Latest / Tech IPO Conversations with Fexingo: Public Software Companies and Capital Markets / Why Tech IPOs Are Hiding Their World Models
Transcript
- Lucas: Look, if you have been tracking the tech IPO space this quarter, you might have noticed a strange silence. I am not talking about the quiet on earnings calls, where everyone is finally giving us the unit economics they promised. I am talking about the silence around the 'world model' promise. For the last eighteen months, every pre ipo tech company was showing off these all-encompassing AI agents. But now, with AMD buying Fei-Fei Li’s World Labs for eight point two billion dollars, the public markets seem to be pulling back. Luna: It feels like they are hiding something, or maybe just realizing that showing the trick makes it less impressive. Do you think investors are tired of the hype? Lucas: I think they are tired of being sold a vision that cannot be audited. Let me give you a concrete example. Look at Modal Labs. They just closed in on a seven hundred fifty million dollar round at a fifteen point seven five billion dollar valuation. That is massive private money. But notice what they are doing. They are not going public yet. And the ones that did go public, or are trying to, are stripping out those flashy world model demos from their quarterly reviews. Why? Because a world model is hard to value when it is still a black box. Luna: So they are trading transparency for optionality? If they show the full model, they risk explaining away the magic before they have monetized it. Lucas: Exactly. It is the classic information asymmetry play, but amplified by AI. When you buy a traditional SaaS company, you look at churn and retention. You know what the product does. When you buy a company claiming to have a world model, you are buying a probability distribution. And right now, the market does not know how to price a probability distribution without seeing the underlying code or at least the inference costs. That is why we see so much private debt financing instead. Companies like Nscale are bypassing the IPO grind because the private creditors do not need a demo deck; they need cash flow coverage. Luna: But isn't that risky for the public investor who eventually has to read the filing? They are getting the leftovers after the private rounds are done. Lucas: It is risky, but it is also smart capital allocation. Think about the AMD deal. Eight point two billion for World Labs. That tells you that the hardware layer wants to own the definition of intelligence. The software layer, the pure-play IPO candidates, they are watching this closely. If they go public now and reveal their architecture, they risk being acquired or rendered obsolete by the giants who control the silicon. So they keep the world model close. They let you use the API, they charge you per token, but they do not show you the engine. Luna: That sounds like a recipe for a major correction if the margins don't hold up once the secrecy ends. Is anyone actually forcing them to be transparent? Lucas: The SEC is trying, but they are moving too slowly. Meanwhile, the market is speaking through its actions. Look at the recent performance of Roblox. It dropped sixteen percent in five days. That is not just about gaming revenue; that is about the market re-rating any tech stock that relies on speculative future growth rather than current cash flow. Investors want to see the math. A world model is math, but it is complex math. And most public company CFOs are not ready to explain complex math in a press conference. Luna: So the opacity is a shield against volatility? If they can't explain it, they can't be blamed for missing a target. Lucas: Precisely. It is defensive opacity. But there is another angle here. The cost of running these models is exploding. We talked about AI infrastructure spending reshaping IPOs earlier, but now it is about the burn rate. If you are building a world model, your compute costs are tied directly to your user engagement. And engagement is fickle. By keeping the model private, or at least keeping the details vague, these companies can delay the inevitable reckoning of whether the unit economics work at scale. Luna: It reminds me of the early cloud days. Everyone promised infinite scalability until the bill arrived. Now we have the same story with intelligence. Lucas: And just like cloud, the winners will be the ones who optimize the efficiency, not the ones who build the biggest model. That is why Shopify opening checkout to browser-based AI agents is such a interesting move. They are not trying to build a world model. They are trying to integrate existing models into a frictionless payment flow. That is a specific, auditable value proposition. No black boxes. Just transactions. Luna: See, that feels like a business. This other stuff feels like a science project waiting for a grant to run out. Lucas: Well, science projects need grants, and right now, private debt is acting as the grant. We saw that with Nscale and others using private credit to stay independent. They are borrowing against future IP potential. It works until interest rates shift or the valuation drops below the loan to value ratio. Then the secrecy becomes a liability because no one can verify the asset's worth. Luna: That is a scary spot to be in. Being valued on hope when the lender needs proof. Does this mean we will see more acquisitions of these private tech firms by the big public players? Lucas: I think we already are. The amd world Labs deal is the tip of the iceberg. Public companies with strong balance sheets, like Microsoft and Alphabet, are sitting on piles of cash. They are not rushing to IPO their own internal AI teams because they can absorb the R&D losses internally. They are buying the external talent and the specialized models. So the IPO route for pure-play world model companies is drying up not because the tech is bad, but because the buyers are staying private. Luna: So the public market is left with the second-tier players who couldn't get acquired and aren't profitable enough to go public. That seems like a tough niche. Lucas: It is a tough niche, which brings me to why I want to talk to you guys today. If these conversations are useful for what you're building or running, remember that small support goes a long way. A couple of dollars a month is genuinely what keeps these shows going. You can check us out at buy me a coffee dot com slash fexingo. It helps us keep digging into these structural shifts without relying on ad algorithms. Luna: Yeah, and it lets us keep the analysis sharp and free of sponsorship bias. Thanks for supporting the work. Lucas: Back to the main thread. So if the top tier is bought or stays private, what happens to the mid-market tech IPOs? They still need to list. And they are facing a new constraint: they must prove their AI is additive, not foundational. They cannot claim to be the brain; they have to claim to be the hands. Luna: Hands are easier to measure. I can count how many widgets you sell. I can't count how many 'thoughts' an AI has. Lucas: Exactly. And this is where the distinction between inference and training becomes critical for investors. Training a world model is capital intensive and opaque. Inference is operational and measurable. Companies that focus on efficient inference pipelines, like those optimizing for low-latency enterprise tasks, are far more attractive to public market analysts than those burning millions on training runs. It is the difference between a factory and a laboratory. Luna: So investors should be looking for the factories, not the labs. Who is currently positioned as a factory in the current market? Lucas: Look at Palantir. They have been doing this for years. They don't sell a world model; they sell decision operating systems. The data is structured, the outcomes are tracked, and the contracts are multi-year. That is why their stock has held up better than the pure AI plays. It is not sexy, but it is real. And right now, real is paying a premium. Luna: It’s interesting how the 'sexy' label became a risk factor overnight. What used to be a buzzword is now a warning sign. Lucas: It is. And for the upcoming IPO window, underwriters are going to be very particular about this. They will ask for clear separation between core revenue and experimental AI revenue. If you cannot separate them, you are not ready for the public square. The era of the 'ai washed' IPO is ending. The market in late September twenty twenty-six is demanding granularity. Luna: Granularity is the new gold. Can you hide behind a slide deck anymore, or do you have to open the hood? Lucas: You have to open the hood. And that is terrifying for founders who built their narrative on the mystery of the algorithm. But it is healthy for the ecosystem. It forces innovation to focus on utility rather than novelty. We are moving from the age of discovery to the age of deployment. And deployment requires boring, reliable, transparent engineering. Luna: Boring is good. Boring means consistent earnings. I hope more companies choose boring over brilliant next year. Lucas: I suspect they will. The pressure from shareholders who are worried about interest rates and inflation is too high to maintain the illusion of magical growth. We will see more IPOs that look like traditional software companies with an AI feature, rather than AI companies hoping for a software business. It is a regression to the mean, but it is a sustainable one. Luna: So the lesson is: if it sounds too smart to be true, it probably is. Stick to the tools that do the work. Lucas: Right. And for those of you listening who are navigating this shift, keep an eye on the private debt markets. They are the canary in the coal mine. If the lenders stop lending to the opaque models, the public market will follow within six months. It is only a matter of time before the truth comes out. Luna: Let's hope the truth is better than the pitch deck. Thanks for walking us through this, Lucas. Lucas: Thanks, Luna. Keep your eyes on the unit economics, not the universe models.