Good afternoon. This is Steven Mah, the Tools and Diagnostics team, introducing our next company in the synthetic biology space. It's my pleasure to welcome back Jason Kelly, CEO and founder of Ginkgo Bioworks.
Yeah. Good to be here, Steve.
Yeah, thanks. So let's keep it interactive. You know, if the audience has any questions, just raise your hand and we'll get a mic over to you. Or you can just email me at steven.mah@tdcowen.com. So.
Yeah, I'm actually very happy to do that. It was, you know, intimate setting, so.
Yeah, yeah, keep it interactive.
That's the question.
Yeah, yeah. And you might not even need a mic, so.
Don't need a mic.
Just yell it out. So, yeah, so Jason, you know, for, you know, those new to the story, maybe give a, you know, quick introduction to Ginkgo, and then we'll kind of dig into your business a little bit more.
Yeah. Okay. Yeah. So I can give a little bit of color on what we're trying to accomplish. So I, I think the, the sort of major, like, macro, shift that Ginkgo's trying to pull off is if you think about the way we develop biotech products today, you do it on-prem, in other words, on-premises at your lab, at a place like Merck, one of our customers, or at a place like Bayer Crop Science if you're developing plant traits in agriculture, or Novozymes if you're doing enzymes. And each one of these companies would have a large internal lab. It's gonna have a big expense associated with the physical facility. They're gonna spend a bunch of money on capital equipment, you know, buying from Thermo Fisher and other places, very expensive, analytical equipment and so on.
Then they're gonna have a bunch of scientists, who are PhD trained, like me, very expensive salaries. And those folks will stand at those lab benches using that equipment and pipetting liquids by hand. So they'll be essentially operating at a relatively low throughput in an expensive, inefficient way, to ultimately generate the products that go into the biotechnology industry. And our view at Ginkgo is that the physical lab work associated with all of that should be centralized into large automated facilities like what we've been building down in the Drydock. So if anyone's interested in coming to see it in the next few days, we're about 10 minutes away from here, about 300,000 sq ft highly automated lab, where we think that lab work should be centralized.
And so again, like, to, to give you an analogy, if you look at, say, the semiconductor industry, over the last 25 years, many semiconductor chip designers have gone fabless, and they outsource their fabrication of chips to Taiwan Semiconductor. And so my argument is roughly that biotech companies of the future should all be labless. They should be outsourcing their lab work to our infrastructure at Ginkgo to get it done much more efficiently at larger scale. And so the way we do that today is we basically do service contracts. We have now, Steve, we went Q4 last year, we had about, I think, a little over 90 active cell engineering R&D partnerships. And Q4, sorry, 2022, and Q4 2023, we had 131, right? And we did that without substantially increasing our OpEx at Ginkgo. And that was because we're driving efficiency with scale.
So as we add more of these programs to our automation, we mentioned that last year our unit economics of doing the lab work dropped by 40% to 50% over the course of the year. That's like a factory, right? As you do more work in it, it gets lower cost. That is not what it's like at the labs of a biotech company. There is no reduction in the unit economics of doing the research if you do more of it. That's really our advantage over time. If we do more work in our centralized facilities, they get cheaper. Our business model is very much like an AWS, driven by service fees for doing the work. Depending on the customer, we'll take some type of value share, like a rev share, either a royalty or milestones, on the success of their product.
Yep. Well, since you mentioned AWS, I mean, there was a little bit of a hurdle.
Yeah.
To get people to do it because they don't wanna put their data up into the cloud. You know, there's a, you know, similar analogy here, like, "Oh, I have, you know, proprietary targets, proprietary data. You know, what makes me confident I can share it with Ginkgo?
Oh, 100%. Yeah. And I mean, and if folks remember from that era, you know, like mid to late 2000s, right? Yeah, 100%, Steve, right? It was, "I don't wanna put my data on Amazon servers in Seattle. I don't trust their uptime. I have my own IT department. They're really good at this. Amazon's not good at this," duh, duh, duh. And very similarly, I think we do face that same resistance. And the answer is you sell through it and you keep building scale, which makes your platform better, and it makes it more and more inevitable that people will outsource to it.
Right.
That's it. When it comes to the data, we can talk about this too also with regard to AI. We see an opportunity to actually aggregate, a subset of data from the industry. That is non-competitive. It's not really what's embedded in somebody's drug or, you know, it's not a thing that's gonna influence their ability to protect their asset. But that in aggregate allows you to have the kind of training data that is useful for AI.
Yeah.
And other types of machine learning.
Yeah.
We also think our centralized system affords us the opportunity to build those data assets.
Right. Okay. Yeah. So you're talking about, like, a code base and what you're getting from Ginkgo.
100%. Yeah.
Working with our.
Yeah. We call it our Codebase. But the real challenge is if you look across the biopharma industry, every company has some subset of data they've been generating.
Right.
Very difficult to put that together, both for intellectual property reasons and just for, like, technical reasons. You know, the controls that one company uses are different than the controls another company uses. So the exact experimental conditions were never recorded. All these reasons make it hard for us to create the big pooled data assets that ultimately have driven success of large neural net AI models in things like human language, right? Everyone had this gigantic human language database to train on. Everyone had big image databases to train on. What do we have in biotech? We have the Protein Data Bank.
Yeah.
We have GenBank. That's about it. And so one of the things we're able to do at Ginkgo is start to accumulate actually much larger non-human genomic assets than are in GenBank today, labeled training data, things like that, that we can collect across these projects. And we think hopefully we can solve that coordination problem across the industry of exchanging data.
Okay. Yep. All right. And yeah, maybe let's talk about, you know, you know, one of the things we track is, like, number of validating partners.
Sure.
You obviously, you know, once you sign up a validating partner like Novo, for instance, you know, that probably helps kind of streamline discussions with other.
Yeah.
Comp validating counterparties as well. Is that true?
Yeah, yeah. Well, and maybe I'll talk a little bit about what we've done in the last year, which is a big motion into biopharma, I think, for this conference especially. That's of interest. So if you look at the history of Ginkgo, we actually started in non-biopharma biotech. So initially in the consumer goods space, so things like flavors and fragrances, like industrial biotechnology. Then we went into things like animal feed, like companies like Cargill or ADM. Then we worked in the ag industry. Syngenta, Corteva, and Bayer are all customers of ours. And it's really been in the last 2.5 years that we've started to substantially grow in biopharma. We went from, I think, $1 million of revenue in 2020 to $44 million last year, from biopharma alone.
So this is a brand new area that is newer for us. It is by far the biggest market for biotechnology research. So you might ask, "Why didn't we start there?" and the answer is we are an outsourcing business, right? Like, we are convincing you to go labless for some portion of your research. And if you have a really good infrastructure, that's a harder sell.
Yeah.
It was actually easier for us to sell outsourced services 8 years ago when my infrastructure wasn't that good to a fragrance company. Whereas 6 months ago, I did a drug discovery deal with Pfizer. That's a deal I could never have gotten 8 years ago.
Right.
Okay? And it's only because we've been growing the infrastructure, which allows us to basically get more data per research dollar substantially year-over-year. We, like we said, we doubled our efficiency last year. That it's because of that that we can now move into the biopharma space. But now that we're here, this is where you're what you're gonna see by far the most new deals for us.
Yeah. Got it.
I think it's more and more in biopharma. And absolutely the fact that we've done deals with Novo Nordisk, Merck, Pfizer, Boehringer all in the last year.
Right.
It is extraordinarily helpful for us to get new people on the platform.
All right. That's helpful. So, you know, we kind of tracked the SynBio deals across the various industries, you know, therapeutics, you know, across all, you know, SynBio is about a third of all deals today.
Today. Yeah.
Yeah, today. Is that kind of similar to for Ginkgo?
Oh, obviously.
About a third?
Yeah. Today it's about a third of our business.
Okay.
But it's gonna go through the roof compared to the others.
Yeah.
In the near years, in the next 2 to 3 years. That's just because the research budgets of the biopharma companies are just so much bigger.
Yeah. They're massive.
Then ag and industrial. Yep.
Yeah. So what, what's sort of like your best guess over, you know, I don't know, next, you know, three years or so of. Percentage of your business being, you know, biopharma or therapeutics related?
I bet it's 80%.
80%?
Yeah.
In next three years?
Yep.
Okay. All right.
Yeah. More than you would think. That it's just and again, I would encourage you to think of, like, what Ginkgo's really trying to pull off as that shift from on-prem to cloud, like, laboratories for this work.
Yeah.
It's, like, a big, big part of it.
Right. Right. Okay.
By far the most on-prem spending is in biopharma. There's just so much more to get after there.
Right. Right.
Does that make sense?
Yeah. That makes sense. And I guess, you know, your, a lot of your existing partners as well expand the scope of their relationship with you as well. So that's a way to kind of grow the therapeutics business as well, right?
Yeah. One biopharma company, you know, will have, like, you know, there's research and development, so there's R&D.
Yeah.
Forget the development for a second, which is all the clinical trials. Just at research budgets of $2 to 5 billion at some of these companies. I mean, one company.
Yeah.
Right? You know, like, my, my total cell engineering revenue last year was like a bit like $140 million-ish, right? You know, so, like, I have a lot of room to go.
Yeah.
Like, even one company.
Yeah.
Making a substantial bet on going labless.
Right.
Would be a huge accelerant to my current revenue.
Right. Yeah. I guess now it's a fraction of a percent, I guess, right now.
Yeah. No one's doing that.
The penetration's almost nothing.
That's a joke. Yeah.
Yeah. Okay.
Yeah.
Yeah. And let's talk about, like, partner mix. I'm talking about in terms of, you know, we've been talking about large, large pharma.
Yep.
What about, like, emerging, you know, biotechnology companies?
Yeah.
You know, obviously they're focused on cash preservation.
Yeah.
Capital markets are a little bit weak still. So, you know, how do you as from a.
Sure.
Like, a new program perspective balance taking.
So, yeah.
You know, companies which are emerging versus, like, you know, larger players?
Yeah. So, so the small companies should actually be an easier sale for me, but they are currently harder for macro reasons.
Right.
So the reason they should be an easier sale is for the same reason the very first customers on AWS were startups. They don't have labs. Like, they are just starting their company. So they could instead of going investing all that money in laboratory infrastructure, they could just go cloud native from the beginning. Okay? So I'm actually not just making it cheaper for them to do lab work relative to their current infrastructure. I'm saving them the upfront capital cost of the $5 or 10 million to build it in the first place.
Right.
So that's actually, like, a better sale. The trouble is the market went like this, you know, in terms of funding early-stage biotech. So all of the current people, like, basically were just focused on keeping the lights on and not as many new companies were getting started, period.
Right.
And so it's been, so that's why you see over the last year and a half, actually, the majority of my growth in biopharma has all been with big guys. Now, that's a good thing. Like, they have huge research budgets for me to tap into. Nothing wrong with that. It's, but it's actually a harder sale.
Right. Right.
Right? If that makes sense. So I do look forward to a looser venture capital market.
Right.
'Cause I'm hopeful, like, on the fundamentals, I have a better sales opportunity to the small guys.
Right.
I should sell to them.
You think that was a primary reason why, you know, kind of like the new program adds kind of slowed, oh, you know, and it's still growing.
That's a good point.
At a decent pace, but maybe not at the.
As what I would hoped.
Yeah.
Yeah. We were hoping for more program counts. The bigger and this is just in terms of number of new customers on our platform that Steve's asking about. That, the big challenge there has been these early markets we started in, like industrial biotechnology. And you know this, like, a lot of the what they is thought of as synthetic biology is often just.
Yeah.
People mistakenly just call it non-pharma biotech.
Yeah.
But things like animal-free meats, the cosmetics, the nutritional ingredients, and all these other applications of biotechnology, again, interest rates go up and venture capital for that space has gone to absolutely zero, right? Like, it's really tough if you're in those spaces right now. So again, I'm hopeful that you see some change there, but they got hit an order of magnitude harder than even the small biopharmas did.
Yeah.
So I think that that was probably more of a hit than I was hoping for. But, you know, that's what it is.
Yeah. Yeah. All right. That's fair enough.
Yep.
You know, in terms of your, you know, kind of.
Like, where it really would have killed us is if we hadn't been able to sell into the big guys.
Right.
Right? So, like, this was something that I was, like, stressing a lot more about 18 months ago.
Right.
But the reality is, now you fast forward 18 months, we were able to actually get over the bar for the big players.
Right.
Well, then this starts to some degree less valuably now because I'm inside these places that have $3 or 4 billion budgets, and I've only so far sold them a $30 million project.
Right.
Right? Like, so, and once you're in, it's a hell of a lot easier to add, right?
Right.
So, you know, I think to some degree, even though, like, I'm hopeful startups come back online, I think you'll just see me doing a lot more at the large biopharmas in the near years.
Right. So, expansion of, you know, into different, different areas.
We already are doing it. We added a second deal with Merck within six months.
Right.
We have yeah. Anyway, some others that we're working on right now. So, like, it is it's pretty easy to add.
A little easier to paper as well from a contract perspective as well.
Okay. Sort of. So yeah. So let me tell you another thing I would love to see happen, shift in terms of how I do business. And again, you know, oh, I'd love to have all this worked out already, but to me, it's a new market.
Right.
Okay? Right? Like, people do not outsource this today, right? I'm not offering them an animal study or a, you know, med chem library from WuXi or something, right? Like, I'm trying to basically do high-end genetic engineering of cells. Like, that is our that's the corner of biotech that we think should be outsourced, right? So, the way that we sell it today is I negotiate a 2 to 3-year research project with the customer, and we are arguing about the ninth quarter exact research plan before we've signed as part of trying to get the contract signed.
Yeah.
Extremely annoying. Okay? And so that is a big friction to actually getting a deal done. Meanwhile, in their internal R&D department does not plan out nine quarters and have a big argument of exactly what experimental work is gonna happen in the ninth quarter. They kind of go based on what they learned last quarter and then choose their next thing. And they're kind of headed in a direction.
Right.
But it doesn't have to be planned out to the nines, all right? So what I would love to do, and I'm working on this, is to have more of like a subscription model.
Oh.
So like, when you sign up for, like, a cloud service provider, you basically commit to $10 million of Google Cloud spend or something, all right? You don't know what you're gonna use it on, but because you committed to $10 million, you get a certain pricing.
I see.
If you commit to $30 million, you get a cheaper pricing. You're just basically trying to budget.
Right.
How much cloud is my IT team gonna use this year, not what they're gonna use it for? That's where I would like to move my contracts to.
Okay.
On the product.
Like a licensing mechanism.
Like a subscription, right? Like, you are subscribing, you know. You are basically buying, you know, and/or like a take-or-pay on a service contract. Like, you are buying. If you wanna buy $10 million, it's this—this set of prices. You wanna buy $30 million, it's this set of prices. But we don't have to negotiate a technical plan. Your team.
Right.
We'll order in partnership with us what you want from the automation to get what you want done that quarter. You just have to believe that of your $3 billion research budget, you should allocate $20 million of it to Ginkgo's infrastructure this year.
Got it. Got it.
Your team will decide what to do with it, right? Like, that has a lot of advantages for me, mainly in speed of closing a customer.
Right. Right. Yeah. I and some of your peers as well. I mean, Schrödinger had a.
Yeah.
Similar model.
Yeah. They do like a SaaS.
Software, like.
Yeah. 'Cause they're software.
Yeah.
Right. Yeah. So I'm trying to do this in the lab space.
I see.
Okay? That, yeah. But you've got to.
Conceptually the same again.
Conceptually the same.
Different.
Yes. In some ways.
All right.
More painful.
Yeah. All right. Let's, we're rapidly running out of time. Let's pivot over. So, you know, you've added new technologies with some acquisitions. Circularis.
Yes.
You added the rack automation.
Yes.
You know, can you point out any particular areas where you're getting more therapeutic interest? Is it like gene therapy? Is it, you know.
Yeah.
mRNA, vaccines?
Sure.
You know?
Yeah. So, like, for example, we did the CO Pfizer. The part of what went into that was the fact that we had acquired a company that had circular RNA technology called Circularis. We acquired a company called StrideBio that had interesting AAV capsids for targeting different organs with gene therapy. We just completed the acquisition of Proof, which has pretty interesting, like, gene editing technology. Well, actually, we acquired it 6 months ago. We just announced it. And so we will acquire IP in certain modalities that we can put on top of our automation that offers a more complete product to customers in that area. Okay? So, like, that we will keep doing.
And so if there's a small biotech in your portfolio or otherwise that you know that is selling a lead asset but has a bunch of technology associated with it that's just gonna get shelved when they sell the asset, Ginkgo would buy that technology. Okay? So keep us in mind. We have done that as we did with StrideBio. Somebody else bought their drug, but we bought the capsids. All right? Because these, to us, are the frequently stranded intellectual property and data that actually could benefit the whole industry but is basically tied to a single drug asset somewhere. So a capsid that does a good job targeting the lung or something is net valuable actually to a whole bunch of potential drugs, but it's sitting at a company that's pursuing one particular drug.
The whole thing lives or dies with that, transacting it is a pain in the ass. It's much better to have that sitting at a platform like Ginkgo where I could outlicense it to everybody in the industry, right? That, that's like found value. Does that make sense?
Yeah. Yeah.
And so those are the types of things we've been buying. And then on the automation front, yeah, absolutely. The Zymergen stuff's killer. Really good. We have flexible robotic automation. If you have not seen it and you wanna see it, I'm telling you, it's 10 minutes away. It's really neat to see. But it, you know, it's starting to feel like a data center in there. We just have. Like, basically robots connected by magnetic track maglev tracks that are sending plates all over this room. Like, it's really neat.
Yeah. We did a channel check with a user actually when it was former Zymergen.
Yeah.
When they were a standalone company before you bought them. And yeah, they swear by it. So it really had been enabling flexible automation technology for them.
Yeah.
Yeah. So that's, that's great.
I see floors full of them. It's gonna be great.
So speaking of that, that's gonna be a focus of, you know, your Biofab1. And maybe.
Yeah.
Let's talk about, you know, the foundry capacities. I mean, you're adding programs, maybe not as fast as you wanted, but, you know, certainly.
That's great.
Some of the bigger, yeah, certainly pretty good growth. Also maybe some bigger partners as well. Talk about the capacity of the foundries.
Yeah.
You know, you guys' right size, CapEx. I mean, you're, you know, and talk about BioFab a little bit.
Yeah. So I mean, we're in a pretty good spot. Like, I would say my overwhelming focus is just adding more customers on the platform. We think we can do that while maintaining OpEx, like, flat to down. And so that's because we have a team that's investing in making all the automation infrastructure more efficient.
Okay.
Right? And so that's been working, right? Like, we really are like, we run a lab like a factory. And so the more work we put on top of it, the more efficient it gets. And if we just look at the utilization rate of our various equipment, we think we could handle it like 2 to 3x more.
Okay.
Just on the equipment. Okay. That's before you factor in, like, other efficiency gains. So we think we're in a square spot. We do have Biofab1 opening in 2025. It's like a new building. I think that opens up a lot of opportunities. Again, like, back to this idea of, like, subscriptions with partners and things like that. You know, we would love to dedicate portions of that to certain customers, right, and things like that. You know, I can imagine having a floor that's really like robotic infrastructure that's on demand for your team and isn't used by anybody else. You're always top of queue. There's all kinds of interesting things we can think of there. So we're pretty excited about it. But it's, we have some time. So, you know, it's still a bit away before we would be, like, really moving in.
Okay.
Yeah.
Would there be an option and I only mentioned this 'cause, you know, some other companies I cover actually have when they have partnership, they actually allow their scientists?
Oh, I'd love that. Yeah.
To come into their, you know, production facilities, help.
Yeah.
Troubleshoot and kinda get products over the goal line. Is that something BioFab1 would.
Would open up the door for?
Yeah.
Yeah. 100%. Yeah. Well, let me say this. Another big sales challenge for me is today, it's Ginkgo scientists using Ginkgo's automation infrastructure to try to deliver whatever, a better performing RNA construct or a better CAR-T therapy or whatever the particular challenge is.
Yeah.
For you. You know, Biogen, AAV manufacturing, you know, right? Like, it's kind of like outsourced to us to solve a technical problem for you and get it back some year later or something. And you and you're getting to check in along the way, but we're really doing the work.
Yeah.
Some people don't like that.
Like, you're a C.E.O., yeah, you might not like that, right? Like, you're like, you're like, mid-level R&D leader at a company. You're like, "Hey, that's my job," I thought. Okay? Right? I don't wanna outsource that to you. And so, there's another world where and this and again, not to nerd out. I know like, we, we're like tech industry adjacent. So, I will give you a little bit of lesson like, how the tech industry if you go to the AWS website, you Google AWS Microservices, you will see a page that explains.
Yeah.
To software developers that the old way of developing software was monolithic, and it has a whole bunch of consequences about why that's bad to have one whole big system, one big program that runs your whole website. The new way to do it is by combining lots of these little microservices that all run in the cloud. And it's not so much about AWS. There's actually other cloud service providers too. But you, as a software developer, the implication is you ought to learn how to program with microservices 'cause that's the future of scalable computing. And if you're still programming monolithically, you're gonna be a dinosaur. You still have a job if you go down the microservices road, but you gotta learn how to program in that environment. That's the message I would actually rather have.
It's like the era of working with your hands at the lab bench is going to end. You're going to order microservices from centralized laboratory infrastructure. You still need your genius of experimental design, biological know-how, but you're gonna have to learn how to do that in this environment if you wanna be, like, scaled up. And this is especially true because this is the beauty, and I and this is really helping us sell. The story of Ginkgo has always been, it's gonna be beneficial to generate more data for your problem, right?
The criticism has always been, "Ginkgo, if you weren't so stupid, you wouldn't need to do so many experiments. I'm a smart scientist. I know just to do the right one experiment." It just so happens I can do it with my hands, but I always do the right one. You only have to do a lot 'cause you don't know what the right one is. The great thing about what's going on with AI, generative AI, is it's showing the value of large aggregate data assets. Yeah.
So it is a new way to have that conversation. It's been very enabling for us. Okay? And so I think if you're a scientist, the answer is you're not gonna be replaced by someplace like Ginkgo. This is how you're going to do your science in the future. You should learn it. The sooner you learn it, the more valuable you're gonna be if this transition ends up happening. Does that make sense?
Yeah. Yeah. Yeah.
I'd love to get there. We're working on that.
Okay.
Okay? And, Lemmer, I mentioned that subscription thing.
Yeah.
Earlier where you would have all those that's much more in that vein.
Right.
'Cause your team is deciding where to take it next. If we instead negotiate a 2.5-year project, that's my team doing it.
Right.
Okay? 'Cause that's the only reason we had to have that whole conversation. Does that make sense?
Yeah. No. That, that makes sense.
Yeah.
They're gonna have a little bit more control of their own.
Correct.
Project.
Then it's their duty, you know, that also solves a lot of my problems too, right? 'Cause then it's their choices that they're making.
Right.
Right? So if it's a bad choice, it wasn't my bad choice. It was their bad choice.
Okay.
But at least they got to do a lot, they got to try it more cheaply 'cause of my infrastructure, right? So there's a.
But how's it gonna work? Like.
Yeah.
I mean, how are they gonna, like, see the data coming off?
No.
Are they gonna physically be at the cloud?
No. Just log in. You don't need to be there. No.
It's gonna be on the cloud.
Yeah. It's just on the computer, right? Yeah. You order the experiments you want, and the data comes back to you, right? Like, you know, like, that, that's.
I see.
Why not, right? You know, and that's how my internal teams work. So when you're the scientist doing, like, whatever, Merck project, Pfizer project, it's actually a relatively small team.
Yeah.
Logging into their computers like Ginkgo, ordering tons of work from our automated infrastructure, getting the data back, planning the next round of experiments, right? Like, and it's not, you know, as clean as just click, click, click, but you see the path, right? And so, so that's where I would like to ultimately get so that our customer scientists can just use our infrastructure.
Okay. Got it. Got it.
Yeah.
Is that related to, you know, this recent deal you signed with Google Cloud where you kind of reserved?
Yeah.
A bunch of compute.
Yeah. Let me talk about that.
Capacity? Yeah. That'd be helpful.
Yeah. So sure. So let me give two cents on the AI thing and biology. So like, without nerding out on AI too hard, you've got, let's take a human language generative AI model like ChatGPT, all right? It is a neural net that was trained on human language to learn to speak English and be able to receive English prompts and answer with English-word answers. Okay? Right? And so how did we train the neural net? Well, we took what is a neural net? It's a node, a node, and a node connected by lines, right? It's meant to model like an.
This is like for the neuroscientists in the room, somewhat embarrassing. But it's meant to be like a human brain. Okay? Right? So you have a node, a node connected to another neuron. And if this one fires and this one fires, they send a little message down the lines connected to the third. And each line has a weight on it. It has a number. So let's say it's 0.2 and 0.4. They send their little signals down. You add them together. It's 0.6. If it's above 0.5, this node fires, and it's connected to three more nodes beneath it. And you have this was the magic of what Sam did at OpenAI. You put.
1 billion of these nodes into one big neural net. And then you train it on English sentences. What does that mean? You give it a sentence with 10 letters in it or 10 words in it. You leave out the eighth word, and it hits the top of the neural net, and then the neural net goes, boop, boop, boop, boop, boop, boop, sends signals down. Out the bottom, it predicts the missing word. If it's right, you say, "Good job, neural net," and you let it be. If it's wrong, you change the weights, and you do it again, and you see if it got closer. And then you leave out the second word. Then you leave out the fifth word. And then you do it for 1 billion sentences, and you do it enough times.
And this thing, which at no point was the neural net designed to know anything about English, learns English grammar. It learns Shakespearean poetry. It learns everything it has seen because you are giving it enough sentences and asking it to predict the missing pieces, all right? But nothing in there was designed knowing English. The reason I'm explaining this is if you look at a gene, it is read end to end. It's made up of letters in a line. It's like a paragraph. You can leave parts out. You can feed it into a neural net. You can ask it to predict what's missing.
And you can do that billions of times if you have billions of genes. In fact, Ginkgo has billions of genes. We have the largest non-human genomic collection in the world. Okay. So you leave all that stuff out, and you train these models so that they can learn to speak DNA, all right?
Yeah.
And then the other missing piece is you would like to know what that gene does. So why did we get AlphaFold first? That was the first, like, computational design, you know, AI model that actually worked. 'Cause we had the Protein Data Bank, which coupled the structure of a protein to the gene that encoded it. So you had the DNA model speaking DNA, and you had labeled data that said, "This is what this DNA does in the domain of protein structure." You put those together, and suddenly you could predict protein structures that had never seen before.
Yeah.
All right? Same exact idea. You wanna have a huge genome collection, and then you wanna have as much labeled data as you can to train it against designing a capsid, getting a, you know, certain expression level for a promoter in a particular organ. Like, whatever you want, if you accumulate enough labeled data, I think you're going to be able to have Generative AI models that are able to design against that particular challenge. And so.
Yeah.
One of the things I'm excited about and why we did the Google deal was, A, so we had the compute to train where we needed to. But the thing I'm and again, for people in the room that are into this stuff, we are good at generating that labeled data at Ginkgo. And there's a company in the AI space called Scale AI that not everybody's heard of. But ChatGPT, when they train their model after they did all the human sentences, they also paid a bunch of people to put sentences in and say if it was a good or bad sent you know, did it do a good job on the answer? And they actually paid a company called Scale AI to generate all that data.
AI also takes images of from the front of a car for Tesla and says, "Hey, that's a dog. That's a cone. That's a bike." They do that labeling of the data so that Tesla can train their models. Scale AI's business is just like data for hire. I'm happy to do that for people. So, so in this world of AI model building, we just announced our technology network.
Yeah.
Also last week, which is 25 companies that are kind of adjacent to Ginkgo. A number of them are AI companies.
Yeah.
I would love people to be able to focus on building great AI models and just pay me to generate a bunch of data for them. God bless. I and we could even figure out ways to give them access to these data assets I have in-house. Also great. But the reason we did the, the Google deal was just to make sure we were building our own muscles on building these models and understanding how to speak the language.
No.
Of people that might ultimately be able to utilize our infrastructure.
Got it. But some of the foundational models you're gonna be building, you know, how are you gonna monetize that?
Well.
Is that gonna just be?
In the near years, it would just be part of our cell engineering services.
Oh.
So we would include. That's one of the ways we would do a better job designing your RNA, designing your CAR-T stuff, whatever we would do for a customer.
Got it. Okay.
In the long run, same as anyone else would. You could charge for inference. You could, you know, right? Like, you could do that. Or again, other people do that, and I backstop them with data generation. I'm open to it.
Okay.
Right? Like, Ginkgo's really, I think, we're at the point now, Steve, where, like, my cost of data generation is so much better than anybody else's that I just want as many people flowing through it as possible. And I'm more than happy to have a bunch of other companies make money doing that if I could figure out a clean interface. Okay?
Yeah.
That's a little bit the exploration we're doing with the tech network.
Yeah. Yeah. Yeah. No, fair enough. Yeah. Data is king. Yeah. I got it. All right.
But I don't wanna keep it to myself. I wanna.
Yeah.
I wanna, like, be able to generate it for others. Does that make sense?
Yeah. That makes sense. Yeah. All right. Well, we're out of time. I did want to give the audience a chance to ask a question if they wanted to, but I'm full. Definitely feel free. Sorry. The time just flew.
I was excited.
A lot of stuff going on. Yeah. We didn't even get to biosecurity, unfortunately, but,
It's happening.
That'll be another time.