10 Insights from Anthropic CFO Krishna Rao on AI's Future, Compute Strategy, and Platform Dynamics

In a recent episode of Invest Like The Best, host Patrick O'Shaughnessy sat down with Krishna Rao, Chief Financial Officer of Anthropic, to discuss the financial and strategic underpinnings of frontier AI development. Rao offered a rare glimpse into how one of the industry's leading labs thinks about risk, resource allocation, and market positioning. From the 'cone of uncertainty' to the perennial debate between platforms and applications, here are the ten most important takeaways from their conversation.

1. The Cone of Uncertainty in AI Progress

Rao introduced the concept of a 'cone of uncertainty' to describe the widening range of possible outcomes as AI systems become more advanced. The further out you look, the less certain you can be about performance, capabilities, and even safety implications. This uncertainty isn't a bug—it's a feature of a technology that hasn't yet hit its limits. For CFOs like Rao, this means planning for multiple scenarios rather than a single forecast. Investments must be flexible, and financial buffers are essential to absorb shocks. The cone also influences how Anthropic allocates resources: more compute is reserved for experiments that could yield breakthrough insights, while proven paths get steady funding. The takeaway? Embrace the fog of innovation—it's where the frontier is born.

2. Allocating Compute as a Strategic Weapon

Compute isn't just an operational expense; it's the scarcest resource in AI. Rao explained that Anthropic treats compute allocation as a strategic decision that sits at the intersection of research goals, capital efficiency, and time-to-market. Not all models need the same level of compute. Some require massive clusters for training, while others need more modest setups for fine-tuning or inference. The key is to match compute type and amount to the expected return on insight rather than simply maximizing usage. Rao likened it to venture capital—invest compute in high-upside research, but also keep a reserve for incremental improvements. This dynamic allocation allows Anthropic to explore the 'cone of uncertainty' without overcommitting to any single path.

3. Returns on Frontier Intelligence Are Not Linear

One of the most provocative points Rao made was about the non-linear nature of returns from frontier AI. As models scale, the marginal benefit of additional compute can be unpredictable. Sometimes a 10x increase in compute yields a 2x performance gain; other times it unlocks entirely new capabilities. This makes it difficult to apply traditional ROI models. Anthropic therefore evaluates investments based on a portfolio of bets, expecting that some will fail and a few will generate outsized returns. Rao stressed that this is not a criticism of scaling laws but a recognition that the frontier is lumpy. Financial planning must account for those lumps—by maintaining liquidity and being willing to pivot when the cones start to narrow or widen.

4. Platform vs. Application: The Crucial Distinction

A recurring theme in the conversation was the difference between building a platform (like a foundation model) and building an application on top of it. Rao argued that Anthropic deliberately positions itself as a platform company, emphasizing the foundational model as the core product. This choice influences everything from compute allocation (more on training/generalization) to pricing (API access, not per‑user software). Platforms have higher upfront costs but can capture value across many applications. By contrast, application-layer companies face commoditization pressure if the underlying platform improves. Rao noted that Anthropic's strategy is to own the platform layer, while encouraging others to build applications—a classic 'picks and shovels' approach in the AI gold rush.

5. The Economics of Model Size and Performance

Rao delved into the trade‑offs between model size, inference cost, and performance. Larger models tend to be more capable but also more expensive to run. For many use cases, a smaller, more efficient model can deliver 90% of the performance at a fraction of the cost. Anthropic therefore offers multiple model sizes and variants, allowing customers to choose the right price‑performance point. Rao emphasized that this is not a compromise on quality but a rational response to the fact that the 'cone of uncertainty' applies to business adoption as much as to research. By providing a spectrum of models, Anthropic can capture demand from cost‑sensitive enterprises while still pushing the frontier with its largest systems.

6. Safety as a Competitive Moat

When asked about Anthropic's emphasis on AI safety, Rao framed it not just as a moral imperative but as a long‑term competitive advantage. In a world where trust is scarce, customers and regulators will gravitate toward companies that can demonstrate responsible development. Rao pointed to Anthropic's 'responsible scaling' framework, which dictates when and how to deploy increasingly capable models. This approach can slow down release cycles, but Rao argued that the trust earned outweighs the speed. He noted that enterprise buyers, in particular, are willing to pay a premium for models that come with clear safety guarantees. In effect, safety becomes part of the product—and a differentiator in a crowded market.

7. The Role of Enterprise Customers in Shaping Product

Anthropic's early enterprise customers have played a disproportionate role in steering product development. Rao explained that feedback from large organizations—especially in regulated industries—has forced the company to prioritize reliability, explainability, and customization. This has led to features like fine‑tuning controls, usage analytics, and compliance tooling. Rao noted that while the frontier models get the headlines, it's the enterprise‑ready packaging that generates recurring revenue. The company maintains a tight feedback loop between customer success and research teams, ensuring that the 'cone of uncertainty' on the commercial side is informed by real‑world usage. This customer‑backed approach reduces the risk of building features nobody wants.

8. Capital Efficiency in a Compute‑Intensive Market

Rao addressed the elephant in the room: the huge capital requirements of frontier AI. Anthropic's funding rounds are among the largest in tech history. But Rao insisted that the company is obsessed with capital efficiency—not just raising money, but making every dollar of compute work as hard as possible. This means rigorous internal cost accounting, negotiating favorable pricing with cloud providers, and investing in software optimizations that reduce waste. He also mentioned that Anthropic has built its own infrastructure stack to better control costs. The goal is to achieve 'frontier intelligence moats' without burning through capital recklessly. Rao's message to investors: we are not trying to spend the most—we are trying to spend the smartest.

9. The Network Effects of the Anthropic Ecosystem

Rao highlighted a subtle but powerful dynamic: as more developers and companies use Anthropic's models, the data and feedback loop improves the models themselves. This creates a kind of data network effect. Unlike traditional platforms where user count directly improves the product (e.g., social networks), here it's the diversity and quality of use cases that refine the model's capabilities. Rao explained that Anthropic actively encourages third‑party integrations and community contributions, knowing that each new application helps push the frontier. Over time, this makes it harder for competitors to catch up, because they lack the same breadth of training signals. The 'cone of uncertainty' narrows for Anthropic as its ecosystem expands.

10. Looking Ahead: The Next Frontier

Finally, Rao offered a glimpse into how Anthropic thinks about the next five years. He expects the pace of progress to remain rapid but uneven—the cone of uncertainty ensures surprises. The company is investing in research that goes beyond current scaling laws, including multi‑modal models, tool use, and long‑term memory. Financially, Rao anticipates that capital requirements will continue to grow, but he sees a path to profitability through enterprise subscriptions and API usage. The biggest unknown, he said, is how regulation will shape the market. Anthropic is actively engaging with policymakers to help shape rules that are supportive of innovation while protecting safety. For investors, Rao's advice: bet on the platform—and be ready for the cone to either expand or contract.

Conclusion: Navigating the Cone

Krishna Rao's conversation with Patrick O'Shaughnessy reveals that managing a frontier AI company is as much about financial strategy as it is about technical prowess. The 'cone of uncertainty' is not something to fear but to plan for with flexible compute allocation, diversified bets, and a clear platform vision. Anthropic's approach—emphasizing safety, capital efficiency, and enterprise readiness—offers a blueprint for other companies operating at the edge of what's possible. As AI continues to evolve, those who can navigate the cone will define the next generation of intelligent systems.