Skip to content

Nate B Jones AI Business

OpenAI Just Filed For Its IPO. The Real Story Isn't The Trillion Dollars.

Nate B Jones argues the trillion dollar valuation is the wrong way to read the coming OpenAI and Anthropic IPOs. The real question is whether the labs can make tokens cheap at scale and build the harness, the work layer of files, tools, permissions, memory, evals, routing and a definition of done, fast enough that companies rent the whole system instead of building their own. He shows why the $200 plans look irrational only if you mistake retail API price for internal cost to serve, why value migrates to the harness once intelligence gets cheap, and why the real lock-in is the harness, not the swappable model. The fork in the road is whether the labs own that harness or their customers do, and he lists exactly which numbers in the S1 filings will reveal which kind of business each lab really is.

Published Jun 14, 2026 11:49 video 20 min read Added Jun 16, 2026 Open on YouTube →

At a glance

OpenAI and Anthropic are both walking toward public offerings, and Nate B Jones argues that the trillion dollar valuation everyone wants to debate is the least useful place to start. The real question, he says, is what public investors are actually being asked to believe, and the answer is two things at once: that the labs can make intelligence cheap enough to serve at massive scale, and that they can build the work layer around that intelligence fast enough that companies rent the whole system instead of building it themselves. He gives that work layer a precise name, the harness, and defines it as everything that turns a raw token into finished work: the files a model can see, the tools it can call, the permissions, the memory, the evals, the routing between cheap and expensive models, and the definition of done. Codex is a harness, Claude Code is a harness, ChatGPT is becoming one, and inside every company the serious AI project is a harness project. The whole IPO bet, in his telling, is cheap tokens plus a proprietary harness equals a trillion dollars, and the fork in the road is whether the labs own that harness or their customers do.

The trillion dollar question is the wrong start

Jones opens by naming the conversation he refuses to have. Both labs are moving toward IPOs, and he knows the whole discussion is about to collapse into one question: are these companies worth the numbers people are putting on them? He thinks that is the least useful place to begin. The trillion dollar number will dominate the headlines, and it will tell you almost nothing about what kind of business is underneath it.

The better question, he says, is what public investors are actually being asked to believe. And the answer is pretty simple. They are being asked to believe OpenAI and Anthropic can do two things at the same time. First, they can make intelligence cheap enough to serve at massive scale. Second, they can build the layer around that intelligence fast enough that companies rent the whole system instead of building it themselves. That is the bet, stated cleanly: cheap tokens and proprietary harnesses equal a trillion dollars.

The real bet, and what a harness is

If that sounds abstract, he makes it concrete with a single distinction that runs through the entire video. A token is raw intelligence. It is the thing you buy by the meter. A harness is everything that turns that raw intelligence into work. He lists the parts deliberately: the files the models can see, the tools they can use, the permissions they have, the memory they keep, the evals that check the output, the routing between a cheap model and an expensive model, and the workflow that tells the system what done means.

Then he grounds it in products you already know. Codex is a harness. Claude Code is a harness. ChatGPT is becoming a harness. And inside companies, every serious AI project is a harness project. That reframing is the whole point of the IPO story. The question is not just whether OpenAI has better models. The question is whether OpenAI and Anthropic can own the work layer that sits above the models.

Figure 1. The distinction the whole video turns on. A token is raw intelligence sold by the meter; a harness is the work layer that wraps it. Jones names the parts deliberately: files the model can see, tools it can call, permissions, memory, evals, routing between cheap and expensive models, and the workflow that defines done. Codex, Claude Code, and an increasingly agentic ChatGPT are harnesses. The IPO question is whether the labs own that layer or merely supply the tokens that feed it.

Why API price is not internal cost

Jones turns to a piece of analysis circulating the day he filmed, which he attributes to SemiAnalysis. It tried to estimate the notional API value of the $200 AI plans from the two labs. The rough claim: a heavy OpenAI user would be getting about $14,000 of value for a 200 buck plan, and a heavy Claude user about $8,000 of value for a 200 buck plan. The obvious reaction is that these companies are lighting money on fire, and he allows that for some users they may be.

But he thinks the sharper read is that API prices are not an internal cost. API prices are retail. They include markup and margin. They reflect the price charged to developers, not necessarily the cost the lab pays to serve the token internally. So the question is not how much API value the user got. The question is what that usage actually cost OpenAI or Anthropic to serve. Those, he stresses, are very different questions, and conflating them is what makes the $200 plan look insane when it may not be.

The $200 plan as subsidy and strategy

Run the math the other way. If the API price carries 70 or 80% gross margins, the internal cost is far below the public sticker price. And if the labs keep improving inference efficiency, model routing, caching, batching, distillation, chip utilization, and everything else that squeezes more intelligence out of the same hardware, then the 200 buck plan stops looking irrational from the outside. It might be a subsidy. It might also be a strategy.

The strategy reading is that the labs are letting power users consume huge amounts of intelligence while they race the cost curve down underneath that usage. They are effectively saying they can afford to serve intelligence close to cost now because they believe the cost of serving keeps falling. And that belief changes the entire IPO frame. If you think the models are hitting a wall and token costs stay high, the business is much harder to run. If you think the labs can keep making inference cheaper, the story gets far more interesting, because the labs are not only selling intelligence. They are trying to make intelligence abundant enough that the real business moves somewhere else.

Figure 2. Why the $200 plan is not what it looks like. The price a heavy user "consumes" is retail API price, which carries markup and 70 to 80% gross margin. The number that matters to the lab is the internal cost to serve, which routing, caching, batching, distillation, and chip utilization keep pushing down. The plan can be a subsidy, or a strategy: serve intelligence near cost now, on a bet that the cost of serving keeps falling. SemiAnalysis pegged the notional value at about $14,000 for a heavy OpenAI user and $8,000 for a heavy Claude user, but notional retail value is the wrong number.

Cheap tokens push value to the harness

This is the key turn. If tokens get cheap, raw intelligence becomes way less defensible. That does not mean intelligence stops mattering. To be clear, he says, electricity matters, bandwidth matters, compute matters. But once an input becomes widely available, the value often moves to what people build around the input. So if intelligence gets cheaper, the question becomes who owns the layer that makes it useful, and that layer is the harness.

That is where the labs' bet becomes much clearer. They do not want to be just API companies forever. They do not want to sell raw intelligence forever, because raw intelligence gets compared and routed and priced down and substituted. They want to sell the work surface. They want to sell the operating layer. They want to sell the thing that makes the intelligence useful before the customer has to understand how any of it works.

Codex is his cleanest example. It is not impressive only because the underlying model is smart. It is impressive because the model sits inside a harness that understands the job. It can see the repo, edit files, run tests, inspect errors, keep track of changes, use the computer, and move through the loop of software and knowledge work. The product is not a model that knows code. It is a system that can participate in general purpose knowledge work. A model gives you intelligence; a harness gives you work. And the IPO question is whether OpenAI and Anthropic can build those harnesses faster than companies can build their own.

Labs have models, companies have context

Companies have one enormous advantage the labs do not: private context. OpenAI does not know how your company works. Anthropic does not know where the real documents live. They do not know which Salesforce fields matter to you. They do not know which approval step is real and which one everyone ignores, who can approve the exceptions, which spreadsheet is a fake source of truth and which one is real, or the internal history that explains why the workflow is broken.

So the asymmetry is sharp. The labs have models, infrastructure, product talent, usage data, and speed. Companies have context. The whole fight is over which side can turn its advantage into the better harness.

	The labs have	Companies have
Asset	frontier models	private context
Plus	infrastructure, product talent	which approval step is real
Plus	usage data at scale	which spreadsheet is the true source
Plus	raw speed	the history behind a broken workflow
Weakness	cannot see inside your company blind	slow, rarely write down "done" slow
The prize	whoever turns its advantage into the better harness owns the work layer

Figure 3. The information asymmetry Jones says decides the whole contest. The labs bring models, infrastructure, product talent, usage data, and speed, but they are blind to how any one company actually runs. Companies own the private context, which Salesforce field matters, which approval is theater, which spreadsheet is the real source of truth, but they are slow and rarely define done. The harness is where one side's advantage gets converted into a durable position.

Forward deployed engineering and the real lock-in

This is why the forward deployed engineering move matters. The simplest reading is that OpenAI is becoming a consulting company, and Jones grants there is something to that, but it is not the deepest point. The deeper point is that forward deployed engineering is how the labs try to overcome the context problem. They cannot know your company from the outside, so they send people inside. They map the workflows, connect the tools, learn which use cases are real, and adapt the product to the customer. They turn the generic harness into a company specific harness.

If that works, the customer is no longer just buying tokens. The customer is reorganizing work around the lab's system, which is much more valuable and much stickier. Once your workflow is rebuilt around OpenAI's or Anthropic's harness, switching gets harder. Even if the model underneath is replaceable, another model cheaper, another better for one task, an open model good enough, your process is now wrapped around one company's way of doing the work. That, he says, is the lock-in. It is not the model.

Renting the harness versus owning it

From a company's perspective, the strategic question is not whether to use OpenAI or Anthropic. Of course you should use them. The question is whether you are renting the harness or owning it. Owning the harness does not mean training a frontier model, and almost no company should do that. It means owning the layer that decides which model gets used for which job: the context, the evals, the permissions, the workflow definition, the review process, and the routing logic.

The consequence is a clean fork. If you own the harness, then OpenAI, Anthropic, Google, DeepSeek, and open source models all have to compete to serve your work, and the labs are suppliers. If the lab owns the harness, the lab becomes your operating layer. That is the fork in the road, and it is the same fork the IPO is really about.

Figure 4. The fork in the road, which is the same fork the IPO is really about. The owner of the harness, the layer that decides which model does which job, holds the leverage. If the company owns it, every lab and open model competes to serve its work and the labs are interchangeable suppliers. If the lab owns it, the workflow is rebuilt around one vendor, switching gets hard, and the lab becomes the operating layer that captures the value. The model underneath stays swappable either way.

Recursive self-improvement as iteration speed

This is also where recursive self-improvement, or RSI, becomes more practical than mystical. The dramatic version is that AI improves AI, intelligence explodes, and everything changes. Maybe, he says. But for the IPO, the practical version is enough. If better models help the labs improve their own products faster, recursive self-improvement becomes an iteration advantage.

Concretely, they can improve code faster, improve evals faster, tune routing faster, optimize inference faster, and compress models faster. They can make the harness better faster. What matters for the business is not whether the model gets smarter in the abstract, but whether the lab can convert smarter models into cheaper tokens and better harnesses faster than customers can respond and build their own.

The bull case and the bear case

The bull case becomes very clean in that world. The labs can manage token costs, compete with open source models on price over time, use their scale to push down inference cost, use their models to improve their own products, and build harnesses so good that most companies decide not to build their own. That is a real thesis, and honestly, he says, they have a shot. Most companies are slow. Most do not understand their own workflows. Most cannot write down what done means. Most will not build routing logic or maintain evals or create a clean internal AI layer. They will just buy the product that works. And if Codex is a sign of where this is going, the labs are getting very good at making products that work.

The bear case is just as clear. If companies learn to own their harnesses and the labs become suppliers of intelligence rather than owners of the work layer, the labs may still be huge companies that make a ton of money, but the valuation changes, because the most valuable layer is no longer fully theirs. The company captures the workflow value, and the lab is left with a token margin. And if token prices keep falling, that is a much less dominant position to be in.

What to watch when the S1s drop

This is what he would look for when the S1 filings are finally released for the two labs. Not just revenue, user growth, cash burn, or the valuation number, which he is sure will be in the trillions. He wants to know whether heavy users are getting cheaper to serve over time, whether gross margin improves as usage grows, whether enterprise customers are buying scalable software or custom deployment labor, whether customers are building real workflows inside the product, and whether forward deployed engineering is a bridge to product or a permanent requirement for the product to work. Those are the numbers that tell you what kind of business this actually is.

Build the harness, not the prompt

For everyone who is not an investor, the practical question is simpler: are you building your own harness, or letting someone else own it? Use the tools, he says. Use OpenAI, use Anthropic, use Codex, use Claude Code, use whatever works. But do not confuse using AI with having an AI strategy. An AI strategy is knowing what work runs where: which tasks need a frontier model and which need very cheap, reliable intelligence. It is owning the context, having evals, having a review path, and being able to swap models without breaking a workflow.

The individual version is the same thing at smaller scale. The valuable skill is not prompting, which he calls thin. The valuable skill is harness building. Can you take a recurring job and define it clearly? Give the model the right context? Connect the right files and tools? Check the output? Make the system better next week? That is where the leverage is, because cheap intelligence is coming either way, and the only question is who knows how to use it.

So the two IPOs are not just stories about whether the companies are worth a trillion dollars. They are the first public test of a cleaner thesis: can the labs make tokens cheap enough and build harnesses fast enough to own the work layer of AI, or will companies use cheaper tokens to build their own harnesses and keep more of the value themselves? Cheap intelligence is the input that makes the token economy possible. The harness is the engine that makes the token economy valuable. Whoever controls the harness holds the dominant position in the token economy of the future. That, he says, is the trillion dollar question he is watching, and the S1s will leak clues, as they inevitably will.

Key takeaways

The trillion dollar valuation is the wrong starting question. The right one is what public investors are being asked to believe: that the labs can make intelligence cheap at scale and build the work layer fast enough that companies rent the whole system.
A token is raw intelligence bought by the meter. A harness is everything that turns it into work: files seen, tools called, permissions, memory, evals, routing, and the definition of done. Codex, Claude Code, and ChatGPT are harnesses.
API prices are retail, not internal cost. The SemiAnalysis estimate ($14,000 of notional value for a heavy OpenAI user, $8,000 for a heavy Claude user, on a $200 plan) measures retail value, not what the token cost the lab to serve.
With 70 to 80% gross margins and falling cost to serve, the $200 plan can be a subsidy or a strategy: serve near cost now on a bet that the cost of serving keeps falling.
When tokens get cheap, raw intelligence becomes undefensible, and value moves to the layer built around it. That layer is the harness, which is why the labs want to sell the work surface, not just the API.
Companies hold the one thing the labs lack: private context. The labs bring models, infrastructure, talent, usage data, and speed. The fight is over which side turns its advantage into the better harness.
Forward deployed engineering is the labs' answer to the context problem. The lock-in it creates is the harness, not the model; once your workflow is rebuilt around a lab's system, switching is hard even though the model stays swappable.
The strategic fork: own the harness and the labs are interchangeable suppliers; let the lab own it and the lab becomes your operating layer and captures the value.
Recursive self-improvement, for IPO purposes, is an iteration advantage: convert smarter models into cheaper tokens and better harnesses faster than customers can respond.
When the S1s drop, watch cost to serve per heavy user, gross margin as usage scales, software versus deployment labor, real in product workflows, and whether forward deployed engineering is a bridge or a permanent crutch.
For individuals, the valuable skill is not prompting, which is thin. It is harness building: define the job, supply the context, wire the tools, check the output, improve it next week.

Chapters

Timestamps are clickable. Click one and the player jumps there and keeps playing while you read.

0:00 The trillion dollar question is the wrong start
0:43 The real bet, and what a harness is
1:28 Why API price is not internal cost
2:36 The $200 plan as subsidy and strategy
3:25 Cheap tokens push value to the harness
4:49 Labs have models, companies have context
5:35 Forward deployed engineering and the real lock-in
6:32 Renting the harness versus owning it
7:15 Recursive self-improvement as iteration speed
8:01 The bull case and the bear case
9:13 What to watch when the S1s drop
9:54 Build the harness, not the prompt

Notable quotes

I know we're all asking the trillion dollar question, but I think the better question is what are public investors actually being asked to believe? Nate B Jones, 0:09

A token is raw intelligence. It is the thing you buy by the meter. A harness is everything that turns that raw intelligence into work. Nate B Jones, 0:39

API prices are not an internal cost. API prices are retail. It includes markup. It includes margin. Nate B Jones, 1:42

A model gives you intelligence and a harness gives you work. Nate B Jones, 4:36

The labs have models and they have infrastructure and product talent and usage data and they have speed. Companies have context. Nate B Jones, 5:18

If you own the harness, the labs are suppliers. If the lab owns the harness, the lab becomes the operating layer. That is the fork in the road. Nate B Jones, 7:00

Do not confuse using AI with having an AI strategy. Nate B Jones, 10:24

The valuable skill is not prompting. Prompting is thin now. The valuable skill is harness building. Nate B Jones, 10:57

Cheap intelligence is the input that makes the token economy possible. The harness is the engine that makes the token economy valuable. Nate B Jones, 11:24

Resources mentioned

Nate B Jones, the channel (AI News & Strategy Daily), for sober, non binary reads on AI markets and adoption.
OpenAI and Anthropic, the two labs walking toward IPOs and the subject of the whole analysis.
Codex, OpenAI's coding agent, his cleanest example of a harness that turns a smart model into general purpose knowledge work.
Claude Code, Anthropic's agentic coding harness, named alongside Codex.
ChatGPT, described as a chat product that is becoming a harness.
SemiAnalysis, the source of the notional API value estimate for the $200 plans ($14,000 for a heavy OpenAI user, $8,000 for a heavy Claude user).
Salesforce, invoked as the example of internal context (which fields matter) that the labs cannot see from outside.
Competing model suppliers in the fork: Google and DeepSeek, alongside open source models.
SEC Form S-1, the IPO registration filing whose disclosures he is waiting to read.
Concepts referenced: recursive self-improvement and forward deployed engineering.

Where it stands

This is a strategy argument, not a stock call, and Jones says as much: he is naming the questions he will ask the filings, not the answer. A few honest footnotes. The standout idea, that value migrates from a commoditizing input to the layer built around it, is a well worn pattern in technology and applies cleanly here; the harness framing is the genuine contribution and it holds up. The softer points are quantitative. The SemiAnalysis valuation of the $200 plans is a third party estimate of notional retail value, exactly the number he then argues is the wrong one, so it functions as a foil rather than evidence. The 70 to 80% gross margin figure is a plausible industry range, not a disclosed number, since both labs are private and their true cost to serve is unknown until the S1s land. And the bull and bear cases are framed as roughly symmetric bets, but they hinge on an empirical question, how fast most companies actually learn to own a harness, that no one can settle yet. The strongest claim is that the lock-in is the harness rather than the model; the claim most exposed to being wrong is any specific guess about which side of the fork wins, which is precisely the open question the video says it is watching.

Full transcript

OpenAI and Anthropic are both moving toward IPOs and most of the conversation is going to collapse into one question. Are these companies worth the numbers people are putting on them? And I think that in some ways is the least useful place to start. I know we're all asking the trillion dollar question, but I think the better question is what are public investors actually being asked to believe? And I think the answer is pretty simple. They're being asked to believe that OpenAI and Anthropic can do two things at the same time. One, they can make intelligence cheap enough to serve at massive scale. And two, they can build the layer around that intelligence fast enough that companies rent the whole system instead of building it themselves. That is the bet. Cheap tokens and proprietary harnesses equal a trillion dollars. If that sounds abstract, I'm going to make it very concrete. A token is raw intelligence, right? It is the thing you buy by the meter. A harness is everything that turns that raw intelligence into work: the files the models can see, the tools it can use, the permissions it has, the memory it keeps, the evals that check the output, the routing between a cheap model and an expensive model, and the workflow that tells the system what done means. Codex is a harness. Claude Code is a harness. ChatGPT is becoming a harness, and inside companies, every serious AI project is a harness project. And that is why this IPO story matters. The question is not just whether OpenAI has better models. The question is whether OpenAI and Anthropic can own the work layer that sits above the models. There's an analysis circulating today that tried to estimate the notional API value of the $200 AI plans from Anthropic and OpenAI. I think it was by SemiAnalysis. The rough claim was that a heavy OpenAI user would be getting $14,000 in value for a 200 buck plan and a heavy Claude user would get $8,000 in value for a 200 buck plan. And the obvious reaction is these companies are lighting money on fire. And maybe for some users they are. But I think the sharper read is that API prices are not an internal cost. API prices are retail. It includes markup. It includes margin. It reflects the price charged to developers, not necessarily the cost the lab pays to serve the token internally. So the question is not how much API value did the user get. The question is what did that usage actually cost OpenAI or Anthropic to serve. And those are very different questions. If the API price includes 70 or 80% gross margins and the internal cost is far below the public sticker price, and if the labs are improving inference efficiency and model routing and caching and batching and distillation and chip utilization and everything else that lets them squeeze more intelligence out of the same hardware, then the 200 buck plan may not be as irrational as it looks from the outside. It might be a subsidy. It could also be a strategy. They may be letting power users consume huge amounts of intelligence while they race the cost curve down underneath that usage. They are effectively saying we can afford to serve intelligence closer to cost now because we believe the cost of serving is going to keep falling. And this changes the IPO frame because if you think the models are hitting a wall and token costs are going to stay high, the business is much harder to run. But if you think the labs can keep making inference cheaper, then the story becomes much more interesting. OpenAI and Anthropic are not only selling intelligence, they're trying to make intelligence abundant enough that the real business moves somewhere else. And that is the key turn. If tokens get cheap, raw intelligence becomes way less defensible. That doesn't mean intelligence stops mattering. To be clear, electricity matters, bandwidth matters, compute matters. But once an input becomes widely available, the value often moves to what people build around the input. So if intelligence gets cheaper, the question becomes who owns the layer that makes it useful and that layer is the harness. This is where the OpenAI and Anthropic bet becomes much, much clearer. They do not want to be just API companies forever. They do not want to sell raw intelligence forever because raw intelligence is going to get compared and routed and priced down and substituted. They want to sell the work surface. They want to sell the operating layer. They want to sell the thing that makes the intelligence useful before the customer has to understand how any of it works. Codex is the cleanest example. Codex is not impressive only because the underlying model is smart. It is actually impressive because the model is sitting inside a harness that understands the job. It can see the repo and edit files and run tests and inspect errors and keep track of changes and use the computer and move through the loop of software and knowledge work. The product is not just a model that knows code. The product is a system that can participate in general purpose knowledge work. That is a huge difference. A model gives you intelligence and a harness gives you work. And the IPO question is whether OpenAI and Anthropic can build those harnesses faster than companies can build their own. Because companies have one enormous advantage the labs do not have: private context. OpenAI does not know how your company works. Anthropic does not know where the real documents live. They do not know which Salesforce fields matter to you. They don't know which approval step is real and which one everyone ignores. They don't know who can approve the exceptions. They don't know which spreadsheet is a fake source of truth and which one is the real source of truth. They don't know the internal history that explains why the workflow is broken. The labs have models and they have infrastructure and product talent and usage data and they have speed. Companies have context. That is a powerful information asymmetry and the whole fight is over which side can turn its advantage into the better harness. This is why the forward deployed engineering move matters. The simplest version is: oh, OpenAI is becoming a consulting company. I do think there's something to that but it's not the deepest point. The deeper point is that forward deployed engineering is how the labs try to overcome the context problem. They cannot know your company from the outside. So they send people inside. They map the workflows. They connect the tools. They learn which use cases are real. They adapt the product to the customer. They turn the generic harness into a company specific harness. And if that works, the customer is no longer just buying tokens. The customer is reorganizing work around the lab system. That's much more valuable. It's also much stickier because once your workflow is rebuilt around OpenAI's or Anthropic's harness, switching gets harder. Even if the model underneath is replaceable, another model might be cheaper, another model might be better for one task, an open model might be good enough, but your process is now wrapped around one company's way of doing the work. That is the lock-in. It's not the model. So from a company's perspective, the strategic question is not should we use OpenAI or Anthropic. Of course you should use them. The question is, are we renting the harness or are we owning the harness? Owning the harness does not mean training a frontier model. To be clear, almost no company should do that. Owning the harness means owning the layer that decides which model gets used for which job. It means owning the context, the evals, the permissions, the workflow definition, the review process and the routing logic. It means OpenAI and Anthropic and Google and DeepSeek and open source models are going to have to compete to serve your work. If you own the harness, the labs are suppliers. If the lab owns the harness, the lab becomes the operating layer. That is the fork in the road. And this is also where recursive self-improvement becomes more practical than mystical. The dramatic version of recursive self-improvement or RSI is that AI improves AI, intelligence explodes and everything changes. Well, maybe. But for the IPO, the more practical version is enough. If better models help OpenAI and Anthropic improve their own products faster, then recursive self-improvement becomes an iteration advantage. They can improve code faster. They can improve evals faster. They can tune routing faster. They can optimize inference faster. They can compress models faster. They can make the harness better faster. And that is what matters for the business. Not just whether the model gets smarter in the abstract, but whether the lab can convert smarter models into cheaper tokens and better harnesses faster than customers can respond and build their own. So the bull case for OpenAI and Anthropic becomes very clean in that world. OpenAI and Anthropic can manage token costs. They can compete with open source models on price over time. They can use their scale to push down the cost of inference. They can use their models to improve their own products. And they can build harnesses so good that most companies decide not to build their own. That's a real thesis. And honestly, they have a shot. After all, most companies are slow. Most companies don't understand their own workflows. Most companies can't write down what done means. Most companies won't build routing logic. Most companies won't maintain evals. Most companies will not create a clean internal AI layer. They will just buy the product that works. And if Codex is a sign of where this is going, the labs are getting very good at making products that work. But the bear case is also really clear. If companies learn to own their harnesses and the labs become suppliers of intelligence rather than owners of the work layer, they may still be huge companies. They may still make a ton of money, but the valuation changes because the most valuable layer is no longer fully theirs. The company will capture the workflow value in that scenario and the lab is stuck with a token margin. And if token prices keep falling, that is a much less dominant position to be in. And that's what I would look for when the S1s are finally released for Anthropic and OpenAI. Not just revenue, not just user growth, not just cash burn, not just the valuation number. I'm sure it will be in the trillions. I would want to know whether heavy users are getting cheaper to serve over time. I would want to know whether gross margin improves as usage grows. I would want to know whether enterprise customers are buying scalable software or custom deployment labor. I would want to know whether customers are building real workflows inside the product. And I would want to know whether forward deployed engineering is a bridge to product or a permanent requirement for the product to work. Those are the numbers that should tell you what kind of business this actually is. But if you're not an investor, the practical question is even simpler. Are you building your own harness or are you letting someone else own it? By all means, use the tools, use OpenAI, use Anthropic, use Codex, use Claude Code, use whatever works. But do not confuse using AI with having an AI strategy. An AI strategy is knowing what work should run where. It's knowing which tasks need a frontier model and which tasks need very cheap, reliable intelligence. It's owning the context. It's having evals. It's having a review path. It's being able to swap models without breaking a workflow. That's the company version. The individual version is the same thing at a smaller scale. The valuable skill is not prompting. Prompting is thin now. The valuable skill is harness building. Can you take a recurring job and define it clearly? Can you give the model the right context? Can you connect the right files and tools? Can you check the output? Can you make the system better next week? That is where the leverage is because cheap intelligence is coming either way. The question is who knows how to use it. So the OpenAI and Anthropic IPOs are not just stories about whether these companies are worth a trillion dollars. They're the first public test of a cleaner thesis. Can the labs make tokens cheap enough and build harnesses fast enough to own the work layer of AI? Or will companies use cheaper tokens to build their own harnesses and keep more of the value themselves? Cheap intelligence is the input that makes the token economy possible. The harness is the engine that makes the token economy valuable. So whoever controls the harness has the dominant position in the token economy of the future. And that is the trillion dollar question I'm watching. And yes, I do think we'll get clues to that when those S1s leak, as they inevitably will for OpenAI and for Anthropic. Stay tuned. And of course, I'll be digging in as soon as we get more information. Cheers.