00:00It's so interesting because I think in aggregate, we see a lot of the most established firms
00:08raising funds with some regularity. This is the 18th fund at some scale. So it's an early stage
00:13fund, $1.5 billion. What do we infer from that? Well, Ed, thanks for having us on the show. And
00:20as you mentioned, we just launched our 18th fund. It's a billion and a half dollars dedicated to
00:24early stage AI entrepreneurs. Greylock over the last six decades has been partnering with
00:28entrepreneurs when companies get started, companies like Airbnb, Facebook, Palo Alto Networks.
00:33And we're excited with this new fund to back a new generation of AI entrepreneurs. And as you know,
00:38we've partnered with many of the early AI leaders dating back many years before AI was obvious.
00:43So we rewind the clock to 2019. We backed Base 10. Base 10 today is the leader in AI inference
00:48with,
00:49you know, serving some of the leading application companies like Cursor, Abridge, and others. We
00:53backed Crest AI, one of the leaders in customer service AI. If you call United or Marriott,
00:58today, you're talking to a Crest AI agent. And of course, we're partners with companies like
01:02Anthropic and OpenAI, powering the foundation model layer that's driving this whole economy.
01:06But at the same time, I think one of the benefits of our history is we've seen many of these
01:10technology waves. And when we think about the AI wave and you connect it back to, let's say,
01:14the mobile wave, we're in the equivalent of 2008 or 2009. We're one to two years after the iPhone
01:19moment. And I think we're going to look forward. And many of the defining AI companies of this
01:23generation have yet to get started. And this new fund is all about finding those entrepreneurs
01:27and helping them build those companies. And I think it's interesting what those entrepreneurs
01:30are doing and what kind of companies that they're forming. You, as an example, so you're investors
01:36in OpenAI and Anthropic. You got into Anthropic at Series F and the later rounds. Those are the
01:43leading frontier labs. With this new 18th fund, you are not necessarily going after the model layer.
01:50What is it you're focused on? We think of it as three layers, and we will continue to invest in
01:55all three. So there's the model layer. And as you mentioned, we're investors in both Anthropic and
01:58OpenAI. We're seeing new companies that get built around new data domains or new approaches. And we
02:04continue to look for and will back new companies at that layer. There's a very big opportunity in
02:08infrastructure. The entire stack is going to get rewritten around agents. Today, we have companies
02:13like Base 10 and Inference, Brain Trust and Observability, Snorkel and Data. But we believe we're going to
02:19see a whole new agent cloud get created. And just like we saw happen with the rise of the cloud
02:24and
02:24the hyperscalers in companies like Datadog and MongoDB and Snowflake, we're going to see similar
02:29purpose-built services for these new workloads. And that's an amazing opportunity for entrepreneurs.
02:34And then, of course, on top of that are applications. And within applications, we already have companies
02:38like Creston, Customer Service, or Abnormal, and Cybersecurity. But we're going to see a proliferation
02:43of application companies, especially now that agents actually work. We're six months into the
02:48models that can power long-running agents. And we're seeing amazing results. For example,
02:53we're investors in a company called Resolve AI. They build agentic on-call engineers.
02:57In the last six months, at companies like Coinbase, DoorDash, Fireworks, engineers don't have to wake up in
03:04the middle of the night when there's an incident because the Resolve agent can autonomously handle that
03:07incident. That's an example of agentic AI having huge impact. And we're just getting started in that
03:12era of technology. We need to talk about Moonshot and Kimi K3. The public markets is where you see
03:21the drama. But what is your interpretation of that? And why a $20 billion startup from China
03:29topping a benchmark and people looking at their economics has caused this reaction?
03:35Well, I'd start by saying it was an amazing release yesterday. Largest open weights model
03:40to ever be released, 2.8 trillion parameters. As you said, really strong initial results.
03:44And they also showed a lot of interesting techniques from an efficiency perspective and how they built
03:48that model that's really impressive. And I think it's great that models like that, models like the
03:53one that came out from Thinking Machines the day prior, continue to give vibrancy to the open source
03:58ecosystem, which is an important part of our overall AI economy. At the same time, I think the reaction over
04:04the last 24 hours is perhaps a little bit premature. And I would say, you know, a couple of points.
04:08It
04:08reminds me a little bit of when the deep seek moment happened last year. And when you look at
04:12this new Kimi model, there's maybe a few things, you know, the audience should consider. The first is
04:17benchmarks are imperfect, right? It's, you know, you can take a model and make it very, very strong at a
04:21particular set of benchmarks. I don't think it's until that models had time to percolate in the real
04:25world that we can get a true sense for the trade-offs that the models incurred and what its real
04:29-world
04:30performance looks like. The second is this whole discussion around cost. I think the discussion
04:35misses the point. It's very focused on token cost. Yes. But we think about things more in terms of
04:41task cost. Not every token is equal. And so I'd make two points as it relates to the, you know,
04:45cost profile of Kimi. The first that's actually interesting is Kimi K3 is much more expensive on
04:50a per-token basis than Kimi K2, which is sort of contra to the narrative. And we do not know
04:55what it costs to train Kimi K3. We have an idea on K2. I just want to put that out
05:00there.
05:00Absolutely. That's absolutely correct. So the per-token cost is more, but even more importantly,
05:05it's not particularly token efficient. And so what that means is for a given task,
05:09it actually uses many more tokens than an open AI or anthropic model. And you're seeing that already
05:14in the early cost benchmarking. And so I think this conclusion that it's going to lead to price
05:18erosion for the frontier is perhaps a bit premature. In this case, the frontier, let's say it's
05:22anthropic and open AI, right? You know, some people are making the argument, well, hold on a
05:27minute. If a $20 billion valuation Chinese startup, we don't know the training cost, but they're
05:32basically saying $3 per million tokens on the input side, $15 per million tokens on the output
05:39side. If they can do that, why are we valuing anthropic at nearly a trillion dollars? Like what's
05:45the moat that anthropic has? I think both anthropic and open AI have multiple moats. The first is they
05:52have very significant revenues. These are the fastest growing companies in human history.
05:56And those revenues are not just on the back of their models and their amazing API businesses,
06:00but their first party products, right? And I think relative to the last time I was on the show,
06:04you look at the success that anthropic has had with cloud code, cloud co-work, most recently,
06:08cloud tag. It and open AI are full stack AI companies.
06:13Can I just say one thing? I think for two years, and it's been too long since you've been on
06:17the show,
06:17but we've basically assumed the best model wins. Is it as simple as that? You seem to be saying it's
06:22not. Well, it depends on how you define the best model, right? And I think I would be, again,
06:28I'd be careful to jump to the conclusion that Kimi is now competitive with the best. Certainly on some
06:32of the benchmarks that they release, it's competitive with, I'd say, one generation prior. And of course,
06:38we don't know what's coming out from anthropic and open AI in the coming weeks and coming months,
06:42but it's been rumored and reported that there are significant new releases coming out from those
06:46models. So we have a particular checkpoint in time that we're comparing it to. I think we're going to
06:51see a lot of really exciting releases in the coming months. But maybe, Ed, if I can, I want to
06:55make one
06:55broader point, which is if you just think about the overall size of the token economy and where we're
07:00going, right? In 2023, OpenAI's API was processing per day about 30 million tokens. In March of 2026,
07:08they announced they were doing 15 billion. So just in less than three years, a growth of 50 times.
07:14I think when we are talking in 2030, the overall token economy will be two orders of magnitude larger
07:19than it is today. And so there's going to be plenty of opportunity for the best leading closed frontier
07:25models to grow, for the open source economy to grow, and for the application layer to grow. I think
07:29the only mistake one can make is underestimating the size of this overall wave.
Comments