TikTok Ban, Data Centers in Space… and What's a Vector? | E2073

Today's show: Alex sits down with two CEOs from companies freshly added to the TWIST500. We get a masterclass from Bob van Luijt of Weaviate, on vector databases helping AI and LLMs do what they do. An then Philip Johnston from Lumen Orbit explains his vision of building data centers in space! * Timestamps: (0:00) Alex kicks off the show. (0:58) Supreme Court ruling on TikTok and key market numbers (3:49) Interview introduction with Weaviate and Lumen Orbit CEOs (4:06) Deep dive into vector databases with Weaviate's Bob van Luijt. (8:23) Lemon. TWiST listeners get 15% off your first 4 weeks of developer time at https://Lemon.io/twist (11:35) Exploring deep learning, vector indexing, and Weaviate's open-source significance (18:03) Northwest Registered Agent. For just $39 plus state fees, Northwest will handle your complete business identity. Visit https://www.northwestregisteredagent.com/twist ⁠today. (19:32) Discussion on vector databases use cases and retrieval augmented generation (RAG) (27:17) Weaviate's business model, revenue streams, and growth (28:12) Vanta. TWiST listeners automate your SOC2 and get $1,000 off at http://www.vanta.com/twist (32:52) The future of AI, agentic architectures, and enterprise adoption (40:50) Lumen Orbit's vision for space-based data centers (42:54) Philip Johnston on the technical aspects of Lumen Orbit (49:12) Satellite demonstrators, technology challenges, and VC interest (53:26) Addressing chip obsolescence and cost advantages in space (56:21) Deployment strategies and competition in the space data center market (1:00:12) Strategies for staying ahead of competition and future plans for Lumen Orbit * Subscribe to the TWiST500 newsletter: https://ticker.thisweekinstartups.com Check out the TWIST500: https://www.twist500.com Subscribe to This Week in Startups on Apple: https://rb.gy/v19fcp Check out Weaviate: https://weaviate.io/ Check out Lumen Orbit: https://www.lumenorbit.com/ * Follow Bob: X: https://x.com/bobvanluijt LinkedIn: https://www.linkedin.com/in/bobvanluijt * Follow Philip: X: https://x.com/johnstonphil LinkedIn: https://www.linkedin.com/in/johnstonphilip/ * Follow Alex: X: https://x.com/alex LinkedIn: ⁠https://www.linkedin.com/in/alexwilhelm * Follow Jason: X: https://twitter.com/Jason LinkedIn: https://www.linkedin.com/in/jasoncalacanis * Thank you to our partners: (8:23) Lemon. TWiST listeners get 15% off your first 4 weeks of developer time at https://Lemon.io/twist (18:03) Northwest Registered Agent. For just $39 plus state fees, Northwest will handle your complete business identity. Visit https://www.northwestregisteredagent.com/twist (28:12) Vanta. TWiST listeners automate your SOC2 and get $1,000 off at http://www.vanta.com/twist * Great TWIST interviews: Will Guidara, Eoghan McCabe, Steve Huffman, Brian Chesky, Bob Moesta, Aaron Levie, Sophia Amoruso, Reid Hoffman, Frank Slootman, Billy McFarland * Check out Jason's suite of newsletters: https://substack.com/@calacanis * Follow TWiST: Twitter: https://twitter.com/TWiStartups YouTube: https://www.youtube.com/thisweekin Instagram: https://www.instagram.com/thisweekinstartups TikTok: https://www.tiktok.com/@thisweekinstartups Substack: https://twistartups.substack.com * Subscribe to the Founder University Podcast: https://www.youtube.com/@founderuniversity1916

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  Hey everybody, welcome back to Twist. This is Alex. Today is Friday. We have a pretty special show for you today. I'm talking to two founders of two of the most interesting private market companies in the world. But before we jump into that, I have a couple of news items because I don't want you to go into the weekend behind on all things that matter.

• 0:19Speaker 01

  This week in startups is brought to you by lemon.io. Hire pre-vetted remote developers. Get 15% off your first four weeks of developer time at lemon.io/twist. Northwest Registered Agent. Starting your business should be simple. With Northwest Registered Agent, you can form your entire business identity in just 10 clicks in 10 minutes. From LLCs to trademarks, domains to custom websites, they've got you covered. Get more privacy, more options, and more done. Visit northwestregisteredagent.com/twist today. And Vanta. Compliance and security shouldn't be a deal breaker for startups to win new business. Vanta makes it easy for companies to get a SOC 2 report fast. Twist listeners can get $1,000 off for a limited time at vanta.com/twist.

• 0:58Speaker 00

  First up, right as we were going to record today, the Supreme Court of the United States upheld a law that would force a divestiture or ban of TikTok in the US on January 19th. This is enormous news. I've only had a chance to skim what SCOTUS wrote, but the gist is that if you were hoping that TikTok was going to be saved at the very last moment, probably not. We're going to see some chaos next week. It's going to be a big story. Also, don't forget, of course, next week is the inauguration, so it's going to be packed. Stick close to Twist. Now, on the news front, a couple of things for you. I have three numbers that you need to know. The first one is 65.4 million. The second one is 150 million. And then the third is 12 and a half billion. What are we talking about?

• 1:43Speaker 00

  Well, the first one, $65.4 million. That is how much money former accounting czar Bench had in liabilities when it failed. A question in the market when Bench went from alive to dead seemingly overnight is, why did that happen? Well, I don't think people understood the liabilities the company had racked up. Now, TechCrunch reports that much of the money is actually owed to the National Bank of Canada. So figure that out as you want to, Canada, but also in Employees, investors and executives are also owed money. This is a mess. It's still been unraveled, but shout out TechCrunch for getting even more on the bench saga. Number, number two, $150 million. That's how much money Crypto Wallet Phantom raised this week.

• 2:26Speaker 00

  And it raised that 150 at a $3 billion valuation, so a multi unicorn price tag. The crypto company claims 15 million monthly active users and more active traders and trading revenue than wallets for MetaMask and Coinbase Wallet combined back in November and December. All that's to say that people are really using it. What's going on? Well, after the most recent election and Bitcoin reaching $100,000 per token, people are once again bullish on all things crypto. There's been a vibe shift, as we say. And when that happens, well, consumers get more interested in crypto, crypto on ramps, crypto exchanges, crypto wallets, and then companies like Phantom see a nice bump in their usage, and then the investors show back up.

• 3:08Speaker 00

  Crypto loves to go up, crypto loves to go down. Right now, we are on one heck of an upswing. And our final number for today is $12.5 billion. That's how much money Insight Partners just raised for a number of new funds. This includes, as you might expect, a flagship fund and also money for, quote, "a dedicated buyout co-invest fund." So expect a multi-strategy approach from Insight with this new amount of money. To put the $12.5 billion into context, I just want to say that we have been seeing venture concentration lately. The number of funds that are raising money seems to be going down, and the larger funds seem to be doing the best. So the rich are getting richer and the emerging managers are struggling. That's why when I see a $12.5 billion raise from Insight, I'm like, eh, yeah, I can see that.

• 3:54Speaker 00

  That fits actually pretty well with the news cycle we have seen. All right, now that we are caught up on the news, I want to do two interviews. One with the CEO and co-founder of WeV8, and then with the co-founder and CEO of Lumen Orbit. Why these two companies? Well, they are the newest additions to the Twist 500, our ever-growing list of the 500 most important private market companies in the world, but also because I think they're incredibly interesting companies that highlight where we are seeing. A lot of innovation in the market today. So first up, co-founder and CEO of Weaviate. It's Bob Van Lout. Weaviate is a startup that's a bit in the weeds of the AI revolution, but I really do think it can become an absolute household name in the world of technology in short order. We'll talk vectors with Bob and then we're going to jump into an interview with Lumen,

• 4:39Speaker 00

  which is all about space. Let's go. So by now you know all about L L Ms and GPUs, but to understand how many AI apps are actually built today, you're going to need to understand vectors and vector databases. They are a critical enough part of the modern AI stack that there are a number of startups working on them, including Vespa, Pinecone and Weaviate. Now I've known about Weaviate for some time, including back before ChatGPT changed everything. And given how critical vector databases are to AI apps today, and how early Weaviate was, along with strong vectors and an innovative open source model, those are the reasons why I added the company to Archlist 500. Now, today to walk us through why vectors matter,

• 5:24Speaker 00

  I have Bob Van Lout, the CEO and co-founder of Weaviate. Bob, welcome to the show.

• 5:28Speaker 02

  Thanks for having me, Alex. It's great to, great to see you again. That was quite some time ago. So this is, this is awesome.

• 5:33Speaker 00

  I know. So for people who don't know, way back in the day, I think it was like 2019, 2020. So way before ChatGPT, you and I sat down and you were incredibly patient and explained a vector databases to me, which I retained for probably about two weeks and then it flew out of my head. Cause I don't do what you do, but it really does seem like the market has come towards, towards Weaviate. And you've become a company name that I feel like I see quite often when I'm doing AI research. So Bob, I was thinking today, you and I could do a little bit of a, a class, if you will, and start by walking people through vectors and why they matter. And then I think we'll talk about Weaviate and the future, but I think a little background is going to help.

• 6:15Speaker 02

  This sounds wonderful and I look forward to it. So.

• 6:18Speaker 00

  All right.

• 6:19Speaker 02

  Let's go.

• 6:19Speaker 00

  What I'm going to do is I'm going to try to explain a series of concepts as they build on one another. And then I'm going to lean on you for some confirmation. So first of all, a vector, it's a mathematical thing. It's a numerical representation of data that provides both magnitude and direction. So essentially, how far away and in what direction?

• 6:39Speaker 02

  That's correct.

• 6:39Speaker 00

  One out of one, nailing it. All right. Now, vector embeddings are essentially assigning vector values to words and/or sentences. So essentially, it's taking data and then assigning those vectors to those individual pieces of unstructured or structured data.

• 6:56Speaker 02

  That is correct as well. And to add to that, the reason why that's so interesting is because if we ask ourselves the questions, how can we make sense of any type of data that's unstructured? Be it language, be it images, be it audio, can be anything. What method can we use to do something valuable with them? And the answer to that question is, if we organize them in space, and we do that by assigning vector embeddings, we can work with it and we do that by distance calculations. And we're probably going to double click on what that means, but that's why they've become so valuable to work with unstructured data.

• 7:34Speaker 00

  And when we talk about things in space, Bob, I pulled up a graphic here that I think shows people a little bit of what we're talking about here. This shows the proximity of several different data points. And essentially, my understanding here is that because we would have vector embeddings for these different concepts, wolf, dog, cat, we can see the distance between them and see that they are relatively close to one another, which means that they are related.

• 7:56Speaker 02

  That is correct. And the way we do that, and that's why it gets so exciting, is that the researchers who put in the work, and by the way, we're not talking about recent work. This is done way back. The only issue was that we didn't have machine learning to train. We'll, we'll get to that.

• 8:12Speaker 00

  Yeah.

• 8:12Speaker 02

  But there was that these researchers asked themselves the questions, how can we somehow say something about language in any way, shape or form that it relates to each other?

• 8:24Jason Calacanis?

  All right, founders, let's be real. Finding great developers is tough, especially when you're trying to run and scale your startup and raise money. All of this leads to you having slow product velocity. But here's the good news. I've got a tip that's going to save you time. It's going to save you money and a ton of headaches. You need to check out lemon.io. Lemon.io has thousands of on-demand developers who can help you. They've done the work to find and vet developers who are experienced, who are results-oriented, and who charge competitive rates. Great developers can be hard to find and integrate into your team. Lemon.io handles all of that for you. Startups choose lemon.io because they only offer handpicked developers with at least

• 9:07Jason Calacanis?

  three years of experience and who are the best of the best. Just 1% of their candidates are accepted. And if something goes wrong, lemon.io will find you a replacement developer ASAP. A bunch of launch founders have worked with lemon.io and they've had great experiences. So here's your call to action. Visit lemon.io/twist and find your perfect developer or even a tech team in just 48 hours or less. And Twist listeners get 15% off the first four weeks. Stop burning money, hire developers smarter, and visit lemon.io/twist.

• 9:39Speaker 02

  And so, for example, if you think about cities, right? If you go from New York to Boston, and you want to say something about the relation between New York, Boston, and let's say, San Francisco, you could say, well, Boston is closer, just in miles, to New York than it is to San Francisco. That's a piece of information. So what these researchers came up with for language, and this is super exciting, they said, what we can do is we can count words in a sentence. So we can basically say, if we, for example, take the word group Eiffel and Tower, so the Eiffel Tower, probably somewhere in that sentence, we're going to find Paris and not... Uh, Madrid to make something.

• 10:18Speaker 00

  Sure, yeah.

• 10:19Speaker 02

  Right. That is what it does. So that is what they came up with. Like, we can make distances by counting the distances of words in sentences. But for example, when we deal with images, same question. So how can we say something about images? Wait, an image is based on pixels and pixels have a color. So if we have a Granny Smith apple, we're probably gonna see a lot of green in certain situations in that uh image.

• 10:45Speaker 00

  Yes.

• 10:46Speaker 02

  That's how it's done. So it doesn't matter what the modality is, can be anything, but the researchers think, how are we gonna say something about distance? And that's what they compress and capture in that embedding.

• 10:58Speaker 00

  And this is a very cool toy, unless you have the ability to use machine learning to do what's called vector indexing, which is going through a large data set and then applying vector embeddings, the numbers, essentially distances, to those data points. And Bob, this is when it starts to feel to me less like technology and more like magic, because I get everything you're saying, but to me, having a machine learning model know how to assign the vector numbers to the discrete data points is where I struggle a little bit. So I was hoping you could just kind of double click on that so we could all understand better.

• 11:35Speaker 02

  Yes. So let's go back to the first example of the cities, right? And I assume that the audience can picture an Excel file, you know, so picture an Excel file. So we wanna say something about, let's start with three cities. So we start with San Francisco, New York and Boston. So we have in every column, we start with the headers, we have San Francisco, we have New York, we have Boston. And we do that in the first row as well. And we're basically gonna say, the distance from Boston to Boston is zero, the distance from Boston to New York is X, and the Boston from San Francisco, and so on and so forth. If we wanna do this for every city in the world, you get a pretty big Excel file.

• 12:13Speaker 00

  Yeah.

• 12:14Speaker 02

  Right? A pretty big Excel file. So and now if you said, now let's have a machine calculate any random distance between cities, it takes quite some time to do that.

• 12:22Speaker 00

  Yeah.

• 12:23Speaker 02

  But now imagine that you do this with words and that you create this Excel file for every single word that you can find on the web. Long story short, you can't do it. The time it takes to go through that huge Excel file that you created, it just, it doesn't work. It's so the academic research and the philosophy behind this is like, I believe, almost 100 years old. We just couldn't do it because it was just, we still today, we do not have the computing power because it gets slower in a linear fashion. Every word that we add gets a little bit slower. So now the idea came, and this was in the wake of deep learning then 15 years ago. It's like, what if we make that Excel file? But rather than just calculating everything brute force,

• 13:08Speaker 02

  we're going to train a model to predict what that distance is. And that worked. And people were like, whoa, this actually works. And that was the unique thing that happened because now all of a sudden we can, so I started my career with a thing from Stanford called Glove.

• 13:25Speaker 00

  Okay.

• 13:25Speaker 02

  Trained it on Wikipedia. And now all of a sudden, we could just do that. I could just do that on my laptop. And that was, there was a breakthrough.

• 13:31Speaker 00

  The where I lose you slightly here is the way that I understood vector indexing was that these are essentially the pre-calculated distances between different vectors. But you just said that in the Excel sheet model, it becomes very difficult to do because it scales linearly in terms of complexity. The prediction element that you just said, or the probability element that you just said, can you, can you double click on that and how deep learning allows us to predict or project distances and be more efficient?

• 13:59Speaker 02

  So if you take a sentence like, uh, what? The Eiffel Tower is in Paris, right? That's a good example. Let's just, let's make it even easier, like Eiffel and Tower. Just those two words. So what you do is like you encounter that word group in a, in a sentence. And then back in the day, you had to calculate all the corpus of Wikipedia. How often does that happen?

• 14:21Speaker 00

  Right.

• 14:21Speaker 02

  But now we train a massive model. What do you think the distance is between Eiffel and Tower? And we call that co-occurrence.

• 14:29Speaker 00

  Okay.

• 14:29Speaker 02

  Co-occurrence in a sentence. And then the model started to predict, well, I predict that the distance between the word Eiffel. And the word tower is one. And that is where it became really good at. So the more we train it, the better it got at it. There's some caveats there, but sure, just for the sake of argument. And now all of a sudden, rather than going to the whole data corpus of Wikipedia that we had in our mental Wikipedia page to say like, is actually the often the distance in the Eiffel Tower one? We now had a model that could predict this way faster than brute force calculating that by going through a whole Excel file.

• 15:05Speaker 00

  So it took something that was essentially impossible, going through the whole Excel file of everything, and made it not only faster, but possible. So it was kind of a double win and that unlocked quite a lot. Now, we start with vectors, which is numerical distance and magnitude and direction. We have vector embeddings, which are numerical values associated with individual bits of data. Vector indexing, essentially figuring out how far apart they are. All of that is stored inside of a vector database. And this brings us to EV eight because you guys have made an open source vector database.

• 15:37Speaker 02

  Yes. So because now, when we had, so 10 years ago, when we had these embeddings and we started to work with them, we were like, okay, this is great. We now have these embeddings, we can do stuff with them. But then the thought, what do you do with stuff that is in a space? You calculate distances. You want to make, you want to know that there's a relation between Eiffel Tower and Paris and those kinds of things. Or if you have documents and you want to store documents or the images, those kinds of things. And then the thing was, hey, wait a second. The vector embedding was a very, very obscure data type.

• 16:10Speaker 00

  Yeah.

• 16:11Speaker 02

  All of a sudden, it comes into prominence because of this machine learning thing where even now, if you take the modern models today, like from Hugging Face and you would open them up.

• 16:20Speaker 00

  Hmm.

• 16:21Speaker 02

  It's vector embeddings all the way down. And you need to store them somewhere and you need to have them somewhere. And we were like, hey, wait a second. There's no database purpose built to deal with this new data type. And just not to nerd out too much on databases, but because the, the thing is that there tends to be this thing happens in the database industry, which is a rather large industry for people listening, like is it a large, it's a very big. Industry.

• 16:47Speaker 00

  Oracle's a big company, I hear. Yeah.

• 16:49Speaker 02

  Uh, yes, yes, but um, look at in the NASDAQ, right? So you see a lot of database companies are on the NASDAQ and uh, including Oracle.

• 16:56Speaker 00

  Yeah.

• 16:57Speaker 02

  And so what happens is like, if a new data type comes into prominence, so for example, think back in the NoSQL day with documents, graph databases, and so on and so forth. That tends to be like a new wave of database companies. And we're part of that. So when I started this, I was not aware of this market dynamic. I learned that by doing it. But that's why we saw that, hey, new data type, who does this? Uh, nobody. Ah, we can.

• 17:21Speaker 00

  Right.

• 17:22Speaker 02

  So that's kind of how you jump into the, that opportunity. Because that new data type emerged, what we didn't know back then, that ML would turn into what we now call AI and how big it would become. But yet I could claim I would, that I foresaw that, but of course I didn't.

• 17:37Speaker 00

  No, sometimes you're early and then suddenly the tree drops a lot of, you know, fruit down upon you. And that's called uh, product market fit by both preparation and luck.

• 17:48Speaker 02

  Yes, there's a famous quote from, I think it's from Mark, and the reason where he says, you know, you have market fit if, uh, and I might be paraphrasing, but he says like, if the market puts two fingers up your nose and pulls them towards you, that's what happened, what AI did for us, right? You were like, whoa.

• 18:04Jason Calacanis?

  Hey founders, you want to build the next great billion dollar business, right? Well, you're gonna need two things, a killer idea and a properly set up company. I can't help you pick the idea, but I can help you get your business started the right way. With Northwest Registered Agent. With Northwest Registered Agent, you can form your business for just thirty nine dollars plus state fees, one of the most affordable options out there. No hidden fees, no upsells, just a simple, straightforward process. And here's how it works. In ten clicks and ten minutes, your LLC is officially formed and ready to operate. Northwest handles all the paperwork, fast, accurate and hassle-free. So you can focus on building your business, not navigating all those crazy legal forms. Thousands of entrepreneurs trust Northwest Registered Agent

• 18:50Jason Calacanis?

  because they make business formation affordable, efficient and stress-free. And with their expert team standing by, you'll always have.

• 18:56Speaker 02

  Have the support you need. So here's your call to action. Don't let paperwork hold you back from your entrepreneurial dreams. Get started today at northwestregisteredagent.com/twist. That's northwestregisteredagent.com/twist. For just $39 plus state fees, your business can be up and running in no time at all. Get more value, more convenience, and more peace of mind only with Northwest Registered Agent.

• 19:20Speaker 00

  That's really funny. The way that I heard that was you have product market fit when customers are ripping the product out of your hands, but that same idea. It's when the market is coming to you and going more, more, more. Okay.

• 19:31Speaker 02

  Yes.

• 19:32Speaker 00

  But let's talk about why the market cares about what you've built. So the way that I understand it is vector databases unlock several things that make AI applications essentially useful. One of which is RAG or Retrieval Augmented Generation. And for folks who are not familiar, the way that I understand RAG is you take an LLM. You have your own data and it allows you to have your own data interplay with the LLM without needing to retrain the original model.

• 19:56Speaker 02

  Yes. So there's, there's one, just one piece of history, I think, that's relevant here.

• 20:01Speaker 00

  Please.

• 20:01Speaker 02

  And that is one of the issues we had. So what we did was if we stored data in the database with these vector embeddings, we looked at the words, for example, in a paragraph of data. So you might remember that I built you a little prototype.

• 20:15Speaker 00

  I do.

• 20:15Speaker 02

  Back in the day. And how we did that was we looked at the individual words, I mean, of the database, looked at the individual word, assigned all these vector embeddings and then said, what's the center of this paragraph? And that's how it was placed in vector space.

• 20:29Speaker 00

  Yes.

• 20:29Speaker 02

  The challenge with that was that the more words you had in it, the bigger your paragraph became, the more centered in the center of vector space, losing its meaning. That was.

• 20:38Speaker 00

  Essentially, it became too generic.

• 20:40Speaker 02

  Yes. So, and the reason for that was that the individual vector embeddings don't know anything about the next vector embedding in the, in your paragraph, right? And now there was a paper released that was called, uh, uh, essentially the famous transformers paper.

• 20:59Speaker 00

  Yeah.

• 20:59Speaker 02

  Who solved that problem? So rather than looking at that co-occurrence, does Eiffel sit next to Tower? The model play, and I'm, you know, I'm, I'm doing air quotes here.

• 21:09Speaker 00

  Yeah.

• 21:09Speaker 02

  It played a game of telephone. So it's like, hey, next to Eiffel, we have Tower and Tower we have In, and next to In, we have Paris. And that's how it started. And now all of a sudden, the model kept the context. Or it could predict the next token in that in that string of vector embeddings.

• 21:28Speaker 00

  And that is generative AI.

• 21:31Speaker 02

  Yes, exactly. So what happens is if you have a sentence, so the Eiffel Tower is in... Takes all these individual words that is translated into a so-called token. Token is just an ID, just as just an ID, so it turns it on. It requests from the model the vector embedding for all these individual IDs. Creates a string of all these vector embeddings. And then it says, "Model, please, based on these vector embeddings, predict the next one." It returns a vector embedding. We do similarity search on it. It says, "It's token 512." We say, "What is token 512?" Well, Paris.

• 22:07Speaker 00

  Eiffel Tower in...

• 22:10Speaker 00

  Paris.

• 22:10Speaker 02

  Paris, yes. So that's how it does it. And so that existed, by the way, this model. So we had something in inside WeV8. We called it generative search where we did that.

• 22:20Speaker 00

  Yeah.

• 22:20Speaker 02

  But it was not really adopted. And then there was this company, you might have heard of them, they're called OpenAI. Oh, yeah, yeah, it rings a bell. Yeah, yeah, yeah. And they created, they's like, "How are we gonna show this to the world? How are we gonna do this?" Because I have friends who work at OpenAI, work by the... And they were asking, we, all of us were like, "How are we gonna make people aware of this?" And they just did this smart thing with the chat interface. And good for them, because now, all of a sudden, people were like, "Whoa, this is amazing." And the immediate follow-up question was, "How do I do that with my data?"

• 22:53Speaker 00

  Yes.

• 22:54Speaker 02

  And we were like, "Hello."

• 22:56Speaker 00

  Good news. Well, so, I mean, you were there with the technology to help people take their structured and unstructured data and get it into a vector database so that it could be used in an AI context. But did, did WeV8 need to do any work on the models themselves that actually did the vector embeddings and assigned all these values? Or was that technology already i- in the market and off the shelf, if you will?

• 23:20Speaker 02

  Okay, so the, the answer is no. So if you look at the, the original research paper, the Retrieval Augmented Generation research paper, it was more sophisticated to what we do today. So what that paper argued was, you know, you can make these big models with information. What if you can somehow make them smaller? So think about like a understanding the language, but... And with zero knowledge, but then it knows that if I need that knowledge, I need to augment what I generate by retrieving something, right? So retrieval.

• 23:52Speaker 00

  Augmented.

• 23:52Speaker 02

  Generation.

• 23:53Speaker 00

  Yeah.

• 23:53Speaker 02

  And that was, that was the idea. So the first, we call that internally, we call this primitive RAG. So that was just use the vector database to retrieve the unstructured data, pipe that over to the, to the model and generate the output that already created a lot of value for people.

• 24:08Speaker 00

  Uh huh.

• 24:08Speaker 02

  But now there are two interesting things happening. So one is like, there's a lot of work happening to more intertwine the models and uh, the database together.

• 24:18Speaker 00

  Okay.

• 24:19Speaker 02

  Funny side note, that's where the name WeV8 comes from, from weaving the model and the database together. But the second thing that's happening that people are like, hey, wait a second, if we do this RAG stuff, so we have a query, run to the model, get it from the database to the general answer, that's a one-way street. What if we just pipe that back into the database? We call that an agent.

• 24:40Speaker 00

  I was go, aha, I got, this is where I was going with this. So actually, I'm going to stop you there and say, let's just make sure we understand what vector databases are used for today. And I'm literally pulling from your, um, your sales material on the WeV8 website. But uh, similarity search, hybrid search and enabling RAG are what I, what I might consider the key kind of corporate use cases for vector databases today?

• 25:05Speaker 02

  Yes, that, that is correct. But there's, if we, there's another way, another perspective to, to, of course, that's how we, you know, tell the world what, what you can do with WeV8, but I'm assuming that the, the people, I mean, I'm 99.9% sure that people listening to this are tech enthusiasts, right? So.

• 25:21Speaker 00

  Yeah.

• 25:22Speaker 02

  What is interesting is that of course, it's a technological innovation, of course. But what cannot be underestimated is the developer experience. So one of the things that happened there as well is that the way that you can build these kind of applications that you just mentioned, where the functionalities need the hybrid search and the offloading and that kind of stuff. But what's more exciting is that you can now build this in like five lines of Python code, combining the model and WeV8. And that is of course the new thing in this new paradigm. AI infrastructure.

• 25:54Speaker 00

  So essentially, we've the eight's vector database handles the the tricky bits of vectors for you. The major models handle all the tricky bits of making a large language model. And then me, the developer with a bucket of data and a open AI API key, just to pick one provider, maybe an anthropic API key. I can very quickly go bing, bing, connect them together, weave them perhaps. And then out comes an application and I look like a genius internally and I get a raise and I'm the CTO.

• 26:22Speaker 02

  Exactly. And that is the so the what's happening now, a lot of work that's happening now is that the the barrier to entry for developer is going down fast.

• 26:32Speaker 00

  Yes, which means that the aperture for what can be built is getting wider.

• 26:36Speaker 02

  Yes, exactly. And I always like to say that because we often talk about this technology of like, how does it work on the hood? What is the functionality? And I I appreciate all that, but I also want to say how important it is to help the developer, because you know, there's like a lot of genius developers walking around, you know, who know how to do these things, but there's also a lot of people, just if you're just, you know, out of college and you work for a company and you want to build an Iraq application. Yeah, that stuff's not easy. So we're doing a lot of work to help people just to get started with five lines of code.

• 27:06Speaker 00

  Yeah.

• 27:06Speaker 02

  And that cannot be underestimated. There's a lot of work happening there in these new AI infrastructure companies to help not some developers, but all developers to build these kind of applications.

• 27:17Speaker 00

  Great segue. We're going to talk business model, then we're going to get to agents really quick. So the thing about we've the one element that I like about the company is that it's an open source piece of software with services attached to it. And these seem to come in two varieties. One is essentially a serverless managed instance. And then you also have what I would call enterprise partnerships with major cloud providers. I think on the website, you have AWS GCP and also Azure, so the big three, which means that I can go essentially if I'm an Azure customer and I can spin up we've eight vector databases on my existing cloud infrastructure. Apart from that, is there another element to the business model that founders listening should understand? Or are those the two main planks today for we've eight?

• 27:59Speaker 02

  Those are the means. We also have something called a BYOC, which stands for bring your own cloud. But the nice thing is, and that's your point, is that thanks to the open source model, there's a wide variety of deployment options. So that's correct. Yeah.

• 28:12Speaker 00

  Yeah.

• 28:13Jason Calacanis?

  Nailing product market fit is every founder's top priority. But once you've got your product dialed in, you need to focus on selling it, especially to big customers. You know, we call them lighthouse customers. Some people call them whales. To land those deals, your security compliance has to be rock solid. Certificates like SOC 2 or ISO 27001 are the keys to building trust and unlocking these opportunities. But they take time and energy, pulling you away from building and shipping a beautiful, great product. And that's where Vanta comes in. They're your partner. Vanta is the all-in-one compliance solution, helping startups like yours and mine get audit-ready and build a strong security foundation quickly and painlessly. How do they do it?

• 28:54Jason Calacanis?

  Well, Vanta automates the manual security tasks that slow you down, helping you streamline your audit. And the platform connects you with trusted VC SOs to build your program, auditors to get you through your audits quickly, and a marketplace for essentials like pen testing. So here's your call to action. Whether you're closing your first deal or gearing up for growth, Vanta makes compliance easy. Join over 8,000 companies, including many Y Combinator tech stars and launch startups who trust Vanta. Simplify compliance and get $1,000 off at vanta.com/twist. That's V A N T A dot com slash T W I S T.

• 29:30Speaker 00

  When it comes to thinking about the revenue mix at WeV8, you have stuff aimed at smaller developers and stuff aimed at larger corporate customers. What's driving the majority of revenue growth for WeV8 today?

• 29:41Speaker 02

  Nothing has changed there when it comes to traditional infrastructure companies. Enterprises, they pay the most from a qualitative perspective. Startups are more quantitative, right? So lots of startups, but they have smaller bills. Enterprises, they pay more, right? So that's very, there's nothing new under the sun there.

• 30:01Speaker 00

  Totally. That's what I figured, but always worth making sure that something unexpected isn't happening. You know, we were joking earlier about the market coming to you and being a little bit early and ahead of a major wave. I presume WeV8 is growing very quickly. What can you tell us, Bob, about the company's growth last year and what you're shooting for this year?

• 30:19Speaker 02

  So indeed, we have an open source model. So what's important to bear? In mind, if you build infrastructure, that takes time and investment to actually build.

• 30:26Speaker 00

  For sure.

• 30:27Speaker 02

  So we started to monetize halfway through 2023. And we started doing that with our serverless offering, which is just shared resource and those kind of things. That went very fast. And then we started to also get the, yeah, because it's the startups, right? So they they're relatively smaller, but they they work, you know, they they just want to build these new kind of things.

• 30:49Speaker 00

  They adopt quickly.

• 30:50Speaker 02

  Yes, but when we released that, the when we got to our first million dollars was like, that was just in an instant. We're just all these kind of developers. But then something interesting happened because we also got the first enterprise requests in. But the problem was I didn't have any enterprise sellers. I mean, I now have like built last year a whole team. And I remember, so the anecdote that I have there, that was like the first ever enterprise contract. They asked for an SLA, so I was like, how do I get to an SLA? So I just downloaded one from the web. So these developers, they already built, they were ready to go live. And I was talking to procurement.

• 31:26Speaker 00

  Oh, no, not procurement.

• 31:29Speaker 02

  This guy sends me an email and he says like, hey, come on a call. So just I joined him on a call and he held up like the the procurement. He said, what's this SLA? I always say, what's this? I said, that's a SLA. He said, you have to promise me one thing. He said, I'm going to sign this, but never ever show this document to anybody else. And then I was like, now it's the time to spin off our enterprise skills. And now we have an amazing team with enterprise sellers. And so we saw a lot of growth last year, right? So and that was because of the enterprises going live as well. And that and and that was also related to the fact that a lot of these applications went to production. So that is like, that's what's what happened, basically. And so that's all gone well.

• 32:12Speaker 02

  It's it's it's the good old old fashioned, old fashioned story of open source infrastructure project. Gets adoption, wants to scale up, starts bottom up. Enterprise sellers come in. It's it's that's again, a very classical story, but it's excited to see this, to see this happening. And it's like a. There are even database companies, and I will not mention them, but on their earnings calls, we are mentioned as like, what do you think about these new kind of players? So it's amazing what's happening right now, and it's a, I couldn't be, you know, prouder than I am today. So it's like, we're like, we just crossed a hundred people. So it's just fantastic.

• 32:52Speaker 00

  Yeah. So let's, let's talk about the future then, because you mentioned agents, and this is the thing that everyone's been, been talking about. I, I think if I read the phrase agentic AI one more time, I'm going to lose it. I think that what we've ended up with is the word agent meaning too many things at once, and therefore meaning nothing. Now, earlier, you set up a very specific mental model of how agents work, which was, if I may, feeding information back into a model after a RAG process, but I may have gotten that wrong. So Bob, from the WeVIA perspective, what is an agent? And is it a real thing or a marketing term?

• 33:25Speaker 02

  So what an agent is, an agent does something with your data, right? So rather than just presenting it, you give it a prompt to do something with it. So let me give you this simple example. We have this concept of generative feedback loops. So when data goes into the database, there's an agent that looks at it, and you could give the prompt, every data object comes in, that comes in, needs to be written in American English. So if you send in a million in American English and one in that's in Spanish, it will store it as American English, right? So that's an example of a very tiny, simple agent. So is it a real thing or is it a marketing term? The answer to this question

• 34:04Speaker 02

  is yes. So it's absolutely a real thing, but of course, it's also a marketing thing. And I don't mind that too much, because it's like a, yeah, we somehow need to explain to the world what's, what's happening, right? And then the market kind of consolidates on certain language and that kind of stuff.

• 34:22Speaker 00

  Yeah.

• 34:23Speaker 02

  But it's definitely a thing. So the steps that we've been taking was like from vector search to RAG, which was a one-way street. To the agents where you create these these feedback loops. So what we see is that the majority of customers started with vector search. So we have a lot of these e-commerce kind of things. We now have like a lot of Ask AI features that we're that we're powering. Uh, let me come with an example. So the search in the sort of Ask AI feature in supplement, for example, that's a great example of that. So that runs on we've yet with a model, then that's RAG. And now we see that people start to build the new things with these agents. The reason why that's so interesting was that if you go back in time to vector search, then

• 35:06Speaker 02

  the devil advocate's argument for what we're doing was like, yeah, this is all great, but isn't this just search?

• 35:12Speaker 00

  Right.

• 35:13Speaker 02

  And that was kind of, yeah, yeah, it was kind of true, right? So, and but then all of a sudden, this new unique use case emerged in RAG. And now this new use, these set of use cases, because of what we like to call these agentic architectures are emerging. And all of a sudden validating more and more and more and more the existence of what we call these these vector databases. So that's why it's so important to us and that's why we talk about it so much. Because we believe that that's a unique value that not from a technology perspective, but from a business perspective that we bring to the world.

• 35:44Speaker 00

  So, I mean, just to be clear, without vector databases, there is no agentic AI. Period. Full stop.

• 35:50Speaker 02

  Well, okay, so you know, if you wanna, if you wanna open a bottle of beer, you can also use a spoon to do that, right? So it's like a, if you want it open, you can always open it. But ideally you just use beer opener, right? So it can open.

• 36:02Speaker 00

  Yeah.

• 36:02Speaker 02

  And that's what the vector database does. It just, it just makes it easier to do that. It's just, it's built for it. And that's why people adopt it and use it. That goes back in time. You you mentioned like the the the famous old school database in existence is Oracle.

• 36:16Speaker 00

  Hmm.

• 36:17Speaker 02

  There's probably somebody listening to this podcast, sitting there angry, it's like, I can do everything in the world with my old database. Sure.

• 36:26Speaker 00

  Yes.

• 36:26Speaker 02

  But turns out that the majority of developers just wanna use the tooling that is built for those specific use.

• 36:31Speaker 00

  Exactly. And this is the old famous uh, comment on Hacker News when someone announced Dropbox and they're like, oh, that's just a quick data pull and I can build that myself. No one's going to use it. Well, it turns out it's a $10 billion company.

• 36:42Speaker 02

  Yeah, but if I may say something about there is, we use that that one, we use that often internally at we've yet. Exactly that that thread on Hacker News because, but sometimes people like hardcore developers, like the the the best, the the the crown of the crown of developers sometimes forget is that not everybody is like them. And that is a large group of developers that just wanna build great software for their own business, for the company they work for.

• 37:10Speaker 00

  Sure.

• 37:10Speaker 02

  And they need the help, they need the tooling to do that. And that's where the developer experience plays such an important role. And that can, in my opinion, cannot be underestimated.

• 37:21Speaker 00

  Uh, Bob, can you leave us with just your perspective on where the AI industry is going this year from the kind of like enterprise app perspective? You talk to a lot more people than I do who are probably willing to share more than they're willing to share with me. So in conversations with customers, partners, rivals and so forth, what do you see happening this year that we should be looking forward to?

• 37:42Speaker 02

  So what I really hope this year that we will see is the paradigm shift that we're expecting from AI. And let me explain what I mean with that. Is that the biggest problem in data today, since day one, we're storing data, is master data management. The, if I, excuse my French, the shit and shit out paradigm has not changed. But thanks to the agentic architectures, for the first time, we can have these models, they can have a corner called opinion on your data. And for the first time, as I always like to say, we can turn chicken shit into chicken salad based on the data. That is the biggest issue we've seen. I hope you don't mind me saying, otherwise you have to cut this out.

• 38:22Speaker 00

  No, we have a, we have a beep, we'll just put that in.

• 38:26Speaker 02

  Yeah. So, but this is, this is the first time where we're seeing that the biggest problem in data management in general, master data management, that there's a solution at the horizon. And that's enabled through these agentic architectures. And that's why I'm so excited about it. Am I excited about all that other stuff? Of course. But this is not, this is the paradigm shift.

• 38:45Speaker 00

  Why does fixing that crap in crap out paradigm change the world? Because to me, we've gotten this far with that chicken shit to chicken salad. Why do we need this? And what will that, what will that change for the industry?

• 39:01Speaker 02

  The issue is like, yes, we've gotten very far, but we're not where we wanna be. So I recently spoke to a CTO of, of a large company and he said, if my leadership asks me, like, how many? Products that we sell globally last year, I cannot answer the question. I have the data, but it's so such a mess. And I cannot have humans try to fix it. It data is coming in faster than than the people can fix it.

• 39:24Speaker 00

  Yeah.

• 39:25Speaker 02

  I I cannot solve this problem. So this is a huge problem, but let's not forget. That is solving a system problem, very important. But it will also open the door to new businesses and new ways for people to new new products, new startups, new ideas based on this new paradigm. And what that is, I don't know, because if I would have known, I probably would have be running those startups as well. But the that is the the what I'm so excited about is going to solve that age old problem of bad data and is going to open the door to new products, new solutions, be it in the enterprise, be it in startups, like everywhere.

• 40:01Speaker 00

  Well, that is a great note to end on because you just told people that might be looking for an idea, what to build next, what to go forth and and do. That's an enormous possible new opening.

• 40:11Speaker 02

  Yes, and also this is the time. So do you remember those days that people like there was like, you can build these mobile apps? Or these mobile, do you remember that people were like playing around with this thought?

• 40:22Speaker 00

  Vivid.

• 40:23Speaker 02

  This is it for AI now, start now, not next year now. It's that's how excited am I I am about this because this is the start of the new paradigm. So this is the time to build.

• 40:35Speaker 00

  I am so excited for 2025. I think it's going to be busy as heck. And I'm also really glad that by having you back on, I got to go back through all my notes about what is a vector? And I don't live in your world every day, Bob.

• 40:49Speaker 02

  Thanks so much for having me.

• 40:50Speaker 00

  All right, Bob, thank you so much.

• 40:52Speaker 02

  Thank you.

• 40:54Speaker 00

  And next up, co-founder and CEO of Lumen Orbit, it's Philip Johnston. Lumen Orbit is a bet that in the future, we are going to put our digital brains in space instead of inside on the ground floor of the local gravity well. If you care about how big a rocket we can shoot up into space and how much it can carry. Well, you're going to absolutely love what Lumen Orbit wants to do with lower launch costs and greater launch capacity. Let's talk. If there's one thing we talked about a lot in 2024, it was the need for more compute to help power our AI future. Be it new chips from Nvidia, new data centers around the world. It was a recurring topic for a reason. We're doing a lot with AI and it's very compute intensive.

• 41:37Speaker 00

  But when we think about what we are doing to power that compute today, it's worth keeping in mind how much energy goes into the process. So if we take a look at this chart right here, this is a map of the US breaking down each state by how much of their total power consumption data centers use today. Now, there's a handful of states in the 10 to 15% range. There's one state that's 26%, but most are between 1, 4, 5. However, things are going to change. Here's some data from Bain showing how much the anticipated gain is in energy demands for data centers over the next couple of years. It's going to be exponential, probably. And therefore, we're going to need new solutions to help make sure

• 42:18Speaker 00

  that we can power all the compute we need to build the AI that we want. So what are we going to do to find all of this power? Well, we could turn to fusion. Some people like Sam Altman think that that is going to be a near term solution. Some people want to get back to building more nuclear reactors. But there is this thing called the sun that we can also use to pull out energy. Now, I'm sure you're thinking about solar panels and great farms here on earth. But what about solar panels up in space powering data centers that are up in the sky? Well, that's what one startup is doing. So I want you to please welcome the show Philip Johnston, the co-founder and CEO of Lumen Orbit. Philip, how are you?

• 42:55Speaker 05

  Hello, Alex. Thanks so much for having me. A huge honor to be on the Twisted 500, so I appreciate it.

• 43:00Speaker 00

  Yeah, it's a fun project on twist500.com. If anyone wants to go take a look, we're trying to find the, basically the startups that are going to have the biggest impact on the world, both in terms of changing how we work and also making a lot of money. And I think that Lumen Orbit could be one of them just because of the sheer scale of your vision. And that's where I want to start, Philip. So when I think about data centers, I think about large nondescript buildings that I drive by and I think, is that a warehousing facility or a data center? Hard to tell from the outside. You clearly have a different vision. So let's start with where does this idea come from? And then I want to talk about progress made so far.

• 43:34Speaker 05

  Yeah, so we are. Building, as you mentioned, very large data centers in space in order to be able to take advantage of the abundant energy, the ability to rapidly, to passively cooling space and the ability to scale. So the idea came from, we initially were looking at space-based solar and when you have very low launch cost, there's a, there's an argument to be made that that starts to make sense even with this huge efficiency loss transferring data, uh, energy from space to, to earth. It's not a new concept. That's been around since like the seventies or sixties. But you know, I think in twentieth time, the forecasts are that half of all terrestrial electricity consumption will go into data processing. So if we can find a cheap way of getting the data centers to space, instead of having this 95% efficiency loss,

• 44:15Speaker 05

  using microwaves to transfer the energy down, we can just use all that data in the, all of that energy in space. And that's really, really where the idea comes from.

• 44:23Speaker 00

  So the video we just showed, I, I think details the scale of what you're thinking about. Cause we're not talking about a, a little bit of compute power in space. And we're not talking about a couple of solar panels. We're talking about a, I think it's a four kilometer per side square block of solar panels, which I presume is thousands and thousands and thousands of cells. To me, it feels a little science fiction to say that we're going to do this. So I'm curious, are we at the point in which we've sorted out the technology here? And it's more a question of can we execute this economically? Or is there still some technology risk to what you're doing? That means we still have a lot of things to sort out.

• 45:01Speaker 05

  The only risk remaining is that the launch cost needs to come down by a lot. So I always say this to investors, if you don't believe that the launch cost is going to come down by 10x in the next, you know, five years, we are not a good investment. If you think it's going to come down by a hundred x, we're an extremely good investment. And if like Starship PR, you believe it's going to come down by a thousand x, then we'll be the world's most profitable company. Uh, so.

• 45:24Speaker 00

  May it be a thousand x. I mean, honestly, I mean, I, I'm here for it. Just because I'm curious about this, have you been watching the Blue Origin new Glenn launch delays? That doesn't mean that we're not going to get competing heavy launch vehicles, I hope, right?

• 45:40Speaker 05

  It would be great if we, if we have some competition to Starship in terms of the launch cost.

• 45:45Speaker 00

  Okay. So, uh, passive cooling in space, when I think about space, it's cold, humans can't live up there for clear reasons. Um, but is it easy to actually get rid of heat when you're up in orbit? I'm not actually sure about the physics of that because one thing we know is that data centers do consume a lot of power and make tons of heat.

• 46:04Speaker 05

  Yeah, no, it's a great point. And the core part of the technology that we're developing is a very large, low-cost, low-mass deployable radiator. Um, so I mean, the quick answer is no, it's not super easy. And the reason is if you don't have an atmosphere, you don't have convection or conduction. There's, there's nothing against the laws of physics that stop us doing this. It just requires a very large, uh, they call it black body radiation to radiate in infrared into deep space. We just have to keep this black panel at around 20 degrees C or higher, and that will radiate, uh, a lot of heat in space, 800 watts per square meter.

• 46:36Speaker 00

  Uh, for those folks like myself who look at circuits and outlets and with terror and trepidation, uh, how much is 800 watts? I don't have a good feel for, um, how much energy dissipation that is. Is that a lot? Is that not much?

• 46:48Speaker 05

  I mean, a typical household light is around, you know, 20 watts, depending on if you have LEDs. Another way to look at it is in proportion of the solar panels generation versus the radiator. So one square meter of solar panel in space generates around 200 watts, and one square meter of radiator dissipates around 800, 800 watts. So you need, you need about a quarter the size, the surface area of the solar panel on the radiator side.

• 47:12Speaker 00

  So we're going to need a four kilometer by four kilometer square of solar panels to power, I believe it's a five gigawatt, uh, data center. And then we need a one kilometer by one kilometer square of radiator to

• 47:25Speaker 05

  Yeah.

• 47:26Speaker 00

  dissipate that. Okay.

• 47:27Speaker 05

  Yeah.

• 47:28Speaker 00

  That to me is ambitious, but I absolutely love the idea that if launch costs get down low enough, we can at least get everything up there. Once it's up there, is there a risk that the solar panels are going to get dinged by space debris, micro asteroids? I, I, I know a little bit about space shielding, but you're building an enormous target. It feels like up in space. So how do you, how do you keep a safe from little objects that are zipping around so fast?

• 47:53Speaker 05

  To solve the problem of orbital debris, either you fly very low or you fly very high. So in the first two missions of flying, it was called Velo, very low orbits around, anything below 400 kilometers is very clean orbit. You don't really get hit by anything there because There's a low levels of upper atmosphere, so stuff de-orbits naturally and within a few months anyway. It's where the ISS flies, for example. The problem with that is you need propulsion. So as the satellite gets bigger, you need more energy going into propulsion so you don't slow down. So then, once the satellite gets large enough, then you can fly very high, around 1,200 kilometers. That also means you're always in the sun, which is great. The problem with that is then you're into the Van Allen radiation belt, and so you need more shielding from radiation. That's also very clean. There's hardly any orbital debris up there because most people are flying in this LEO band of 400 to 800 kilometers.

• 48:38Speaker 05

  So that's really how you go about it.

• 48:40Speaker 00

  But there's a trade-off then between orbital debris and radiation, and it sounds like it's a better solution, a better problem to have in much higher orbits down the road.

• 48:51Speaker 05

  The amount of compute you can have scales with the volume of the satellite and the shielding scales with the surface area. So as the satellite gets bigger, the total amount of shielding you need as a percentage of the mass of the satellite goes to zero. Essentially, if we have a larger satellite, it's flying far, it's fine to fly higher. So for these small puny ones that we're doing in the demonstration, we're flying very low. Once we scale up, it's okay to fly far high.

• 49:12Speaker 00

  And the demonstrators, I believe, start to fly as early as May of this year. So quickly, tell us what the first demonstrator will be, capacity. What are you going to be able to show this year once you get it up into space?

• 49:25Speaker 05

  Yes, it's about a 1 kilowatt, 50 kilogram satellite. It's about 100 times more powerful GPU compute than has ever been flown in space. We'll have the state-of-the-art terrestrial Nvidia chips. And that's really the big difference of the first one with what everyone else is doing. So normally, you would fly radiation hardened, these jets and chips from Nvidia. They're at least 100 times less performant than the state-of-the-art AI training chips that they have.

• 49:46Speaker 00

  So you're going to put that up into the LEO, and then we get to have the fun conversation of how do you get the data up and down? Because just thinking about what you want to do, I get the idea of having lots of constantly on solar power. I understand the cooling effects you can probably take advantage of. There's lots of space in space. There's many things that make sense to me, but you have to get data up and down, which to me sounds very tricky to do at a high speed to allow this to have the kind of throughput I presume you need. So talk to me about how you get information from down here to up

• 50:21Speaker 05

  Yes. On the first demonstrator, we have three ports of connectivity. So we have a small terminal to connect into the Iridium network. That's another constellation that we've been data through. We have an antenna to connect to customer satellites, and then we have also an antenna to connect to ground stations. It's not great connecting to ground station because you have to wait till you pass over one, but it's very, it's fairly slow bandwidth. On the second satellite, Lumen 2, which we've got booked for launch in mid 2026, it's going to be the first commercial offering, have about 100 times more powerful GPU compute again. That one will have an optical terminal, possibly two optical terminals on it, which will allow us to connect both to customer satellites and also to

• 50:58Speaker 05

  directly into the Starlink network, ideally. No, no contracts signed with that yet, but they announced a product called Laser earlier this year, which stands for Plug and Play Laser, which enables satellite customers to connect directly into Starlink.

• 51:11Speaker 00

  I was going to ask, by optical, you meant laser, but I want to double click on something that you said that I didn't know about. You mentioned customer satellites. And that implies to me data from them to your either satellite demonstrator or cluster or data center in space, and the ability or demand for compute between in-space objects. I was thinking about this entirely as terrestrial to space to back again. It sounds like there could be a space-based demand as well in terms of your ability to collect data, crunch it, and send stuff back without it ever needing to go all the way down.

• 51:44Speaker 05

  There very much is, yeah. There's a huge lack of and demand for compute in orbit right now. People are just not solved the problem of putting high-performance terrestrial GPUs in space, and the initial customers will be, you know, military satellites and other types of Earth observation constellations. And then that enables us to build out the expertise and the capabilities. And as the launch cost then comes down over the next five years, we have a commercial service that produces more cash than it costs to build for the next few years. And that transitions into, as launch cost comes down, this service that can move almost all data centers to space from Earth.

• 52:17Speaker 00

  I'm curious about the venture thesis here, because this is going to be hard. This is going to be super duper duper hard. It's going to be capital intensive. You're going to have to deal with everything from not only building your own hardware and dealing with contracts with a lot of very large companies and governments, but also, you know, launch schedules and getting capacity. And it's the opposite of like enterprise SaaS. And so I'm curious, when you're out there pitching VCs, what element of this is resonating the most with them to engender such an amazing reaction?

• 52:44Speaker 05

  We're at the intersection of three trends, which I think to some people, some people view as obvious. The first one is huge demand for energy. The second one is huge demand for commute. And the third one is the launch cost about to come down by 100x. These three trends intersecting, there's an inevitability to what we're doing. It's really just a matter of the time, the timeframe. And when, when what we're doing works, it's going to be, you, you don't have to explain the TAM to anybody. It's like, well, it's, it's a 10, 10 trillion dollar business, basically.

• 53:12Speaker 00

  Oh, no, there's, if you can make this work, the TAM is infinite.

• 53:15Speaker 05

  Yeah, yeah, exactly.

• 53:16Speaker 00

  Yeah, I know, I'm with you on that. But one thing we have seen, and we just came out of CES. So we heard a lot of really great announcements from a lot of companies is that generations of chips improve. I don't know exactly who you're working with, so I'm only speaking for myself here. But Nvidia has, has raved about demand for their upcoming Blackwell line, replacing the kind of venerable H100s that are out there. And to me, if I had spent all the money to send up my data center into orbit, and then Nvidia came out with a chip that was picking a random number here, three times as good. I'm going to be pretty mad. So how do you handle essentially just like chips losing their in-market primacy when they're in space?

• 53:58Speaker 05

  Actually, in orbit, they have a longer lifetime. And the reason is, so, so you have exactly the same problem terrestrially. So we're expecting four-year life of the chips. But terrestrially, I mean, that's roughly the same as, as terrestrially. The problem with, with, with on earth is if you're paying five cents per kilowatt hour for your marginal increase in, you know, electricity consumption. Essentially in space, our marginal electricity cost is zero. Once we get it up there, and that means that running the chips for longer, five or six years, is more economical than on earth, because on earth, there comes a point where you don't want to pay the five cents kilowatt hour, because they're not giving you enough value back. But in space, they'll always be giving you some value.

• 54:34Speaker 00

  Because the power is free, so there's no real downside. Actually, we have a, we have a table here from the Lumen white paper that I was reading before we jumped on, that I think kind of details the economics of this, because I'm sure that some people watching are still thinking, I'd rather just plug an ethernet cable into an AWS data center. Why would I do all this work? Well, as you point out here, the cost of electricity is enormous when we think about the overall cost footprint over, say, a 10-year data center lifespan. And so walk us through the economics of how actually getting up to space can save lots of money.

• 55:07Speaker 05

  Yeah, so maybe I'll talk instead of a 10 year timeframe, I'll talk about the four year timeframe, which is just the life of the chips, because then the economics of it are very clear. Let's say you run a 40 megawatt data center, which is what you can fit in one full starship payload bay. So it's about 100 tons worth of compute, solar and radiators and satellite structure. If you run that for four years on earth and you're paying, let's say, 10 cents per kilowatt hour, which is the average of the data centers paying, that's $140 million just in electricity cost alone. Versus you can launch it. Yeah, it's crazy. You can launch it depending who you believe. Elon's saying it's going to be $5 million for the launch cost, but you know, even assuming it's $10 million or more. Your solar panels are another $5 million and then everything else nets out. So the cost of the chips is the same, cost of the radiators and cooling loops are the same.

• 55:48Speaker 05

  So instead of $140 million for electricity for four years terrestrial, you've got $10 million for launching solar in space. And then, you know, that's the trade off we're doing. And the real beauty of space is you can scale it. So if you want to build a 200 megawatt data center terrestrially, there's not many places in North America where you can draw that amount of power.

• 56:05Speaker 00

  Absolutely.

• 56:06Speaker 05

  It's like two decade lead time to build that type of energy project. Whereas you can just launch five of our modules, put them, locate them physically in the same place in space. And you're running, in a month, you can be up and running a 200 megawatt data center. And you don't have to stop there. You can go up to multiple gigawatts. And that was the point of the video.

• 56:21Speaker 00

  Can you make the four kilometer by four kilometer data center that we showed earlier, just connect to another one and then make it like, do four of those and then have it eight by eight? And then you could do four of those and have 16 by 16? I mean, this is modular, I presume.

• 56:34Speaker 05

  We're looking at a design actually now where instead of having all of that compute in one spot in the middle, we'd have it running along a spine. And then you can just attach modules to the spine where the solar panels and radiators coming out each side. So you can just keep attaching modules. Yeah.

• 56:49Speaker 00

  That's the coolest thing ever. One thing I am curious about though, is once you get out of a payload bay, you know, you go out to space, you come out. To me, this sounds like a tent in a bag. You have to like, take it all apart and put it all together. And I'm not going to lie, Bill, I have no idea how that works in space. So how hard is it to unpack the package you send up and turn it into these functioning data centers? And how often will that go wrong and cause a problem?

• 57:17Speaker 05

  I mean, right now, that's how all satellites function, even ones with large solar panels. You know, obviously, the International Space Station have very large solar panels.

• 57:25Speaker 00

  Yeah.

• 57:25Speaker 05

  So that that problem is relatively solved. In fact, my co-founder, our CTO, was previously designing NASA's lunar path under mission. And he was responsible for deploying very large solar panels and radiators.

• 57:37Speaker 00

  Well, very, very large for current standards.

• 57:39Speaker 05

  For current standards, nothing compared to what's coming, yeah.

• 57:41Speaker 00

  Yeah, because I've seen those unfurl like wings and it's like, I don't know, several hundred square feet, not

• 57:48Speaker 05

  No.

• 57:49Speaker 00

  square kilometers.

• 57:50Speaker 05

  I mean, essentially, what we're going to be doing is roll, unrolling them because because it's going to be so large. You can roll it. Yeah, you have these very thin, flexible solar cells. So that's, that's the longer term way that we'll be deploying is even on limit to we'll do that. Um, but I would just say also robotics in space and for space construction is coming very soon. Um, like it will be probably five years before data centers are being managed by humanoid robots. And it doesn't take much, a big stretch of the imagination to imagine humanoid robots constructing stuff in space. That does sound a bit sci-fi, I agree, but we don't need that yet. We can do everything by unrolling.

• 58:23Speaker 00

  It doesn't sound science fiction at all, actually, because it's much easier to have humanoid robots in space than it is to have humans in space.

• 58:29Speaker 05

  Much easier.

• 58:29Speaker 00

  Humans are so fragile.

• 58:31Speaker 05

  Yes.

• 58:31Speaker 00

  Like, we are just little bags of meat that's shocked by a small amount of electricity and it's water just talking. Like, it's not, it's not good. Uh, robots, we already have pretty good effectures, effectuators, whatever they're called. And, um, they're not heavy. And if we can just charge them, we can have.

• 58:48Speaker 05

  Yeah.

• 58:48Speaker 00

  Oh, man, I'm so excited. This is gonna be, this is gonna be so cool.

• 58:52Speaker 05

  Like, yeah, yeah, I can't remember. I saw this tweet, uh, maybe it was from Mark Andreessen. He's like, I was sitting in the shower the other day. I was just realizing everything is coming, gonna come true, like the space colon is gonna come true, like humanoid robots gonna come true. It's like, everything's gonna come true. And it is, soon it is.

• 59:09Speaker 00

  I'm just so excited about this and we're going to need computing space. We're gonna need, you know, Varda's working on manufacturing. We talked a little bit about launch systems going up. It seems that everything's pointing in one very clear direction, Philip. And that does worry me a little bit. Are there any stumbling blocks that you can see that could dramatically slow down humanities industrialization of, let's just say, low to higher orbits?

• 59:38Speaker 05

  It's all dependent on getting Starship flying frequently, Starship and New Glenn and any other of these types of rockets. If there were to be some, you know, for example, if a large scale war breaks out with China or somebody, that would be a very bad situation, which is why I think Elon's so keen on doing it quickly because we're at a very, you know, tight window of time now where we can actually do this. But I mean, all the physics is proven now, like Starship re-entered. People, even until very recently, people thought Starship wouldn't, the math didn't math and it wasn't even possible physically. But no, we know that everything now is gonna happen.

• 1:00:12Speaker 00

  Well, I guess then the question is, how are you gonna make sure you stay ahead of your competition then? Because the thesis of low launch costs, you know, solar panel in space and a lower total bill of ownership is gonna resonate with a lot of folks because everyone's building and buying data centers. I mean, I can't go a day without reading, "Amazon pledges $5 billion to Tennessee" or whatever. How are you gonna make sure that you guys stay ahead of potentially state-backed or maybe just like Mag7-backed competitors?

• 1:00:41Speaker 05

  I mean, firstly, I do think there'll be a couple of, if all data centers are in space, it's not like just one company is gonna be doing that. And it wouldn't surprise me if we see Starlink and Azure or Piper doing this at some point. All of the big hypescalers are gonna need this capability, Microsoft, Meta, Google. They don't have Oracle, they don't have Space Arms themselves, so they'll need to partner with somebody like us. But I would say the way that any startup stays ahead is we have a moat in our team. We have the most absolutely kick-ass team from SpaceX and all these, you know, MIT grads and all these very smart people working on this. It's very hard to pull together this team immediately and, you know, we're super far ahead now. I think nobody's even close to what we're doing. And the final moat is, you know, is quite capital intensive. And I think we, certainly

• 1:01:24Speaker 05

  from the startup, in the startup world, we're ahead of everybody in that game.

• 1:01:29Speaker 00

  All right. So I'm gonna have you back on the show once the first one goes up later this year so we can talk about it. And may that launch go well and may everything turn on and beep and boop as it should. It's tricky up there. It's funny how fast something goes from "that will never happen" to "oh, really?" to "of course."

• 1:01:45Speaker 05

  Yeah, you wouldn't, so when we put the white Well, we had quite a few folks being like, these guys are crazy. This is never going to happen. And in the last like three months, it's everything seems to have changed. And now it's like an inevitability almost. Um, so yeah, it's true.

• 1:01:59Speaker 00

  Well, I'm, I'm, I'm really excited about it. I can't wait to watch the launch. Please make a lot of noise about it because I'm hoping that it goes well. And you proved that this is possible. Because there'd be nothing more gosh darn science fiction awesome than uh several square kilometers of solar powered data center in space. That just makes my inner nerd sing, Philip. So good luck and thank you for coming on.

• 1:02:18Speaker 05

  Awesome. Thank you so much for having me.

• 1:02:20Speaker 00

  All right, friends, that is Twist for this fine Friday. We are gonna have a packed week next week. There's a couple of things going on at the national level that impact the world of startups. So expect us to be on the move and on the mic. Stick close to Twist. We're on all podcasting platforms. We go live on YouTube and all other digital and social places. My name is Alex, X.com slash Alex. Jason is X.com slash Jason. You are X.com slash my favorite person and I will see you on Monday. Bye.