- 8 ore fa
Categoria
📺
TVTrascrizione
00:00It's all about accelerating. The first thing that we are going to be accelerating is to scale our training compute
00:06capacity. We have been building better and better algorithms and better and better techniques, and the team is ready to
00:13scale the techniques that we've been building, so we're excited to train bigger and faster models.
00:18It's also about building more infrastructure. We are deploying technology that is based on open source models, and that allows
00:26to have sovereignty for our customers, and that actually takes a very significant physical footprint, in particular in Europe.
00:34And then finally, we're expanding more and more our go-to-market and our operations with customers, so in Asia
00:40and in the U.S., as well as in Europe.
00:43So part of the proceeds of this fundraising will also be about building further verticalization in manufacturing, as with Samsung
00:53in particular, but also with financial services, where we have been growing significantly in the U.S.,
00:59and defense and the public sector that, again, are very concerned about AI sovereignty and for whom we have very
01:07good solutions.
01:09I want to talk about one of your investors, because fascinating company Samsung coming on board here in a bigger
01:15way.
01:15Last time you had the big investment round last year, led by ASML, it led to a big commercial partnership
01:23between the two of you working on various different ways to integrate your technology with what ASML was doing on
01:29the manufacturing side, et cetera.
01:30What work are you planning to do with Samsung already underway with Samsung?
01:35Yeah, so a lot of it is confidential, but today's announcement is really about the investment itself.
01:41But as you can tell from what we've been doing with ASML, which has been to embed deeply into their
01:47core businesses, into making their machines work better, into designing better things,
01:52these things are the way we work with customers. And so when we work with manufacturing companies, we embed our
02:00teams, we work on their deepest and dearest data,
02:04and we help them build AI systems that can leverage that IP in a way where they feel confident this
02:09IP remains their own and the systems remain their own.
02:12So this is, of course, something we're looking forward to do with companies like Samsung.
02:17And the fact that they are now at the cap table sitting alongside ASML and NVIDIA as representative of the
02:25semi-world is, I think, telling about the focus we have on industrial AI and on making AI useful for
02:34high-end manufacturing.
02:35So a lot of opportunities in further partnering together there, and that investment is indeed the first step toward advancing
02:44technology in the semi-space.
02:46Arthur, you're getting a lot of commercial traction now, some big partnerships, big deals coming through.
02:51I think you previously said Mistral's annual recurring revenue were on track to exceed a billion dollars this year.
02:57Do any of this new funding, new partnerships mean that ARR may be higher than you previously anticipated?
03:04Yeah, I think we are absolutely tracking toward that target and expect to be beating it if everything happens as
03:12they are trending.
03:14The models that are going to come out of Mistral very soon are very competitive.
03:19And so it means that as a European company, you can rely on a European provider providing European models.
03:25But then the question is rather, as a European company, can you use models from China?
03:32So the first answer is, in general, the Chinese labs are actually not operating and not embedding at all and
03:39not working with enterprise customers outside of China.
03:41I think that's something to keep in mind.
03:43Now, in certain cases, the assets that are being put by labs in China and the US and Europe can
03:51be used and deployed on our infrastructure to serve certain customers that we have.
03:56It is not creating a very strong dependency to the AI labs in that, at the end, the data is
04:04staying with us, the customization is made by our engineers, and the sovereignty that our customers are requiring is there.
04:12Now, the question is, can you rely on long-term support for those models?
04:16Can you know that they are going to be updated over time?
04:19Are you sure that in one year from now, they will still be out there and they will still be
04:24improved?
04:24And the answer, to be honest, is no, because you can't make any kind of assumption on how models are
04:31going to be exported.
04:32At this point in time, we are seeing that the volatility in this space is actually quite extreme.
04:37And so what that means for us is that, yes, we provide everything that is open source and that we
04:42can provide, because at the end, this is beneficial for everyone.
04:45Now, we also need to be a trusted partner for our customers, and they want us to be sure to
04:52certify that in one year from now, they will get access to better models than they have access today.
04:58And the only way we can provide that guarantee is by continuing to train our models ourselves.
05:02So there's really two reasons. One is business continuity, making sure that our customers are getting long-term support.
05:08The other is that we are spinning the open source flywheel.
05:10We are putting assets out there. And that means that every labs in the world that have an open source
05:16strategy wants to make something even better.
05:18And Arthur, just on this open source story, I mean, one of the big stories in the past couple of
05:22weeks has been NVIDIA's acquisition of Hugging Face.
05:24Hugging Face has been such an important platform for major open source players around the world, including yourself.
05:31Are you concerned at all about this acquisition, that it could erode the neutrality of Hugging Face and create more
05:38lock-in to NVIDIA potentially down the road?
05:42No, we're not particularly concerned. I mean, we don't really work with Hugging Face a lot.
05:47So that's a place where we host, where we put our models assets. So it's a great place for sharing
05:54technology.
05:55So far, it is not linked to an inference platform. We actually don't know how it's going to evolve.
06:01But overall, for us, Hugging Face is a place where we share assets. It's a place where every lab is
06:07sharing assets.
06:07I think it's great that if they continue to operate that platform, which is super useful for hosting the weights
06:15of models.
06:16But at the end of the day, our customers are not going through Hugging Face to deploy the technology.
06:21And so, hard to comment on that.
06:25I think it's yet another proof that NVIDIA has a very strong interest in seeing open source models succeed.
06:31And it's also the reason why we work with them on the NemoTron Coalition.
06:36Again, we have very strongly aligned interest with cheap providers to make sure that the foundation for building a application
06:44in the enterprise is an open source foundation.
06:47And so, I think if you take that lens, that's very positive news.
06:51And if we're interested in using open source drilling, we must be able to use these tools.
06:52And that's very straightforward.
Commenti