# From Challenge to Competitive Advantage: How VC Portfolio Companies Leverage NVIDIA to Win Markets

## Abstract

AI venture funding reached $100.8 billion in 2024—37% of all venture capital (VC) investment. Yet, most promising AI startups don’t fail due to weak markets or teams, but from challenges bridging research and production: scaling infrastructure, solving domain-specific problems, and building defensible technical moats. The focus is now on “doing AI right” for real impact. Top VCs now treat NVIDIA partnership as essential portfolio infrastructure—not just for compute, but as a strategic accelerator that helps founders solve hard problems, lead technically, and speed up R&D. This session shares how startups used NVIDIA’s ecosystem to achieve breakthrough results.

## AI Summary

- The CEO of Able discussed the company's focus on making agent tools accessible to business users, emphasizing the importance of consistency and explainability in AI applications.
- David Katz, a partner at Radical Ventures, highlighted the firm's investment in AI pioneers and its approach to forming strategic partnerships with NVIDIA to help startups scale and solve complex technical challenges.
- The CEO of Roboflow shared how the company's participation in NVIDIA's Innovation Lab program enabled the development of real-time vision transformer models, significantly advancing the field of visual understanding.
- Miro Gasparic, CEO of Sensible Biotechnologies, explained the company's mission to optimize mRNA production using AI, and how NVIDIA's compute resources and platforms were crucial in accelerating their research and development.
- The discussion emphasized the importance of building from first principles, focusing on deep, complex, and technical challenges to create sustainable value.
- Advice for startups and investors included the importance of finding patient investors, engaging early with NVIDIA programs, and iterating quickly to adapt to rapid changes in AI technology.

## Transcript

**[00:00:09 – 00:00:11]** Thank you, Sydney, for the intro.

**[00:00:12 – 00:00:15]** As she said, my name is Alice Noland, and we're gonna open

**[00:00:15 – 00:00:20]** with some introductions, give you a brief high level on who these folks

**[00:00:20 – 00:00:24]** are, what their companies do, and then we'll get into a little bit

**[00:00:24 – 00:00:26]** more of the meat of the questions.

**[00:00:26 – 00:00:27]** Fantastic.

**[00:00:27 – 00:00:29]** I'm Orijit. I'm the founder and CEO of Able.

**[00:00:30 – 00:00:32]** Before this, I started something called BeyondCore, which became

**[00:00:32 – 00:00:35]** Salesforce Einstein Discovery, and taught the first AI course

**[00:00:35 – 00:00:38]** in the MBA program at the Harvard Business School.

**[00:00:38 – 00:00:41]** Able is focused on making things like Safe Claw accessible

**[00:00:42 – 00:00:43]** to all business users.

**[00:00:43 – 00:00:45]** So Jensen was talking about how you want to take Open

**[00:00:45 – 00:00:47]** Claw and make it useful.

**[00:00:47 – 00:00:48]** It's not just about security.

**[00:00:48 – 00:00:50]** It's about consistency.

**[00:00:50 – 00:00:54]** How does an agent ensure that two different people asking the

**[00:00:54 – 00:00:56]** same question get the same answer?

**[00:00:56 – 00:00:59]** How do we make sure we explain what happens to the business

**[00:00:59 – 00:01:02]** KPIs when you have teams of agents working together?

**[00:01:02 – 00:01:06]** That's what we solve, and we launch Safeclaw here with

**[00:01:06 – 00:01:12]** Cisco, HP, Dell, seven different partners, Nebius, several others.

**[00:01:12 – 00:01:15]** Thank you. David Katz, I'm a partner at Radical.

**[00:01:15 – 00:01:21]** Radical Ventures is an AI specialized fund that originated

**[00:01:21 – 00:01:27]** out of the research community, the early AI research community

**[00:01:27 – 00:01:31]** in Canada, centered around folks like Jeff Hinton and

**[00:01:31 – 00:01:34]** his disciples and students.

**[00:01:34 – 00:01:35]** Jeff is part of the fund.

**[00:01:35 – 00:01:40]** Together with Jeff and some other pioneers, we co-founded the Vectron

**[00:01:40 – 00:01:48]** Zoo for AI and we are early backers of some of the pioneers like Raquel

**[00:01:48 – 00:01:53]** Erdison and Fei-Fei Li, who's a scientific partner at Radical.

**[00:01:53 – 00:01:56]** People like Richard Socher, Peter Rabeel and others.

**[00:01:56 – 00:02:00]** We will invest as early as helping to incubate companies like we have

**[00:02:00 – 00:02:04]** with Cohere and Recca and Emerald.

**[00:02:04 – 00:02:10]** World Labs and Sentinel will invest right through growth and we've

**[00:02:10 – 00:02:11]** followed on through growth stage.

**[00:02:11 – 00:02:17]** We've also invested fresh capital into new growth companies like we

**[00:02:17 – 00:02:20]** have with Rider and Crusoe Energy.

**[00:02:21 – 00:02:23]** We are also global.

**[00:02:23 – 00:02:26]** We have hubs in Toronto, San Francisco, London, New York.

**[00:02:26 – 00:02:33]** And we love investing in the most ambitious AI founders and have

**[00:02:33 – 00:02:37]** been doing that since 2019 through our institutional funds and for

**[00:02:37 – 00:02:42]** probably a decade earlier as angel investors and as founders as well.

**[00:02:42 – 00:02:45]** I'm super excited to be here.

**[00:02:45 – 00:02:48]** I'm Joseph, co-founder, CEO at Roboflow.

**[00:02:48 – 00:02:51]** At Roboflow, our mission is to make the world programmable,

**[00:02:51 – 00:02:53]** and the way we do that is building tools, platform,

**[00:02:53 – 00:02:57]** and solutions for developers and enterprises to create and deploy

**[00:02:57 – 00:02:59]** systems for visual understanding.

**[00:02:59 – 00:03:02]** So millions of developers and about half the Fortune 100

**[00:03:02 – 00:03:05]** Build with Roboflow for things like Manufacturing Quality

**[00:03:05 – 00:03:10]** Assurance, Logistics Understanding, Supply Chain, Healthcare Analysis.

**[00:03:10 – 00:03:13]** Our core belief is that, in fact, one of the biggest impacts

**[00:03:13 – 00:03:17]** of AI has yet to have happened, and that is applying AI to the

**[00:03:17 – 00:03:20]** real world, and what's increasingly being called physical AI.

**[00:03:20 – 00:03:23]** In fact, if you think about the brain as a biological

**[00:03:23 – 00:03:25]** inspiration for neural networks,

**[00:03:25 – 00:03:28]** About 30% of neurons in the brain are the visual cortex.

**[00:03:28 – 00:03:30]** It's the most of any sense.

**[00:03:30 – 00:03:32]** And if you think about a GPU and the amount of tokens that

**[00:03:32 – 00:03:34]** are processed, there's actually a pretty big disparity of

**[00:03:34 – 00:03:37]** how much compute power is put into problems that are

**[00:03:37 – 00:03:39]** related to visual understanding.

**[00:03:39 – 00:03:42]** And so we think that gap inevitably gets bridged and we build

**[00:03:42 – 00:03:44]** the tooling and infrastructure for that to happen.

**[00:03:44 – 00:03:46]** We are a venture-backed company, and we've had the pleasure

**[00:03:46 – 00:03:49]** of working closely with the NVIDIA team for the last five

**[00:03:49 – 00:03:52]** years since our inception.

**[00:03:52 – 00:03:55]** The other folks that we have the pleasure of building with

**[00:03:55 – 00:03:58]** are also AI luminaries like Jeff Dean and Sam Altman, and

**[00:03:58 – 00:04:02]** backers like YC, Craft Ventures, Google Ventures, and others.

**[00:04:02 – 00:04:05]** So, excited to be here today to share a bit about how we

**[00:04:05 – 00:04:09]** see the future of physical and visual AI unfolding, and how NVIDIA

**[00:04:09 – 00:04:12]** has been a key part of that story.

**[00:04:12 – 00:04:14]** Well, I'm Miro Gasparic, CEO and co-founder of

**[00:04:14 – 00:04:16]** Sensible Biotechnologies.

**[00:04:16 – 00:04:20]** And our vision at Sensible is to enable the full potential

**[00:04:20 – 00:04:22]** of mRNA theropotics.

**[00:04:22 – 00:04:26]** We think that mRNA molecule is the most exciting theropotic

**[00:04:26 – 00:04:31]** modality of the 21st century and could unlock the medicines all

**[00:04:31 – 00:04:36]** the way from cancer treatments to gene therapies for rare diseases.

**[00:04:36 – 00:04:38]** mRNA technology today has some challenges and the

**[00:04:38 – 00:04:41]** biggest challenge is the way how mRNA is made.

**[00:04:42 – 00:04:45]** It's made synthetically using cell-free processes and that

**[00:04:45 – 00:04:50]** has problems on the quality side and on the supply chain side.

**[00:04:50 – 00:04:53]** And at Sensible, we are the world's first company that builds this

**[00:04:53 – 00:05:00]** integrated platform for cell-based mRNA AI-enabled design optimization

**[00:05:00 – 00:05:04]** and manufacturing to unlock the

**[00:05:04 – 00:05:07]** Safe, effective, and affordable and accessible

**[00:05:07 – 00:05:09]** mRNA medicines for everybody.

**[00:05:09 – 00:05:13]** We are based in the U.S., U.K., and in Slovakia.

**[00:05:13 – 00:05:18]** Also venture-backed company, Y Combinator, Blue Yard Capital,

**[00:05:18 – 00:05:21]** Recode Ventures, some of the other biotech luminaries,

**[00:05:21 – 00:05:22]** just co-founder of BioNTech.

**[00:05:22 – 00:05:28]** And we've been working with NVIDIA for over a year and a half.

**[00:05:28 – 00:05:33]** on building our design frameworks for mRNA and our first-in-class

**[00:05:33 – 00:05:34]** therapeutic platform.

**[00:05:34 – 00:05:37]** Excited to be here and share more.

**[00:05:37 – 00:05:39]** Thank you all for making those introductions and giving us

**[00:05:39 – 00:05:43]** a little bit of an overview of what we'll get to dig into.

**[00:05:43 – 00:05:46]** So I want to start with David.

**[00:05:46 – 00:05:54]** So you're our VC on stage with the startups, and I'm really interested

**[00:05:54 – 00:05:59]** to hear your perspective on how you're thinking about that lens

**[00:05:59 – 00:06:02]** of compute versus partnerships and the different kind of partnerships

**[00:06:02 – 00:06:04]** and how you're approaching those.

**[00:06:04 – 00:06:07]** Yeah, thank you for that.

**[00:06:07 – 00:06:10]** So, at Radical, I lead the investment team and our compute

**[00:06:10 – 00:06:13]** practice group, which is part of a program we call our Velocity

**[00:06:13 – 00:06:16]** Program, which is our value-add services team that we help

**[00:06:16 – 00:06:20]** with technical recruiting, compute, go-to-market, public affairs,

**[00:06:20 – 00:06:24]** governance, and other areas, and we have a team structure around that.

**[00:06:24 – 00:06:31]** Compute, we stood up probably five years ago when we discovered

**[00:06:32 – 00:06:34]** very quickly That this was...

**[00:06:34 – 00:06:37]** Anything from existential to emote.

**[00:06:37 – 00:06:39]** Existential, you just can't build what you need to build

**[00:06:39 – 00:06:45]** and accelerate without compute, but emote because if you have the

**[00:06:45 – 00:06:48]** compute, you can attract the best researchers who want to go to where

**[00:06:48 – 00:06:53]** the compute is and you can, with the best talent, continue to create

**[00:06:53 – 00:06:58]** this cycle that's really important of advancement and acceleration.

**[00:06:58 – 00:07:00]** But I think I made a mistake in the early days, and I'm

**[00:07:00 – 00:07:04]** going to sound like a NVIDIA fanboy here because I suspect

**[00:07:04 – 00:07:07]** Jensen got a hold of my notes because there's a lot of overlap

**[00:07:07 – 00:07:09]** between his keynote yesterday and what I'm going to talk about.

**[00:07:09 – 00:07:14]** And that is that in the early days, we used to look at NVIDIA

**[00:07:14 – 00:07:19]** as a chip company that could help our companies win.

**[00:07:19 – 00:07:22]** And I remember even a few years ago talking to the New

**[00:07:22 – 00:07:25]** York Times where I described, and people thought I was crazy, that

**[00:07:25 – 00:07:29]** the subterranean world of AI is actually more exciting than what's

**[00:07:29 – 00:07:32]** happening at the model layer and above, and referred to it almost

**[00:07:32 – 00:07:34]** like a Game of Thrones scenario.

**[00:07:34 – 00:07:39]** But at that time, what I hadn't appreciated was that we are at the

**[00:07:39 – 00:07:44]** front end of a replatforming of the entire globe and the AI economy,

**[00:07:44 – 00:07:46]** and it's gonna hit every industry.

**[00:07:46 – 00:07:51]** And that really makes NVIDIA a platform company.

**[00:07:51 – 00:07:55]** And the sort of five layers that Jensen talked about are

**[00:07:55 – 00:07:57]** incredibly relevant.

**[00:07:57 – 00:08:01]** And so approaching NVIDIA as a platform company opens the doors to

**[00:08:01 – 00:08:05]** so many different ways you can work with them, not just on a compute.

**[00:08:05 – 00:08:10]** And so we typically ask three questions to try to figure

**[00:08:10 – 00:08:13]** out what kind of partnership we want for our companies.

**[00:08:14 – 00:08:17]** The first is, is our company working on something that

**[00:08:17 – 00:08:23]** is like a real technical challenge at the frontier that requires

**[00:08:23 – 00:08:29]** engineering and research assistance to achieve this breakthrough?

**[00:08:29 – 00:08:32]** Can they benefit from that, knowing that NVIDIA has incredible

**[00:08:32 – 00:08:33]** resources in that space?

**[00:08:33 – 00:08:37]** So that's one category of help.

**[00:08:37 – 00:08:40]** I'd say the next one is design partnerships.

**[00:08:40 – 00:08:43]** And so if you are building technology that fits somewhere

**[00:08:43 – 00:08:49]** in these layers of energy, chips, infrastructure,

**[00:08:49 – 00:08:51]** models, applications.

**[00:08:51 – 00:08:56]** And you need to navigate yourself to one of these key partners

**[00:08:56 – 00:08:58]** in that infrastructure stack.

**[00:08:58 – 00:09:02]** NVIDIA can be incredibly helpful in helping to bring you to

**[00:09:02 – 00:09:05]** the data center providers, to the Neo clouds, or whoever it might

**[00:09:05 – 00:09:08]** be, if that's really important to you proving that your technology

**[00:09:09 – 00:09:11]** is gonna work and help scale.

**[00:09:11 – 00:09:14]** And I'd say the last one.

**[00:09:14 – 00:09:17]** Which is a little bit more nuanced is, is what you're

**[00:09:17 – 00:09:21]** building gonna help remove friction from enterprise adoption of AI?

**[00:09:21 – 00:09:25]** And if the answer is yes, you're absolutely gonna have

**[00:09:25 – 00:09:29]** NVIDIA as a partner, and you'll find ways to navigate through

**[00:09:29 – 00:09:32]** that, whether it's the application layer or access to technology

**[00:09:32 – 00:09:34]** that can help you get you there.

**[00:09:34 – 00:09:37]** And then what we do is we overlay those questions on the

**[00:09:37 – 00:09:39]** stack and say, where in the stack

**[00:09:39 – 00:09:43]** Where are you going to actually find that value?

**[00:09:43 – 00:09:49]** And ultimately, what it comes down to is you can get access

**[00:09:49 – 00:09:53]** to compute, as I talked about, which access is getting harder,

**[00:09:53 – 00:09:56]** we're in a compute-constrained world, but not only is access

**[00:09:56 – 00:10:00]** hard, knowing where it's coming from and when it's coming,

**[00:10:00 – 00:10:02]** NVIDIA can really help with that.

**[00:10:02 – 00:10:04]** You can have very close collaborations, like, you

**[00:10:04 – 00:10:08]** know, one of our companies, Leighton Labs, which, you

**[00:10:08 – 00:10:12]** know, just recently announced an acceleration of their molecular

**[00:10:12 – 00:10:18]** biology models by four times in partnership with NVIDIA and Nebius.

**[00:10:18 – 00:10:19]** Like, those kinds of close collaborations, like,

**[00:10:20 – 00:10:23]** how does my model perform better on NVIDIA chips?

**[00:10:23 – 00:10:24]** Very important.

**[00:10:25 – 00:10:27]** But there can be other ones, like, okay, I don't just need

**[00:10:27 – 00:10:30]** compute or need help with working on the models, I don't

**[00:10:30 – 00:10:31]** know what else, but...

**[00:10:32 – 00:10:34]** I need strategic access to compute.

**[00:10:34 – 00:10:39]** I am working like GoHear does quite often in in the field

**[00:10:39 – 00:10:43]** of sovereign AI and I need to know where that compute

**[00:10:43 – 00:10:47]** is as well not just like how to get access to it or in the case of

**[00:10:47 – 00:10:49]** prime intellect they're building a

**[00:10:49 – 00:10:55]** Compute-backed infrastructure for RL, agent training, and inference.

**[00:10:55 – 00:10:56]** If you're compute-backed, you need to know where the

**[00:10:56 – 00:10:59]** compute is, and you need to start speaking with the neoclouds

**[00:10:59 – 00:11:00]** about partnerships for inventory.

**[00:11:00 – 00:11:03]** So that's another example.

**[00:11:03 – 00:11:05]** But if I can just say the last thing, which maybe is

**[00:11:05 – 00:11:10]** the most impactful, is that those things are all great.

**[00:11:10 – 00:11:14]** But if you are working on a technology, and you have

**[00:11:14 – 00:11:18]** the ability to get designed into the reference architecture of

**[00:11:18 – 00:11:23]** this infrastructure build out, that is like home run type of stuff.

**[00:11:23 – 00:11:26]** And that's what we've seen with, like yesterday, there

**[00:11:26 – 00:11:29]** was an announcement, Emerald AI.

**[00:11:29 – 00:11:33]** who was mentioned in the keynote just recently announced a

**[00:11:33 – 00:11:36]** partnership it was either this morning or yesterday

**[00:11:36 – 00:11:41]** about dsx flex and actually being designed into the AI

**[00:11:41 – 00:11:46]** factory architecture so that you can bring flexibility to the grid

**[00:11:46 – 00:11:50]** And we all know energy and AI are on a collision course.

**[00:11:50 – 00:11:51]** And so that's huge for them.

**[00:11:51 – 00:11:56]** And we've seen similar things with Orbital Materials, and they're an

**[00:11:56 – 00:11:59]** AI for materials discovery company.

**[00:12:00 – 00:12:03]** And they're coming up with new solutions and formulations

**[00:12:03 – 00:12:05]** for direct-to-chip cooling.

**[00:12:05 – 00:12:07]** Who can be a better partner?

**[00:12:07 – 00:12:09]** And so those are the kinds of things that we look for

**[00:12:09 – 00:12:10]** in the partnerships.

**[00:12:10 – 00:12:13]** It's super helpful, and I think that some of the things you

**[00:12:13 – 00:12:17]** touched on make a great transition to discussing more in molecular

**[00:12:17 – 00:12:20]** biology and the biotech space.

**[00:12:20 – 00:12:26]** So Miroslav, as SensibleBio went through NVIDIA Innovation

**[00:12:26 – 00:12:30]** Lab, which is a component of some of these strategic opportunities

**[00:12:30 – 00:12:31]** that are available to startups,

**[00:12:32 – 00:12:36]** And VC-backed portfolio companies, can you talk a little bit

**[00:12:36 – 00:12:40]** about, you know, what's that change and what's that unlocked in

**[00:12:40 – 00:12:43]** that exploration of the mRNA space?

**[00:12:43 – 00:12:46]** And even about the data, right?

**[00:12:46 – 00:12:51]** The data is such a core component of how people are able to identify

**[00:12:51 – 00:12:53]** what is now their intellectual property and where some of

**[00:12:53 – 00:12:59]** those opportunities to improve model outputs or accelerated

**[00:12:59 – 00:13:02]** high-performance computing outputs.

**[00:13:02 – 00:13:06]** Well, if you think about it, beefers have to say biology

**[00:13:06 – 00:13:07]** is incredibly complex.

**[00:13:08 – 00:13:09]** I love to use this analogy.

**[00:13:09 – 00:13:14]** If you take a typical kind of mRNA drug candidate, a sequence, there

**[00:13:15 – 00:13:18]** are more ways to kind of engineer that sequence to kind of put

**[00:13:18 – 00:13:23]** these letters together than there are atoms in the universe, right?

**[00:13:23 – 00:13:26]** So iterating one by one and searching for the best one,

**[00:13:26 – 00:13:29]** that's just not possible.

**[00:13:29 – 00:13:32]** What we are trying to do at Sensible is something very unique.

**[00:13:32 – 00:13:36]** We are challenging a 40-year-old paradigm that all the big

**[00:13:36 – 00:13:42]** pharma somehow successfully got to, which is making RNA synthetically.

**[00:13:42 – 00:13:46]** So we are uncovering completely novel biology by taking the

**[00:13:46 – 00:13:51]** cells to make mRNA, and we need to make our own optimal

**[00:13:51 – 00:13:53]** design optimizations.

**[00:13:53 – 00:13:57]** And how this started really was that

**[00:13:57 – 00:14:01]** We wanted to optimize some kind of a design of our technology

**[00:14:01 – 00:14:08]** to make sure that we can get better quality mRNA out of our cells

**[00:14:08 – 00:14:10]** that we are using as producers.

**[00:14:10 – 00:14:13]** And we wanted to do these computational optimizations,

**[00:14:13 – 00:14:16]** and it turned out that just to run one iteration, that would take

**[00:14:16 – 00:14:20]** two weeks, three weeks on CPUs.

**[00:14:20 – 00:14:24]** And then we started working with NVIDIA Innovation Labs

**[00:14:24 – 00:14:27]** and suddenly we've been able to take it down to less than a day.

**[00:14:27 – 00:14:31]** So from something that would be kind of absolutely unviable

**[00:14:31 – 00:14:35]** and unusable in any kind of a practical sense, we were

**[00:14:35 – 00:14:39]** able to get to results that allow us to iterate almost,

**[00:14:39 – 00:14:44]** in biology you could say that's literally iterating in real time.

**[00:14:44 – 00:14:48]** And we think that this is very important because it's

**[00:14:48 – 00:14:52]** very hard to do kind of design from the first principles.

**[00:14:52 – 00:14:54]** You can't do this in some things in biology.

**[00:14:54 – 00:14:59]** You have to just use a bit of a force and brute force

**[00:14:59 – 00:15:01]** to kind of uncover.

**[00:15:01 – 00:15:02]** Those principles.

**[00:15:02 – 00:15:06]** And we started with this one project, but then this started

**[00:15:06 – 00:15:12]** transpiring further into our design optimization manufacturing stack.

**[00:15:12 – 00:15:19]** So we are now building and using NVDS frameworks for optimization of

**[00:15:19 – 00:15:23]** our manufacturing, for optimization of these therapeutic sequences

**[00:15:23 – 00:15:24]** to have better performance.

**[00:15:25 – 00:15:28]** That's all work in progress.

**[00:15:28 – 00:15:32]** Kind of to your point, Alice, I think what really matters is in

**[00:15:32 – 00:15:34]** biotech, the data is a core mode.

**[00:15:35 – 00:15:37]** And for Sensible, the unique thing has been that we are the first

**[00:15:37 – 00:15:43]** company in the world that actually is able to generate the data

**[00:15:43 – 00:15:47]** on naturally made cellular mRNA.

**[00:15:49 – 00:15:52]** And we can then measure its performance, its functionality,

**[00:15:52 – 00:15:55]** and build this kind of a data flywheel where you're screening

**[00:15:56 – 00:15:58]** more and more of the mRNA, get more data on a performance,

**[00:15:59 – 00:16:03]** and then can go back and use this to improve your design, so creating

**[00:16:03 – 00:16:06]** this virtuous feedback loop.

**[00:16:06 – 00:16:10]** And we've built a high-tropical screening pipeline for that.

**[00:16:10 – 00:16:12]** I think that a transition that we've been able to do

**[00:16:12 – 00:16:17]** is to start moving from a publicly available data and

**[00:16:17 – 00:16:21]** our optimization starting there towards now generating proprietary

**[00:16:21 – 00:16:28]** data and using NVIDIA's algorithms and capabilities for this purpose.

**[00:16:28 – 00:16:32]** So that's been kind of really a game changer

**[00:16:32 – 00:16:34]** and the fully integrated

**[00:16:34 – 00:16:39]** A company that aims to capture volume across mRNA, therapeutics,

**[00:16:39 – 00:16:43]** design optimization, manufacturing and development would just not

**[00:16:43 – 00:16:47]** be possible in ages before NVIDIA.

**[00:16:47 – 00:16:49]** Thank you so much.

**[00:16:49 – 00:16:53]** And I think another great transition in talking about

**[00:16:53 – 00:16:56]** design from first principles as you approach some of these

**[00:16:56 – 00:17:04]** systems is to hear from about what Able has done to approach

**[00:17:04 – 00:17:06]** the software stack, right, going from a design of first

**[00:17:06 – 00:17:10]** principles, and how y'all are then able to take advantage of some

**[00:17:10 – 00:17:14]** of the compute that's less hotly demanded, because folks haven't

**[00:17:14 – 00:17:18]** necessarily explored that sector as well.

**[00:17:18 – 00:17:20]** Absolutely.

**[00:17:20 – 00:17:22]** In fact, to create a little bit of controversy, just to

**[00:17:22 – 00:17:25]** understand that everything that we talk about typically

**[00:17:25 – 00:17:28]** in the world of Gen AI right now is not actually designed for agents.

**[00:17:28 – 00:17:32]** So the moment you hear anybody say time to first token, they haven't

**[00:17:32 – 00:17:33]** thought about agents, right?

**[00:17:33 – 00:17:35]** That's a chat-oriented mindset.

**[00:17:35 – 00:17:39]** So, NVIDIA brought us the first, actually, Grace Hopper,

**[00:17:39 – 00:17:40]** way before the GB10.

**[00:17:40 – 00:17:43]** When the Grace Hopper first came out, the Innovation Lab

**[00:17:43 – 00:17:46]** guys basically said, hey, we have this thing you can hit on

**[00:17:46 – 00:17:49]** Launchpad, do you want to just tell us what's interesting about it?

**[00:17:49 – 00:17:52]** We started playing with it, and what we found was...

**[00:17:52 – 00:17:55]** That if you could put the agent tools, the agent memory, the

**[00:17:55 – 00:18:00]** agent coordination on the Grace, the CPU, and run all the models

**[00:18:00 – 00:18:04]** on the hopper, and you really focused on agent optimization,

**[00:18:04 – 00:18:08]** as opposed to model optimization, you get some really good results.

**[00:18:08 – 00:18:10]** And then the GB10 came out.

**[00:18:10 – 00:18:13]** And the GB10 is this dinky little supercomputer.

**[00:18:13 – 00:18:14]** You guys should see it.

**[00:18:14 – 00:18:17]** It's really cool. You should come to our booth, go to any

**[00:18:17 – 00:18:18]** of the NVIDIA ones.

**[00:18:18 – 00:18:20]** But what is interesting about it is people were

**[00:18:20 – 00:18:22]** saying it's really slow.

**[00:18:22 – 00:18:25]** What they did is they would go in and hit it, and the

**[00:18:25 – 00:18:27]** GB10 is optimized for pre-fill.

**[00:18:27 – 00:18:30]** So if you put in a thousand tokens in and take ten tokens

**[00:18:30 – 00:18:32]** out, it'll perform really well.

**[00:18:32 – 00:18:35]** If you put ten tokens in and take a thousand tokens out,

**[00:18:35 – 00:18:36]** it'll perform really badly.

**[00:18:36 – 00:18:38]** It's not designed for chat.

**[00:18:38 – 00:18:41]** And it's actually really good at parallelization.

**[00:18:41 – 00:18:44]** Like, the performance curve is very good when you're running 10,

**[00:18:44 – 00:18:46]** 20 different things in parallel.

**[00:18:46 – 00:18:50]** So, we really had to look at that and say, no, again, this is

**[00:18:50 – 00:18:55]** really perfect for agents because in agents, my pre-fill, the context

**[00:18:55 – 00:18:58]** is much bigger than the output.

**[00:18:58 – 00:19:02]** And again, Nimotron, the reasoning part of it, but the tool calling

**[00:19:02 – 00:19:03]** aspect was really good.

**[00:19:03 – 00:19:06]** The problem with NVIDIA, and it's a positive problem for

**[00:19:06 – 00:19:10]** us, is NVIDIA doesn't come and tell you, this is designed for agents.

**[00:19:10 – 00:19:12]** They will come in and tell you, I have something interesting

**[00:19:12 – 00:19:15]** and new, what can you do with it?

**[00:19:15 – 00:19:17]** And that's where the first principle of thinking becomes key.

**[00:19:17 – 00:19:18]** You've got to test it out.

**[00:19:18 – 00:19:22]** Like, if you're just using a layer on top, it doesn't really matter.

**[00:19:22 – 00:19:27]** So at an NVIDIA webinar we did recently,

**[00:19:27 – 00:19:30]** We were able to show that the exact same agent running

**[00:19:30 – 00:19:35]** on a small little GB10 performs faster than the same agent running

**[00:19:35 – 00:19:37]** on Google Cloud, for example.

**[00:19:37 – 00:19:40]** Even though the model is much faster on Google Cloud.

**[00:19:40 – 00:19:42]** Does this make sense? The end-to-end agent is

**[00:19:42 – 00:19:45]** much faster on the little device because everything

**[00:19:45 – 00:19:47]** is talking synchronously.

**[00:19:47 – 00:19:52]** And last year we showed you could get 30x better GPU utilization

**[00:19:52 – 00:19:56]** if you start thinking this way and use all of that parallelization

**[00:19:56 – 00:20:00]** to get more bang for the buck.

**[00:20:00 – 00:20:03]** The result of this has been our customers are being able

**[00:20:03 – 00:20:08]** to go in and look at 5 billion rows of data, 10 million variable

**[00:20:08 – 00:20:11]** combinations, trying to figure out what's changing in their

**[00:20:11 – 00:20:16]** customer purchase behavior, looking at patient no-show

**[00:20:16 – 00:20:18]** patterns for Baptist Health.

**[00:20:18 – 00:20:22]** But the level of scale of analysis that can be done once you've

**[00:20:22 – 00:20:24]** optimized these agents this way,

**[00:20:24 – 00:20:26]** It's very different than what you would do if you were hitting

**[00:20:26 – 00:20:29]** it with a chat bot and asking questions, right?

**[00:20:29 – 00:20:33]** Because an AI can ask a million questions in the time a human

**[00:20:33 – 00:20:34]** can ask seven questions.

**[00:20:34 – 00:20:38]** So why are you starting from a human-first mindset of natural

**[00:20:38 – 00:20:42]** language querying, start with an agent-first mindset and optimize it

**[00:20:42 – 00:20:43]** all the way to the hardware stack?

**[00:20:43 – 00:20:45]** Hopefully that resonates with you guys.

**[00:20:45 – 00:20:49]** We've got to think, we can't just slap some harness on top

**[00:20:49 – 00:20:51]** of a model and call it an agent.

**[00:20:51 – 00:20:53]** You've got to think from the ground up.

**[00:20:53 – 00:20:55]** How do you optimize it?

**[00:20:55 – 00:20:56]** Great insight.

**[00:20:56 – 00:20:59]** And there's so much that resonates there.

**[00:20:59 – 00:21:05]** And yes, the way that you... We often personify, right, the things

**[00:21:05 – 00:21:06]** that we are interacting with, the things that we're building,

**[00:21:06 – 00:21:09]** and the abstraction layers, as we get further and further

**[00:21:09 – 00:21:11]** from them, we can lose sight.

**[00:21:12 – 00:21:15]** One of the things I'm also really interested in is, Joseph,

**[00:21:15 – 00:21:19]** as I understand it, getting access to some of the Innovation

**[00:21:19 – 00:21:21]** Lab program and...

**[00:21:21 – 00:21:26]** The pivot evolution, probably more accurate, for where Roboflow

**[00:21:26 – 00:21:30]** was able to then explore in terms of the technology that

**[00:21:30 – 00:21:35]** y'all were developing and I would love to hear more on some of

**[00:21:35 – 00:21:39]** that shift and that tipping point and what that unlocked for y'all.

**[00:21:39 – 00:21:43]** Yeah, maybe I can tell this is a bit of a story because GTC

**[00:21:43 – 00:21:48]** last year, we were like sweating bullets like a week beforehand.

**[00:21:48 – 00:21:49]** And I can tell you why.

**[00:21:49 – 00:21:52]** So Roboflow, we make it so folks can deploy systems for

**[00:21:52 – 00:21:53]** visual understanding.

**[00:21:53 – 00:21:56]** And we've always supported a wide array of any open source

**[00:21:56 – 00:21:59]** model that someone wants to bring and use on Roboflow.

**[00:21:59 – 00:22:03]** So it's like ResNets, it's CNNs, vision transformers, increasingly.

**[00:22:03 – 00:22:10]** And at NeurIPS 2024, so December 2024, we gave a talk on the

**[00:22:10 – 00:22:11]** state of vision AI.

**[00:22:11 – 00:22:13]** And why does it seem like transformers have had such

**[00:22:13 – 00:22:17]** a massive impact in language, but they seem to lag in vision?

**[00:22:17 – 00:22:21]** The attention you all need paper is 2017, vision transformers are 2020.

**[00:22:21 – 00:22:23]** What gives? Because early on in machine learning, you

**[00:22:23 – 00:22:26]** had ImageNet, and it was sort of like image first, and then

**[00:22:26 – 00:22:28]** language kind of took over.

**[00:22:28 – 00:22:31]** And we did a deep analysis and we identified that, in fact,

**[00:22:31 – 00:22:36]** vision transformers can be more accurate than CNN-based approaches.

**[00:22:36 – 00:22:41]** And some vision transformers can also be made to be faster.

**[00:22:41 – 00:22:43]** But no one had actually taken the combination of these two

**[00:22:43 – 00:22:45]** things and combined them together.

**[00:22:45 – 00:22:48]** And Roboflow, as you described, for the longest time,

**[00:22:48 – 00:22:50]** supports any sort of open source architecture, but didn't

**[00:22:50 – 00:22:52]** publish any of our own models.

**[00:22:52 – 00:22:55]** And so here we are at NeurIPS in 2024, and we're like, clearly this

**[00:22:55 – 00:22:59]** is possible, like somebody should combine together the benefits

**[00:22:59 – 00:23:01]** of how accurate transformers are, use the large pre-trained

**[00:23:02 – 00:23:04]** context, and someone else should use the capabilities that

**[00:23:04 – 00:23:07]** we've seen for acceleration that can exist if you use a smaller head

**[00:23:07 – 00:23:09]** and smaller backbone for the model.

**[00:23:09 – 00:23:10]** And everyone's like, yeah, someone should do that.

**[00:23:10 – 00:23:14]** We're like, all right, all right, I guess that someone

**[00:23:14 – 00:23:16]** is us, you know, the best way to complain is to make things.

**[00:23:16 – 00:23:19]** And so we heard about the Innovation Lab program that

**[00:23:19 – 00:23:23]** NVIDIA has, and the promise is sort of like, hey, if you have

**[00:23:23 – 00:23:25]** a cool idea that you think some amount of compute can meaningfully

**[00:23:25 – 00:23:29]** change the field and grow the ecosystem, and to David's point,

**[00:23:29 – 00:23:32]** one of the key ways of partnering with NVIDIA is what allows

**[00:23:32 – 00:23:35]** moving the entire field forward, and NVIDIA really does live the

**[00:23:35 – 00:23:37]** value of what grows the ecosystem.

**[00:23:37 – 00:23:40]** And we said, hey, we think we can introduce an Apache

**[00:23:40 – 00:23:43]** 2 model that will be the best in the world and first real-time

**[00:23:43 – 00:23:46]** instant segmentation transformer, is what we promised NVIDIA, right?

**[00:23:46 – 00:23:49]** And they're like, great, well, we can give you a cluster

**[00:23:49 – 00:23:50]** from now until GTC.

**[00:23:50 – 00:23:54]** And so you have December 2024, GTC March 2025.

**[00:23:54 – 00:23:57]** And for a lot of these labs, when you first introduce

**[00:23:57 – 00:24:00]** an architecture, it may be that pre-training alone

**[00:24:01 – 00:24:02]** takes three months.

**[00:24:02 – 00:24:04]** And now there's a lot of release cycles, but that's because

**[00:24:05 – 00:24:07]** a lot of these training runs have been running in advance.

**[00:24:07 – 00:24:10]** So we start from a standing start and start trying out

**[00:24:10 – 00:24:12]** various different iterations.

**[00:24:12 – 00:24:16]** And I remember the weeks leading up to GTC, because we promised

**[00:24:16 – 00:24:18]** that we would be able to announce it at GTC, that we've done this,

**[00:24:18 – 00:24:23]** that we've introduced real-time transformer architectures.

**[00:24:23 – 00:24:25]** We were figuring out which training runs, which heads, and ultimately

**[00:24:25 – 00:24:29]** reintroduced Roboflow Detectioner and Transformer, RF Detter,

**[00:24:29 – 00:24:33]** which is the first real-time instant segmentation model.

**[00:24:33 – 00:24:35]** The reason that matters is when there's companies that

**[00:24:35 – 00:24:38]** are increasingly using larger and larger data centers, where like

**[00:24:38 – 00:24:40]** ZIG, when everyone else is ZAG, people want to run their vision

**[00:24:40 – 00:24:43]** models on the edge, on robots, on manufacturing lines, where

**[00:24:43 – 00:24:45]** you don't have an infinite amount of compute, but you still want

**[00:24:45 – 00:24:47]** real-time accurate performance.

**[00:24:47 – 00:24:49]** And in the evals, it actually retook

**[00:24:49 – 00:24:50]** state-of-the-art for the U.S.

**[00:24:50 – 00:24:52]** Previously, it was Baidu, who spends about$

**[00:24:52 – 00:24:54]** 3 billion a year on R&D.

**[00:24:54 – 00:24:57]** And so this tiny company, Roboflow, has put the U.S.

**[00:24:57 – 00:25:01]** back on the map in terms of beating global competition for some

**[00:25:01 – 00:25:03]** of the most important tasks in AI.

**[00:25:03 – 00:25:05]** And in fact, the thing that gave us the nudge to do that,

**[00:25:05 – 00:25:09]** to give us the willingness to make the bet, was, hey, we know this

**[00:25:09 – 00:25:13]** is gonna be a massive compute cost, we think this is gonna bear fruit,

**[00:25:13 – 00:25:16]** we're not sure, and the innovation lab from NVIDIA was like,

**[00:25:16 – 00:25:17]** try it, just see what happens.

**[00:25:17 – 00:25:19]** Like, the worst that happens is we invalidate the hypothesis

**[00:25:19 – 00:25:22]** and the best that happens is this new ecosystem of products.

**[00:25:22 – 00:25:25]** And to your point, now we've kind of doubled down on that,

**[00:25:25 – 00:25:27]** expanded the head, expanded the task types, expanded the

**[00:25:27 – 00:25:31]** model sizes, and it's become a core reason that a lot of folks

**[00:25:31 – 00:25:34]** choose to build with the tools that we create as not just this Apache

**[00:25:34 – 00:25:37]** 2 architecture, but techniques like neural architecture search, new

**[00:25:37 – 00:25:41]** heads, new backbones, and spoiler alert, now we spend even more

**[00:25:41 – 00:25:44]** on GPUs with NVIDIA because the technology works so well that we

**[00:25:44 – 00:25:45]** want to scale up our experiments.

**[00:25:45 – 00:25:51]** And so the nudge and the benefit of NVIDIA really thinking like what

**[00:25:51 – 00:25:54]** allows the ecosystem to grow and if you have something that you want

**[00:25:55 – 00:25:59]** to try, go for it, gave us the bump to be able to do it even though

**[00:25:59 – 00:26:03]** it kind of came down to the wire with the release at GTC last year.

**[00:26:03 – 00:26:06]** Well, thank you for recounting that story.

**[00:26:06 – 00:26:11]** David, now that we've heard from three great startups

**[00:26:12 – 00:26:15]** and in various verticals, you know, going back to your earlier point

**[00:26:16 – 00:26:20]** around, hey, especially if you're going into enterprise, you know,

**[00:26:20 – 00:26:25]** As you hear these stories, not to put you on the spot,

**[00:26:25 – 00:26:29]** what stands out to you in terms of, you know, what they're doing, kind

**[00:26:29 – 00:26:32]** of like what you would be looking for as an investor, and some

**[00:26:32 – 00:26:37]** of that, like, enterprise audience focus, that B2B, like, opportunity?

**[00:26:37 – 00:26:41]** Yeah, it's interesting, and I'll approach that a little differently,

**[00:26:41 – 00:26:46]** and that is not from the context of investment, but from the context of

**[00:26:46 – 00:26:50]** the compute practice group, and the reason I say that is when we stood

**[00:26:50 – 00:26:54]** up the compute practice group, one of the things that we wanted

**[00:26:54 – 00:26:57]** to do was not try to provide value to our portfolio companies by going

**[00:26:57 – 00:27:02]** and buying a cluster and managing that, but rather demonstrating

**[00:27:02 – 00:27:05]** to our cloud partners at the time.

**[00:27:05 – 00:27:10]** that radical portfolio companies are somewhere in the area

**[00:27:10 – 00:27:13]** of an 80% perfect customer for them and what that means

**[00:27:14 – 00:27:17]** are a they're buying compute and B they're helping to solve

**[00:27:17 – 00:27:21]** a problem that those cloud partners have not been able to solve

**[00:27:21 – 00:27:25]** Because maybe the talent is not sitting inside their companies,

**[00:27:25 – 00:27:29]** but they're building things that will be highly valuable to the

**[00:27:29 – 00:27:31]** enterprise customers of the clouds.

**[00:27:31 – 00:27:34]** And I think we did that really well in how we chose the companies

**[00:27:34 – 00:27:39]** to invest in, and all the stories I'm hearing are super consistent

**[00:27:39 – 00:27:41]** with what I'm talking about.

**[00:27:41 – 00:27:45]** But then we realized something really interesting.

**[00:27:45 – 00:27:49]** As we were going through these motions of procuring, and

**[00:27:49 – 00:27:52]** I think it's billions of dollars now of compute for our portfolio

**[00:27:52 – 00:27:56]** companies, we also found that these partners would be great

**[00:27:56 – 00:27:58]** go-to-market partners.

**[00:27:58 – 00:28:03]** If we can integrate our portfolio companies into those platforms.

**[00:28:03 – 00:28:06]** And so we went on a quest to do that, and we had a lot of success.

**[00:28:06 – 00:28:09]** A lot of our companies are integrated into those platforms.

**[00:28:09 – 00:28:12]** But what we didn't realize or appreciate as much is,

**[00:28:12 – 00:28:14]** those are really big companies.

**[00:28:14 – 00:28:17]** Like, getting those solutions in the hands of the enterprise

**[00:28:17 – 00:28:20]** customers requires a lot.

**[00:28:20 – 00:28:23]** It's a lot of, you know, looking at comp plans of salespeople, and how

**[00:28:23 – 00:28:26]** to interact with marketplaces, and this and that and everything, and

**[00:28:26 – 00:28:30]** all the model wars and chip wars and everything that come with it.

**[00:28:31 – 00:28:33]** Whereas when we started working with NVIDIA, and it's probably

**[00:28:33 – 00:28:37]** been a couple of years now,

**[00:28:37 – 00:28:40]** I think we have about a dozen deals together where NVIDIA

**[00:28:40 – 00:28:44]** has invested in our companies or we've co-invested together.

**[00:28:44 – 00:28:48]** The fit is closer to a hundred percent, and that's because

**[00:28:48 – 00:28:52]** the NVIDIA platforms are open and they're collaborative,

**[00:28:52 – 00:28:55]** as opposed to some of the clouds that run these closed platforms

**[00:28:56 – 00:28:58]** that are very proprietary.

**[00:28:58 – 00:29:01]** And so this comes back to the point I was saying earlier

**[00:29:01 – 00:29:06]** about finding that fit with NVIDIA is actually not hard

**[00:29:06 – 00:29:11]** at all if what you're trying to do is drive enterprise adoption.

**[00:29:11 – 00:29:14]** Where it's really interesting that working with NVIDIA,

**[00:29:14 – 00:29:17]** like we like it when they invest in our companies, it

**[00:29:17 – 00:29:21]** drives like alignment.

**[00:29:21 – 00:29:25]** And we find that NVIDIA has this bias towards action,

**[00:29:25 – 00:29:31]** which is once they get involved, they want to demonstrate value,

**[00:29:31 – 00:29:34]** which has been really helpful.

**[00:29:34 – 00:29:40]** I'd say Howard, Sydney, James's team, and being able to navigate

**[00:29:40 – 00:29:43]** NVIDIA and find the people that are the best partners.

**[00:29:43 – 00:29:46]** They've been just great to work with.

**[00:29:46 – 00:29:51]** And of course, Vishal and Alan on the corp dev team have been

**[00:29:51 – 00:29:52]** incredible on the investment side.

**[00:29:52 – 00:29:55]** You always know where you stand with them, and they're

**[00:29:55 – 00:29:57]** always very helpful.

**[00:29:57 – 00:30:02]** And so that sort of match of what you're investing in, is it actually

**[00:30:02 – 00:30:05]** going to help push the boundaries?

**[00:30:05 – 00:30:10]** And if so, then we find that it's an investable.

**[00:30:10 – 00:30:13]** It's an investable opportunity, and then what we look at is,

**[00:30:13 – 00:30:15]** is it an enormous opportunity?

**[00:30:15 – 00:30:17]** Are you building something that's just a feature, or

**[00:30:17 – 00:30:21]** are you building a giant business like these folks are trying to do?

**[00:30:21 – 00:30:25]** I am intimately familiar with those long enterprise sales

**[00:30:25 – 00:30:28]** cycles that you are referring to and some of those challenges.

**[00:30:28 – 00:30:30]** Absolutely, right, you do.

**[00:30:30 – 00:30:33]** There's an existential challenge that can present of, are you

**[00:30:33 – 00:30:37]** moving fast enough and getting enough to help sustain you

**[00:30:37 – 00:30:41]** so that you have the runway so that you can get to the enterprise?

**[00:30:41 – 00:30:43]** Inception helps a ton with that.

**[00:30:43 – 00:30:47]** Yeah, and Inception is easily one of my favorite teams to

**[00:30:47 – 00:30:49]** work with at NVIDIA.

**[00:30:49 – 00:30:51]** Can I tell a story about long enterprise sales cycles

**[00:30:51 – 00:30:53]** then, very quickly?

**[00:30:53 – 00:30:55]** Since we have seven and a half minutes, what I want to do

**[00:30:55 – 00:31:00]** is a lightning round of hearing...

**[00:31:00 – 00:31:04]** Something that you have, maybe it's recently announced, something that

**[00:31:04 – 00:31:07]** you've been working on with NVIDIA.

**[00:31:07 – 00:31:11]** One sentence, two sentence, we wanna give everyone a chance

**[00:31:11 – 00:31:13]** to go down the row so we can do all four of you.

**[00:31:14 – 00:31:18]** Okay, so Safeclaw is the way you can get these

**[00:31:18 – 00:31:21]** long-running enterprise agents going really quickly.

**[00:31:21 – 00:31:24]** We launched it with NVIDIA, but we also launched it on

**[00:31:24 – 00:31:27]** the DGX Park, and last week we were selling this at Gartner.

**[00:31:27 – 00:31:30]** We had Fortune 500 customers buying at the booth.

**[00:31:30 – 00:31:34]** I've never seen that happen because it runs air-gapped on a device,

**[00:31:34 – 00:31:35]** so they can prototype very easily.

**[00:31:35 – 00:31:38]** They don't need to have the conversation with the seesaw, and

**[00:31:38 – 00:31:41]** once they have prototyped, they can push it out to Nebius, to any of

**[00:31:41 – 00:31:45]** the clouds, to HP, Dell, et cetera.

**[00:31:46 – 00:31:48]** For us, I'd say I'm going to take a portfolio view.

**[00:31:48 – 00:31:52]** If you look at all the companies in our portfolio, if you assembled

**[00:31:52 – 00:31:55]** them together, you probably would have all the software

**[00:31:55 – 00:31:58]** you would need for a cloud.

**[00:31:59 – 00:32:02]** And so one of the things we're most excited about is working with all

**[00:32:02 – 00:32:08]** of NVIDIA's cloud partners to find where are the missing pieces in

**[00:32:08 – 00:32:12]** their offering today where we can plug in in a modular way and help

**[00:32:12 – 00:32:14]** build out that whole ecosystem.

**[00:32:14 – 00:32:16]** Wonderful.

**[00:32:16 – 00:32:20]** One of the things I'm most excited about is we've been

**[00:32:20 – 00:32:24]** firmly of the belief that AI in the real world of understanding

**[00:32:24 – 00:32:30]** real world context is as, if not more important, than only language.

**[00:32:30 – 00:32:34]** And it seems like the pendulum's kind of swung to agree that that's

**[00:32:34 – 00:32:36]** the case on a go-forward basis.

**[00:32:36 – 00:32:39]** And as a company that lives at the infrastructure level for powering

**[00:32:39 – 00:32:43]** all of the, you know, a billion inferences every seven days,

**[00:32:43 – 00:32:45]** or tens of thousands of folks that are training models every week,

**[00:32:45 – 00:32:51]** The explosion of enabling builders and enterprises to deploy visual

**[00:32:51 – 00:32:55]** and physical AI, it's a trend even bigger than us that we're fortunate

**[00:32:55 – 00:32:59]** to be able to accelerate, and a lot of the vision's GPT moment,

**[00:32:59 – 00:33:03]** if you will, of prompt and describe the outcome you want, is arriving,

**[00:33:03 – 00:33:06]** given the quality and increasing improvement of the models.

**[00:33:07 – 00:33:10]** And so this is something that I've been very excited about, and I've

**[00:33:10 – 00:33:14]** been building a company around for five years, and the fact that now

**[00:33:14 – 00:33:16]** It feels like the rest of the world agrees.

**[00:33:16 – 00:33:19]** It feels very much like go time.

**[00:33:20 – 00:33:23]** Well, I'm a bio-engineer and what we care about a lot at

**[00:33:23 – 00:33:27]** Sensible is, well, we can't fully understand the rules of biology,

**[00:33:27 – 00:33:31]** but we want to be able to get to the point where we can tweak our

**[00:33:31 – 00:33:35]** systems to get the best possible mRNA drugs out of our platform.

**[00:33:35 – 00:33:37]** So two things I'm excited about.

**[00:33:37 – 00:33:38]** Thing number one.

**[00:33:38 – 00:33:43]** We recently built up a model that should allow us to really

**[00:33:43 – 00:33:46]** improve the quality of the mRNA we can get out of the

**[00:33:46 – 00:33:50]** cells and we are now experimentally testing that and if that works,

**[00:33:50 – 00:33:52]** that's going to be really huge.

**[00:33:52 – 00:33:54]** And the thing number two is we are actually now building

**[00:33:54 – 00:33:59]** kind of a virtual metabolic model of our producer cells that should

**[00:33:59 – 00:34:04]** allow us to vastly improve kind of the amount of mRNA we can get out.

**[00:34:04 – 00:34:08]** Ultimately, make mRNA cheaper, make medicines more affordable.

**[00:34:08 – 00:34:10]** And we are very glad that this is kind of a three-way

**[00:34:10 – 00:34:14]** work, but for us, using the NVIDIA's compute resources, using

**[00:34:14 – 00:34:19]** our own kind of proprietary cells and work, but doing this in a...

**[00:34:19 – 00:34:22]** And machines and kind of cutting-edge bioreactors from

**[00:34:22 – 00:34:27]** our partners at Sartorius, which is actually NVIDIA's partner as well.

**[00:34:27 – 00:34:33]** So these are two super exciting projects we are working on,

**[00:34:33 – 00:34:36]** and we are looking forward to seeing the results soon.

**[00:34:36 – 00:34:40]** It's great to hear all the news that y'all have present

**[00:34:40 – 00:34:44]** or on the horizon, and I love thinking about part of the

**[00:34:44 – 00:34:49]** way that you got there as well in the explorations and what

**[00:34:49 – 00:34:51]** opportunities have been unlocked.

**[00:34:51 – 00:34:53]** With our last three and a half minutes,

**[00:34:53 – 00:34:57]** What advice do you have for the folks in the audience,

**[00:34:57 – 00:35:00]** whether they are thinking about finding some of the strategic

**[00:35:00 – 00:35:05]** advantage and technical moat for themselves, whether it's VCs that

**[00:35:05 – 00:35:10]** are thinking about the potential opportunity, their portcos?

**[00:35:10 – 00:35:15]** Again, three minutes left for the whole talk, so anyone

**[00:35:15 – 00:35:18]** want to go first?

**[00:35:18 – 00:35:21]** Well, for the VCs, we are about to raise our first

**[00:35:21 – 00:35:23]** institutional round.

**[00:35:23 – 00:35:26]** The reason we didn't take money from VCs is if you really

**[00:35:26 – 00:35:29]** want to do something very deep and it'll take some time, it was

**[00:35:29 – 00:35:32]** hard for us to find the investors who are going to be patient.

**[00:35:32 – 00:35:36]** So we took 50 mil, but from angels, not from VCs.

**[00:35:36 – 00:35:38]** Now that we have proven it, it's a different story.

**[00:35:38 – 00:35:41]** But I think there is a disconnect here.

**[00:35:42 – 00:35:44]** Shiny stuff is easy to get funded.

**[00:35:44 – 00:35:47]** Boring, hard stuff actually creates the value.

**[00:35:47 – 00:35:50]** That's what we are focused on.

**[00:35:50 – 00:35:53]** Yeah, we love boring hard stuff.

**[00:35:53 – 00:35:58]** And I mean, it is the backbone of what we're building on top of.

**[00:35:58 – 00:36:03]** And as I said earlier, we've always found the subterranean

**[00:36:03 – 00:36:06]** world of AI to be more interesting.

**[00:36:06 – 00:36:10]** And so I think there's different kinds of investors for different

**[00:36:10 – 00:36:12]** types of businesses.

**[00:36:12 – 00:36:16]** But I do agree with the point that selling

**[00:36:16 – 00:36:20]** What you are building to the wrong investors is going to

**[00:36:20 – 00:36:25]** end badly for you and particularly understanding the time horizon, but

**[00:36:25 – 00:36:29]** I would suggest that if what you're building is something that's really

**[00:36:29 – 00:36:33]** complex, non-trivial, technical challenge at the frontier, there

**[00:36:33 – 00:36:37]** are VCs that will fund you and will have patience for doing that.

**[00:36:38 – 00:36:43]** And if it's something that you think is tracking with things

**[00:36:43 – 00:36:48]** that NVIDIA cares about deeply, I would engage early with some of

**[00:36:48 – 00:36:52]** the programs offered by Inception.

**[00:36:53 – 00:36:55]** But there are lots of programs out there.

**[00:36:55 – 00:36:58]** I would just be sure to speak to different investors and

**[00:36:58 – 00:37:02]** speak to other founders more importantly than VCs and so

**[00:37:02 – 00:37:07]** everyone can find a company or a founder that you respect or

**[00:37:07 – 00:37:09]** they're working on a hard problem.

**[00:37:10 – 00:37:12]** And don't just look at who their investors are and call

**[00:37:12 – 00:37:13]** them, call the founders.

**[00:37:13 – 00:37:18]** And we always tell all companies that if you want to understand more

**[00:37:18 – 00:37:20]** about Radical, don't speak to us.

**[00:37:20 – 00:37:22]** Our pitch is going to be our pitch.

**[00:37:22 – 00:37:25]** Speak to our founders and ask them about the value we add.

**[00:37:25 – 00:37:29]** And we do very well by that.

**[00:37:29 – 00:37:32]** I can probably best speak to getting started building something.

**[00:37:32 – 00:37:35]** I see incredibly smart friends seeing the models move incredibly

**[00:37:35 – 00:37:38]** quickly and then have paralysis because they're like, oh,

**[00:37:38 – 00:37:39]** the model's going to be able to do that, and so they almost

**[00:37:39 – 00:37:41]** never get started.

**[00:37:41 – 00:37:44]** So my advice is, one, think about the things that will never

**[00:37:44 – 00:37:46]** change and kind of hold those true.

**[00:37:46 – 00:37:48]** So for example, people will always want a more accurate model,

**[00:37:48 – 00:37:50]** probably want a faster model, probably want a model that works

**[00:37:50 – 00:37:54]** in their context, yet also build things that you're almost ready

**[00:37:54 – 00:37:56]** to rebuild every three months.

**[00:37:56 – 00:37:58]** And here I would look at a company that's the size and

**[00:37:58 – 00:38:02]** scale of NVIDIA, where most of the keynote was about OpenClaw,

**[00:38:02 – 00:38:05]** which didn't even exist 12 months ago, six months ago.

**[00:38:06 – 00:38:09]** So if you're building and you have this multi-year product

**[00:38:09 – 00:38:10]** roadmap, tear it up.

**[00:38:10 – 00:38:13]** Probably don't want a product roadmap longer than three months,

**[00:38:13 – 00:38:14]** maybe three weeks in some cases.

**[00:38:14 – 00:38:17]** And especially if you're just getting started, build, talk

**[00:38:17 – 00:38:19]** to users, and get going.

**[00:38:19 – 00:38:22]** Perfect, enemy the good, start shipping.

**[00:38:22 – 00:38:24]** Are you nervous that your business has changed since

**[00:38:24 – 00:38:26]** being up here with that clock in?

**[00:38:26 – 00:38:29]** Probably has.

**[00:38:29 – 00:38:33]** I think that it's easier today to start a company than ever before.

**[00:38:33 – 00:38:35]** But what I would encourage, and this is what we have done,

**[00:38:35 – 00:38:38]** is to not only look on this kind of first-order problems,

**[00:38:38 – 00:38:42]** but on the second-order problems and question things that people

**[00:38:42 – 00:38:45]** think have been solved.

**[00:38:45 – 00:38:48]** Classic synthetic cell-free mRNA technology has delivered

**[00:38:48 – 00:38:52]** billions of mRNA vaccines to the people across the globe.

**[00:38:52 – 00:38:55]** Everybody thought this is solved until we started talking

**[00:38:55 – 00:38:58]** to pharma companies and hearing that no, this is not solved, it

**[00:38:58 – 00:39:02]** worked for this, but it might not work for many other applications.

**[00:39:02 – 00:39:07]** And I think that challenging kind of established dogma

**[00:39:07 – 00:39:10]** and thinking about not only on a surface, but kind of

**[00:39:10 – 00:39:15]** going to the second and a third order is where ultimately a lot of

**[00:39:15 – 00:39:19]** exciting businesses that will truly transform our future will be built.

**[00:39:19 – 00:39:21]** And I encourage you to do that.