Every few years, a technological transition creates a new generation of extremely valuable companies. The internet created Google, Amazon, and Facebook. Mobile created Instagram, Uber, and WhatsApp. Cloud computing created Snowflake, Datadog, and Stripe. Artificial intelligence -- and specifically, the emergence of foundation models -- is creating the next generation, and at Milestone AI Ventures, we have been deploying capital aggressively into this transition for the past three years.
But the question that founders and co-investors ask us most frequently is not whether AI is a major transition -- that much is clear. The question is where exactly in the AI stack the most durable, defensible, and valuable companies will be built. That is the question this piece is designed to answer.
The Foundation Model Layer Is Commoditizing Faster Than People Think
Let's start with the uncomfortable truth that shapes our entire investment framework: the foundation model layer -- the large language models, diffusion models, and multimodal models that have captured most of the public attention and a substantial fraction of venture capital over the past three years -- is commoditizing faster than nearly anyone anticipated when GPT-3 launched in 2020.
When OpenAI launched GPT-4 in early 2023, the capability gap between it and all other models was enormous. Today, GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, Llama 3, Mistral Large, and a dozen other models are all within shouting distance of each other on most standard benchmarks. The capability frontier is still moving forward, but the gap between frontier and "good enough" is closing rapidly as open-source models improve and model training costs decline. Within two to three years, we expect that a 90th percentile enterprise AI application can be built on an open-source model fine-tuned on proprietary data -- no API dependency required.
This does not mean that foundation model companies are bad investments. OpenAI, Anthropic, and Google DeepMind are building genuinely remarkable technology and will continue to push the capability frontier. But for seed-stage investors like us, the question is whether a new foundation model company started today can build sufficient defensible differentiation before the commoditization wave arrives. In most cases, we believe the answer is no -- unless the company has a truly novel approach to a specific modality, domain, or capability where existing models are genuinely inadequate.
Where the Durable Value Actually Lives
If the foundation model layer is commoditizing, where does the durable value in the AI stack actually live? At Milestone, our investment thesis converges on three distinct areas that we believe will generate outsized returns over the next decade.
1. Data Infrastructure and the Intelligence Stack
The most underappreciated reality of deploying AI at enterprise scale is that the hard problem is almost never the model. The hard problem is getting the right data to the model, at the right time, in the right format, with adequate governance, observability, and lineage tracking to satisfy enterprise compliance requirements. Every Fortune 500 company we talk to reports spending 60 to 80 percent of their AI deployment effort on data infrastructure -- and then struggling to maintain it as their AI applications evolve.
This creates enormous opportunity for companies building the data infrastructure layer for AI. Vector databases, embedding pipelines, retrieval-augmented generation systems, data labeling and curation platforms, real-time feature stores designed for AI workloads, and AI observability platforms are all areas where we are actively investing and where we believe durable, high-margin businesses can be built. Unlike foundation models, these infrastructure companies benefit from deep workflow integration and high switching costs -- once a Fortune 500 company has standardized on a specific vector store or data pipeline architecture, the cost of switching is genuinely prohibitive.
Our portfolio company DataNexus is a clear example of this thesis. They are not building a foundation model -- they are building the infrastructure that makes foundation models useful at enterprise scale, specifically the vector database and semantic search layer that enables retrieval-augmented generation for billion-document corpora. Their business is defensible not because of model weights or training data, but because of the depth of their integration into enterprise data systems and the compounding advantage of their query optimization research.
2. AI-Native Vertical Applications with Proprietary Data Flywheels
The second major investment category we are focusing on is AI-native vertical applications -- companies that are building the dominant AI-powered workflow for a specific, large enterprise function. The key word is "AI-native": we are not interested in companies that have bolted AI onto an existing SaaS product. We want companies that were designed from the ground up around AI capabilities, where AI is not a feature but the core value delivery mechanism.
What makes these businesses defensible is the proprietary data flywheel. When a company deploys an AI system to manage, say, clinical documentation for a hospital network, that system generates an enormous amount of highly specific training signal -- clinician feedback, annotation data, outcome correlation data -- that makes the model dramatically better at that specific hospital network's specific workflows. This creates compounding data advantages that are essentially impossible for a competitor to replicate without deploying their own system at scale, which requires months of sales cycles and implementation effort.
We have seen this dynamic play out in our portfolio with Cognex Health. Their AI diagnostic system is now deployed at 40 hospital systems, and each deployment generates training signal that improves the model's performance at other hospital systems. Their accuracy advantage over the best foundational medical imaging model is now 15 percentage points on their benchmark datasets -- not because they have better model architecture, but because they have three years of proprietary clinical feedback data that nobody else has.
3. AI Agent Infrastructure and Orchestration
The third area where we are investing aggressively is the emerging infrastructure layer for AI agents and multi-agent systems. As language models become capable enough to reliably execute complex, multi-step tasks, the fundamental unit of AI deployment is shifting from single-turn inference (ask a question, get an answer) to agentic workflows where AI systems take sequences of actions over extended time horizons.
This shift creates a new category of infrastructure needs: agent orchestration frameworks, tool integration platforms, agent memory and state management systems, multi-agent coordination protocols, and safety monitoring systems specifically designed for agents that take autonomous actions with real-world consequences. This is an extraordinarily early market -- most companies we talk to have not yet deployed production agentic systems -- but the infrastructure layer is being built now, and the companies that establish the standard frameworks will benefit enormously from their first-mover position.
The Investment Criteria That Define Our Decisions
Understanding where we invest is only half of the equation. The other half is understanding what we look for in the specific companies we back within these categories. After reviewing more than 400 pitches and making 28 investments, we have developed a clear framework for evaluating AI seed-stage opportunities.
Technical founding team with research-grade credibility is non-negotiable. The AI field moves at research speed, and the companies that win will be led by people who can read papers, run experiments, and make genuine technical contributions. We want founders with the credibility to hire exceptional AI researchers and engineers, and the judgment to evaluate technical claims from competitors, employees, and potential acquirers with appropriate skepticism.
A clear path to proprietary data advantages is essential for application-layer companies. We spend a significant fraction of our diligence on understanding how a company's product generates training data that improves its AI system, and how that training data is protected from being used to train competitor systems. The companies that cannot articulate a compelling data flywheel story rarely survive the commoditization of the underlying model layer.
Market timing awareness -- specifically, a crisp understanding of why now is the right time to build this particular product -- is often the most differentiating factor between good and great pitches. The AI capabilities required to build many of the products in our portfolio did not exist two years ago. The best founders understand exactly which recent capability developments made their product possible and why the window for establishing category leadership is limited.
The Macro Environment for AI Investing in 2025
No investment thesis operates in a vacuum, and it is worth addressing the macro environment for AI investing as we enter 2025. The enthusiasm that drove AI valuations to extraordinary heights in 2023 has given way to more sober assessment of which companies are generating genuine revenue versus impressive demos. We believe this is healthy for the ecosystem -- it rewards companies that have solved real customer problems with real willingness to pay, and it creates opportunities for seed-stage investors like us to back companies at more reasonable valuations than were available eighteen months ago.
Several macro trends make us particularly optimistic about the AI investment landscape in 2025 specifically. Enterprise AI budgets are growing rapidly after two years of cautious pilot projects. Regulatory frameworks for AI are crystallizing in ways that create new compliance and governance requirements -- which create new markets for companies in our AI safety and infrastructure thesis. And the cost of training state-of-the-art models continues to decline, which lowers the barrier to building differentiated AI applications and accelerates the commoditization we described earlier.
We are investing from Milestone Fund II, a $125M vehicle raised specifically to capture this opportunity. Our fund size is deliberately constrained -- large enough to write meaningful checks and maintain the portfolio concentration required to generate outsized returns, but small enough that we are disciplined about investment selection rather than deploying capital broadly. We believe the next three years represent the single best opportunity to invest in seed-stage AI companies that we will see in this decade, and we are approaching it with the same intensity that our founders bring to building their companies.
A Note on What We Are Not Investing In
Equally important to our positive thesis is clarity about what we are passing on, because it shapes where we spend our time. We are not investing in companies whose core technology is a fine-tuned version of a publicly available foundation model without significant differentiation in training data, user experience, or integration depth. We are not investing in AI applications in markets that are too small to support a venture-scale business, even if the AI application is genuinely impressive. And we are not investing in teams that are primarily marketing people or business development executives without a strong technical co-founder -- the AI companies that win are built by people who deeply understand the underlying technology.
We are also increasingly cautious about companies building in categories where Google, Microsoft, or Amazon have publicly announced competing products. The hyperscaler risk in AI is real and growing, and founders need to have a clear and credible answer to the question of why they can win in their specific market even when a company with unlimited compute, distribution, and engineering talent enters with a competing product.
Conclusion: The Decade Ahead
The foundation model era has created a genuinely unprecedented opportunity to build transformative AI businesses. But the opportunity is not uniformly distributed across the AI stack. The most valuable companies of the next decade will be those that build durable data advantages, deep workflow integration, and infrastructure that other AI systems depend on. Foundation model capability will increasingly be a commodity input to these businesses, not their core differentiator.
At Milestone AI Ventures, we are oriented around this conviction in every investment decision we make. We believe the founders who understand this distinction -- and who build accordingly -- will create the defining AI companies of the 2020s and 2030s. We are privileged to be partnering with them from the earliest stages, and we are deploying capital with urgency because we believe the window for establishing category leadership in many of these markets is measured in months, not years.
If you are building in any of the categories described here, we would like to hear from you. The best way to reach us is through the contact form on our website or by emailing founders@mstone-ai.com directly.
Dr. Sarah Chen is the Managing Partner and Co-Founder of Milestone AI Ventures. She previously served as Research Director at Google DeepMind and holds a Ph.D. in Machine Learning from MIT. The views expressed here are her own and do not constitute investment advice.