Artificial Intelligence

The AI funding moment is driven by three intersecting forces: unprecedented capital concentration, the commoditization of foundation models, and the urgent need to find differentiation at the application layer. Understanding these forces is essential for founders deciding whether to build, pivot, or raise in this landscape.

The market itself is massive and accelerating. According to Grand View Research, the global AI market was valued at approximately $391B in 2025 and is projected to reach $3.5T by 2033, representing a compound annual growth rate (CAGR) of 30.6%. This isn't theoretical growth—it's being funded right now. The OECD reported that AI startups captured 61% of global VC in 2025, totaling $258.7B out of $427.1B in total venture capital. That reallocation from "AI is interesting" to "AI is the default category" happened in a single year. For context, in 2024, AI accounted for roughly 45% of VC dollars. The shift was dramatic and structural.

Mega deals are the story within the story. Deals above $1B represent roughly half of total AI investment value by dollars, even though they represent a small fraction by count. Mega-rounds—Anthropic's $13B, xAI's $10B, SoftBank's $40B commitment into OpenAI, Mistral AI's $1.5B—account for 73% of total AI investment value. This creates a funnel effect. While 2,911 pre-seed funders and 3,725 seed funders exist, only 403 Series C funders and a subset of 818 growth equity firms can deploy into later-stage AI. The implication: if your AI company doesn't reach escape velocity with infrastructure (models, compute, data) or a defensible application layer by Series B, you're competing for capital in an increasingly crowded mid-market. The mega-deal infrastructure is separate from the venture ecosystem—it's a different game with different players.

Infrastructure captured $109.3B of the 2025 total, making it the largest subsector by capital despite attracting fewer funders (23 infrastructure-specific investors in our database). This concentration reveals the calculus: building foundational models and inference infrastructure requires venture-scale capital but attracts only specialized investors. OpenAI, Anthropic, xAI, and a handful of others have captured outsized share. Meanwhile, 30 funders back Generative AI companies—the layer above infrastructure. The pricing power, compute efficiency, and differentiation live at the application layer for most founders, not in building new foundation models. The exception is vertical-specific foundational models (medical imaging, code generation), where specialized data and domain knowledge matter.

The generative AI wave is now three years into maturity, which changes the investment thesis. The 2022-2023 period was marked by "anything with GPT in it gets funded." By 2025, that changed. Founders now face questions: Are you building a feature or a product? Can you achieve unit economics with current model costs? What's the switching cost for your customer? The early-stage AI universe (pre-seed and seed) is still capital-rich, with 6,636 funders combined, but series A and beyond have gotten selective. Check sizes remain robust ($250K median floor, $2.375M median ceiling), but conviction has tightened. Founders are no longer betting on AI adoption—that's happened. They're betting on specific applications with defensible advantages.

Enterprise adoption is the funding catalyst most often overlooked. AI infrastructure spending is happening inside enterprises right now—not as future plans. Companies are burning through consumption-based pricing on API calls, evaluating fine-tuning ROI, standing up new infrastructure teams, and building agent systems for internal automation. That spending doesn't always flow to AI startups; it flows to cloud providers and the public models. But the friction points—data integration, model customization, compliance, cost optimization—are where specialized startups extract value. This is why Vertical AI (industry-specific models, compliance-aware systems) is underfunded relative to opportunity. Only 4 funders focus on Vertical AI, but the revenue potential is enormous because enterprises will pay for specialization.

Talent is the second-order funding constraint. The pool of engineers who can train models, optimize inference, or build complex multi-agent systems is tiny relative to capital available. Salaries at frontier labs (OpenAI, Anthropic, Google DeepMind) have created a talent vacuum. Founders building outside the mega-labs face a different talent market than other sectors—they're competing globally against well-capitalized incumbents. This tilts advantage toward founders with research pedigree, existing networks in academic labs, or the ability to offer meaningful equity upside. Some of the strongest AI teams have formed around departures from foundation model companies, but those windows are temporary. The implication for founders: your competitive moat is partially who's on your team, not just what you build.

Compute economics are non-negotiable. Inference costs for large language models have fallen 90% in two years but remain material for consumer-grade applications. Training costs for new models are astronomical—measured in millions to tens of millions for competitive proprietary models. This creates a bifurcation. Startups that can (a) fine-tune public models, (b) use smaller specialized models, or (c) operate on-device are viable. Startups that require training new models from scratch must either raise massive capital or be acquired by someone with compute. The winner's curve favors leverage: use free or cheap infrastructure (open-source models, cloud APIs), optimize for specific use cases, and build revenue fast. The median check size of $250K to $2.375M makes sense in this context—it's enough to find product-market fit without requiring you to solve foundational research.

Open source versus closed source is reshaping the game. Meta's Llama models, Mistral, and others have democratized access to capable base models. This is deflationary for anyone building language-model-only businesses but inflationary for anyone building infrastructure around those models (specialized fine-tuning platforms, inference optimization, safety layers). Startups built on top of closed APIs (OpenAI, Anthropic) face supply-side risk—the API provider can always undercut your pricing or launch a product that replaces you. Startups leveraging open-source models benefit from commodity economics but must differentiate on integration, domain knowledge, or operational efficiency. Our data shows 10 funders focus on Machine Learning broadly (not generative AI specifically), suggesting that ML infrastructure and MLOps remain underfunded relative to consumer-facing AI.

Global competition is intensifying. The US dominates with approximately 75% of global AI VC, but that share is shrinking. The EU27 captures 6%, the UK 5%, and China 5% of global AI investment, though China's official figures may understate state-coordinated funding. The UK, despite Brexit friction, is building a serious AI hub (222 funders in our database). India (183 funders) is emerging as both a training ground for AI engineering talent and an application layer for cost-sensitive use cases. The geographic spread in Superscout's data (1,246 funders in Europe, 869 in Asia, 1,003 global players) shows capital is flowing beyond the US, but concentration remains. For founders outside the US, the path to top-tier capital often still requires a US base or US-facing product.

The agent paradigm shift is the frontier. Most 2024 AI funding went to retrieval-augmented generation (RAG) and prompt-engineering tools. By 2025, the thesis moved to autonomous agents—systems that can plan, use tools, learn from feedback, and operate with minimal human intervention. Our data flags only 6 funders with AI Agents as a core focus, despite agents being discussed in every board meeting. This gap exists because agent economics are still unclear. The value proposition (less human labor) is clear; the business model isn't. Agents that work well might be cheap to run (leading to commodity pricing), or they might require domain expertise to deploy (creating service-attached business models). Founders exploring agents are funding the research phase right now, betting that standards will emerge (like ReAct or similar planning frameworks) and that enterprises will pay for agents with proven ROI on specific tasks. Early traction in autonomous customer service and supply chain optimization is real, but scale is still proving out.

Regulatory risk is factored into later-stage checks but often overlooked in early-stage decision-making. The EU AI Act came into force in 2025, creating mandatory compliance for high-risk AI applications (hiring, credit decisions, autonomous systems). The US lacks comprehensive federal AI legislation but has sector-specific rules (FDA for medical AI, FTC for consumer AI). China is tightening controls on generative AI and LLM deployment. For founders, regulation is either a moat (if you build compliance infrastructure) or a tax (if you need to comply). Regulatory fragmentation favors centralized, well-capitalized teams that can navigate multiple jurisdictions. Smaller teams often underestimate this cost. The best founders are already thinking about how AI Act compliance or US executive order requirements become product features, not just legal overhead.

The next 18-24 months will clarify which bets pan out. Foundation model scaling (whether LLMs or multimodal systems) continues to follow Chinchilla-style scaling laws, but at exponential cost. Diminishing returns in benchmark performance are already visible; the next leap requires either new architectures (sparse models, mixture-of-experts) or new training methodologies (agent-based learning, world models). Application-layer winners will be companies that (a) found defensible use cases where AI genuinely changes unit economics or customer experience, (b) locked in customers with switching costs (data, integration, workflow), and (c) built revenue fast. The founders raising at Series A and beyond in 2025-2026 will face a tighter, more disciplined investor base. Series C and growth equity investors are asking hard questions about unit economics, customer concentration, and defensibility. The cheap capital days (2023-2024) have ended. The good capital days are just beginning.