As we spend more time pontificating about $1T+ in annual lab revenues, I can’t help but wonder if we’re naively conflating value creation at the labs with value capture. I say this because many of the arguments made tend to imply takeoffs that are primarily relevant to STEM and/or verifiable domains, an area where selling tokens is a fundamentally bad way to capture value.
This leads me to believe that frontier labs should focus far greater effort on making discoveries themselves as opposed to being the instrument that helps everyone else make them.
A couple things for us to get in order before we begin:
I believe this argument has far weaker potency within non-verifiable domains (i.e. traditional white collar work). This contention is constructed to apply solely to verifiable domains and STEM, which happens to be where reinforcement learning with verifiable rewards (RLVR) has models hill climbing the fastest. Do I ascribe high probability that we’ll see OpenAI or Anthropic wholly pursue this path? No. For many reasons outlined in the following but also because white collar automation is almost certainly the bigger pie. However, for lagging frontier labs or neolabs, this feels like the direction the wind is already beginning to blow.
I’m focusing on share of surplus, not absolute revenue. Between white collar automation and some routine research workloads, Jevons dynamics could still make token demand enormous relatively speaking, therefore labs could still earn large “finite” revenue while capturing a vanishing fraction of the value created. Both are not mutually exclusive.
Where does the value in STEM go?
Okay now that we’ve prefaced this thought, just humor me for a minute and imagine this hypothetical, but very real, scenario. A discovery team at Merck is running a campaign against a target that’s eluded them for the better part of a decade. Somewhere in the process sits a frontier model, let’s say GPT-5.6 Sol for the sake of the argument. It proposes a binding candidate that no chemist on the team had previously considered. Merck’s wet lab confirms its validity. Ten years and three trial phases later, it’s a drug doing $25B per year at its peak.
Now let’s just think about who got paid here. Merck holds the patent, the trial data, the relationship with the FDA, the manufacturing, and the sales team. Therefore also their $25B. OpenAI, who built GPT-5.6 Sol holds what is essentially an invoice for let’s call it $50k in tokens, which is priced the same as anyone else’s tokens via the API. Meaning it’s indistinguishable from some arbitrary startup that’s using the same model to generate marketing boilerplate.
And right now, the working assumption across this entire discourse (which is implicit in how the labs describe themselves and implicit in how the market values them) is that the companies that build the country of geniuses in a data center will own what the geniuses produce. Or better put, the lab whose model cures cancer will own the value of the cure.
Yet in this hypothetical scene I just laid out, the geniuses actually got paid like a utility and the delta between the value that’s on the table and the number on the invoice is probably in the realm of six OOMs. The obvious response to this is that it’s a pricing problem, and pricing problems eventually get fixed. So, for the sake of this argument, let’s walk through a few options.
Let’s say you’re Dali Rajic, the CRO of OpenAI and this hypothetical Merck scene is keeping you up at night. You have a few tools at your disposal so you think through each one:
You can try royalties. You draft a term sheet so OpenAI lays claim to a % of revenue on any compound your model(s) materially contributed to. Problem is that you hit two walls immediately. First being an attribution problem. Was the model the but-for cause of the discovery? One could argue that Merck’s proprietary assay data shaped every prompt and their chemists ran a thousand of the model’s other ideas into walls before this one. Not to mention, their wet lab did the only verification that counts. And trying to litigate “share of invention” against a big pharma legal team will not work when they’ve spent the last several decades perfecting this exact fight. The second wall is much simpler, which is that Merck’s procurement team is not sentimental. The moment your terms include a royalty, the discovery workloads end up migrating either to a competitor that didn’t ask for one or to some open weights model that’s 90% as good and asks for nothing at all.
You could try outcome-based pricing. Fine, you skip the royalties and charge on results. The problem is that results may take a decade to resolve across multiple trial phases, model generations, and potentially even a deprecation of the model that did the actual work. So no contract is realistically surviving that, and you’ll be writing biotech timescale agreements on software timescale products.
Maybe you try value-based pricing tiers instead. You charge discovery customers more per token. But to the API, the $25B prompt is again indistinguishable from a graduate student’s homework. So you can’t identify the golden query ex ante (which is the whole point of it being a discovery) and any tier you build gets arb’d through some subsidiary account later that week.
Finally, you could try exclusivity. License the model to just Merck at an eye watering price. But now you’ve effectively destroyed demand and you’ve handed every other pharma company a reason to fund your competitor.
The point in laying this out is that within all four proposed solutions, they all fail for one of two reasons. Either someone else can substitute for you or nobody can precisely say what you contributed. Which boils down to substitution or attribution. Or in other words, this isn’t about pricing.
From Merck’s point of view, when evaluating models to utilize, yes there’s likely a clear “best” option in terms of capabilities. And maybe that affords you finding a drug candidate three months sooner which also means three months of peak drug revenue. However, the problem for the labs, and in particular the one with the SOTA model is that companies like Merck won’t pay them for the value they create. They’ll pay them the smallest amount that makes choosing them rational i.e. the price of the next best option plus a sliver for the difference.
And it’s not like this is a new concept as this is how everyone’s wages get set. Even the plumber isn’t paid for what a fixed pipe is worth to your home, they’re paid based on what the next plumber charges. Meaning rarely does anyone get paid what they create, they get paid what it costs to replace them.
The reality is that instrument providers rarely ever get paid for the value that they create. Just take a look beyond the AI world at companies like Zeiss Optics, Thermo Fisher, or Illumina. All of which have been the creators of incredibly important tools that provide massive value to the biotechnology and pharmaceutical space, yet they capture little value relative to what their instruments have helped develop.
There are exceptions though. Just take a look at ASML. Their EUV lithography machines are used in the development of every leading-edge chip, of which ASML captures spectacular economics on the basis that there are zero viable alternatives. Point being, ASML is the sole source and has decades of accumulated process knowledge which would take a herculean effort to emulate. The exception tells us the actual law here though, which is that instrument makers do get paid but only when what they make is irreplaceable.
And if we’re talking about models and their tokens, they are becoming far from irreplaceable. I mean the gap between the first and second best model alone has dwindled over the past couple of years with the state-of-the-art typically alternating every month or so (sometimes even less). Right now, frontier labs can only monetize the quality gap between them and the nearest competing (or open weight) model. Especially in areas like math or coding where we thought there might be some kind of moat. I recall many thinking that Anthropic’s moat was in coding and they were the definitive winners there, but lo and behold, GPT-5.6 Sol in Codex is more than a formidable alternative to using Fable 5 in Claude Code.
What strategy the labs should pursue
In a world where the most probable outcome is that we’ll see the commoditization of models, then the labs need to either verticalize within these verifiable/STEM domains, or they have to find a way to sell their discoveries. That is if frontier labs want to focus on capturing a larger share of value that they create in these fields as opposed to just absolute revenue generation. That being said, while I’d argue that both of these options are better than purely focusing on just the token business, they do have their pitfalls.
Let’s go back to this hypothetical OpenAI and Merck scenario where OpenAI now decides to swing for the fences, verticalize, and try to become a pharmaceutical company like Merck. Beyond the issues of building up all the physical assets that Merck holds, a glaring issue before we can even think about the difficulty of such, is simply where is the capital even going to come from to fund this endeavor? There is already sufficient debate ongoing about whether there’s enough capital to sustain the ecosystem’s ever-growing compute commitments. And while the API business is currently the golden goose for many of these labs, all of the returns from that business are going directly to paying down their compute commitments. A different debate for a different time, but my point is that if we’re concerned about capital for compute, then we certainly don’t have enough capital to sustain that plus now the incredible amount of capital it will require to build a big pharma vertical business.
Oh, and did I forget to mention that not only would you (OpenAI) have to do this for pharmaceuticals, but you’d also have to do this for every single domain in which you want to capture superior economics in. A little bit later I’ll explain how I think the labs will actually go about capturing value given what is a near endless amount of domains to choose from, but it becomes clear quite quickly that this strategy of verticalizing is untenable, potentially even for just one domain, let alone all other relevant ones.
Furthermore, I just do not see a world in which instead of using Rolls-Royce turbines on a Boeing 787 Dreamliner, we have OpenAI turbines on the OpenAI commercial jet. It’s an absurd example, but one could argue that this isn’t even the most preposterous statement that you could come up with.
As a sidebar, this leads me to believe that there is a world in which SpaceXAI is the best positioned frontier lab. Given what seems like an increasingly likely merger between SpaceX and Tesla, Elon would then have control over a frontier lab that has verticals as an automobile business, robotaxi business, energy business, humanoid robotics business, satellite business, rocket business, neocloud business, and semiconductor business. Which in a world where it’s valuable to be a holder of physical and/or digital assets, it’s hard to argue that Elon isn’t best positioned.
But back to our main point. The second option the labs could pursue would be to sell discoveries. This works, however, your ability to sell them is highly dependent on the IP-regime of the domain you’re working within. In pharmaceuticals or biotech, this strategy works great because the field is highly dependent on the IP-rights to specific drug compositions, hence why big pharma either licenses patents or acquires smaller biotech companies once they have cleared FDA trials. However, in areas like mathematics for example, your ability to convert IP into value is far more ambiguous as solving a specific mathematics problem does not grant you the right to the value that is created from said solution. Merely trying to patent and sell the solution to Navier-Stokes is just not possible.
Selling discoveries isn’t a bad idea, but it does force us to create a value ladder of what domains have the strongest IP-regimes versus what ones do not. This is what would ultimately determine the shape of how the labs could recognize value from them. I’d argue there are three likely buckets with three different strategies for capturing value. The first is where the IP-regime is strongest, in which you directly patent the discoveries your models make and license them out to companies already operating in the domain. In moderate IP-regimes, labs could spin up a NewCo around each discovery and hold a notable equity stake. And in weak IP-regimes at scale, labs could make venture investments in relevant startups or strategic stakes in advantaged incumbents.
Coincidentally enough, there is already a business that exists and runs a very similar structure. That business being Alphabet. They already have something along the lines of an IP engine with Google DeepMind being an example, NewCos such as Isomorphic Labs or Waymo in which they hold notable equity stakes, and then they have Google Ventures which holds smaller equity stakes across beneficiaries in the fields the broader platform will benefit from.
In either scenario where the labs decide to verticalize or sell discoveries, I’d argue there’s far greater upside and value capture opportunity amongst both options compared to merely selling tokens.
But what about RSI?
A pushback that I anticipate to receive is that my argument only holds so long as we stay on the current trajectory. However many would argue that timelines are continuously being brought forward and the current rate of change we’re on is increasing week by week. Therefore, this argument is moot because it’s working on an assumption that we know is likely to be false, if not proven false very soon. I believe many would point towards developments such as recursive self improvement (RSI) as a scenario in which the token selling business could continue to prosper. Meaning no need to either verticalize or sell discoveries.
Let me just lay this out quickly.
The common rebuttal to my argument is likely that achieving RSI would be a way to avoid this drastic need for shifting the business model in STEM markets. To put it simply, in a world where RSI is achieved in the near-to-medium term, this presumably would be a clear path for labs to avoid a situation in which models get commoditized. Now this would only benefit either a single lab that reaches this milestone first or a very small cohort of labs that achieve RSI at a similar time. Reason for this is because years of R&D could be condensed into months, which quickly creates a compound effect, thus the lab or labs that reach RSI will see an increasing lead over their competitors that will likely expand at a growing rate as well.
This dynamic will ultimately allow whoever reaches RSI to not only develop an increasingly better model than their competitors, but to also maintain token pricing power over the market and sell tokens at a premium. Just like they are in the status quo. One could make the argument that this would extend across any market the labs address, whether that be the white collar automation segment of the market or the STEM segment of the market.
And while I do think this is a reasonable argument to make (given I myself am RSI-pilled) and that RSI will result in more capable models meaning more discoveries within our reach, there is one core reason why RSI will likely only put negative pressure on the current STEM business of just selling tokens.
The reason for making the RSI argument as a savior for selling tokens is that it potentially gives a couple of labs a way to avoid token commoditization. However, I would point out that the biggest issue for the labs in STEM markets isn’t that tokens are going to get commoditized, it’s that models are using far less tokens than we anticipated to solve what are deemed incredibly difficult problems. Just look at two of the more notable recent data points. OpenAI’s internal Astra model making ten notable advancements in mathematics and theoretical computer science which humans had not made in quite some time. All for just $2,000 of total token spend at Sol API pricing. Further, an unreleased Anthropic model while attempting to solve the Riemann hypothesis made a tangential discovery in just a few days and with a few thousand dollars of compute.
The point is that these models are in one sitting effectively one or two-shotting these kinds of difficult problems and are spending far less compute than at least I had anticipated. So at the end of the day then, it doesn’t matter if you (a frontier lab) charge $5/1M tokens or $10/1M tokens. The likely outcome is that as models get smarter, the number of tokens it will require them to make these kinds of discoveries will become less and less. Not to mention that these models themselves will become more token efficient as well. All of this will lead to the shrinking of the STEM addressable market over time, and again, will force a reconsideration of the business model.
For a second let’s just look at the other outcome where we don’t reach RSI (which I find increasingly unlikely). Many would say “well what about Jevons?” Maybe we do see the degradation of token economics that leads to commoditization, but cheap discovery means vastly more discovery attempts, so falling tokens-per-discovery gets offset by exploding volumes. Playing the Jevons game works in the white collar automation segment of the market because even in a world where tokens are commoditized, labs are willing to play the game because the TAM is so large that there is plenty of headspace to continue growing. However it doesn’t in STEM for the simple reason that discovery volume is throttled downstream by verification.
Let’s just go back to our pharmaceutical analogy. The world cannot run 1,000x more drug campaigns just because tokens get cheaper. Maybe in the future, but the current system whether that be wet lab throughput, trial slots, or regulatory bandwidth cannot absorb this kind of increase. Chat and code can because they don’t have such a throttle given consumption is far more elastic. But discovery consumption is inelastic because the vast majority of the relevant fields have physical bottlenecks and thinking is not one of them. So instead of those efficiency gains translating into exploding volumes, it leads directly to shrinking token revenue.
Regardless of RSI or not, it’s likely that the STEM token business has a limited ceiling. Thus the optimal strategy remains for leading frontier labs to pursue what I’d consider more aggressive value capture within each relevant vertical that said lab has an interest in maintaining superior economics in. RSI at this point merely decides how fast the labs have to admit this.
Why I believe this is relevant to discuss
Do not get me wrong, I remain bullish on labs such as OpenAI and Anthropic and am becoming more positive on labs such as SpaceXAI which I thought would remain in a laggard position. Because again, for the labs, this argument is only relevant to one segment of the overall TAM for them.
While I am a subscriber to the belief that we’ll see tremendous white collar automation in the future due to the proliferation of AI, I do not believe that we are likely to automate most of white collar jobs in the next five years. Or put differently, I think there is far greater probability that we automate 100% of cognitive work in STEM R&D than we automate 100% of white collar work in that five year timeframe. Thus, while the TAM for white collar (and blue collar) automation is far greater than that of STEM R&D automation, it is the latter that will experience the greatest impact in the near-to-medium term. Meaning we should at least focus some attention there prior to events occurring which will force us to contemplate it anyways.
Additionally, it’s not that the labs are unaware of this predicament that they find themselves in. I do believe that they are conscious of this, but it is hard for them to move on this for the same reasons I outlined earlier. If we just listen to interviews with the lab CEOs, most of them will either admit or imply that selling tokens is not the terminal business model. However, knowing the reality and moving on that reality are two different things, particularly when the API business is printing 80% gross margins that fund everything else.
Not to mention that investors have made their preferences on strategy perfectly clear, which is to focus on selling tokens, consumer chat, and code harnesses. When investors got nervous about OpenAI’s side quests, the lab folded in teams including their OpenAI for Science division, which had barely been created prior to that point. Further, while I’d love to say that Anthropic leaning into biology and healthcare is the counterexample, I wouldn’t hold my breath. Firstly, the timing feels like they’re trying to manage public sentiment in a year where AI needs some goodwill. But even their recent protein design work, while impressive, resulted in them packaging everything into a free workbench product for scientists. Which means they’re still thinking about this through an instrument-selling lens, not a discovery-selling lens.
So what does this all mean? It means that regardless of who you are, lab or not, it is beneficial to hold complementary assets whether those be digital or physical ones. For the labs, it’s why they value their respective chat applications and their code harnesses so much. Because that is the product, not the model. And it does not take long to find digital and physical asset holders that will benefit in the medium-to-long term whether it’s pharmaceutical companies like Merck or hyperscalers like Google.
Which brings me back to where we started. Simply put, the lab whose model cures cancer can still own the value of the cure, but that means they will have to sell the molecule instead of the meter.
If you enjoyed the content and want to support the page please feel free to subscribe, like, and/or share this article with others! And as always, comments and feedback are much appreciated.