Martin Casado on Where the Value Is Going in AI

Martin Casado frames the current AI moment around a single binary question: either the frontier labs win everything, or they do not. He describes AI as the biggest wealth unlock he has seen in his career, larger than anything he anticipated even after the 1990s, and argues that the traditional venture capital belief that the asset class cannot deploy large amounts of capital is being disproven in real time. The clearest evidence is that small teams can now productively deploy enormous amounts of capital and convert it directly into capability and growth faster than any prior technology cycle. He cites a widely used multimodal model built by roughly 20 people at a cost of approximately two billion dollars as unprecedented in the history of engineering, while noting that whether putting ten dollars in returns nine dollars on the other side remains uncertain.

On lab dominance, Casado presents data showing labs have held roughly 95 percent of the market measured by dollar-weighted revenue across three years of generative AI. OpenAI and Anthropic combined have raised approximately 220 billion dollars, more than the entire downstream ecosystem combined. Labs maintain pricing power by staying even a small margin ahead on the frontier, because frontier models operate in competitive equilibria where being slightly better captures disproportionate share. Auto catalytic effects, meaning using AI to build or improve AI faster, improve unit economics for frontier labs, though Casado distinguishes this from recursive self-improvement in the literal sense, which he says is not currently occurring. Labs also own the majority of GPU supply during a period of constrained availability, which he expects to ease around 2028.

The argument against total lab dominance is that the surface area of AI applications is expanding rapidly beyond code and language reasoning, many domains require a services arm and closer customer relationships, open source models are performing well, and a maturing ecosystem is forming around serving them. His predictions are that dollar-weighted, large labs will capture roughly 80 percent of the model market going forward, while token-weighted, roughly 60 percent of usage will go to long-tail and open source models. He also expects increasing value to accrue to application layer companies, which will erode some margin share from labs.

Casado explains Open Router as a two-sided marketplace aggregating access to many models under a single API with analytics and visibility, and describes it as the brand leader in model aggregation. He argues that models are stickier than commonly assumed, partly due to procurement dynamics like pre-purchased credits. The primary practical gain from smart routing today is cost reduction while holding quality constant rather than selecting the qualitatively best model for a given question. Determining which model will best answer a specific question is itself an AI-complete problem, making routing circular, and when a new frontier model launches it tends to be Pareto efficient across prior models, which would concentrate routing to a single model anyway. He frames Open Router's core value as demand aggregation on the two-sided marketplace, not routing arbitrage. He explains the Stripe acquisition of Open Router through cultural alignment, arguing that both companies treat tokens and payments as forms of value within a market-oriented worldview.

Casado describes the Cursor acquisition by xAI as what he believes is the largest private merger and acquisition deal ever for an independent venture-backed startup outside of Elon Musk's X transaction, valued at approximately 60 billion dollars. He says Cursor represents the fastest growth he has seen in ten years of investing and twenty years in the valley. The strategic logic is that Cursor brings data and distribution while xAI brings compute, and coding is widely viewed as the path to AGI or broader computer user intelligence. Casado argues that Cursor succeeded by treating software engineering improvement as a product problem rather than a model architecture problem, which he sees as a rare and deliberate choice, contrasting it with research-heavy firms that struggle to engage core teams on anything outside new model architecture.

On subsidized token access, Casado says labs subsidize developer usage because enterprise is where margin is actually made, and free tier access functions as a low-cost customer acquisition channel. Only the top five percent of free tier users tend to generate losses, and platforms use policy levers to control that cohort. He also describes sophisticated operations out of China that arbitrage subscription AI plans by signing up, draining all tokens within days, canceling, and recouping prorated refunds to resell access cheaply.

For infrastructure investing, Casado says the primary form of inquiry is founder market fit, meaning the intersection of a specific founder and a specific market, rather than evaluating either in isolation. He notes that the path a founder takes to a startup gives them specific sensitivities and earned knowledge that others lack. He advises founders and investors to avoid zero-sum thinking about moats and defensibility in the near term and instead focus on what is strategically important in the world being created, noting that value is accruing across all layers of the stack simultaneously, including Nvidia, model companies, app companies, inference companies, services companies, and media companies.