Navigating the Cone of Uncertainty: Anthropic CFO Krishna Rao on AI Compute, Platform Strategy, and Frontier Returns

In a recent episode of Invest Like The Best, Patrick O’Shaughnessy sat down with Krishna Rao, Chief Financial Officer of Anthropic, to explore the financial and strategic dimensions of frontier AI. Rao offered a rare glimpse into how one of the world’s leading AI labs thinks about uncertainty, resource allocation, and the long‐term value of intelligence. This article distills the key insights from that conversation, covering the cone of uncertainty, compute allocation, returns to frontier intelligence, and the platform vs. application debate.

The Cone of Uncertainty in AI Development

Rao introduced the concept of a “cone of uncertainty” to describe the wide range of possible outcomes in AI progress. Unlike traditional technology roadmaps, where the next steps are relatively predictable, frontier AI is marked by fundamental unknowns. “We don’t know exactly when we’ll hit certain capabilities, or what the economic impact will be,” Rao explained. This cone widens as we look further into the future, but narrows near term based on observable trends in model scaling and hardware improvements.

This uncertainty has profound implications for financial planning. Anthropic must balance aggressive investment in compute and talent with the reality that returns may come in bursts rather than smooth curves. Rao emphasized that the company uses probabilistic scenario planning—modeling best‐case, base‐case, and worst‐case scenarios—to allocate capital without betting the entire company on a single outcome.

Why Uncertainty Is Actually a Good Thing

Rather than seeing uncertainty as a risk to be minimized, Rao framed it as a source of opportunity. “If everything were predictable, there would be no edge,” he noted. The cone of uncertainty means that early movers who can navigate ambiguity can capture outsized returns. Anthropic’s strategy is to build a robust financial foundation that allows it to stay in the game long enough for the cone to narrow in their favor.

Allocating Compute: The New Strategic Resource

Compute—the raw processing power used to train and run large models—has become the most scarce and valuable resource in AI. Rao described it as “the new oil” and explained that allocating compute effectively is perhaps the single most important financial decision at Anthropic.

The allocation process involves multiple dimensions:

• Training vs. Inference: Deciding how much compute goes into training the next generation of models versus serving existing models to customers. Rao noted that the balance shifts rapidly as models improve and new use cases emerge.
• Research vs. Product: Compute must be split between fundamental research (exploring novel architectures) and product development (ensuring reliability and safety).
• Short‐term vs. Long‐term: Short‐term compute can generate immediate revenue, but long‐term compute builds the capabilities that will define the next frontier.

Rao likened the process to venture capital portfolio management: “We make many small bets to inform a few large ones.” Anthropic uses internal metrics—such as “compute efficiency” (improvement per FLOP) and “capability growth per dollar”—to guide these decisions.

Returns to Frontier Intelligence

One of the most debated topics in AI investing is whether the enormous costs of frontier model development will yield proportionate returns. Rao argued that history suggests the opposite: “The returns to the most advanced intelligence have always been superlinear, and we see no reason that will change.”

He pointed to three sources of returns:

1. Direct Revenue: Licensing models and API access already generate significant income, and Rao expects this to grow as enterprises embed AI into core operations.
2. Platform Effects: The most capable models attract the best developers and data, creating a virtuous cycle that compounds advantages.
3. Societal Value: Many applications—from scientific discovery to personalized education—are public goods that will eventually be monetized through broader economic growth.

Rao cautioned against short‐term thinking: “If you only look at next quarter’s margin, you miss the fact that the next leap in capability could unlock entirely new markets.” He cited the example of GPT‐3, which created a new ecosystem of developers that didn’t exist before.

Platform vs. Application: The Strategic Choice

A critical strategic decision for any AI company is whether to focus on being a platform (providing the underlying infrastructure and models) or an application (building end‐user products). Rao explained that Anthropic has deliberately chosen the platform route.

“We believe the greatest value will be captured by the layer that becomes the default operating system for AI,” he said. By offering powerful, safe models through an API, Anthropic enables thousands of applications to be built on top, each solving specific problems. This approach has several advantages:

• Scaling: One platform can serve millions of users across countless use cases, creating massive economies of scale.
• Safety: Concentrating expertise in a few highly capable models allows Anthropic to invest heavily in alignment and safety research—something that would be diluted across many applications.
• Flexibility: As the technology evolves, a platform can pivot faster than applications that are tied to specific user interfaces.

However, Rao acknowledged that the application layer is where many users first encounter AI. Anthropic supports a vibrant ecosystem of partners who build applications using its models, and the company periodically launches its own apps (like Claude) to demonstrate capabilities and gather feedback. “We don’t see platform and application as mutually exclusive,” he noted. “But our core bet is that the platform will be the enduring source of value.”

Implications for Investors

For those watching the AI sector, Rao offered a lens through which to evaluate different companies. Platform players need massive upfront capital for compute and research, but they can achieve high margins at scale. Application players may reach users faster but risk being commoditized if the platform layer becomes dominant. “The cone of uncertainty applies here too,” Rao concluded. “The only way to win is to place a big, informed bet and then adapt as the cone narrows.”

Conclusion

Krishna Rao’s interview paints a picture of a company that is thoughtful about both the enormous potential and the genuine risks of frontier AI. By embracing the cone of uncertainty, treating compute as a strategic asset, believing in superlinear returns, and choosing a platform strategy, Anthropic is positioning itself for the long haul. For anyone interested in the financial logic of AI, this conversation provides a masterclass in balancing vision with pragmatism.