Do AI Labs Have "Power"?

I’ve been listening to the Acquired podcast for the last year or so, and they talk a bunch about the “7 Powers Framework,” sourced from a book by Hamilton Helmer1. I haven’t read the book, but I have listened to the podcast! And the idea is there are two kinds of “successful” companies: the first is normal companies that can totally succeed by being a good company that sells a quality product to willing customers and brings in more money than it costs. These normal companies though are not particularly differentiated, and one expects that over time those companies will be in competition (or someone new will enter the market) and drive their profits down to the whatever is the smallest steady state that keeps them operating. Most businesses are this. Most businesses should be this!

Then there are some businesses that can achieve “persistent differential returns.” This means that they make more money than their competitors (differential returns) and that they can keep doing it, it’s not a fluke (persistent.) These businesses are much much rarer. VCs basically only want to invest in these kinds of companies, and Acquired, as a podcast, basically only talks about them. And per Helmer, there are only really 7 ways that you can do that:

In some cases, if you get big enough, you can take advantage of your scale to make your per-unit cost way cheaper than any competitor’s, and so you can either charge less than them (and sweep up all their business) or charge the same as them (and keep more profit, because your per-unit cost is cheaper.)

A competitor has trouble catching you because they can’t compete with your costs until they’re as big as you… which they can’t do because their costs are higher, etc etc.

In some cases, a thing becomes more valuable the more other people use it, similarly making it hard for smaller competitors to catch up. Social networks (and other networks, like telephone networks) are the paradigmatic example of this, where I want to join the network that already has people, and the more people that join the more useful the network is to me and more I want to join.

You can also have two-sided network economies, like marketplaces or ad exchanges or search engines (the more people use Google, the more websites want to make sure their stuff is available on Google, and the more stuff is on Google the more people will use it, etc etc etc.)

A competitor has trouble catching you because your network is more valuable than theirs, and that self-reinforces the growth of your network and stalling of theirs.

There is something that you can do that is different from your competitors, and crucially they would not be able to do your thing without abandoning or betraying something about how their business already works. Disney can counterposition itself against other studios by being family friendly, and other studios can’t fully compete because that would require them abandoning the non-family-friendly markets that they currently make money on, and once you have 1 R-rated movie on your streaming service suddenly you’re not family friendly at all2

Or for one that’s not branding/content related, I watched a talk a while back from Grafana, an observability platform. They talked up how they’ve been investing heavily in using AI tools in order to figure out what log line and events can be skipped so they can lower your Grafana bill. This is counter positioning Grafana against other observability vendors (let’s say Datadog to have a name), because in order for them to compete with Grafana, Datadog would have to introduce tools that would lower customer’s Datadog bills. Datadog is probably not going to do that unless they absolutely have to! Lowering customer bills is bad for their existing business model, which means the insurgent can use that to counter position! Most successful classical “disruption” is often a form of counter positioning.

A competitor has trouble catching you because they can’t do what your doing without betraying either some part of their business model, or some part of what you’re doing differently.

Once customers are using your product, it’s too much cost or effort for them to switch to another. A lot of enterprise products work this way (it’s going to cost you millions of dollars to leave AWS or Salesforce, so you might as well stay) but so too for consumer products (the more Sonos speakers you buy, the more expensive it’ll be to replace them all to switch to another brand.) And it can combine with others, especially network effects (signing up for a new social network is free, but if all your friends are on the old one it’s a huge effort to either make new friends on the new, or convince your existing friends to switch as well.)

A competitor has trouble catching you because they can’t successfully get your customers pay the large cost it takes to switch away from you to them.

In a case where a customer is choosing between two identical products, they would be willing to pay more for your product because there’s something else about your brand that attracts them. This can tie to counter positioning, such as Apple counter positioning itself as privacy conscious compared to Google, which Google can never directly compete with because it would jeopardize their ad product; that privacy focus is now a well-known part of Apple’s brand3.

But branding it doesn’t necessarily have to follow from counter positioning. And counter positioning power is not the same as branding power, in that counter positioning you are actually providing a product or pricing benefit in and of itself. Branding is going a step further, where there is no literal benefit to the product, just the power of the brand.

Luxury goods are probably the ultimate example of this, where you pay more for Grey Goose vodka than Kirkland because ~*~*~branding~*~*~4

A competitor has trouble catching you because customers are irrationally and emotionally choosing you over them, and it’s hard to reason someone out of an emotional choice, and hard to build an emotionally powerful brand overnight!

You have access to something that other companies just straight up don’t have. Maybe you have the most productive oil well or diamond mine, or a huge stockpile of a literal resource, or the most prized locations with the best foot traffic, or control over particularly lucrative IP, or the company perks and culture that are the most attractive to a particular type of high-value worker (Google historically was uniquely able to attract academics and researchers because of its research-friendly culture.)

A competitor has trouble catching you because they do not have access to the valuable resource you have.

There is something about the way you do things that is strictly better than anyone else, and difficult for them to copy. Toyota in the 80s would be an example on the manufacturing side, or Pixar in the original 90s/2000s run on a creative and content side. They just did things differently, and that different way of doing things was strictly better (either in terms of efficiency like Toyota or quality like Pixar.)

A competitor has trouble catching you because they either don’t know your secret sauce for how you do things, or it’s really hard and they can’t or won’t replicate it.

So book report done, here’s the actual meat of this post: infinity bajillion dollars is being poured into the major AI foundation labs, and that infinity bajillion is on the expectation of one or more of them being able to return infinity bajillion times some big number back in the form of profits. That means you must be expecting that one or more of these labs will be able to create……. persistent differential returns!!!!

So the followup question is:

Aka: do any of these apply (or are likely to apply in the future) to big frontier AI foundation model companies? 👀

In this context, when I say “big frontier AI foundation model companies” or labs or whatever, I’m mostly talking about OpenAI, Anthropic, xAI, Facebook, Microsoft, Google; to a lesser extent Amazon or Mistral or Safe Superintelligence or Thinking Machines etc etc. Companies whose business model is “turn venture capital investment (or other profitable business lines) into training of big general purpose AI models that you can monetize by selling access.”

Perhaps! There are two major costs to running big AI models: the cost to train it initially, and the cost for “inference” every time it is invoked. In both cases in theory the more customers you have, the more you can amortize your costs and have it cost less on a per-user basis.

For inference though, it’s pretty weak. Part of that is because there is a significant marginal cost in power or cooling or depreciation of chips to every individual query, so you might get some economies of scale, but not enough to cover everything. The big server-based frontier AI labs also have to contend with open-source and on-device models, which can charge zero for inference if it’s running on the user’s device. You might be able to gain some inference economies of scale compared to a competitor server-based frontier lab, but that will never be cost-competitive with something running on the user’s iPhone.

Training is different though: training is a massive fixed cost, where you have to spend billions of dollars for a single training run, and then you can use that same model over and over again for inference. This means training is perfect for amortizing the big upfront cost across a customer base over time. So in this case, again all else being equal, the lab with the most customers should have scale economy power regarding its training of models.

Not really. You using ChatGPT does not really influence my use of it in any meaningful way, and in my experience ChatGPT et al is not any better or more powerful for me for the fact that you also use it. For B2B stuff perhaps you get a little bit of this in the sense that it is better for an entire company to be using the same tool rather than everyone BYO AI model, but that’s pretty weak and basically is an argument that anything that is sold to businesses now has “network effects.”

You may be tempted by looking at the biggest of the three areas where AI has so far found product-market fit: coding tools5. Claude Code totally took over Copilot and Cursor and Codex in the last year, wasn’t that an example of network effects? But here is the difference between network effects vs I guess what you could call bandwagons: a thing can become popular, and a thing can take advantage of free marketing from its users to get even more popular, but for it to be a network effect the addition of nodes in the network has to make the product itself better. Each user that joins a social network makes it a better and more useful social network for everyone else6.

Me using Claude Code does not make Claude Code itself a better product for you in any meaningful way. The closest is the more people use one tool the more 3rd party tooling there will be built up around it, but 1) that’s pretty weak, and 2) the quirks of general purpose language models is that by and large any third party tooling built for eg. Claude will work pretty well with OpenAI models as well and vice versa. Gemini will happily read and incorporate your Claude.md file!

So I’d say that, unless there’s an entirely new paradigm of true “multi-player” AI that I can’t even imagine right now, no network effects.

The major AI labs do not counter-position between themselves basically at all. There is counter positioning happening in the market, but it’s between companies (and China) focused on open-source or on-device models, rather than big server-side ones. Facebook used to be very counter positioned when it was focused on bringing frontier-lab level quality to open-source models! But it has abandoned that to be an also-ran in the server-side frontier model/enterprise sales game.

Arguably Apple is the only major tech company still counter positioned, in that it is (in a face-saving way, not an intentional way) shifted from trying to be a model developer itself to a partner for the model companies and a platform for excellent smaller on-device models. Spoiler for the rest of this post, but if it turns out that there is limited “power” to being a frontier AI lab, this might not be a bad call!!!!!

The closest I can see since Facebook abandoned their open-source counter positioning is 1) Google starting to get big into the chip building business, counter positioning it against the pure model and software companies, and 2) Anthropic declining to work with the DoD. For Google’s case, it’s entering an extraordinarily expensive new market that blows up the (theoretical) software margins you could get from being a software-only lab, and in Anthropic’s it’s forcing other labs to potentially forgo a bunch of revenue if they want to chase Anthropic’s “good guy” (or at least….. ever so slightly better guy, in this one specific respect) image.

But yeah so far pretty weak. Facebook used to have it, they gave it up (and of course unclear if it would’ve worked in the end anyway!)

So far the switching costs seem to be zero; there is nothing stopping you as a user of any AI model from chasing the best deal you can find, especially in contexts where the underlying models are the same quality. Even if the models aren’t, then that’s not really a switching cost, that’s just having a competitive product. “I don’t want to switch from iOS to Android because I like iOS more” is not switching costs, “I don’t want to switch from iOS to Android because even though I like Android more, I’ll have to re-buy all my apps, or get a new number, or leave and rejoin all my group chats” etc are.

Google is maybe trying to make a bit of a play for this on the consumer side with recent announcements of its Gemini “always on” assistant that is deeply integrated into your other Google products and Google account context (reading your emails and calendar and photos, etc.) In this case it is not that Gemini itself has a high switching cost, but the combo of Gemini with everything you’ve been doing in Google already for a decade+ might be high enough to prevent you from switching. So… maybe on the consumer side, but TBD (and zero so far on the business side.)

To some degree you have anti-branding for xAI and Grok, what with the Mecha-Hitler and the nonconsensual imagery, and you got a little bit of it with people mass deleting ChatGPT to switch to Claude post the Pentagon shenanigans (see again: no switching costs!) So there’s a little bit of branding here, but is it enough to give any particular company persistent differential returns? I would argue no! Again maybe other than Grok, people seem willing to broadly tolerate any company’s brand if it’ll get them a better or cheaper result. Grey Goose this is not.

This is the one piece that feels the closest, in two real and two potential ways.

The real way is: there is a limited pool of skilled AI researchers, and if you can poach and keep them then you can have access to a genuinely differentiated cornered resource! The reason this seems real to me is that people are paying an extraordinary amount to make sure they are not cornered out of this resource! Now that said, no one has managed to stably corner this market, with people bouncing pretty freely between all the main labs and the neo-labs like Safe Superintelligence/Thinking Machines etc7. But this feels like an active fight at the moment.

Similarly, there feels like there is an active fight over control of compute resources (chips and RAM, and the data centers to run them.) No one has managed to effectively corner this, but that’s certainly an area where people are fighting.

On the theoretical side then, I’d first put up intellectual property. As of now, all of the main models are trained on deeply immoral theft of training data from the commons8. As much as everyone just assumes “eh what’s done is done we can’t deal with it” that is not a foregone conclusion! The law is unsettled and especially if AI gets more and more broadly unpopular, the legal realism of “Would a judge ever rule against a rich and powerful company?” could run up against the legal realism of “Would a judge ever rule for a hated and unpopular company?”

And so in a world where intellectual property rights become salient to whether or not you are allowed to train a model or serve a model you’ve already trained, and where by all accounts the more and better data you have, the better your model is, cornering the market for training data could provide a genuine differentiation between models and companies.

Finally, the (gross) holy grail that these losers are clearly all gunning for: if you can be the one to crack using AI to train and develop AI in a recursive way that does not involve humans and does result in continued forward progress, potentially that can be your differentiator. This last one can only be a true “cornered resource” if either 1) no one else ever figures it out, 2) this gives you a long-term gap in capabilities (idk, >1 year?) or, 3) it combines with other elements from this list. If discovering the AI recursion loop 3 months before the next lab just means you’re perpetually 3 months ahead, to me that alone is not enough of a quality gap to ensure that no one will ever choose eg: the cheaper 3-months behind version or the 3-months behind version that’s better at some random quirk you haven’t prioritized or whatnot.

So this seems plausible: an active fight to try and corner researchers and compute, and a theoretical world where you might try cornering IP or recursive self-improvement.

This I suppose ties to the cornered resource, in that 2/4 potential resources are ones that directly impact the process of how you develop AI (researchers and potential recursive self-improvement.) But it is also possible that maybe some lab cracks a secret of “this is the better way to do it” in a way that is unreplicable outside of the lab that discovers it.

I think this is the thing that 3-ish years ago the labs assumed would provide a them power and a persistent lead. Some lab would discover something, and that would put them insurmountably ahead. That has certainly not happened yet! Models continue to be commodities!

And going forward I am skeptical of this world, or of this world in the absence of also cornering the above resources. Partly based on the last few years and just assuming “well absent new evidence, assume it’ll keep working the way it’s worked so far, with no major unreplicable discoveries and therefore no process power”. But also because your researchers can always leave or start their own companies, researchers love being able to publish9, and China can always distill your models and just live ~6 months behind whatever you design10. But who knows, maybe a genuinely new paradigm breaks this current semi-stalemate!

So all of those together, I would say that for the big mega labs, they have no network economies, counter positioning, or branding. They currently have no process power, but potentially could if there’s a new unreplicable breakthrough, and there’s no switching cost but Google is trying to leverage the switching cost of its other products to boost the usefulness of its AI ones. There are some scale economies on the training side, and there is active fights over cornered resources that no one has cornered yet.

Therefore as it stands now….. there is no way you would expect anything other than fierce competition and for that competition to drive down profits to the minimal sustainable level, absent something changing. You should expect frontier models to be commodities with commodity pricing. And that’s what you currently see!

And looking at the state of the world and state of these powers, the likeliest version of a world where that could change seems to me like some combination of:

• “some people give up, the more that give up the more resources the survivors can corner, the more resources they can corner the better their products can be, and they can get persistent differentiated returns that way”, with

• “some people give up, the survivors can get economies of scale on training, and that prevents any new entrants from joining and catching up.”

That said, in this world if the profits are primarily the result of the cornered resource, if that resource is anything other than the “we invented the robot to do research for us”, then it is more accurate that the winning AI lab(s) would be renting or buying access to the resource: they are the ones buying the chips, they are not the source of them; they are hiring the researchers, they are not the researchers themselves; they are paying for the intellectual property, they are not the IP creators. So even then I would expect a significant amount of profit to be going not to owners and investors of labs, but to the the actual sources of the resource(s) being cornered (the chipmakers, the researchers, the IP holders.)

And if the profits are from scale economies, that really only works if you can get down to one main lab, or at least a main lab with a large lead over the others. If there are at least two basically equal labs (or one plus Google making infinite money from search willing to subsidize their lab) then you would expect that to negate most scale economy benefits.

And of course, the other elephant in the room is: humans exist! If your product is fundamentally “pay us to replace people” then there is also an upper limit on the amount you can charge before it switches to “it’s cheaper or better to just pay a human.”

…..So it really does kinda seem like, from an investor perspective, everything is resting on either inventing recursive self-improvement, some other major process power breakthrough, or some sort of exogenous collapse that leaves only one lab standing.

Which I mean…. Sure! But I guess I see why people are pretty skeptical of “being a frontier lab” as the next great world-changing business model!!!!