Here’s How Anthropic and OpenAI Are Building Economic Moats

The upcoming IPOs of leading AI labs like Anthropic and OpenAI will put the question of their durable competitive advantages at the forefront for every public-equity investor. We’re on the more bullish side, as we think leading AI labs are building moats, as stated in the first installment of our three-part series about moats and monetization research.

We see that OpenAI and Anthropic benefit from two potential sources of economic moats, which can help them outlast competitors over the long term. The first is network effects (driven by data and product improvements), and the second is a structural cost advantage built on growing training costs and gigawatt-scale compute discounts.

How OpenAI and Anthropic Are Building Network Effects

AI frontier labs’ businesses naturally lend themselves to network effects. As more users interact with a company’s models, it can leverage this to improve those models over time. For users without enterprise agreements that prohibit these labs from training on their interactions, every interaction feeds into the development of next-generation models. As models improve, we see a clear pattern of increased usage, which in turn generates more data.

The network effect we observe at the leading labs runs through a six-station flywheel, with early adoption of coding products a clear demonstration of this dynamic. We expect this dynamic to ripple across other model use cases and personas as adoption of this technology broadens beyond the initial software development use case.

In the coding context, we have seen that improved models lead to greater use of coding tools such as Codex and Claude Code. This increased usage has allowed both Anthropic and OpenAI to collect more data and leverage it to improve their models for those specific use cases. Alongside that, these firms have also leveraged usage data to create new tools and products (Claude Design, financial services and legal research plug-ins, Microsoft Office add-ons). These improvements, both in the model layer and in the tooling around it, keep users on the platform and attract new users.

One dependency we are watching within this flywheel is the growing role of third-party reinforcement learning environment providers, sometimes described as AI gyms. The major labs appear to rely heavily on these vendors to source task data and evaluation environments, suggesting they do not hold all relevant enterprise-grade training data internally, partly because enterprise agreements restrict training on customer interactions. If the highest-value post-training data ends up owned by third parties selling to every lab, the proprietary data leg of the flywheel would be weaker than it appears, and we would revisit our network-effect argument accordingly.

This particular flywheel is not a classic direct network effect (such as at Facebook or Instagram), wherein each user makes the product mechanically more valuable to other users. Rather, we see the network effect forming around the model businesses as a data-and-product flywheel, serving as a feedback loop these model firms can use to iterate faster on both model and tool development.

We are seeing clear adoption of these tools, and the results are also quantitatively present. Anthropic’s net retention rate, which is a measure of how much extra existing clients are spending on your platform, is north of 500%, well above the gold standard in software, which is around 120%-140%.

There is a clear link between the acceleration in Anthropic’s product velocity in recent months and its usage, with both better products and more usage feeding off each other. In our view, part of this acceleration reflects the labs increasingly using their own models to write code and run research—an early shade of self-improvement that compounds in favor of whichever lab holds the best models at any given time.

Cost Advantages: We See Four Barriers to Entry

The second moat source these model companies are busy constructing is a durable cost advantage. While financial disclosures remain limited, we know that building large language models is expensive. Keeping costs down while building and providing these models at scale is a tough nut to crack. This dynamic opens the possibility that these firms possess cost advantages over smaller peers and potential new entrants.

Anthropic and OpenAI are using algorithmic advancements, custom silicon, and multiyear compute deals to quantitatively reduce their cost to serve these LLMs, putting them at a clear cost advantage over other firms. Overall, we see this advantage as being built using four reinforcing channels.

Massive Upfront Training Costs

Training costs for models increase materially with each generation. As a result, a challenger entering today must not merely match the leaders’ current spending, but also have a novel method to “catch up” on the accumulated learning these model companies have gained by iterating through various model architectures and methodologies. We attribute this dynamic to the whittling of the model field from many dozens of model companies claiming to be at the leading edge back in 2023 to just a handful at the frontier in 2026.

Compute Volume Discounts

Leading model companies secure better compute pricing than their smaller peers. The leverage is both volume and commitment term: capacity bought on multiyear reserved terms is priced at roughly $2.25 per B200 GPU (graphics processing unit) hour, compared with $16.11, the list price at the hyperscalers. The labs are the only buyers able to commit at that duration and scale.

Applied to a gigawatt of capacity—or, loosely, 465,000 Blackwell-class GPUs—that gap is the difference between a $9 billion and a $66 billion annual compute bill. It is also worth noting that the $9 billion figure sits close to the $10 billion compute budget used in our earlier stylized income statements, meaning the labs operate at gigawatt scale and buy on terms no subscale entrant can access.

Talent Costs

AI, unlike much of software development, lends itself well to the “superstar effect.” The underlying reason is that much of AI progress and development depends on developing effective training and post-training processes. To do so, you need outstanding researchers who demonstrate exceptionally high marginal returns to talent. A good example is Meta’s hiring spree to build its research team, with some researchers offered tens of millions. These top researchers are locked up in a handful of labs, leaving smaller labs increasingly unable to attract them.

In fact, much of the movement we have seen in talent has been between large AI labs, rather than exiting the small group altogether.

Custom Silicon

All major labs are investing in custom AI accelerators—hardware components designed to speed up AI and machine learning processes. As we have seen with Amazon’s AMZN Trainium and Alphabet’s GOOG/GOOGL tensor processing units, these accelerators can be much more efficient for training and running inference on AI models, providing AI labs with a permanently lower cost of goods sold than labs or companies building or running models on off-the-shelf GPUs.

Beyond cheaper compute on an absolute basis, we believe the leading labs also extract more from each unit of compute they own, with utilization, batching, and goodput optimizations further lowering cost per token. Several of these optimizations scale directly with the absolute size of a lab’s fleet, deepening the advantage as these businesses grow.