This is the second of four posts discussing the state of AI. The previous post provided an analogy to help describe the nature of LLMs and the power of scaling laws. This post delves deeper into the industry’s competitive dynamics.
OpenAI’s release of its chatbot in November 2022 has come to be known as the ChatGPT moment – the point when the world realised how capable large language models (LLMs) had become, and how capable they were likely to become. Three and a half months after ChatGPT’s release, Anthropic, founded by a group of ex-OpenAI employees, launched Claude, a rival product emphasising the safety aspects of its model.
ChatGPT went on to capture the world’s imagination and become a leader in the consumer sector, with a billion users and >$24B in annualised revenues, while Anthropic established itself as the leader in the enterprise space, gaining the trust of large enterprises and ramping its revenues to a $30B+ run rate. Can these two LLM leaders maintain their dominance?
They face three broad categories of competitive risk: commoditisation, vertical integration, and the emergence of radically new technologies. In this post we will examine the potential risks of commoditisation (and address the other two categories of risk in subsequent posts).
There is already strong competition among a handful of leading LLM providers, as well as from several challengers – most notably from China, where enormous pools of technical talent exist. These challengers also have the advantage of being able to use the models of the leading LLM providers to improve their own models, legally or illegally training on the leading models, and thus “copying” their knowledge base (in a process called “distillation”.) Could this competition take meaningful market share away from the leaders or significantly erode their margins? There are several factors that work in favor of today’s incumbents.
First, the leading LLM providers have found effective ways to compete with lower-cost alternatives such as the open-weight models from China (but also from Europe and from Meta). While these models exert downward pressure on pricing, the LLM leaders have responded to such threats successfully to date by reducing prices for older model generations while maintaining premium pricing for their most cutting-edge models. As a result, they have been able to better compete with budget and open-weight alternatives while keeping users on their platforms. The LLM leaders have also shown an ability to rapidly adopt new innovations. For example, when DeepSeek introduced its first reasoning model in January 2025, the world was astounded by its capabilities and impressed by the efficiency innovations it had introduced. Since then, the leading LLM providers have adopted many of the architectural improvements DeepSeek first introduced, such as Mixture of Experts methods which splits up the LLM into multiple smaller sub-networks, reducing the model’s footprint and thus the cost of delivering answers. Indeed, adopting other companies’ incremental innovations can cut both ways – just as new entrants can “copy” or learn from the incumbents, the incumbents can also learn and adapt to the techniques introduced by the new challengers. This makes it more difficult for those challengers to sustain a lasting edge. At the same time, to maintain their lead, the incumbents continue making step-change improvements in their best models while offering prior generation models at lower cost.
Second, there are significant barriers to entry in this market, largely due to the staggering requirements of scale in the industry. As a result of the scaling laws mentioned in the previous post, the leading model providers are spending more than a billion dollars on infrastructure to train their current models, up 10x from just a couple of years ago, and expected to grow another 10x in upcoming generations. Any challenger would need comparable financing to catch up, without the user- or revenue base the incumbents already enjoy. Furthermore, the business dynamics of LLMs differ fundamentally from those of traditional software, which has near-zero marginal costs and historically allowed well-funded newcomers to undercut incumbents on price. LLMs have significant marginal costs: each query consumes substantial compute resources, and operating large-scale inference infrastructure efficiently requires enormous capital investment. In that way, developing an open weight model is only the first part of the cost equation – once a company has a model, it still needs to turn that into a running system that can deliver intelligence in real time at scale (even if the cost is passed on to the customer). All of this requires enormous amounts of capital, compute, and power. This creates a structural advantage for large players even if model capabilities become commoditised. The dynamic is comparable to cloud infrastructure: Amazon’s AWS offers a largely commoditised service, yet Amazon has built a massive and highly profitable business leveraging scale, facing only two major competitors, Google and Microsoft. A “commoditised” LLM industry may end up resembling cloud services, where a handful of at-scale players dominate the market with healthy margins, leaving little room for new entrants.
Besides the actual dollars required, any potential competitor would also need significant scale in: (a) talent, (b) data and (c) distribution.
a) Talent
In the previous post, we noted that tech behemoths like Amazon, Microsoft, and Apple have largely avoided competing directly with the LLM providers, preferring instead to partner with them, and to dedicate their own AI talent pool to a plethora of other internal needs, such as developing applications on top of those models. These behemoths certainly have the capital to compete, so their choice is at least partly driven by the talent pool and expertise they would need to acquire or reallocate just to get in on the race. In support of this argument, we can point to Meta, the one titan which did decide to vie for a position in this race, and which reportedly had to offer pay packages in the hundreds of millions to the researchers it was trying to poach, despite having a substantial in-house research effort already in place. It is also instructive that only two companies have managed to come from behind and emerge as serious contenders in the race for consumer chatbots, and both have had deep research pedigrees to draw on. The first, Google, has been an AI pioneer for decades and could draw on its enormous AI research division to launch Gemini. The second, xAI, could tap into Telsa’s deep AI talent pool and history of model development, as well as Elon Musk’s role as a founder of OpenAI to attract the necessary personnel. The fact that such a deep pool of talent is required to compete inevitably makes it more difficult for newcomers to enter the market.
b) Data
Each of the LLM providers has significant proprietary data to help build and solidify its competitive moat. OpenAI and Anthropic have obtained that data by being first movers and incorporating their user interactions into the feedback loops needed to improve their models. Gemini was able to draw on a wealth of data across Google search, news and YouTube, and xAI has had a stream of real time and historical tweets to draw on.
c) Distribution
Both Gemini and xAI also had existing distribution channels they could leverage in trying to catch up with OpenAI in the consumer chatbot race. Any new competitor would need to reach millions of consumers or tens of thousands of businesses to compete. xAI and Gemini were able to do so because of their existing reach to consumers. Similarly, in the enterprise space, the partnerships Anthropic and OpenAI have with Amazon and Microsoft also strengthen their moats. As evidence of the switching costs of enterprise AI integrations, we can point to the recent differences between Anthropic and the US Department of Defense (DoD), where Anthropic refused to allow the government to use its models for certain autonomous applications. This resulted in the DoD deciding to replace Anthropic as a supplier. In this case, it was estimated that it would take the DoD 6 months to switch away from Anthropic. This gives us an indication of the stickiness of enterprise LLM integrations.
Few companies have the depth of talent, the breadth of differentiated data, and the distribution capabilities to enter this race with a significant advantage. This makes it more likely that the scenario painted above will persist – i.e., that there will be a handful of LLM leaders dominating the market for providing AI chatbots for consumers and APIs to businesses.
Third, each leading model will continue to be differentiated, even in a commoditised market. As noted above, each of the leading models has been built with emphasis on specific input datasets, be it the consumer interactions of xAI, OpenAI and Gemini, or the enterprise experiences of Anthropic. As a result, not only does each company have different strengths and weaknesses, but each model also exhibits distinct characteristics and personality. Power users of LLMs today typically use multiple models concurrently. They learn the nuances and characters of each model and tend to ask each LLM different kinds of questions, playing to the strengths of each, much like a business executive might brainstorm with different advisors and team members, each bringing their own expertise and style.
This differentiation is unlikely to disappear soon, because it is embedded in the specific training regimes each company imposes, based on the data they have, as well as their approach to the training. In our first post we compared LLM training to the schooling of humans. Extending the analogy, we can think of each LLM as having received a different type of education and upbringing. Each LLM is shaped by the culture and values of its schooling experience. Notably, Anthropic formalises this by embedding an AI Constitution within its model. This document influences the training regimen of the LLM and gives Claude its own personality and set of values. Even if other LLM companies do not encode their values as explicitly, their institutional culture and values inevitably permeate the systems they build. This leads to significant differentiation between the models.
Fourth, leading-edge models will likely continue to dominate the market. We have seen this pattern in multiple high growth industries, notably in the semiconductor industry, where customers have consistently demanded the most recent and most performant technology, and they have been willing to pay a premium for it. We are also seeing this pattern each time ChatGPT or Claude release new versions of their models. Customers consistently pay a premium for new releases and each time the increase in performance drives more demand. This pattern will likely persist for two main reasons.
a) Each step change in model capability unlocks new usage domains.
For example, in 2024 LLMs were unable to complete complex coding projects, and as a result many programmers only used them to complete a few lines of code. By the end of 2025, there was a marked improvement in coding capability with the latest models. These capabilities allowed models to perform multi-step coding tasks reliably, and in turn motivated more programmers to use the latest models more, on more complex tasks driving up demand for tokens – and the most expensive ones at that. This underscores Claude’s astounding growth from a $9B run rate at the end of 2025 to $19B in March of 2026 and surpassing $30B in April.
b) On the internet we have seen that even incremental changes in performance and speed can drive enormous increases in usage.
For example, Google research shows that a 0.1 second delay on mobile apps can reduce conversion rates and page views by 8%, an enormous jump for such a small timing difference. This result has been consistent across various products on desktop and mobile. As such, we would expect even small differences to make a disproportionate difference in the use of LLMs, be they for consumer or enterprise apps where smoothness of interactions and the incremental improvements in output quality could have a disproportionate impact on final results. These examples seem to indicate that even if the gap between the leading-edge models and the challengers narrows the leading edge will still dominate the market.
Fifth, technical commoditisation would not necessarily imply commoditisation of trust. As LLMs are entrusted with more complex and sensitive tasks, brand reputation and reliability will matter more. Consider a programmer using the popular Cursor tool today. Cursor allows developers to choose the model they want to write code with – be it Claude, Gemini, or ChatGPT. This choice has serious consequences. A weaker model may introduce subtle bugs that require significant effort to resolve, wiping out the expected efficiency gains from AI. Worse, a weak model may misunderstand code in ways that only become apparent much later, creating what is known as technical debt: bad code whose true cost emerges painfully over time. As a result, once a developer learns to work with a specific model – understanding its capabilities, limitations, and reliability – the costs and risks of switching providers increase substantially.
If you are asking a model simple questions with true or false answers, you might choose the model that delivers the most correct answers at the lowest cost and greatest speed… but if you are asking it to complete complex, qualitative tasks like coding, which are difficult to verify, trust becomes far more important. The importance of trust can also be seen in the adoption rates of the models released by the leading LLM providers. As mentioned above, Anthropic and OpenAI charge a premium for their most advanced models while discounting their older or less capable models –despite their higher prices, in areas like coding, users seem to prefer the most capable models because these ultimately provide the best performance – delivering results users can trust which ultimately also result in the highest ROI.
This same dynamic is expected to spread to other sectors (beyond coding) as AI systems evolve into business agents that take actions on users’ behalf and earn their trust to manage increasingly difficult and sensitive workflows.
In this context, choosing an AI agent is less like choosing between Coke or Pepsi at lunch and more like choosing a surgeon, lawyer, psychiatrist, or business partner. When an AI manages workflows, negotiates transactions, or executes business decisions, minimising mistakes and reducing risk becomes paramount. In such situations, users are likely to favour providers they have learned to trust.
This is the second of four posts about the current state of AI. In the next post, we will delve more deeply into the risks and opportunities of vertical integration in the industry.
Author: Salman Farmanfarmaian & Richard Rimer
If you’d like to explore any of these themes further, or discuss how you can take advantage of these opportunities, please get in touch with us at info@bedrockgroup.com.
Important Legal Information
The views presented in this article are Salman ans Richard’s own views and might not necessarily reflect the investment approach of Bedrock Group.
The information and opinions contained in this article are for background information and discussion purposes only and do not purport to be full or complete. No information in this document should be construed as providing financial, investment or other professional advice. The views and opinions expressed herein are based on current market conditions and are subject to change without notice. No representation is made that any forecast or projection will be realised.
Information included in this document is intended for those investors who meet the definition of Professional Client under the Swiss FinSA regulation as well as Professional Client or Eligible Counterparty under the UK Financial Conduct Authority.
Investment Risks
The value of all investments and the income derived therefrom can fluctuate due to market movements and you may not get back the amount originally invested. invested.
Limitation of Liability and Indemnity
Bedrock expressly disclaims liability for errors or omissions in the information and data contained in this document. No representation or warranty of any kind, implied, expressed or statutory, is given in conjunction with the information and data. Bedrock accepts no liability for any loss or damage arising out of the use or misuse of or reliance on the information provided including, without limitation, any loss of profits or any other damage, direct or consequential. You agree to indemnify and hold harm less Bedrock and its affiliates, and the directors and employees of Bedrock and its affiliates from and against any and all liabilities, claims, damages, losses or expenses, including legal fees and expenses arising out of your access to or use of the information in this presentation, save to the extent that such losses may not be excluded pursuant to applicable law or regulation. Any opinions contained in this presentation may be changed after issue at any time without notice.
After spending a week inside Silicon Valley’s AI ecosystem, David Ephrati argues that today’s AI moment is not a repeat of the internet wave, but something materially larger. This article distils seven observations he believes carry lasting relevance for investors.
Gain firsthand insights from Silicon Valley’s top visionary founders, investors, and industry leaders on the AI, robotics, and venture capital trends shaping the valley.
In this article, we explore some historical examples of the types of companies that have benefited from new disruptive technologies - notably Coca-Cola's transformation as a result of refrigeration - and draw some conclusions about the potential impact of AI, highlighting some of the companies set to be potential winners.