The numbers are staggering. Global AI investment surpassed $340 billion in 2025, representing a 47% increase over the prior year. Enterprise AI adoption crossed the 70% threshold at Fortune 500 companies. More than 4,200 AI-focused startups raised their first institutional round in the United States alone. By every quantitative measure, we are living through the largest technology investment cycle in history.
And yet, in conversations with limited partners, fellow investors, and the sophisticated founders who cross our transom daily, I encounter the same anxiety: is this sustainable? Are we at a peak? Where is the signal in this noise?
At Swarm Capital, we have spent the better part of 2025 doing what early-stage investors are supposed to do: thinking deeply, investing carefully, and ignoring the noise. What we have found is both reassuring and clarifying. The headline figures obscure a more interesting story — one that has profound implications for founders, investors, and anyone trying to build intelligently in this environment.
The Infrastructure Build Is Different This Time
Every major technology platform cycle has a characteristic arc. Infrastructure gets built. Applications get built on top of that infrastructure. The infrastructure companies often generate the most durable returns, while the application layer sees both the highest highs and the most spectacular failures. This pattern played out in cloud computing, mobile, and social — and it is playing out again in AI.
What is different about the current AI cycle is the speed at which the application layer is catching up to infrastructure. In the cloud era, it took roughly five years after AWS launched EC2 (2006) before application-layer companies began generating meaningful revenue at scale (2011-2013). The foundational models that power today's AI applications were largely available — in usable commercial form — by 2022 or 2023. By 2025, we are already seeing application-layer AI companies reaching $10M, $20M, even $50M ARR within three to four years of founding. The compression is unprecedented.
This matters for investors because it fundamentally changes the risk calculus. A seed-stage AI company founded today does not need to wait for infrastructure to mature before it can build a real business. The rails exist. The models exist. The talent has diffused broadly enough that strong teams can be assembled outside of a handful of zip codes. The question has shifted from "can this be built?" to "can this team build it and win the market?"
Three Structural Shifts Investors Are Missing
Most AI investment commentary in 2025 focuses on the obvious: large language models, generative AI, model cost curves, and the incumbent versus startup dynamic in foundation model development. These are real and important topics, but they are also already priced into market consciousness. At the seed stage, we are more interested in the structural shifts that are less discussed but will generate outsized returns over the next decade.
1. The Vertical AI Platform Wave
The first wave of enterprise AI was predominantly horizontal — tools and platforms that could, in theory, be deployed across any industry vertical. ChatGPT, Copilot, and similar products fit this description. They are genuinely useful, widely adopted, and will generate substantial revenue. But the next wave — the one where we are making our highest-conviction bets — is vertical AI: companies that go deep on a specific industry, build proprietary domain data assets, and create AI systems that outperform general-purpose solutions by a significant margin on the problems that matter most to that industry.
Healthcare is the most advanced example. AI companies with proprietary clinical datasets, validated against real-world outcomes data, and integrated into existing clinical workflows are demonstrating accuracy rates that general AI models cannot match. Similar dynamics are playing out in legal (contract intelligence and legal research), financial services (underwriting and fraud detection), and manufacturing (quality control and predictive maintenance).
The defensibility of vertical AI companies is substantially higher than horizontal AI companies, because the data moats are industry-specific and take years to build. A well-capitalized competitor with access to general-purpose models cannot easily replicate five years of proprietary clinical data collection and validation. This is where we want to be investing.
2. The Agentic Turn
The second structural shift is the transition from AI-as-assistant to AI-as-agent. This distinction matters enormously. An AI assistant helps humans do tasks more efficiently. An AI agent completes tasks autonomously, often in ways that require fewer human touchpoints and create fundamentally different economic value.
We are still in the early days of the agentic turn, but the early indicators are compelling. Companies building agentic AI systems for back-office automation, customer service, code generation, and document processing are not just incrementally better than their predecessors — they are displacing entire workflow categories that previously required dedicated human teams. The total addressable market for agentic AI is not the software tool market; it is the global services market, which is orders of magnitude larger.
The investment implication: companies that can build reliable, auditable, human-in-the-loop-where-needed agentic systems for high-value workflows are building real businesses with strong unit economics. The enterprise buying motion for "autonomous agent that replaces ten FTEs" is qualitatively different from "AI tool that makes ten FTEs 20% more productive." We are investing accordingly.
3. The AI-Native Data Stack
The third shift is less visible to outsiders but deeply important: the entire data infrastructure stack is being rebuilt for AI workloads. The data architectures that worked well for business intelligence and analytics are not well-suited to the needs of machine learning pipelines, vector embeddings, retrieval-augmented generation systems, and real-time inference at scale.
This is creating enormous opportunity for seed-stage companies building AI-native data infrastructure. Vector databases, streaming pipelines optimized for ML feature stores, data quality platforms designed for training data validation, and governance systems built for AI model lineage are all in their early innings. The companies winning in these categories today have small revenue bases but are accumulating the relationships and product depth that will make them indispensable as enterprise AI adoption continues to accelerate.
What the Capital Concentration Means for Seed Investing
One of the most significant dynamics of the 2025 AI investment landscape is the extreme concentration of capital in a small number of large foundation model companies. OpenAI, Anthropic, xAI, and a handful of others have collectively raised hundreds of billions of dollars. This concentration has two important implications for seed investors.
First, it has dramatically reduced the cost and complexity of building AI applications. When state-of-the-art language understanding, code generation, and multimodal reasoning capabilities are available via API at declining cost curves, seed-stage founders can build products that would have required massive research infrastructure just three years ago. The capital efficiency of the application layer has improved dramatically as a result.
Second, the concentration of capital in foundation models means there is less competition from well-capitalized players for the best seed-stage application layer companies. The major foundation model companies are not competing with our portfolio companies — they are, in many cases, the infrastructure that our portfolio companies are building on. This is a more favorable competitive environment for early-stage investors than existed in, say, the cloud computing era, where AWS was simultaneously building cloud infrastructure and launching competing application products.
The Correction That Has Already Happened
It is worth noting that AI investment has not been uniformly frothy. The AI application layer saw a significant correction in mid-2024 as investors became more discriminating about AI companies that had achieved revenue growth primarily by riding the initial novelty of generative AI rather than by solving genuine enterprise problems. Companies with high customer churn, poor gross margins from API costs, and undifferentiated positioning from general-purpose models saw valuations reset sharply.
What survived — and what is now raising rounds on increasingly strong terms — are the companies that built genuine differentiation. Strong proprietary data. Workflow integration depth that creates switching costs. Domain expertise that translates into better model performance for specific use cases. These are the companies generating net dollar retention above 120%, maintaining gross margins above 70% despite AI infrastructure costs, and demonstrating that their AI-native products are measurably better than incumbent alternatives.
At Swarm Capital, the correction of 2024 was one of the best things that happened for our portfolio selection process. It raised the bar for what kind of AI company can survive and thrive, and it gave us clearer signal for what we should be backing with conviction at the seed stage in 2025.
Our Outlook for 2026 and Beyond
Looking into 2026, we are most excited about three investment themes that we believe will generate the best risk-adjusted returns for seed investors over a five-to-seven year horizon.
First, AI for regulated industries. Healthcare, financial services, legal, and government represent the largest addressable markets for AI but have historically been among the slowest to adopt new technology. The combination of dramatically improved model capabilities, declining regulatory uncertainty around AI governance frameworks, and the competitive pressure that incumbents now feel from AI-native competitors is accelerating adoption in ways that create real opportunity for category leaders.
Second, AI-enabled automation of knowledge work. The 40 million knowledge workers in the United States alone represent a massive market for AI tools that augment human expertise. We are not talking about replacing humans — we are talking about allowing one highly skilled human to do work that previously required five people, or to operate across domains of expertise that would previously have required a specialist. Companies building in this space with enterprise-grade security, compliance, and auditability are building large, durable businesses.
Third, AI infrastructure for trust and reliability. As AI moves into high-stakes enterprise and regulated applications, the need for AI systems that are auditable, explainable, and provably reliable creates entirely new infrastructure categories. Companies building evaluation frameworks, red-teaming tools, model monitoring systems, and compliance automation for AI are in the early innings of what will be a multi-billion dollar category.
For Founders: What This Means for Your Pitch
If you are a founder building an AI company and approaching the seed stage, the current environment has several practical implications for how you position your company and approach investors.
Lead with differentiation, not the technology. Every early-stage investor understands that large language models exist and that fine-tuning, RAG, and agentic systems are possible. The question is why your approach is better for the specific problem you are solving. What data do you have access to that competitors do not? What domain expertise does your team bring that improves model performance? What workflow integration creates switching costs?
Show evidence of enterprise pull. In 2025, the bar for what constitutes a compelling seed-stage AI company has risen. Investors want to see at least three to five paying enterprise customers or committed pilots with clear paths to expansion. Revenue-generating proof points matter more than product demos, no matter how impressive the demo. If you can show early evidence of strong net dollar retention — customers expanding their usage over time — that is one of the most powerful signals in the current market.
Be honest about your AI cost structure. Gross margin has become a first-order concern for AI company investors. If your product is primarily powered by API calls to third-party foundation models, you need a credible story for how gross margins improve over time — through fine-tuning to reduce inference costs, through the development of proprietary models for high-volume tasks, or through the layering of high-margin services and workflow features on top of the AI core. Investors who have been burned by AI companies with 40% gross margins are now asking these questions at the seed stage.
The opportunity in AI investing has never been greater. The technology is more powerful, more accessible, and more economically transformative than any previous technology wave. The correction of 2024 has cleared the deck of companies without genuine differentiation, leaving a cleaner competitive landscape for the founders who are solving real problems with real technological advantages. At Swarm Capital, we have never been more excited about the companies we are seeing at the seed stage. The moment is now — for the right founders with the right ideas.
