The Agent Arbitrage: What 2025 VC Actually Funds

The top 10 AI deals of 2025 totaled $84 billion. Follow that money to its edge and you find something the headlines miss: the agent layer is where value crystallizes.

Cursor—an AI code editor—raised $2.3B at a $29.3B valuation. It doesn't train foundation models. It wraps them in an execution layer that makes developers 10x faster. The market values execution capability, not model capability.

This is the arbitrage hiding in plain sight: $84B poured into infrastructure creates a value capture vacuum. Foundation models are converging toward commodity. The agent layer—software that executes, not converses—is where differentiation lives. We've been tracking this thesis for months. The 2025 funding data confirms it.

The Infrastructure Paradox

The numbers suggest infrastructure dominance:

Every company in the $5B+ club is infrastructure. Not one is an application. This looks like dominance. It's actually vulnerability—the same production gap that separates demos from deployed systems.

Meta's $14.3B Scale AI stake reveals the anxiety. When model builders vertically integrate into data infrastructure, they're admitting the model layer alone isn't defensible. You have to own the supply chain to survive.

Where Value Actually Lives

Compare the infrastructure giants to what's happening at the execution layer:

Cursor raised $2.3B at a 12.7x revenue multiple—comparable to the best SaaS companies at peak ZIRP. But unlike SaaS, Cursor's value compounds with model improvements it doesn't pay for. As Claude and GPT get better, Cursor's product gets better. The infrastructure layer absorbs the R&D cost; the agent layer captures the value. This is agent economics in action—cost per completed task drops while capability rises.

The Energy Constraint Creates the Agent Opportunity

For the first time in modern venture history, nuclear startups appeared in top-tier funding: X-Energy ($700M), TerraPower ($650M), Base Power ($1B). This isn't a climate story. It's an AI constraint story.

When Nvidia's venture arm invests in a nuclear company, the thesis writes itself: compute is energy-constrained. This constraint creates the agent opportunity.

The infrastructure layer's energy constraint is the agent layer's moat. Build agents that accomplish tasks efficiently, and energy constraints work in your favor.

Defense Tech: Agents With Guns

Defense tech's emergence as a core venture pillar isn't separate from the agent thesis. It's the clearest proof of it. Anduril, Helsing, and Shield AI are building agents that execute in the physical world—autonomous systems that make decisions and take actions without human-in-the-loop.

Anduril's "Arsenal" initiative aims to mass-produce autonomous drones and interceptors. Shield AI's "Hivemind" software enables aircraft to operate without GPS or communications—making decisions in contested environments. These aren't chatbots. They're agents that execute under constraints where latency kills.

The pattern extends to robotics. Figure AI ($39B valuation) and Physical Intelligence ($5.6B) are building agents that execute in physical space—humanoid robots for warehouses and factories. The common thread: software that does, not software that talks.

The Execution Spectrum

The funding data reveals a spectrum of agent companies, from digital to physical:

Every company in this chart wraps foundation models in execution capability:

• Cursor ($29.3B valuation): AI that writes code. Developers describe intent; Cursor executes.
• Figure AI ($39B valuation): Humanoid robots for warehouses. General-purpose physical execution.
• Wayve ($1.05B raise): Embodied AI for autonomous driving. SoftBank-backed.
• Sierra ($4B+ valuation): Enterprise agents that handle customer service end-to-end.

Conversational AI is table stakes. Execution AI commands the premium. The funding data proves it.

DeepSeek's Efficiency Challenge

One data point challenges the infrastructure-heavy thesis: DeepSeek.

The Chinese lab reportedly operates on a fraction of the capital its US competitors consume. DeepSeek-R1 matches frontier model performance while training on allegedly fewer resources. If true, this undermines the "capital wall" narrative.

The regional picture reinforces this tension:

The US bet is capital intensity. China is betting efficiency can win. Europe is betting on specialized verticals. The agent layer benefits regardless of which strategy dominates—agents are model-agnostic.

The Operational Takeaway

The 2025 funding data suggests a clear strategy for startups that don't have $10B:

The stratification is real. The top 100 rounds captured the majority of capital. The long tail of SaaS startups face a liquidity drought. But the agent arbitrage creates a path forward for companies that can't raise $10B—and don't need to.

Appendix: Top 25 Rounds of 2025