AI Money, Space Computers, and the Startups You've Never Heard Of

Lauren: Welcome back to Tech Insider Weekly. I'm Lauren. And Derek, I have to say this week's stories made my head spin a little bit. Derek: Mine too. We've got billion dollar AI deals, a government lawsuit over model access, and I kid you not, a startup cleaning apartments for free. Lauren: That last one. We'll get there. But OK, so get this. The WSJ this week profiled a $13 billion AI startup. Pitching itself as the cheaper alternative to OpenAI and Anthropic, the budget option in a space that's been burning money at a historic pace. Derek: And while that's happening, Menlo Ventures, which bet early on Anthropic, just closed $3 billion in new funds. Crunchbase reported it's the biggest raise in the firm's 50-year history. Lauren: 50 years. Derek: 50 years. Lauren: Wow. Derek: So the Anthropic bet paid off enough to go... go bigger than they ever have. Lauren: And then there's SpaceX. CNBC reported a computing deal with open source AI startup Reflection worth up to $6.3 billion. SpaceX's Colossus data center is now basically a commercial AI compute platform, Derek: 6.3 billion for compute. That's not a deal. That's a land grab. Lauren: right? And we've also got Groq raising $650. 50, Qualcomm reportedly closing in on a $4 billion chip acquisition (Reuters flagged that one), and circling back, Derek: The Apartment Cleaners Lauren: the apartment cleaners. It is exactly what it sounds like and it says something real about where some AI startups are at right now. Derek: The glamour of the industry Lauren: All right, first up, the battle shaping up at the top of the AI stack. Okay, so get this. There's a startup that just raised $13 billion and their entire pitch is, we're the cheap option. Derek: That is such a Silicon Valley sentence. Lauren: Right? The WSJ had a piece on this Thursday. The company is positioning itself as a cheaper alternative to OpenAI and Anthropic and enterprise buyers are actually biting. Derek: Wait, back up. $13 billion is, that's not cheap company money. That's almost what some mid-sized banks are worth. It's worth. Lauren: And yet the pitch is, hey, you don't have to pay OpenAI prices. We'll get you there for less. Derek: So what does cheaper actually mean technically? Are we talking inference costs, fine-tuning, API pricing? Lauren: Probably all three, honestly. The enterprise AI buyers right now are getting sticker shock from the incumbents. If you're running millions of API calls a day, even a 20% discount is a massive line item. Derek: That's a real number when you're at scale, and it tracks. I've seen companies burn through AI budgets faster than their cloud budgets, which nobody expected two years ago. Lauren: Which is why the timing on this Menlo Ventures news is so interesting. Derek: Oh, the $3 billion fund. Lauren: Three billion. Crunchbase reported it today, largest capital raise in Menlo's 50-year history, and a big part of why anyone's paying attention is their Anthropic bet. Derek: So they backed Anthropic early. That bet looks incredible right now. right now and they come back to market saying give us more money we know how to pick ai winners Lauren: Exactly. And Bloomberg had Menlo's Venky Ganesan on yesterday. He made the case that we're still very early in figuring out which AI companies actually have staying power. Derek: hmm i mean i buy that the foundation model race isn't over but three billion is also that's a statement Lauren: It's a statement that VC is all in. No hedging. Derek: Subtle. Lauren: Very subtle. But then, and this is where things get weird, an AI startup actually sued the U.S. government this week. Derek: Wait, what? Lauren: Business Insider reported it today. A legal tech company called Legion filed suit over a government order that's keeping Anthropic's new models, Fable 5 and Mythos 5, away from foreign nationals. Derek: So the government stepped in and said certain Anthropic models can't go to foreign users. And a startup that depends on those models is suing because they can't serve their own customers? Lauren: That's the read. And let's get real for a second. This isn't just a legal footnote. If a government order can cut off access to a model overnight, Derek: Wow. Lauren: that's an infrastructure risk that almost nobody priced in. Derek: Right, right. We talk about AI supply chains like it's all about chips and data centers, but the model access layer is a choke point too. Lauren: And who controls that choke point? Apparently, the government can, without much warning. Derek: That's the part that should make enterprise buyers nervous. You build your whole product on top of a third-party model, and then one executive order later... Lauren: Your product doesn't work in half your markets. Derek: Yeah, that's a bad day. Lauren: So you've got a $13 billion startup selling itself as the scrappy alternative of VC firm raising its biggest fund. an ever off an AI bet, and now a lawsuit that reveals just how fragile model access can be. Derek: The money is enormous. The moats are still being dug. Lauren: And the control question, who actually owns the AI supply chain, that's not settled at all. Derek: Which makes you wonder, if model access can be locked down at the software layer, what happens when someone decides the answer is to just own the physical compute outright? Lauren: Like build your own infrastructure so nobody can pull the plug on you. Derek: Exactly. And some people aren't just buying servers to solve that problem. Lauren: Oh, some people are thinking considerably higher than that. Derek: How much higher are we talking? Lauren: Oh, you'll see. Okay, so speaking of compute control, SpaceX just signed a deal with an AI startup called Reflection worth up to $6.3 billion. Derek: 6.3 billion with a B. Lauren: With a B. And wait for it, Reflection is open source. Derek: Which is the part that should stop everyone in their tracks. Open source AI companies aren't exactly known for needing billions in compute infrastructure. Sure. Lauren: That's a nice story, but show me the numbers. And the numbers here are wild. CNBC's Deirdre Bosa reported that SpaceX has turned its Colossus data center into a commercial computing platform. They've already landed deals with Anthropic, Google, Cursor, and now Reflection. Derek: So Elon is basically renting out the world's most expensive server farm. Lauren: That's one way to put it. And the Reflection deal specifically... Reuters reported it gives the startup access to Colossus II, SpaceX's second-generation data center. Derek: Okay, but why does an open source startup Lauren: need that scale of compute. Like the whole pitch of open source is that the model is free. Where's all that power going? Derek: That's exactly the question. A few theories floating around, one is that training open source frontier models has gotten brutally expensive. You're not just releasing weights, you're competing with GPT-4 class performance. Another is that Reflection may be building proprietary services. This is on top of an open source core. Lauren: Right, the classic open source on top business model. But $6 billion is a lot of maybe we'll figure out monetization. Derek: Right? I've seen startups torch a lot of money on a lot of things, but this one makes me do a double take. Lauren: Here's what I keep coming back to, though. SpaceX turning Colossus into a commercial platform is actually the bigger story, because they're not just a launch company anymore. or their infrastructure. Derek: And the WSJ covered the deal, too, framing it as one of the largest infrastructure deals in AI space in recent memory. $6.3 billion in compute access is not a rounding error. Lauren: And there's a layer beyond this that I can't stop thinking about. Time magazine had a piece about startups literally racing to put AI data centers in orbit. Not near orbit, in space. Derek: Wait, actual orbital data centers? Lauren: Yeah, the argument is latency for certain applications drops if you're processing in space and beaming results down. Also, you sidestep a lot of land, power, and water constraints that are choking ground-based data centers right now. Derek: Okay, that is strange and also kind of makes sense. The power problem alone for AI data centers is enormous. Lauren: The timing matters because the window to get there before the big players claim all the orbital real estate... Date might be short, it's the same land rush logic, just vertical. Derek: So you've got Reflection buying time and a data center that's physically inside a SpaceX facility while other startups are trying to get the data centers into space entirely. Lauren: The acceleration is a little dizzying. Derek: Just a little. And all of this raises one pretty sharp question. Who actually builds the chips running inside any of these data centers, orbital or otherwise? Lauren: Which is exactly where things get interesting. Interesting, because while Reflection is buying time in Colossus 2, the companies building the actual inference chips that make any of this run are raising their own serious war chests. Groq just pulled in $650 million. Qualcomm is reportedly closing in on a $4 billion acquisition. Derek: The hardware layer is heating up just as fast as the model layer. Lauren: And that's where we're going next. Derek: All right, so from orbital compute, slip this to the chip layer underneath all of it. Lauren: Because someone has to build the hardware actually running these models, and this is where things get wild. Derek: Go. Lauren: Okay, SiliconANGLE reported that Groq, the inference chip company, just raised $650 million to scale its cloud platform. And for the non-engineers in the audience, inference chips are... are a completely different problem than training chips. Derek: Break that down. Lauren: Training is how you teach a model. You do that once, it's expensive, it's slow. Inference is what happens every time someone actually uses the model. Every query, every response, that's the real speed bottleneck. Derek: So Groq's bet is that fast inference at scale is where the money actually is. Lauren: Exactly. And if they're right, the company running the fastest The fastest chips wins every enterprise contract that needs real-time AI. Derek: Half a billion dollars is a big bet on speed. Lauren: It is, but look at the demand side: those Colossus-scale data centers we just talked about, they still need chips that can keep up. Derek: Okay, and then there's the Qualcomm story. Lauren: Yeah, Bloomberg reported and Reuters picked it up: Qualcomm is reportedly nearing a four billion dollar deal to acquire an AI chip startup called Modular. Derek: Wait, Qualcomm, the phone chip company? Lauren: That's the one, and this is a big repositioning move for them. Derek: Wow. Lauren: Modular builds software infrastructure for AI hardware, essentially making it easier to deploy models across different chip types. Derek: So Qualcomm isn't just buying chips, they're buying a strategy. Lauren: Right. They want to own more of the stack, not just the silicon but the software layer that sits on top of it. Derek: Let's get real for a second. Does this actually threaten NVIDIA? Because... Because that feels like the subtext here. Lauren: I wouldn't frame it as a direct attack yet—NVIDIA still dominates training. But inference at the edge, on devices, in cars, in hospitals—that's where Qualcomm plays—and Modular helps them get there faster. Derek: Quarter million dollars faster. Lauren: Yeah, exactly. And you've got Groq doing it from the cloud side. Qualcomm pushing from the edge. Two different bets on the same underlying problem. Derek: Which is that the model is only half the equation; the hardware running it is the other half. Speaker 3: Right. Lauren: That's what all this money is actually buying-the ability to run AI fast everywhere without choking on compute costs. Derek: And apparently six hundred and fifty million and four billion dollars are the entry fees for that conversation. Lauren: Minimum. Derek: Okay, so we've gone from orbital data centers to inference chips; but here's the thing about all these massive capital bets. Bets. Not every founder story ends with a billion-dollar acquisition. Some of them end differently. Lauren: Oh, very differently. Derek: Coming up, zombie unicorns, a founder who learned working tech isn't enough, and an AI startup that started cleaning apartments for free. Lauren: The dream. Derek: Switching gears to ground level, from billions in chips to startups that are barely breathing. Lauren: The Economist had a piece this week calling them zombie unicorns, companies that were once valued above a billion dollars and are now just stuck. Big number on the label, no path forward. Derek: And quietly, there are a lot of them. Rising interest rates change the math and suddenly all these valuations that made sense in 2021 just Just don't anymore. The Economist framed it as a slow-moving crisis, which, yeah, that tracks. Lauren: Slow-moving but very real. Nobody rings a bell when a unicorn becomes a zombie. Derek: Right, and that's what makes it dangerous. Companies can operate in that purgatory state for years, burning cash, not growing, not dying. Investors don't write down the valuation, founders don't admit the company's stuck. Everyone just waits. Lauren: Hmm; until they can't. Derek: Until they can't. And then there's the other side-founders who built something that actually worked and still couldn't scale it: Inc covered Figur8 founder Nan-Wei Gong. Lauren: This week, she built wearable tech that brought real data to musculoskeletal care. The technology functioned, Derek: Wow. Lauren: and she still hit a wall. Derek: What was the wall? Lauren: Healthcare's distribution maze, hospitals, payers, reimbursement codes, working tech doesn't buy you a path through any of that. The company got acquired by Myant, which she's framing as the next chapter, not a failure, but the lesson is sharp. Solving the technical problem is table stakes. Speaker 3: Yeah, you can build the thing. Scaling it is a completely different job. Lauren: Right. Different skills, different team, different playbook. Speaker 3: Okay, so the apartment story. Lauren: Yes, Futurism reported on this yesterday. An AI startup out of runway started sending college-educated people to clean New York City apartments for free camera-on collecting data. Speaker 3: So they're training a home robotics model? Lauren: That's the bet they need real world domestic data and they don't have money to pay for it the normal way. So you scrub the bathroom, they get the training set. Speaker 3: I mean, is that clever or is that desperate? Lauren: Both. I go back and forth. The clever read is they found a creative way to generate proprietary data when they couldn't raise. The desperate read is they're cleaning apartments. Speaker 3: With Stanford degrees. Lauren: With Stanford degrees. But honestly, I've seen stranger. What can stranger pivots survive? What kills me is that this is the same week we're talking about $6 billion computer deals and $3 billion venture funds. That's the spread of this moment, from orbital data centers to someone mopping a floor in Brooklyn to collect training data. Speaker 3: And somehow both stories make sense right now. Lauren: That's the AI economy in 2025. The gap between the top and the floor has never been wider and people... And people are hustling at every level of it. All right, that's a wrap on today's episode. And honestly, what a week to cover AI. Speaker 3: Right? SpaceX's AI infrastructure, government orders cutting off model access overnight. This stuff is moving fast. Lauren: The thread running through it all is the people building on top of someone else's model are carrying more risk than they've priced in. That's the real story. Speaker 3: And the money is following that risk. Menlo's $3 billion raise. Groq 650 million, everyone's betting the infrastructure layer is where the leverage is. Lauren: Exactly. If today's episodes spark something for you, subscribe wherever you listen, drop us a review, it genuinely helps, and tag us on social if there's a founder we should get on the show. Speaker 3: New episodes every Wednesday. We'll see you then. Lauren: Thanks for listening, everybody. Don't build on a foundation you don't control. Speaker 3: Words to live by.