Krishna Rao - Anthropic's CFO on Compute, Scaling to $30B ARR, and the Returns to Frontier Intelligence - [Invest Like the Best, EP.471]

Anthropic's Strategic Approach to Compute Procurement and Allocation

Compute is the lifeblood of Anthropic's business, making procurement one of the most consequential decisions. Buying too much risks bankruptcy, while buying too little prevents serving customers and staying at the technological frontier. This challenge is magnified by the "cone of uncertainty."

The "cone of uncertainty" refers to the difficulty of predicting demand in an exponentially growing business. Small fluctuations in growth rates lead to vastly different outcomes over time, making it hard to forecast compute needs one to two years out, especially for humans who tend to think linearly.

Anthropic addresses this challenge through a highly disciplined and flexible compute strategy. They utilize three distinct chip platforms: Amazon's Trainium, Google's TPUs, and Nvidia's GPUs. These resources are used fungibly across model development, internal product acceleration, and serving customer demand.

The company has invested years into building an orchestration layer and custom compilers, optimizing each chip to its best purpose and achieving what they believe is the most efficient compute usage among frontier AI labs. This flexibility allows dynamic allocation and re-prioritization of resources based on ROI discussions.

If you buy too much compute, you go out of business. If you buy too little compute, you can't serve your customer and you're not at the frontier. Same thing.

Anthropic Continuously Improves Compute Efficiency in AI Models

While new AI models offer increased intelligence, similar to upgrading from a sedan to a sports car, they surprisingly achieve greater compute efficiency. Unlike the common car analogy where higher performance often means worse fuel economy, successive generations of models like Opus 4 to 4.7 are multiple times more efficient at processing tokens.

This continuous efficiency improvement provides a dual benefit. Firstly, customers receive more capable models for inference at potentially lower costs. Secondly, Anthropic's internal research and development, particularly reinforcement learning, becomes more efficient because it relies on inference within a sandbox.

The research team constantly drives these advancements, deploying dynamic efficiency improvements even between major model releases. This ensures that models are always becoming more efficient over time, optimizing both customer-facing capabilities and internal development workloads.

The returns to being at the frontier are really high.

Frontier AI Unlocks Significant Enterprise Growth

Staying at the forefront of AI intelligence, by immediately adopting new models like Opus 4.7 or GPT 5.5, yields significantly higher returns than relying on older, cheaper versions. The speaker notes that intelligence is multidimensional, going beyond benchmark scores to include enhanced real-world capabilities.

Newer models improve the ability to handle long-horizon tasks, utilize tools, and complete agentic tasks much faster. For instance, an employee completing a task in a day versus a week can be seven times more effective. These efficiencies translate directly into significant business value.

For enterprises, consistently releasing new models unlocks Total Addressable Market (TAM) in unique ways and enables more use cases. The speaker's company saw its run rate revenue increase from $9 billion to over $30 billion in just four months, directly attributing this exponential growth to these leaps in model intelligence.

Enterprise customers are highly motivated to invest heavily in newer models and more tokens because they see concrete benefits. Each new model generation allows them to do more, better, and more efficiently. This cycle reinforces the core thesis that the returns to frontier intelligence are not slowing down, especially in the enterprise sector.

Especially in enterprise, the returns to frontier intelligence are not slowing down.

Anthropic's Models Accelerate Their Own Development and Augment Human Talent

Anthropic heavily invests in both compute resources and skilled talent to harness its own AI models, such as Claude Code, for internal development. This long-standing strategy has significantly accelerated the company's progress, including the rapid creation of new products and features.

The company reported 30 product and feature releases in January alone, a pace directly attributed to leveraging their advanced models in conjunction with human expertise. This synergy enables faster access to underlying intelligence and streamlines development cycles.

Despite the growing involvement of AI in its own creation, Anthropic primarily operates as a research lab. Human researchers lead experiments that push the boundaries of model capabilities, while the models assist in the process.

The core philosophy emphasizes that top human talent is essential for setting strategic direction, guiding new discoveries, and effectively managing the AI development lifecycle. This collaboration ultimately accentuates and accelerates the capabilities of their existing research and engineering teams.

I think of it as accentuating and accelerating the talent that we already have.

Unyielding AI scaling laws demand exponential business adaptation

Anthropic internally validates AI scaling laws by tracking models across development phases, from pre-training loss curves to post-training reinforcement learning, and crucially, by integrating customer feedback. They use customer pain points as direct training targets, confirming that current observations indicate no slowdown in the underlying scaling laws.

The continuous, exponential growth of AI capability makes traditional linear business planning, like quarterly forecasts, obsolete. Companies must instead embrace dynamic scenario planning and maintain a very low bar for updating their current perspectives, as what was true even a month ago might no longer hold.

This rapid evolution means businesses need to constantly re-evaluate market opportunities. For instance, Anthropic observed a significant jump in coding capabilities with Sonnet 3.5 and 3.6, which subsequently drove adoption, usage, and revenue. This pattern now serves as an analog to anticipate similar shifts in other parts of the economy and business.

But from what we see, the scaling laws are not slowing down.

Anthropic's Strategic Compute Acquisition and Dynamic Deployment

Anthropic actively seeks both immediate and future compute capacity through diverse strategic partnerships. This includes a near-term collaboration with SpaceX for its Colossus facility in Memphis and significant long-term commitments, such as a five-gigawatt deal with Google and Broadcom for TPUs starting in 2027, and another multi-gigawatt, over $100 billion commitment with Amazon for Trainium, with much of this compute landing soon.

The company employs a dynamic assessment process to evaluate compute opportunities. This involves analyzing price-performance over time, deployment timelines, location, the specific type of compute, and how efficiently it can be utilized internally. This systematic approach applies to both short-term acquisitions and large-scale, long-term commitments.

Anthropic evaluates various chip platforms, including different generations of TPUs and Trainium, each with unique price-performance characteristics. The deployment strategy is highly granular, matching specific compute types—like CPUs for reinforcement learning or leading-edge hardware for the fastest model training—to particular use cases and internal needs, driven by customer demand.

Despite a perceived shortage across the industry, Anthropic has developed the capability to rapidly consume and deploy a wide array of compute resources. While integrating heterogeneous compute was once challenging, the company can now spin up and deploy almost any type of compute very quickly across various use cases, reflecting increased operational fungibility.

It's starting at different times with different capabilities, and we're very dynamically comparing that compute. It's price performance over time, that's really, really important to us, when it lands and what we think we can do with it internally in the business.

Anthropic balances foundational AI platform development with strategic application building

Anthropic primarily operates as a platform provider, akin to early AWS, focusing on offering a reasoning engine and intelligence through its Claude models. This strategy aims to allow customers to build significant value on top of their platform, benefiting from tools like prompt caching, virtual machines, and the Claude agents SDK. The company believes this horizontal approach will accrue substantial value to both Anthropic and its partners.

Despite its platform-first focus, Anthropic also develops its own applications on the same platform. These internal applications serve two main purposes: demonstrating the future capabilities of their models (like Claude Code, which was initially ahead of model capabilities) and showcasing value to the ecosystem that others might emulate. Examples include vertical-specific solutions for financial services, life sciences, and security.

The company emphasizes a collaborative, partner-oriented approach for these vertical applications, working with the ecosystem rather than competing directly. This strategy allows Anthropic to add value and perspective in specific areas, ensuring that the platform's intelligence can proliferate effectively across various customer products and services.

The rapid pace of AI development, with capabilities evolving in months rather than years, can sometimes create apprehension among potential customers who fear competition from Anthropic. The company addresses this by engaging in early access programs, collaborating closely with customers, and actively listening to their desired capabilities, aiming to accelerate their partners' success.

most of what we're building is platform. We think that there's so many examples of where a platform Platform can accrue a lot of value, but the customers who are building on that platform actually accrue even more value.

Anthropic balances stable pricing for accessibility with robust returns on compute investment.

Anthropic maintains a stable pricing strategy for its models like Haiku, Sonnet, and Opus, a counter-intuitive approach given the industry trend of rising compute costs and high demand. The company's goal is to make its intelligence accessible to a wide range of businesses, from startups to large enterprises, fostering broad ecosystem proliferation.

This strategy was exemplified by a price reduction for the Opus family of models. When Opus was found to be underutilized relative to its capabilities, lowering its price made it more accessible. This led to a significant increase in consumption, demonstrating the Jevons Paradox, where increased efficiency or lower cost leads to greater demand and overall usage.

Anthropic's financial model focuses on the robust return across its entire 'compute envelope,' rather than just the incremental variable cost of serving individual tokens. This holistic view encompasses all compute-intensive activities, including inference, long-term model development, and internal acceleration for new products, all of which contribute to current and future revenue. This allows them to balance delivering significant value to customers with achieving strong returns on their substantial compute investments.

We think of it as what is the return on that full envelope of compute? We feel really good about where we are from that perspective, and we're balancing delivering value to customers with also seeing a really, really strong return on that compute ourselves.

Anthropic cultivates deep, multifaceted partnerships with major cloud and chip vendors

Anthropic maintains extensive collaborations with major hyperscalers like Amazon, Google, and Microsoft, alongside chip vendors such as Broadcom and NVIDIA. The company uniquely operates its models across all three major cloud providers and leverages their diverse chip platforms.

These partnerships extend beyond simple procurement. For example, Anthropic's teams are deeply embedded with Amazon's Annapurna Labs, working closely on Trainium. This involves joint capacity planning and optimizing compute utilization.

The multifaceted nature of these relationships encompasses developing core technologies, securing compute capacity, serving models, and distributing them to customers. Each partner contributes unique strengths, enabling Anthropic to effectively utilize different aspects across the entire ecosystem.

Our teams are deeply embedded with the Annapurna Labs team. We are good users of Trainium, we've spent a lot of time and energy and worked closely with the team, and that's something where we plan capacity together.

Anthropic's finance team extensively uses Claude for automation and strategic insights

Anthropic's finance team leverages Claude Code as a "digital coworker" for a wide range of tasks, moving beyond traditional coding applications. This internal usecase helped shape the development of more general co-worker AI capabilities, extending agentic software development to all knowledge work.

Specifically, Claude is used to produce statutory financial statements for all legal entities, which are then human-checked. The team also uses a real-time platform called Ant Stats, featuring over seventy finance-specific skills in a common repository. This allows for rapid data analysis and report generation, shifting focus from sifting through data to understanding drivers and implications.

A key application is the Monthly Financial Review (MFR) skill, which generates reports that are 90-95% ready. This dramatically reduces the time spent on reconciling numbers and instead allows the team to focus on strategic implications, such as reinvesting in the business or dynamically allocating resources. What once took hours to produce a weekly report now takes 30 minutes, speeding up insight delivery to business leaders.

Notably, senior members of the finance team, including the head of tax, are among the biggest users of Claude, applying it to automate tax policy engines and other complex workloads. This demonstrates that AI tools are transforming how experienced professionals work, not just entry-level staff, leading to increased overall productivity and enabling the company to tackle more work with highly talented individuals.

If we're not super users of this, if we're not pushing the limits of it, how can you expect customers to do that?

Anthropic's Exponential Growth and Evolving Investor Understanding

Anthropic faced skepticism from investors early on, particularly during its Series D fundraising two years ago. Questions arose about the viability of combining AI safety with building a large business, the returns on frontier models, and the perceived small size of their sales force compared to traditional enterprise software companies. Investors struggled to fit Anthropic into existing paradigms.

The company consistently exceeded investor expectations, proving the high returns to frontier intelligence. By the end of 2024, Anthropic's business scaled to nearly a billion dollars in run rate revenue. Their unique approach of investing in AI safety research, like interpretability and alignment science, created a downstream benefit: it built trust with enterprise customers, including nine of the Fortune ten, who increasingly rely on Anthropic for sensitive workloads.

Internally, Anthropic's revenue predictions consistently outpaced initial linear forecasts, shifting perceptions towards an exponential growth mindset. While initial growth targets like reaching a billion dollars seemed to defy 'laws of physics' and 'law of large numbers' arguments for enterprise adoption, the business demonstrated that rapid customer movement and underlying exponentials made such growth possible.

A key concept for investors to grasp is the fungibility of compute at Anthropic. Unlike traditional businesses where R&D and COGS are distinct, Anthropic can repurpose compute resources daily for both model development and inference. This unique flexibility maximizes the return on massive compute investments, driving both short-term and long-term revenue, though it challenges conventional cost separation paradigms.

We've raised seventy-five billion dollars since I joined the company. We have another fifty billion dollars that'll come in in the future from the Amazon and Google deals that we, that we inked last month.

Addressing AI's Public Perception and Governance Challenges

The AI industry faces a significant public perception issue, with the technology's generic concept being less popular than Congress among the general populace. This indicates a challenge in effectively communicating AI's value to those outside the tech industry who do not yet grasp its potential benefits.

Despite this, the industry remains highly optimistic about AI's transformative capabilities. Examples include accelerating drug development for both common and rare diseases, enhancing healthcare delivery, and raising the standard of living in developing nations. A key goal is to ensure economic gains from AI are distributed globally, not concentrated among a few.

For the public to trust and embrace AI, the industry must adopt a balanced communication strategy. This means openly articulating both the immense opportunities and the inevitable risks and "bumps on the road." An honest assessment, rather than solely positive messaging, is crucial for building credibility.

Addressing these challenges requires proactive collaboration between commercial entities and governments to develop comprehensive solutions and frameworks. This ongoing dialogue about potential downsides and mitigation strategies, combined with transparency, is essential for responsible AI development and deployment.

AI, just the generic concept, being less popular than Congress amongst the general populace... when you really think about it, we need to solve this problem.

Anthropic Adopts Phased Release for Highly Capable Mythos Model

The release of Anthropic's Mythos model marked a significant moment, with many observers expressing concern about its advanced capabilities, particularly in cybersecurity. This sentiment highlighted the public's growing anxieties about powerful AI and prompted Anthropic to publicly address its responsible development.

Krishna Rao clarified that Mythos is incredibly capable across various dimensions, but showed a significant spike in its cybersecurity performance. For instance, while a prior model identified 22 security vulnerabilities in an open-source codebase, Mythos uncovered 250. This dual capacity, both beneficial and potentially harmful, necessitated a careful approach to its release.

In response to these advanced capabilities, Anthropic adopted a phased release strategy for Mythos, consistent with its mission and principles. The company focused on ensuring the model's use for positive, defensive applications, such as identifying and patching code vulnerabilities, rather than for offensive purposes. This strategic rollout is considered a potential template for future highly capable AI models.

Furthermore, Anthropic acknowledges the increasing role of external oversight, including government regulation. The company prioritizes building strong relationships with regulatory bodies, believing that regulation has a critical part to play in navigating the responsible deployment of cutting-edge AI technologies.

This one kind of makes me scared.

Anthropic's Distinctive Culture Emphasizes Collaboration, Transparency, and Mission

Anthropic's unique culture is largely shaped by its seven co-founders, who serve as direct examples. The company takes cultural fit extremely seriously, conducting a rigorous culture interview process. Even highly intelligent and skilled candidates will not be hired if they fail to meet this cultural bar.

The culture fosters an incredibly collaborative environment, actively discouraging individual 'fiefdoms' or the need to take sole credit. A strong sense of humility prevails, with an internal reminder that 'our competitors are incredibly capable and success is far from guaranteed,' which drives a focus on continuous progress rather than excessive celebration.

Intellectual honesty and rigorous debate are central to the company's operations. Employees are encouraged to question ideas and express diverse viewpoints, leading to productive dialogue and eventual collective alignment on decisions. Transparency is also paramount, with CEO Dario Amodei regularly sharing insights and openly answering unfiltered questions from employees.

This distinctive and transparent culture, coupled with a strong mission alignment, is a primary reason Anthropic has been able to attract and retain top talent in the industry. The vast majority of its co-founders and early employees remain with the company, showcasing its ability to compete for skilled individuals even when not always offering the highest compensation packages.

We take culture extremely seriously. We do a culture interview, and it's not some pro forma thing we do to check a box. Somebody could be flying colors on everything else and really the smartest person you've met in this role, we won't hire them if they don't pass the cultural bar.

Anthropic's Vision for Virtual AI Collaborators

Anthropic's next frontier in AI development focuses on creating "virtual collaborators" designed to revolutionize enterprise knowledge work. These AI agents are envisioned to possess deep contextual understanding within an organization, integrate with specific tools—whether homegrown or purchased—and have robust memory capabilities to learn from past interactions and mistakes.

The goal is to enable these virtual collaborators to work on ideas and projects over very long time horizons, moving beyond simple task completion. This approach aims to significantly accelerate productivity in knowledge work, a sector estimated to be worth $40 trillion annually worldwide.

The company believes this requires not just generically smart AI, but intelligence tailored to specific use cases. Success in the coding domain with tools like Claude Code and CoWork, which has seen rapid adoption, serves as a blueprint for how these virtual collaborators could impact broader knowledge work. CoWork, in particular, has demonstrated even faster growth than Claude Code did at a comparable stage.

I think it is towards this vision or this goal of like a virtual collaborator. And so think of this as something that has context within your organization, that can use all of the tools that are specific to you... that has memory and the ability to effectively learn from mistakes... The ability to work over a very long time horizon on not just a task, but an actual idea.

Personal Evolution for Leaders in Hyper-Growth

Krishna Rao discusses the personal challenge of scaling leadership in a company experiencing unprecedented, exponential growth. He highlights how traditional product development models are being replaced by a 'fleet of agents' that necessitates everyone becoming a manager, pushing leaders beyond linear thinking to adapt to rapid, often daily, changes.

To navigate this environment, Rao emphasizes thinking in first principles and maintaining intellectual openness. He recounts how an early vision for the company, shared by Chief Compute Officer Tom Brown, initially sounded 'crazy' but largely materialized, illustrating the need to embrace radical, forward-looking ideas.

Rao's leadership strategy also involves hiring 'partners' rather than direct reports. He seeks individuals from diverse backgrounds, including hyperscalers and private equity, who will challenge his thinking and bring unique perspectives. He values their ability to operate at different levels of detail, filling the gaps where he cannot be at a 'five hundred feet' view for everything.

Additionally, he finds parallels in past high-pressure situations, such as leading Airbnb's financing during the pandemic, to inform decision-making in rapidly changing scenarios. Despite the demands, he makes a point to regularly acknowledge the unique opportunity and appreciate working with his team on such a significant problem.

I'm not really hiring you as like a direct report of mine. I'm hiring you as a partner, and I want you to treat it as a partnership.

Identifying potential risks to the pace of AI progress

Krishna Rao identifies potential factors that could slow the rapid pace of AI development and adoption. One significant risk is the challenge of integrating AI capabilities into large organizations. Human teams within these entities often operate with long-standing tools and practices, making rapid change difficult.

This organizational inertia means that real-world use cases and the practical application of AI might struggle to keep pace with the swift advancements in AI models themselves. If the spread and integration of these new technologies hit a wall or significantly decelerate, it could directly impact the revenue growth associated with AI.

Another critical risk mentioned is the potential for scaling laws to slow down. These foundational principles have driven much of the current progress in AI, and any deceleration here could have far-reaching implications for future capabilities.

The capability and the use cases are playing catch up to where the model is.

AI's Potential to Revolutionize Biotechnology and Healthcare

Krishna Rao expresses strong optimism regarding AI's profound potential to transform outcomes in biotechnology and healthcare. He envisions a future where the technology significantly accelerates drug discovery and development, leading to faster cures for diseases that are currently considered untreatable.

The application of AI and R solutions is already streamlining the drug development process. These tools are helping to rapidly complete time-consuming tasks like paperwork and clinical study reports, thereby reducing the overall timeline for bringing new treatments to market.

Beyond administrative efficiencies, the most significant impact of AI is expected in the deeper stages of drug discovery. Human researchers often struggle with the intricate complexity of molecules and proteins, where even minute structural changes can have substantial implications. AI is positioned to manage this complexity, enabling more effective and rapid identification of potential new drugs.

We may live in a world where you're diagnosed with a disease that is not curable, but in your lifetime that cure can be found much more rapidly and you might not die of that disease.