What's the trillion dollar opportunity for AI?
Some more thoughts on the state of AI in 2025 and why we need AI governance
TL;DR
This newsletter is about the opportunities and challenges with building and deploying applications built on top of foundation models. It looks Sequoia's thesis for where value will accrue in AI development, the relevant counter-arguments and the importance of governance for AI engineers.
Here are the key takeaways:
Sequoia believes that the value of AI will accrue in the application layer. This includes all the companies building applications on top of foundation model provided by the frontier developers.
This is done through the process of AI engineering. This includes a range of techniques for augmenting general capabilities of foundation models specific domains or use cases.
Sequoia seems to be of the opinion that there is an opportunity to harness the power and capabilities of LLMs to build tools that solve problems for others. And if AI engineers manage to do this well, then there is latent demand to tap into and generate lots of revenue from.
However, there are a few major difficulties with this as a business model:
The foundation model developers themselves are, and may continue to, creep into the application layer themselves, bringing with them greater access to compute infrastructure and VC funding.
The physical infrastructure for AI is scarce and therefore distributes control over the development of models and systems away from model developers and, crucially, AI engineers.
Dealing with the physical infrastructure problem is that most difficult. The barriers to entry for those wanting to build their own foundation models from scratch are so high.
But to compete with model developers building applications using their own models, AI engineers may need to find ways to make themselves meaningfully distinguishable from model providers. This might involve taking advantage of proprietary data to power products that serve a valuable niche.
However, to ensure longevity, AI engineers will also need to focus on AI governance. This is important for three key reasons:
It helps deal with the complexity of AI engineering
They can use legal compliance as a USP
It becomes easier to build a community that trusts them
If AI engineers can better deal with the complexity of LLMs, ensure legal compliance by default and be attentive to user needs, then they could deepen their moat.
The opportunity
Sequoia, a venture capital firm based in Silicon Valley, has a thesis for how AI diffuses across society.
By diffusion, I mean the implementation and use of AI for various domains and tasks across different sectors.
The argument from Sequoia regarding why diffusion has increased and will continue to do so has a few building blocks to it:
Better capabilities. Reasoning models promise to be the next big advancement in AI development. The excitement here concerns the scaling of reinforcement learning with verifiable rewards. By training models on math and coding datasets using reinforcement learning, they also show the ability to perform other tasks better. I explore this a bit more in How machines (might) take over the world and also 3 thoughts on o1. Sequoia contends that further advances in reasoning models is key to a higher rate of AI diffusion.
Falling costs. As I mentioned in How machines (might) take over the world, the cost of running LLMs is falling. According to research from the Stanford Institute for Human-Centered AI, "the inference cost for a system performing at the level of GPT-3.5 dropped over 280-fold between November 2022 and October 2024." Falling costs encourages more experimentation with and implementation of LLMs into wider applications and systems.
The underpinning tech stack is there. Several foundational technologies have developed and matured enough over the years to enable AI diffusion. Furthermore, each wave tends to be additive, meaning that each wave is more significant than the previous, like a snowball effect.
So where will the value accrue as a result of this diffusion? The application layer, or the 'white space' as Sequoia articulates it. This is the front-end part of the AI landscape that users interact with, whether those users are individuals or enterprises.
This application layer consists of companies building applications and systems on top of foundation models. OpenAI was one of the first to really propel this type of AI system with ChatGPT. As Keach Hagey states in her book The Optimist: Sam Altman, OpenAI, and the Race to Invent the Future, OpenAI stumbled across this approach back in 2022 when considering ways to deploy its GPT models in a more convenient manner rather than simply making them available via an API:
The core technology behind ChatGPT had been available for two years, and the updated model has been plugged into the API for nearly a year. In theory, anyone could have made ChatGPT themselves at any time by putting a chat interface on the model OpenAI was selling access to. But there was something special about the chat interface. "The raw technical capabilities, as assessed by standard benchmarks, don't actually differ substantially between the models, but ChatGPT is more accessible and usable," John Schulman [one of the founders of OpenAI] told MIT Technology Review.1
Today, there are several companies that have taken this approach, building on top of AI models to provide a variety of applications and systems for different use cases.
But why does Sequoia think that the value of the latest AI hype will accrue at the application layer?
The thinking here is that the general capabilities of foundation models can be further augmented by other entities to tailor their capabilities for a particular use case. The main process for achieving this is AI engineering, which itself includes a range of techniques for engineering applications and systems with foundation models. This includes prompt engineering, building retrieval augmented generation (RAG) systems, fine-tuning and even building autonomous AI agents.
The so-called 'LLM wrappers' have more to them than it may seem, according to Sequoia:
Application layer AI companies are not just UIs on top of a foundation model. Far from it. They have sophisticated cognitive architectures that typically include multiple foundation models with some sort of routing mechanism on top, vector and/or graph databases for RAG, guardrails to ensure compliance, and application logic that mimics the way a human might think about reasoning through a workflow.
You can think of foundation models as the motor engine and the wrapper as everything else that is needed to build a functioning car:
Sequoia seems to be of the opinion that there is an opportunity to harness the power and capabilities of LLMs to build tools that solve problems for other individuals or companies. And if AI engineers manage to do this well, then there is plenty of latent demand to tap into and generate lots of revenue from.
The challenges
This path for AI diffusion that Sequoia paints is a seemingly smooth one. But there are some challenges.
For one, the foundation model developers themselves are, and may continue to, creep into the application layer themselves. And this is not just limited to chatbots. For example, both OpenAI and Anthropic have recently announced AI coding agents underpinned by their respective foundation models.
It is possible that foundation model providers continue to encroach on territory currently occupied by its customers, and they will have several advantages when doing so. This includes access to massive compute infrastructure and lots of VC funding. And this could be devastating for the AI engineers:
Two years after ChatGPT’s debut we still have…ChatGPT. Not because founders aren’t trying, but because incremental consumer features die the moment OpenAI or Google wedges them into the base model, and because consumers habit-stack: once you have a single chat box anchored on your phone, every new one feels like toggling search engines circa 2004. The economics look even harsher for venture-backed “GPT wrappers”—Meeker’s slide on collapsing inference costs is good news for end-users and hyperscalers, lethal for startups that mark up tokens for margin.
In fact, OpenAI seems intent on doing more in the application layer with its models. According to an internal document revealed during the anti-trust proceedings against Google by the Department of Justice, the model developer plans "evolve ChatGPT into a super assistant that knows you, understands what you care about, and can help with virtually any task."
Accordingly, the LLM wrappers could be viewed as "features waiting to be absorbed."
Another counterpoint to Sequoia's predictions for AI diffusion is the reality of the bottlenecks. As Dave Friedman puts in his post:
Modern AI is bottlenecked not by clever apps, but by:
Power generation and grid interconnects
Physical land and real estate for data centers
Cooling systems and water usage rights
Multi-year permitting timelines
Supply-constrained GPU inventory with geopolitical exposure
These constraints don't blend to agile cycles, MVPs, or blitzscaling.
Software is elastic. AI is entropic.
In other words, the infrastructure for AI is scarce and therefore distributes control over the development of models and systems away from model developers and, crucially, AI engineers. This impacts the longevity of those businesses building AI applications and systems on top of foundation models.
The solution
How do AI engineers overcome these challenges?
The latter challenge regarding AI infrastructure is the most difficult to overcome. The barriers to entry for those wanting to build their own foundation models from scratch are so high. It requires an amount of resources that can only be expended by the bigger players, namely the frontier model developers.
But the former challenge relating to encroachment could be more realistic to tackle.
Firstly, there is some doubt over how much encroachment would actually happen. The more that model developers want to get into the application layer, the wider they spread themselves, potentially complicating their business model beyond what is manageable. Additionally, model developers may end up hampering the very activity that contributes to their main revenue-generating method; API access to their models. So maybe the incentive to build every application with their models and become the platform for everything is actually quite low.
And then separately, it seems that AI engineers will need to find ways to make themselves meaningfully distinguishable from model providers to survive long-term.
In this regard, I think Cursor AI, the AI-powered code editor, presents an interesting case study. Anysphere, the company behind the app, is the fastest growing startup ever after raising $900 million in its latest funding round. This brings the total capital raised to $1 billion, with a valuation of $9.9 billion and $500 million in annualised revenue. But is the hype justified?
For AI engineers to survive, they may need to harness the aspects of their product that nobody else has.
Cursor AI is a code editor with a variety of AI features. These features utilise the capabilities of foundation models, including Claude from Anthropic. For example, it includes an auto-complete function where the app suggests code edits to users as they are writing, and by pressing tab on their keyboard they can directly incorporate these suggestions into their scripts.
The core aspects of a product like Cursor AI include:
Building a UI to incorporate the foundation models into the editor for ease of access and use (in this case a fork of VSCode, another popular code editor from Microsoft)
Implementing tools to allow models to read files and make direct edits to a users' file (which enables the auto-complete function for example)
Develop a highly detailed system prompt to query the foundation models with to ensure that the responses used in app's features are tailored to help users code (see here the system prompt used for Cursor AI)
But in addition, Cursor AI does have proprietary data that it uses to build its own models. Its privacy policy states that, unless users opt-out, Cursor AI will collect "prompts, editor actions, code snippets, files in the repository, and edits made to this code" and use this data to improve its AI features. Michael Truell, CEO of Anysphere, also recently spoke in an interview with Bloomberg about how Cursor AI is powered by an ensemble of models that includes its own custom models.
Therefore, in addition to its UI, internal tooling and prompt engineering, Cursor AI can also leverage its proprietary data to continue improving its product and building its moat as an AI-powered coding assistant (which I would say is one of the more promising use cases for AI right now). Maybe this is enough to secure its long-term future.
However, this longevity also depends on another important factor that all AI engineers will need to contend with: AI governance.
Why should AI engineers care about governance? Three key reasons:
It helps deal with the complexity of AI engineering. LLMs are programmed implicitly rather than explicitly, which makes them unpredictable and therefore difficult to work with. Good AI governance goes beyond mere legal compliance by implementing measures that support quality management of AI systems and help ensure successful engineering projects.
They can use legal compliance as a USP. Many AI engineers will forgo legal requirements on the premise that the potential success of the product outweighs legal risks. But in fact, by dealing with legal compliance as early as possible, AI engineers avoid this bottleneck and outlast competitors.
It becomes easier to build a community that trusts them. AI governance is about managing the risks and benefits of AI systems, and therefore is about focusing on the needs and interests of users. By being attentive to this, AI engineers can foster a community with their users that creates a positive feedback loop for iterative improvement of its products.
If AI engineers can better deal with the complexity of LLMs, ensure legal compliance by default and be attentive to user needs, then they could deepen their moat. The can survive the current wave of diffusion and potentially take advantage of AI's trillion dollar opportunity.
Keach Hagey, The Optimist: Sam Altman, OpenAI, and the Race to Invent the Future (W. W. Norton & Company 2025), p.269.