I will soon be joining Anthropic, and so this is my last opportunity to write down some thoughts on the AI lab business model before I can be accused of spilling any inside information.

Although large language model (LLM) providers like OpenAI and Anthropic have spearheaded a new AI revolution, it is still not clear at this point that they are good businesses. The question on the minds of investors is whether these pioneering labs will reap profits, or whether instead other parts of the AI ecosystem (hardware vendors, cloud platforms, or end-user companies) will capture the lion’s share of value. This article surveys the competitive landscape and suggests some strategies that the labs could pursue to secure good economics for themselves.

(While questions of AI safety its broader societal impact are important, I will leave these topics aside here and focus purely on the financial angle.)

I. No technical leader

The status quo is that labs are spending considerable resources training successively better models. For example, OpenAI has raised $60bn, Meta is spending as much as $60bn per year, and xAI has raised $12bn. Their reward for deploying these resources is to spend just a few months at the top of the benchmark leaderboards before being supplanted by a competitor. Despite the increasingly secretive development process of LLM technology, no lab has been able to take a decisive technical lead. Why does this happen?

Perhaps the most important factor is that the labs are making many parallel discoveries. Technological breakthroughs often appear simultaneously as soon as prerequisites are in place - consider Newton and Leibnitz, or Tesla and Marconi. Even in the case where ideas aren’t rediscovered independently, many of these labs are based in California, where employment law that is hostile to non-compete agreements allows employees - and hence ideas - to flow freely throughout the community.

Once public, ideas spread like wildfire. For instance, the Transformer architecture itself (from Google’s 2017 paper “Attention is All You Need”) was rapidly adopted by everyone even in domains far outside the original machine translation usecase. More recently, techniques like LoRA or FlashAttention provide big efficiency gains and are implemented widely mere months after invention.

Working in favour of parallel discovery is that LLMs remain fundamentally simple technology, so you might expect that a lot of the improvements that make them better are also relatively simple. Consider that a complete specification for a leading open-source LLM such as Lllama can be written in a few hundred lines of pytorch. Torch itself can be implemented in less than a thousand lines of code. Validating this intuition, Dario Amodei has said that Anthropic has “$100 million secrets that are a few lines of code”.

II. Competitors abound

Barriers to entry in the industry do not seem to be high - witness, for example, the quick ascendancy of xAI. The technical progress made by incumbent labs can be replicated by making the right hires. Once you have this knowledge, actual model training costs are quite moderate. Notoriously, the Deepseek R3’s main training run cost only $5m. Anthrophic’s CEO, Dario Amodei, reports that Claude 3.5 Sonnet cost “a few $10M’s to train”. Putting these numbers in the context of the multi-billion-dollar funding rounds by the labs, it seems that most of the spend must be on smaller-scale experiments and researcher time - which seemingly only serves to create IP with a very short shelflife.

This means that deep-pocketed investors have continued to fund new entrants, fuelling the Red Queen’s race in LLM tech. They do this because of desire for a regional champion (Mistral), concerns about AI safety (Anthropic), to avoid dependence on a limited group of suppliers (Meta), a desire to “democratise” access to AI (Stability), or perhaps just ego (xAI).

A related factor keeping barriers to entry low is that of model extraction, where you train your own model on output from another. As early as 2020, a paper found that you could get 90%+ agreement with a victim BERT model for only a few hundred dollars worth of tokens generated by the victim. In 2023, Stanford replicated GPT 3.5’s instruction tuning on a Llama base model using just $500 worth of tokens. By some estimates, a leading LLM can be replicated by training your own model on just 1T output tokens from the victim model: the cost of these tokens for GPT-4o would be just ~$5m. Although there is only weak evidence for it, some suspect that DeepSeek’s frontier models partially owe their excellent benchmark performance to being trained on GPT-4o output.

The implication is stark. An LLM lab might invest tens of millions in training a cutting-edge model, only to have others reap the benefits by training cheap knock-offs using its outputs. This dynamic is reminiscent of generic drugs copying a pharmaceutical formula: the R&D investment is large, but once the product is released to the public, copycats can emerge at marginal cost.

LLM labs are not blind to this threat. They impose API terms to forbid using their model outputs to train competitors. But enforcement is difficult and moreover, labs face a dilemma: the more widely they distribute and monetize their models (which they need to do to recoup costs), the more they enable potential distillation by third parties.

III. Strategic dynamics

In classic tech strategy, one way to gain power is to commoditize your complement – make the product or input that you don’t sell cheap and abundant, so that demand (and pricing power) increases for the part you do sell. In the AI ecosystem, we see multiple players applying this strategy against each other:

  • LLM Labs vs. Compute Providers (Nvidia, et al.): Training and inferencing cutting-edge models requires vast compute, much of which is supplied by Nvidia’s GPUs (or similar AI chips). From a lab’s perspective, this dependency is dangerous – if one vendor holds a near-monopoly on hardware, they can extract a lot of the value via high chip prices. Nvidia earns an extraordinary gross margin of more than 70%, even while selling cards at MSRP far lower than the grey market price. LLM labs are thus highly motivated to commoditize the hardware side. We see Anthropic, for example, making sure its models run on a variety of hardware: they use Google’s TPU v5e chips “at scale” for training Claude, and they also partner with Amazon to use AWS’s custom Trainium accelerators.

  • Hyperscalers (Cloud Platforms) vs. LLM Labs: The big cloud providers – Amazon AWS, Microsoft Azure, Google Cloud – view AI models as valuable content that can drive compute consumption on their platforms. But if one model lab becomes too powerful (able to dictate high prices or exclusive access), it threatens the cloud’s control. So hyperscalers try to commoditize the models themselves. One approach is backing multiple AI labs and open models: e.g. Amazon has invested in Anthropic, while also partnering with smaller startups and supporting open-source Hugging Face models on its Bedrock service. Microsoft has a special relationship with OpenAI, but it also hosts alternatives (like Meta’s LLaMA) on Azure, and develops its own “Phi” model inhouse. Google has its in-house models (PaLM, Gemini, etc.) but also invested in Anthropic to cover a second source. The cloud wants customers to view models as swappable commodities available through its API marketplace and commoditize the LLM model layer, so that the real money is made on cloud infrastructure usage. This is classic complement commoditization: make the AI model cheap (or even free) so that customers use more cloud compute and storage.

  • AI Customers (Enterprises, Tech Companies) vs. Both Labs and Cloud: Large consumer-facing companies like Meta, and many enterprise users of AI, have a stake in not becoming overly dependent on outside AI providers. For them, the AI model (and even cloud) is a input to their core business – a complement to what they do. Their incentive is to commoditize AI capabilities so that they can integrate intelligence into their products without paying monopoly rents. Zuckerberg has described Meta’s open-sourcing of LLaMA in this light: by releasing a high-quality model to the public, Meta hopes to undermine the proprietary advantage of OpenAI et al and ensure a competitive pool of models. If everyone has strong LLMs, no single supplier can overcharge Meta or dominate the market narrative – and Meta itself can benefit from the community improvements on these models.

In summary, there is a strategic tug-of-war in the AI ecosystem:

  • Labs want cheap, ubiquitous hardware and cloud (to lower their costs).
  • Clouds want many labs and models vying for attention (to keep models replaceable and drive cloud usage).
  • End-users want many readily available models and cloud options (to avoid dependence on any one supplier).

Each player tries to commoditize the other layers. This strategic dynamic will heavily influence profit pools. If the hyperscalers succeed, the labs could see their models become low-margin components, with most profit accruing to the cloud platforms selling the compute cycles. And if open-source efforts backed by users succeed, no one supplier may have big margins – AI might become an infrastructure commodity that end-users deploy at cost.

IV. A way out?

So, LLM labs spend huge resources just to keep pace technically with a growing number of competitors, all while facing intense commoditization pressure from the other actors in the ecosystem. Is there any reason to believe that these can become good businesses?

One point of hope is that while OpenAI’s API margins are being compressed, as of early 2024 they were still at a relatively healthy 55% level. More recently, DeepSeek claimed an 80% gross margin. Revenue growth is also extremely healthy - OpenAI expects revenue to triple in 2025. Anthropic is growing even faster, albeit from a lower base.

How can these margins and revenue growth levels be sustained long term, given the neck-and-neck nature of the technical race and wide field of competitors? Ideas would include:

  • Capture more context: in the most common LLM usecase, the only context the model has about the problem is whatever appears in the prompt. Features like Claude Projects changes this, because it creates a persistent workspace that the LLM can read/write to, and this context creates a cost to the customer of switching from one provider to another. You can imagine lots of other ways in which you could increase switching costs - one reason it’s hard to switch from Google to Bing is because Google has learnt from your habits to build up a good model of the kinds of results you want to see, and it’s clear that LLMs will also be able to learn preferences like this.
  • Obtain a training data advantage: a key input to the RLHF process is labelled human preference data. This data is expensive to obtain - Scale AI, a key provider of this data, makes about $2bn/year in revenue which directly comes out of the R&D budget of the labs. If a lab can operate a consumer website at scale they can enlist their own users to help generate it: for example, ChatGPT.com will frequently asks which of two candidate responses the user prefers. This cheap source of training data could be a durable moat for OpenAI.
  • Build a marketplace: marketplace businesses that aggregate demand from buyers and sellers have excellent economics. The AI labs are the natural place for sellers of tools (e.g. case-law databases, route planning software, websearch) to reach customers who want to use these tools to enhance the capabilities of their AI assistants. There is an emerging ecosystem of tools here with the model context protocol and I would expect this to evolve into a full-fledged “App Store for AIs” at some point.
  • Build a regulatory moat: many labs are pushing for governments to impose information controls and minimum security standards on developers of advanced AI. Their stated motivations for this are to advance the national security interests of their host countries, but it would also go some way to lessening the diffusion of their own research to their competitors at home and abroad, allowing them to capture more of the value of their own research spending.
  • Vertically integrate: if it is indeed inevitable that most of the value will be captured by customers, you could move upstream and try to use your LLM tech to disrupt an existing industry. e.g. you could run a purely LLM-based legal practice or consultancy business, have your army of superhuman coders build the next Microsoft, or run a movie studio where all the content is designed and created by AI.
  • Specialize: in a related point, it may be worth trying to capture more value from particular high-value verticals by providing a specialized AI product that e.g. has access to professional databases. This is the approach being taken in the legal world by Harvey.
  • Bundle: per Chris Dixon’s excellent explanation of the economics of bundling, it can improve both consumer and producer surplus when a bundle aggregates goods where buyers’ variance of preferences for those goods is high. This suggests an LLM product that bundles access to lots of underlying data products (news, case law, scientific publications, textbooks…), and is able to route a user’s query to whichever is needed to answer the query de jour. Think something like a Bloomberg Terminal but for natural language.
  • Hope that economies of scale emerge: current model training costs seem to be relatively low, but frontier model costs are expected to continue to scale by 2-3x per year. If costs scale faster than revenue growth, you’ll inevitably end up with a monopoly or oligopoly of large firms that can amortize their training costs across large numbers of customers.

I think it’s worth paying particular attention to the “drop in remote worker”, as envisioned in e.g. Aschenbrenner’s Situational Awareness. A number of labs (Anthropic, OpenAI, Manus) are working towards an AI agent that you would be able to onboard to your company’s intranet just like any remote employee, and which could then e.g. start to code its way through your JIRA backlog, come up with new product designs, coordinate with suppliers to negotiate costs, or prepare reports on business strategy. This has a number of advantages over the current business of selling tokens:

  • Natural moat: after onboarding an agent from a new lab, you will need to invest substantial effort in training it to use your internal systems in just the same way that you would need to train a new human employee, and agents will be able to share their knowledge with each other. Thus, the lowest friction choice when onboarding an extra agent will be to use an extra instance from whichever provider you are already using.
  • Huge TAM. Software development seems like it is only maybe 5 years from being substantially automated, and according to BLS data in the US alone 2023 spending on information technology wages was in excess of $200bn/year. A lot of this is up for grabs by a sufficiently capable agent model – to say nothing of the opportunity internationally or in verticals other than software. Sequoia asks the “$600bn question” about where AI labs will find their revenue, but the market opportunity here is large enough that a computer use agent could plug this $600bn hole all by itself.
  • Lowered risk of model extraction - because the remote worker will likely run on the lab’s hardware and only connect via remote desktop/another collaboration tool to the customer’s intranet, the only output the customer will observe is the action taken by the model. Any sort of internal chain-of-thought or reasoning tokens can be hidden. Thus there is only a risk of model extraction if you can buy enough example data from the lab to do a full RL training run, which is unlikely to be economical.
  • Virtuous cycle of AI demand: increasingly capable AI systems will cause more and more business processes to be purely automated, with no human in the loop at all. These fully automated processes are more efficient than partially-manual ones because e.g. they don’t take holidays, are available 24/7, and can quickly scale capacity to meet demand fluctuations. This then increases the incentives to automate any process that interacts with the newly-automated one, because Amdahl’s law implies that a larger fraction than before of the time required to complete a task is being spent within the manual process. Thus, AI deployment within businesses has a self-accelerating nature to it, and net dollar retention should be extremely high in consequence.

In short, while there are certainly challenges to the AI lab business model, it does seem like there are a lot of strategic options open to them. It seems unlikely to me that the labs will end up being completely commoditized - so maybe there is some hope for my Anthropic stock options after all :-)