couple of weeks with GLM 5.2, which is, I'd
05:05say, a frontier-level model that is actually usable
05:09and isn't just doing well on the benchmarks.
05:11When people are using it, they are feeling that, you know, at the frontier capability.
05:16For us, it's just like, how do we make these models very, very easy to use?
05:20And it doesn't really matter what size they are.
05:22We work with small models, big models, massive models.
05:25What we are trying to give our customers the ability to run these models when they don't necessarily have access
05:32to the infrastructure teams of the frontier labs doing it for them.
05:35So let's go back to what you were talking about, Apov, cost, but also utility and capability.
05:42You know, the other case study in this open source debate was Meta kind of changing its narrative, saying, well,
05:48actually, we are committed to open source,
05:50but at some point we need to be realistic about the benefit of closed models.
05:57What is the point on the capability part?
06:00And actually, take a moment, explain what Base10 does, why it is such an important player in that respect.
06:07Sure.
06:10Look, yesterday it was DeepSeek.
06:12Today it's GLM.
06:13It'll be something tomorrow.
06:14And as I said, it's seasonal.
06:16And that battle at the model here is not – we are nowhere near the endgame.
06:21What Base10 does is it helps customers like Cursor, like Abridge, like OpenEvidence harness the power of the model underneath,
06:29combine that with your unique advantage as an enterprise, your workflow, and deliver something that is best for that specific
06:36workflow.
06:36Now, it's not just one model.
06:37It's typically a collection of portfolio of models that are fine-tuned to deliver that, and that compounds over time.
06:43And so Base10 is successful as these models get better.
06:47It benefits us.
06:48And ultimately, in the service of the app layer, which is the customers that we discussed, who are delivering AI
06:55to the customer.
06:56Jensen's been repeating a lot recently.
06:59Jensen Wang, the CEO of NVIDIA.
07:00Sometimes you have to remind the audience just in case.
07:03He wants to see his teams burning tokens.
07:10How does that apply to your customers, right?
07:12So Paul has just explained really clearly the value that you're adding, but the economics are really, really important to
07:21some of these companies.
07:23Yeah, look, I think using tokens is obviously just what's synonymous with, hey, I want us to be using AI
07:32everywhere.
07:33Yeah, it's in their interest for that to be the story right now.
07:36And I think for us, what we see is customers kind of go through the same curve over and over
07:42again, is that they go very aggressively.
07:45They start using AI everywhere.
07:47They start to see gains, but they don't necessarily do it profitably.
07:51And then they need to figure out how to do it profitably.
07:53And that's when they come to open source models.
07:55That's when they come to Base10.
07:57That's when they come to post-trained models on Base10 to be able to do it better, faster, and cheaper.
08:02And that's when you decide to get both intelligence everywhere, but also unit economics that makes sense for your business.
08:10What I've been trying to reconcile is Altimeter is invested in the big frontier labs as well.
08:19And those big frontier labs plus SpaceX, depending on how you define what they are, would also say that they
08:25have a role to play in that market.
08:27Where are we in this cycle of whether they're just in the NBA, very big players, where does Base10 fit
08:33in that?
08:34Look, a lot of ink has been spilled on this debate of open versus closed.
08:38Let me just take that head on.
08:40The closed models, the frontier models, are very, very good.
08:43And they're very good for the use case that require highest intelligence, highest reasoning, new use case discovery.
08:50If you didn't want to think about just one model as a service, they're very good, and they will always
08:54be very good for that.
08:58For a class of customers, that's not a long list.
09:00As Jonathan Ross said, there's about 35 companies that drive 99% of all inference.
09:05The top five are the labs that you just listed, top four, top five.
09:08The next 30, that's people like Cursor, like Open Evidence, like Abridge, like Harvey.
09:13And for them, if you're not playing the game of the emperors and you're thinking about how do I deliver
09:20something that is not rented from somebody else but mine that I can compound with, that's where post-trained open
09:27source can help.
09:28And so we posted this chart this morning about how Harvey, for example, achieved frontier capabilities post-training an open
09:37source model.
09:38But obviously with a much better cost and control frontier as an example of what the next 30 are thinking
09:44about and then the next 1,000 are thinking about.
Comments