Transcription
Most people are still talking about OpenAI the way they talked about Apple back in 2008, as if the whole story is who has the best device. Heading into 2026, I think that's the incorrect frame for the entire conversation around AI. So, in this executive briefing, I want to talk about the real question, the one that's strategic.
If you assume that you have a multimodel stack baked in, which I talk about all the time, a lot of leaders are now getting, then ask yourself, what is OpenAI trying to become in 2026? And what happens to everybody else if they succeed? And what happens to everybody else if they fail at that plan?
The cleanest description I've seen is this. OpenAI is behaving like a company operating under significant constraints, not necessarily like a company that has a single coherent product strategy to execute against. This has been really true in the last couple of months. The tension is kind of fundamental at this point given their success.
ChatGPT is being optimized as an engagement container for a billion people, only 5% of whom are willing to pay. While the market's willingness to pay is shifting toward delegation engines, systems that enterprises can purchase where you hand off work and walk away. A lot of the Codex line of direction and strategy seems to me to be headed that way, where these are designed to be fully working autonomous agents, very, very high-quality inference. You'll pay for the inference, but you'll get excellent results. And if you prompt it properly, it will give you fully finished enterprise work product. Maybe that's code initially. I would not be surprised to see that branch out into other places given recent launches in late 2025 from OpenAI.
So, as a strategic diagnosis, this tells you what OpenAI is defending, what it's postponing, and it implies where the tradeoffs are going to be when the system is pressured. So let's dig into that a little bit more.
To understand OpenAI's 2026 strategy, it helps to stop thinking in terms of the product as a singular entity and start thinking in terms of seats, because effectively, I think the right analogy is that OpenAI is running an airline with scarce inventory. In this case, it's like you have an airline that's running a popular route from New York to London, and you just cannot get enough seats on that airplane. In this case, the compute is the scarcity, and they have to allocate seats on that jet between the consumer who's not willing to pay a whole lot, by and large where defaults have to be cheap, where they have to be fast, and the enterprise seat, where outcomes and governance are demanded. You have a lot of standards here. And then there's the investor and capital seat on the plane, where the only real question is, do you have enough cash runway and deals in place on compute to keep the machine flying until you get to cash flow positivity, until you get to profitability?
And so the key thing I want to call out is that OpenAI currently has a distribution advantage. Now, Google can push Gemini and is pushing Gemini through search, through Android, through Chrome, and they're growing faster than OpenAI at this point, but it's still true. OpenAI has the king of distribution advantages in the AI space. But to keep it that way, OpenAI is now in a position where they have to defend that territory. They have to earn and retain all of their users while growing at the margins in a market that increasingly has people who have already picked AI systems other than OpenAI. So now it's not just, "Can I introduce you to AI? Pick one up." It is, "Can I introduce you to AI that's OpenAI's AI? Hey, don't use Gemini." That's a different proposition.
In that world, compute is both a unit economics constraint for consumers and also a capacity constraint for enterprise, because you have to think of it as like, the consumer cares and is price sensitive. Maybe you tip over more consumers into paid if you can serve compute more cheaply, serve intelligence more cheaply. But from an enterprise perspective, you don't necessarily want the cheap intelligence. You want to burn tokens. Sam has said in a recent interview he has enterprises knocking on his door and saying, "We can ingest a trillion of your tokens. Please give us a trillion of your tokens." There's a capacity constraint at that scale where it's like, how do we develop the compute to serve that kind of capacity to enterprise?
This is the conversation that people are missing when they're talking about model quality, because OpenAI's most important shipping service is not the weights in the model. It's actually the allocation of compute. It's where they route queries from consumers. It's what are your defaults on your chat surfaces. It's what are your plan limits, and which experiences do they make easy for you versus which they stay hidden. And you can see that fundamental compute constraint leaking into a bunch of their recent product choices, right? Rollback of slower reasoning by default in ChatGPT 5.2 is arguably an assessment that for free users, the cost and latency is not worth it, and users prefer, frankly, faster, dumber models that are cheaper to serve. And this just underlines the thesis that chat is largely a saturated use case. The free user base is going to be happy with dumber models, is going to shape public perception of what AI is capable of, like it or not. And that's the world that we all live in, including OpenAI.
I saw, um, a survey, I think it was in the last couple of days, that said that 67, 66% of people believe that an AI's answer is either a retrieval from a database or simply reading a prescripted response. Two-thirds. And these are, these are people who use AI. This is why the free user base is having challenges understanding the capacity of AI. We are still in the fundamental product dilemma of what happens when you scale the power of your product by 10x in 2 years, but your chatbot looks the same, and people just cannot figure out how to use it better, and they don't have the mental models to do that. And increasingly, the behavioral evidence suggests that OpenAI is not finding it economically useful to serve that audience, 950 million people who are on the free plan, a high-grade intelligence.
The plan is clearly to use that compute in two big plays in 2026. I think it's pretty clear. Number one is the ongoing deep inference research that will be needed to push out extremely intelligent models for science and medicine, which they're aiming at really aggressively, and to push out a lot of very thoughtful, high-quality inference tokens and make them available to enterprise. Both of those are paid allocations, and the science and medicine one in particular aligns strongly with the long-term research vision that OpenAI has. I know that we talk about OpenAI as a company, and it is, but it started with a nonprofit sense of mission, and I think we are incorrect if we don't believe that that DNA is still strong, especially in the research part of the company. People believe in AGI. They believe in it as if it is something that is worth doing on its own for the benefit of humanity. That is the level of passion they bring. That's frankly what they need to bring to do a task that hard. And in that world, they are going to be interested in focusing on the medicine use cases, the science use cases, the physics use cases, the things that advance humanity.
And I have been in enough organizations to tell you it is not necessarily true that leadership sets the roadmap. In many cases, when you have high-powered research and engineering organizations, research and engineering shape the roadmap, because if you are working on something that your engineers and your researchers actively think is antithetical to what the business is supposed to be doing, they'll just disagree and tell you they don't want to do it, and you can't replace them. So, you'll end up working on what they want to work on, which is usually the harder, more interesting problem. And I don't know, but I suspect that there is a strong democratic component where researchers are leaning into working on interesting problems at OpenAI. And those interesting problems are leaning the company toward science, toward medicine, toward heavy inference, super intelligent use cases that go way beyond what you need in chat. And this is why I've said chat is in many ways a side play for OpenAI, even though they have the biggest distribution advantage on the board right now.
So here's where 2026 gets really interesting. OpenAI is trying to win three different games at the same time. Three different chess games, right? They're trying to win the Frontier Lab chess game. They're trying to win the mass consumer platform chess game. And they're also trying to win the enterprise productivity chess game. And the required tradeoffs there conflict, and they conflict around compute. This is a three-game problem set, and it predicts the organizational behavior that you'd expect. You are going to have what we hear described as code red reallocations. And I think Sam was correct to say maybe that was overblown in the news, because to me, what it read like is less code red drama and more, "We need to reallocate because we have a potentially dangerous chess position on one of our boards." In this case, it was on the mass consumer board.
And when you are trying to reallocate resources and compute between three different games at once, you are going to have difficulty explaining the narrative as a whole, because the narrative is three-pronged. It can feel incoherent at times, because the company is repeatedly reprioritizing to protect the core usage habit loop that they need across all three. Like, to be a winning frontier lab, people need to use your product. To be a mass consumer platform, people need to use your product. And to be enterprise productive, people in the enterprise need to use your product, too.
So, if you've felt some whiplash in the last couple of quarters and you've wondered, "What is OpenAI emphasizing from quarter to quarter? What's shifting?" I think this is the underlying cause. The company that doesn't own the distribution truly cannot treat a consumer habit as optional. Keep in mind, Google owns distribution in a way that OpenAI does not. Apple owns distribution in a way that OpenAI does not. This is exactly why OpenAI would like to get into the device game. They would like to own distribution, because without distribution, their current footprint advantage, the distribution they have, is earned by the consumer habit loop. It's not taken for granted the way Tim Cook can take the iPhone for granted.
Now, adding capital. Now, a lot of leaders will handwave and say, "This is the AI bubble. They can just raise money." I think it's not quite that simple. I agree they can raise, but I think that increasingly in 2026, we need a case for long-term profitability, and investors are going to start to expect it. From the conversations that I've seen in the public spaces, interviews Sam has given, other reports that we've seen on OpenAI, I think that the core flywheel is likely, the core story around profitability is likely that enterprise inference is the long-term profit engine. It's those business class passengers that make the airline profitable. It is business class that is going to make OpenAI profitable.
Compute scarcity does remain the binding constraint for the next few years, and they are betting that the enterprise paying for those heavy token usage for inference, the high-quality tokens that they need to do heavy work, that is going to fund, at least in part, continued frontier model training to support even higher quality inference for enterprise. And if you combine that with one or two big raises and an IPO bridge, you can get across the capex gap and get to profitability. That's essentially the bet. Some of the math actually pencils out there. I know it's really become fashionable to say OpenAI is going to hit a cash wall, etc. It's not really that clear.
If you think about it, Reuters reported OpenAI is in preliminary discussions to raise up to $100 billion at a valuation of, take your pick. I've heard anywhere from $750 billion to $830 billion, alongside rumored IPO preparation that would value the company as high as a trillion, with a possible filing in the second half of next year. This is not background noise. This is capital strategy driving product strategy for all of us at OpenAI, because compute remains the bottleneck that determines what they can ship to whom and to when.
As a reminder, they have said repeatedly that they are not shipping their best models to us, the public, or to enterprises, because they are compute constrained, and their best models internally are compute-intensive, and that just remains a barrier. Recently, Sam Altman told Big Technology that enterprises have been clear about how many tokens they want to buy. I think I referenced that one. And OpenAI is going to, as he put it again, "fail in 2026 to meet enterprise demand." That is a high-quality problem to have, because that single sentence is a bridge between the consumer demand, the reality that AI is here, that people are desperate for high-quality tokens, and that when that scarcity persists, you're going to have to keep making those allocation decisions in ways that shape our pricing, our defaults, the policies OpenAI has, how a billion consumers experience this technology, and also how good the underlying models serve to enterprise are. Will we live in a world where Codex is available only for at high power, only for select engineers at most enterprises, because Codex plans are expensive, not because OpenAI wants to constrain it, but because the compute itself is constrained?
I have wondered if part of the reason why Codex has leaned into the coding use case, and yes, they absolutely, you can use coding for Codex for non-coding use cases. I have done it. I recommend it. I love it. It is used that way at OpenAI. They recommend that too. But if you are in that spot and you're compute constrained, if you're selling the enterprise plan and you're sitting on the sales team in OpenAI, you may have less options on how far those plan limits can go for non-tech, if you remain, as Sam says, compute constrained through 2026. What do enterprise plan limits and what does pay-as-you-go look like in that?
And it's not like there's a free lunch elsewhere. Anthropic is notoriously compute constrained. Google is definitely working on getting to the point where they have enterprise-scale product offerings, but a lot of what they're bringing to the table is tied into the Google Office and Google productivity suite, similar to Microsoft in their models tied into the Microsoft productivity suite. And Google's also tied into the Google Cloud footprint. And so each of these players has different incentives around their unit economics that are shaping where constraints appear.
I do want to take a moment to talk about usage here, because I think this story gets a little bit uncomfortable for people who assume that the current consumer dominance for OpenAI automatically translates to a durable advantage for the company. I talked earlier in this video about the idea that, yes, OpenAI has a distribution edge today with a billion people, but they don't control distribution with a device the way Google does and the way Apple does. Conversion can be structurally difficult in a world where you've already hit 5% paid at scale. A Reuters information story cited internal modeling suggesting roughly a 60-70% upside, upside to 8-12% paid conversions by 2030, which to me, having worked in consumer businesses, feels really reasonable. If you get to 8-12% paid conversions, you have a phenomenal product. It is not a knock at all.
And so if OpenAI is looking for new monetization streams over the top for the 95 to 92% of consumers who will not pay, what does that look like? And how does that shape usage behavior? So let's talk about perhaps shopping assistance that can open up commissions and ads. Maybe separate from the chat so you don't contaminate the chat with ads, but you have ads elsewhere. Maybe looking at spaces where consumers can essentially agree to pay attention with their time and in turn do get useful work back from the agent. When conversion remains hard, the way we're talking about moving from 5% to 8.5% over 5 or 6 years with plenty of hard work with great products, and you have to maybe monetize over the top with ads, your incentives are tough, especially in a company that has a passionate mission for a larger, more intelligent future that may not fit well in the chat, right? Because the company can be simultaneously pushed to defend engagement, to experiment with monetization, and also to continue to sustain the habit loop that you need to keep enterprises knocking on the door for those tokens. It's, it's a fragile place to be, more fragile than people might think.
And that distribution pressure I talked about is showing up in growth rates. I think I mentioned earlier, Gemini's grew 30% from August to November. And ChatGPT apparently grew about 5%. And so Gemini's faster growth is something that is going to be more of a story if it continues into 2026, and we start to see a situation where there are two dominant players, where OpenAI remains very dominant over a billion, but perhaps Gemini starts to hit those billion-person numbers as well.
So, why does this matter for all of us heading into 2026, assuming that we already have a multimodel stack as I've been preaching? Because even in a multimodel world, even if you're in an enterprise and you've set this up so you can swap your models in and out because you don't want to be dependent on one player, the default interface layer sets the mental model for your employees, for the stack, for the people you work with. And the mental model determines whether AI is a toy, a tool, or an operating system inside your business.
And so to me, I think we still are coming back to what I talked about at the beginning, where the chat box itself is illegible. If ChatGPT's mental model for a billion people, and Gemini's to some extent too, remains either a chatbot I ask questions or a nice friend who makes me images, then the product is hiding tremendous capability breadth. It's diluting the peak value people believe they can extract from it. And that does include work implications. And it means that your people at work are going to underuse it, undervalue it, and ultimately not sustain usage.
This is jumping over a bit, but you'll notice that Microsoft ran into this with Copilot. Microsoft is cutting Copilot sales targets because people who pushed the button to adopt it, CTOs largely, are seeing their people not use it. I don't believe that's only a Copilot problem. That is a larger problem with the way we enable chatbots at the enterprise level. People's mental models are sticky. Mental models don't stop at the office door. If you have a mental model of AI from your phone, guess what? It's the same mental model you bring to AI at work.
This is why the enterprise seat can be misleading. Leaders may get premium treatment on their executive seats or whatever, but adoption is driven by thousands of employees who, regardless of the seat that you may buy them, are doing the default. And this is why, whether you're using Copilot or Claude Enterprise or ChatGPT, if all you're doing is having your employees try it out, they are mostly just going to rewrite their emails with it. If the default teaches chat for quick answers, and if that's the default in the consumer world, you get shallow usage without sustained effort. And if you're able to get to the point where you're an AI-native organization, you will be able to teach teams to delegate work and come back to outcomes. That is the bridge that organizations will need to cross to move from that shallow usage pattern. But to do that, you have to deeply engage your teams in ways we've never had to do for traditional software, because they have this mental model that's very sticky from their consumer devices. You have to convince them, regardless of what you use at home, this is how you work with AI at work.
And this teases up the crux. OpenAI's strategy only truly works if they're able to escape this engagement trap and become an outcome engine for enterprise. And so if distribution advantage plus a compute constraint is where we are living now, we're on a jet plane that is like constrained by seats, but it's a popular plane. It's a popular route. The winning way forward is to own extremely high-quality outcomes for the enterprise that drive those enterprise seats like crazy. Basically, you want to be in a position where your experience in business class on this jet is so good, you're just going to get everybody to sign up for your airline to fly to London.
So, that means this explains a lot, I think, of where they're going with Codex. You need to be able to run your tasks really efficiently for long periods of time. You're going to want the ability to run your tasks in parallel. You're going to want to return to finished work with a very predictable quality bar. You're going to want to wrap it in enterprise governance. You're going to want to have enterprise-grade code review and QA, which Codex really leans into. It goes even further than this.
If enterprise inference is really driving the funding engine, then a first-class delegation layer where you allocate the compute essentially, is how you convert to paid outcomes at scale. And this is where, like, there's this weird relationship between the decision to shift consumers onto a cheaper, faster model and the decision to allocate high-quality tokens to enterprise. They might look separated, but one, they're switching compute across, and two, the habit loops are very entangled at the enterprise level. And there are some interesting feedback consequences to choosing to give people cheaper models to play with and then expect them to magically know what to do when enterprises have fancier models at work.
That's why all of this matters for us heading into 2026. OpenAI, of course, is not just another model vendor in the portfolio. It is the company that made the ChatGPT moment. It is the company that is most aggressively trying to become the default layer where work begins, while simultaneously financing a massive compute buildout that, frankly, it's a meaningful chunk of the broader economy if it comes into play. And it ends up being downstream of whether this whole approach of buying business class seats actually holds up.
And so when we talk about the AI bubble, part of why I don't necessarily buy it is I agree with Sam. I don't see a shortage of demand from enterprises for high-quality inference tokens. I see a shortage of human capability in using those tokens. And I think that's a massive question for 2026, and we've talked about that here, but the demand is there. And so I think that's where OpenAI has a case for a financing flywheel that ends up into positive cash flow territory, ends up into profitability, ends up in the IPO space.
So if you want the overall takeaway here, 2026 is the year OpenAI needs to prove it can turn compute scarcity and capital and the consumer habit piece into enterprise outcomes. And it has to do that without letting the pressure of monetization driven by compute constraints deform or twist the product into the incorrect shape. And so the implication for all of us, for leaders, for builders, people who are rank and file, all of us in that are employees, it's that multimodel isn't really the end of the strategy, it's just your starting condition. That the real question that we're wrestling with as we go into 2026 is who owns delegation, who owns governance, and who owns workflow outcomes on the top of those models. Basically, if we are going to have really fancy, strong inference, how do we make sure our people are there so that they can own the allocation of the model? They can own the workflow outcomes we're able to drive. You can delegate effectively to models. That is the question we all have to answer in return as we assess OpenAI's strategy.
I hope this conversation on OpenAI strategy has been useful. I've written some other pieces on how we need to scale up as teams. I think they're very relevant, and I think that OpenAI's strategy will continue to pose a strategic question for us as to how we scale up our people to meet the enterprise demand for high-quality inference that is driving this entire product strategy. Best of flock.