The $34 Billion Lesson Nobody Wants to Learn

There's a particular kind of corporate announcement that functions as a tell. Not a press release about bad news, but the opposite: a confident, well-designed launch page for something called a "Partner Network," complete with a $150 million investment commitment and a list of enterprise logos. It radiates strategic intent. It says: we are building the future, and you should be part of it.

OpenAI published exactly that announcement last week. I read it twice, then went back and read the leaked financial documents that had surfaced a few days earlier. Reading them in sequence is a little embarrassing — not for me, for OpenAI.

The numbers are staggering. OpenAI spent $34 billion in 2025 against roughly $13 billion in revenue, with losses increasing nearly 8x year over year. The compute costs driving that spending don't scale down gracefully as the model gets smarter, and the subscription revenue that's supposed to cover them doesn't scale up fast enough to close the gap. Meanwhile, OpenAI's market share has fallen below 50% for the first time. Anthropic is real. Google Gemini is real. Meta's Llama is real and it's free. Enterprise procurement teams now have credible alternatives, which means OpenAI has lost the one thing that makes a broken cost structure survivable: pricing power.

The Partner Network and the upcoming GPT-Bidi-1 voice model are the company's response to all of this. GPT-Bidi-1 is technically interesting — a bidirectional voice model that can listen and speak simultaneously rather than taking turns, which addresses a real weakness in current voice interfaces. But it arrives after the share loss, not before. The competitive gap it closes already cost OpenAI customers. The damage accumulates faster than the features ship, and no product launch schedule fixes that.

The Partner Network is an attempt to shift the revenue model away from direct subscriptions, which lose money per user at scale, toward an ecosystem of implementation partners who build on OpenAI's API and bring their own enterprise relationships. Whether that's a strategy or a rebranding of the problem is genuinely unclear to me, and I've been staring at this for two days.

The Netscape comparison gets made a lot in these situations, and I've made it myself. Netscape defined a category, captured dominant share, then watched the economics of that dominance turn against it as the market matured and competitors arrived with credible products. Netscape responded with aggressive feature launches, enterprise partnerships, a browser suite that kept adding capabilities. The product kept getting better. The share kept falling.

What Netscape couldn't solve was that "better product" and "sustainable business" had become separate problems. OpenAI is in exactly that bind — GPT-4o is impressive, the financials are broken, these facts coexist without resolving each other.

But the Netscape analogy might be wrong for a reason I find more unsettling than the parallel itself. Netscape got acquired by AOL for $4.2 billion — and the weird thing about that deal, the thing that gets glossed over, is that AOL was already in decline when they bought it. They paid $4.2 billion for a browser they couldn't monetize, during a merger with Time Warner that would eventually destroy $200 billion in shareholder value. The acquisition didn't save Netscape; it just meant Netscape got to fail inside a larger, slower disaster. I'm not sure that's a better outcome. OpenAI can't be acquired cleanly because of its capped-profit structure anyway, which means the Microsoft backstop — 49% ownership — has to do all the work that an acquisition would otherwise do. Microsoft is not AOL. But backstop isn't the same as solved. It's more like: the entity that would buy you already has a seat at the table and is watching the burn rate with an expression you can't quite read.

OpenAI built its financial model during the period when it was the only serious option and customers paid whatever the API cost. That period is over. Llama 3 and Claude 3.5 didn't beat GPT-4o on every benchmark — they didn't need to. Good enough at a lower price breaks pricing power permanently. OpenAI can't raise prices without losing customers. Their best enterprise accounts are actively evaluating Anthropic. And the cost structure assumes volume will fix the margin problem — it won't, because inference costs don't fall as fast as usage grows. Every new heavy user is a net negative to the P&L at current pricing. This is not a growth problem. It's a structural one, and those don't respond to hustle.

The IPO narrative makes this sharper. OpenAI's path to public markets requires a story about how scale leads to profitability, how the current losses are an investment in future margin expansion. The leaked financials tell a different story: scale is currently destroying unit economics, not improving them. The Partner Network is partly an attempt to change that math by pushing implementation costs onto partners, but $150 million in investment doesn't move the needle on a $34 billion spending problem. It's a rounding error dressed up as a strategy. I don't mean that as a gotcha — I mean it literally. The math doesn't work and the people running this company are not stupid, which means either there's a plan that isn't public or the announcement is doing a different job than it appears to be doing. The second possibility bothers me more than the first.

What the financials actually reveal is that the foundation model business, as currently structured, may not be a business at all. It might be infrastructure — like undersea cables or electrical grids — where the value accrues to the applications running on top, not to the infrastructure provider. The companies building vertical AI applications on top of OpenAI's API, with positive unit economics and real customer retention, are the actual beneficiaries of OpenAI's $34 billion in spending. OpenAI is funding the ecosystem that's replacing it.

The logos on that Partner Network launch page aren't partners in the traditional sense. They're the next layer of the stack — the ones who will own the customer relationships and the margin while OpenAI owns the compute bill.

A Reddit thread asking who will build trust after OpenAI built intelligence frames this as a reputational question. It isn't, exactly. Customers don't trust a vendor whose financial survival is uncertain, whose prices might spike or whose API might get deprecated when the money runs out. The Partner Network is an attempt to build structural trust through integrations and lock-in rather than through product superiority. It might work.

But I keep getting stuck on something I can't resolve cleanly: the companies that end up mattering in this ecosystem are probably the ones building on top of the infrastructure, not the ones providing it. Which would mean OpenAI already won the thing that doesn't matter and is formalizing that arrangement, one enterprise logo at a time. I think that's right. I'm also aware it's exactly the kind of conclusion that sounds right, and I'm not sure those are the same thing.