The Future of AI Storage: DNA Data, DeepSeek & the $500B Stargate Project
The hosts look at how to meet AI's growing demand for data storage, after the announcement of the $500 billion Stargate project and the release of DeepSeek's R1 model. Laila Rizvi, Director of Strategy and Research at WebsEdge, joins to discuss why storage has lagged behind AI. Jeff Nivala, Co-Director of the Molecular Information Systems Lab, a partnership between the University of Washington and Microsoft, explains how digital data can be stored in synthetic DNA.
Transcript
Welcome to Agents of Tech, the podcast where we explore the groundbreaking innovations and critical ideas shaping our future. I'm Autria Godfrey here in Washington, DC.
And I'm Stephen Horn in London.
Today we'll be exploring how to meet AI's ever increasing demand for data storage. As AI advances at breakneck speed, pushing traditional data infrastructure to its limits, AI not only needs large amounts of data for training, but also for greater data infrastructure for data storage and memory. Stephen, I don't know how closely you've been paying attention to what's going on across the pond over here, but we have been on an AI roller coaster. We have the Stargate project that was just announced by the Trump administration, and it's a new company which intends to invest 500 billion dollars, that's a capital B, over the next four years building new AI infrastructure for OpenAI here in the United States. Stargate will be building the physical and virtual infrastructure to power the next generation of advancements in AI. Tremendous amounts of money are coming in.
We're talking Arm, Microsoft, NVIDIA, Oracle and OpenAI are all key initial technology partners. So truly everyone at the table is coming together here. Of course they're going to be dealing with DeepSeek AI. That's a startup backed by a hedge fund, High-Flyer Capital Management, and they've released a version of their AI chatbot, R1, that they say can perform just as well as competing models such as ChatGPT, but at a fraction of the cost. So that's got a lot of the AI players over here taking note. In fact, Mark Zuckerberg has reportedly launched a war room where he's got many of his AI employees trying to figure out how they can mimic the success of DeepSeek at of course, the DeepSeek costs. So lots of exciting things happening in the AI world over here.
It's really quite hard to tell, isn't it, at the moment, sort of where this whole thing is going. But you know, exciting for us Autria, we've got Laila joining us later so she'll be able to sort it out.
We're also going to be talking to Jeff Nivala, Co-Director, Molecular Information Systems Lab, which is a partnership between the University of Washington and Microsoft, about his work developing methods for storing digital data in DNA.
You can start to imagine data centers of the future, data forests and these sorts of things.
Well, I'm absolutely delighted, Autria, to be joined by Laila. Laila, welcome.
Thank you. I'm delighted to be here.
Laila, for those of you who don't know, is the Director of Strategy and Research at WebsEdge, the company behind the podcast. Laila, thank goodness you're here. All right. So what's going on?
So we've got this issue with data storage where the field of data storage has lagged behind that of AI innovation and technological innovation, where for the past four decades we've just been making transistors smaller and smaller and smaller. And that's resulted in kind of reasonable increases in, the ability for us to store data more efficiently. But we've kind of hit a limit with that because you can't, you can't get any smaller than where we are now. So I think it's a very interesting space where I think we're going to have to think really big and, innovatively and look to other domains to try and solve this problem, which hasn't really been solved by kind of engineers and computer scientists yet. And I am a little bit, you know, I'm a bit reserved when it comes to pop sciency kind of ideas, but I think you've got to think big to make it big.
I need to push you on this DeepSeek stuff. I mean, where are we? I mean, are we 50 times more efficient or not?
So I think what DeepSeek has done, obviously, is showing that you can achieve the same result using different strategies and less compute. So rather than working on a giant scale, you can kind of call to these smaller, smaller parts of the model. So I think it's shown us that we don't need as much compute or the price of compute isn't that high, but I still don't think it takes away from the data problem and the fact that we're still going to continue to generate tons of data from everything we do, all of our technology, all of our existing technology, as well as AI and the use of AI, the things that that's generating as well as memory.
So we talk a lot about agents, and that's very hyped that we're going to have these agents, that agent. But in order for agents to function well, they need to be able to speak to what is a repository of knowledge. They need to be able to access it. So when it comes to using things like AI in, cars, autonomous vehicles, in other use cases, the AI needs to be able to access information quickly, right? So the data storage problem and the data retrieval problem is still a problem. So I don't think something like DeepSeek necessarily has solved every single problem that we've ever had. There's still a long way to go. And I do think data is key to it.
Perfect. So Autria we are indeed going to need alternative forms of data and data storage.
Yes. I would like to ask Laila one more question, in why she thinks that DeepSeek was able to create this AI foundation for a fraction of the cost. Do you think that it's as good AI as we're seeing for, you know, with the $100 million companies?
I think it's great. I think the reason, and I mentioned this to Stephen earlier today as well, is that when you're constrained on resources, you're able to do amazing things. And I think it's the same premise by which OpenAI was able to leapfrog DeepMind and Google, that they have limited resources. So you're forced to innovate with the small amount of resource that you have. And I think that is just fundamental to all progress that we've ever seen in terms of technological innovation.
And so now to dig a little bit deeper is Dr. Jeff Nivala joining us, the Co-Director for Molecular Information Systems Lab, in a partnership between the University of Washington and Microsoft.
If you're enjoying our show, please don't forget to subscribe to Agents of Tech. This really helps develop the show and shine a light on what Laila was just talking about. The cutting edge research. So thank you all.
It's our pleasure now to have Dr. Jeff Nivala, Co-Director for the Molecular Information Systems Lab, joining us here. This is a partnership between the University of Washington and Microsoft. Thanks for your time today.
Yeah of course. Great. Great to be here.
Jeff, can we kick off, Everybody's talking about data storage at the moment. Data and data storage. Can you talk us through your work with synthetic DNA for data storage and why you've taken this particular path?
The rise of modern technology, sort of, we're going through the information age right now, things like IoT, cloud services, the internet is generating tremendous amounts of data at a scale that humanity has never seen before. And so we're starting to essentially exhaust the limits of conventional storage technology such as tape and in solid state drives and and these sorts of technologies. So exploring new technologies, like DNA data storage is really attractive because compared to those traditional storage media, DNA is incredibly dense. So orders of magnitude, more dense, you can store a lot more information in less space. And it's also incredibly durable. So when stored in the right conditions, DNA data storage can have lifetimes of many thousands to potentially millennias of lifetimes.
Jeff, I'd like to get your take on, you know, the big announcements here recently, like the Stargate project. You know, this this huge $500 billion investment into AI in the United States over the next four years. Your thoughts on, I guess, really the excitement of that and the basic question how on earth do we create the storage that this type of investment might generate?
So what we do know about AI, it's incredibly powerful, but it's incredibly, data hungry, as we are exhausting our limits, our ability to store data using conventional, methods, being able to, potentially store data much more efficiently by learning from how biology stores information, like using DNA or other molecular forms. Potentially this would be a more sustainable approach to be able to power, all the information storage that we're going to need for this next generation of AI technology.
So to that point, you're working obviously, at the intersection of biology and computing. I think for most people, that's two fields that, seem very different. What drew you to combine these areas and what surprises have you found in trying to make biology work as a computer storage system?
Yeah. So, I mean, I, I see biology really as a technology, sort of the original technology. And so if we look around us, all the major amazing things that biology can do today, you know, that it's evolved, you know, over billions of years to store information, to process information. And these are all the sorts of things that we build electronic systems to be able to do as well. And so I think the more that we can learn from biology to, build better computing systems, I think that's sort of the future of sustainable computing.
Some surprising things that I've learned is that it can be incredibly difficult to to engineer, using biology to act like a computing system because, you know, biology, like I said, is sort of evolved over billions of years to do the things that it wants to do. And when we try to make it do what we want it to do, surprising things can happen. Because you're sort of trying to override the inherent, what it was originally programmed to do. Everything that you build, with biology, it's sort of in a black box, right. So it has a lot of interactions that you don't know about until, you discover something is not working as behaved or something is not behaving as, as designed. And you need to figure out, okay, what's what's going on in this system that I don't understand yet.
You know, when we think about data storage, we don't we don't automatically think about DNA and we don't automatically think about biology. We think about big server racks in huge concrete buildings with the air conditioning, you know, units all around them. So I think it's really interesting the way that you said that you saw biology almost is the first, you know, technology. And I wonder, do you think that this means we need more, if you pardon the pun, cross-fertilization. You know, we need more computer scientists, more biologists or geneticists actually working together on this.
So certainly it's that interdisciplinary nature that's going to make this, this technology, reach its full potential. And so that's what we're trying to do with the Molecular Information Systems Lab. It's really bringing together experts in molecular biology, experts in computer architecture, programming algorithms and these sorts of things that really fertilize across, across fields. That cross-fertilization is, as you said, because really, no one can be a domain expert in all of these areas, right? It really takes these broad groups of backgrounds to be able to build technology, for real applications.
So how does this work? How do you encode digital data in DNA? Can you boil it down for us to kind of explain the actual mechanics in terms of the process?
So, so it starts, with taking your digital bits, right, which are traditionally represented as zeros and ones. You can essentially encode any, any information, using binary as it's called. So you know, text, images, videos, so you can take the zeros and ones, these digital bits and you map them into essentially sequences of As, Ts, Gs, and Cs. And you do that, That's called the encoding portion of it. So you have this mapping scheme that goes from the digital bits to DNA bases, and then you actually then synthesize or physically make these strands of DNA. And now you actually have this information now preserved within these DNA molecules that now exist physically. And those can be stored away. Right. And like I said, so you can store these away under the right conditions, this information will, will be preserved within that DNA for, for a very long time, using essentially minimal energy expenditure to keep it safe. And then when you want to retrieve that information back out, you use, DNA sequencing technology that essentially you put these molecules into a DNA sequencer, it'll read back these the sequence of the As, Ts, Gs and Cs within those DNA strands. And then you can use software that can decode those sequences back into the digital bits that represent your information. And then you've essentially gone end to end, from digital bits to DNA, back to digital bits again.
Jeff, I want to kind of build upon that. And talk about why we're using biological infrastructure as opposed to electronic infrastructure. And this is mainly because, I mean, you obviously mentioned that DNA was kind of the first technology, but throughout history when we've, I guess, tried to mimic, biology such as, you know, flight, we built planes which didn't necessarily mimic the wings of birds, we didn't try and build muscle necessarily. So how come you have, kind of gone down the route of, bio utilization as opposed to biomimicry because, you know, sequencing and PCR, is quite laborious.
What we're really doing is sort of at the very beginnings of using, biomolecules for computing. The easiest thing to do is directly borrow what nature has done. Right. And so using, you know, sort of additional DNA based storage or using traditional, you know, sequences of DNA, molecular structures that biology has used, we're using, sequencing and synthesis technologies borrowed from the life sciences. So I think that's gotten us to this stage. But there's still a lot of, technical barriers to overcome to make it mainstream. And I think this is where your point, comes in is as now we sort of need to think about not just directly borrowing things from nature, but how do we, sort of now take it to the next level, going beyond what biology has done and building, using technology as the substrate. But now optimize it, to get to that next level and, and be able to not compete with electronics necessarily, but complement it.
Current DNA storage is positioned as cold storage. Do you think DNA could be used as a storage which would allow for more frequently accessed data? I know I recently came across a paper from Peking University where they were using, epigenetic methylation to kind of speed up that process. But, do you have any views on that and how we can progress from going from cold storage to more kind of dynamic storage?
Yeah, for sure. So I think to progress beyond cold storage to dynamic storage, we're going to have to sort of reinvent new technologies for reading and writing of DNA or potentially other molecules, beyond what's been developed for the life sciences. The metrics that you might want out of a DNA sequencer to sequence a human genome, might be very different than what you might want when you're trying to sequence, you know, information that's encoded with synthetic DNA. You know, you could deal with a higher error rate if it decreases the latency. Right. The research that I'm really interested in developing is new technologies that are optimized for reading and writing biomolecules with the applications directly for for computing and data storage.
I think the idea of DNA as a universal constructor, in the regard that it can essentially take knowledge and turn it into matter is pretty extraordinary. And I guess on that point, do you see potential for us to build something which acts like DNA as a universal constructor in the respect that in the DNA or the synthetic DNA is able to have active properties like self-replication, where it not only stores the information but has the capacity to replicate it and build something from it.
Potential is all around us in what DNA has the capability of doing, in not just storing information, but being able to replicate it? Having things, molecular mechanisms that encode or are able to do things like error correction as you make copies of it. So amazing things, in addition to, you know, being blueprints that can then build molecules that are much more complex than the DNA itself. So I think if you look at sort of, you know, the full stack of what, exists in biology, going from storage all the way to information processing, it involves not just DNA, but, you know, things like RNA, proteins. And so I think as you're moving beyond storage and you think about more like more towards molecular computing, I think we're going to have to extend beyond DNA because it's relatively limited in what it's going to do. And I think, things like protein engineering and protein design, will allow us to do more sophisticated, computing, similar to the way biology is able to do so much, you know, sense and response and process information in the way that it does.
So how do you make what you're working on that scalable? And how do you make it more accessible to non-experts?
Well, a big part of it is, building upon these technologies that are originally designed for life sciences. Now we also have to rethink the design for information storage and computing applications. And so that's how we get these, potentially get these read and write more scalable in particular the technology that I've been developing for a long time, it's called nanopore technology. So it's a way to sequence DNA or other biomolecules that's, incredibly, portable, low cost can plug into your laptop, even a smartphone and be able to sequence DNA, and potentially other molecules have information stored in them. So I think going towards more ubiquitous molecular storage nanopore technologies is going to be, a very promising technology going forward. And the way that we may also make it more accessible is by automation, you know, everything that we're doing right now, sort of in the lab and people are bumping into each other with pipettes and whatnot. Being able to have automation technology such as, you know, microfluidic systems, digital microfluidic systems, that can be miniaturized, low power and be able to, you know, essentially move all these droplets of water around, essentially is what we're doing to carry all this information in an automated, compact way, is is what we need to do to get to that next level.
So kind of on that point, like where things stand right now, how could other people maybe working in your field, leverage your work in order to contribute to the field? And when do you think the average person could start using DNA based storage?
To the first question there, folks getting more involved? I mean, I think what's great about this, you know, molecular information storage and DNA based storage and information processing, is that it really, it's such a broad field that, you know, if you have expertise or interest in biology, in particular molecular biology, you can start to get involved with it that way. Or if you're coming from the computing side, you know, learning more about biology and how you might be inspired to build better systems, with this new substrates that biology offers, there's a lot of open source sort of encoding and decoding tools that folks can play around with. As far as to explore how information is practically encoded into DNA sequences, as far as when will it practically sort of every day lives, I mean, I think for the archival storage or this cold storage, I think that's going to be within the next 5 to 10 years. I think we will start to use that at a much larger scale.
There's already, you know, startups that are commercializing the technology, right now as, as far as using DNA for storage, going beyond that, you know, I think there's a ways to go to make it more practical as far as competing with traditional electronics as, as we think about computing. But I think computing can be really broadly defined in that there are, you know, we can bring molecular computers into, into spaces, and applications that traditional computing cannot do. I think for healthcare and just, you know, new mechanisms for understanding biology, molecular computing is already starting to have applications. And I think that's just going to continue to grow from there.
I do have one final question for you. Over the next 5 to 10 years, when you see this becoming much more practical and applied, do you see DNA based storage as being one of the key solutions to the data storage dilemma?
I do I mean, I'm sort of inspired by nature. If you just look around us, all the information, that's contained, in living life forms, it's all around us. It's able to stably, sustainably propagate and coexist with, with other life, in harmony, essentially with life on Earth, you know, looking 100, 500 years from now, you know, you can start to imagine data centers of the future, data forests and these sorts of things. And, so I think that's sort of the imaginaries. I like to think about, it's a long ways away, but it's, you know, it's what drives me.
All right. Really fascinating technology. Thank you so much for your time today. This has been so, so interesting. And we really appreciate it.
Thank you so much, Jeff.
Thank you.
Thank you.
Well, it's not very often, is it, that you finish an interview with a picture in your mind, right? You know, and I can't get that out of mind now with forests of data. Right. So I started off today with these concrete built data centers all around the world that are just going to be, you know, a blight on our environment. And now we've got a lovely forest in my mind for data. Wouldn't that be great?
Yeah. It feels good to look towards nature. I'm kind of with Jeff on that because like DNA, it's one, obviously, in everything and I feel like I'm stating the obvious, but if you actually look at it, it's just four letters which can store like it's so data dense in that respect in a way that it can, I think one gram of DNA can store 10 million hours of HD video.
Before we did this interview, had you ever even imagined DNA as a data storage device?
My experience of dealing with DNA is what Jeff was talking about. Pipetting in a lab. You're there till 11 p.m., and it's frankly a nightmare. So no, I didn't. I didn't think about it. But his point about being able to use biological computing in places which electronic computing can't be used just blew my mind.
You know, in addition to the high demand for space in order to store all of this, there's also the ongoing conversation about sustainability and the drain that this might have on the power grid. And are we going to able to power this? And so, you know, I think that these are really interesting and dynamic and creative ways to try to find a solution to that problem. Really fascinating content today. But that is it for this episode. If you enjoyed our discussion today, don't forget to like and subscribe, because that really helps us improve the show and curate our content. Until next time, thank you so much and goodbye.
Goodbye.
Until next time and goodbye.
Agents of Tech is brought to you by WebsEdge. Our special thanks to our studio engineers Adam Deen and Cody Huynh, editor Abbie Harries, and our producer, Cath Sheehan.