This Issue at a Glance
Krishna Rao (CFO of Anthropic) delves into the podcast on the decision-making logic behind computing power as the company's "lifeline." He introduces the concept of the "cone of uncertainty" and reveals that Anthropic achieves interchangeable use across three chip platforms—Trainium, TPUs, and GPUs—dynamically allocating computing power to model development, internal use, and customer demand through daily meetings. Core judgment: The returns on frontier intelligence continue to rise, especially in the enterprise sector—Anthropic pushed its annualized revenue from $9 billion to $30 billion in Q1 2026, a leap driven precisely by the intelligence transition of its models.
Compute Allocation: Three Platforms Interchangeable, Dynamically Allocated Daily
Krishna Rao argues that compute is Anthropic's "lifeline" and "canvas," and the company achieves the highest compute utilization efficiency in the industry through the interchangeable use of three chip platforms.
- Three Platforms: Anthropic simultaneously uses Amazon's Trainium, Google's TPUs, and NVIDIA's GPUs, and has spent years building an "orchestration layer" that enables different chips to be used interchangeably. Rao notes: "We are likely the most efficient compute users among frontier labs, and this did not happen overnight."
- Allocation Mechanism: The company holds daily compute allocation meetings to dynamically allocate resources among three uses—model development (with a minimum guarantee that is never breached, even if it impacts customer service), internal use (to accelerate product and model development), and customer service. Rao emphasizes: "This is not a zero-sum game, but a highly collaborative culture."
- Flexibility Advantage: Different generations of chips (e.g., TPU V5E, V6, V7; Trainium 2, 3) are assigned to the most suitable workloads. Rao states: "We can use one chip for inference in the morning and switch it to model development in the afternoon or evening—this is impossible in traditional software companies or factories."
Frontier AI Returns: Enterprise Adoption Continues to Climb, Model Capabilities Shift Across Multiple Dimensions
Rao points out that returns on frontier AI continue to rise in the enterprise sector, and model capabilities are multidimensional—encompassing not only IQ scores but also long-horizon tasks, tool use, and agentic capabilities.
- Revenue Surge: Anthropic pushed its annualized revenue from $9 billion to $30 billion in Q1 2026. Rao explains: "This shift is driven by both leaps in model intelligence and the products built around them."
- Multidimensional Intelligence: Rao argues that model intelligence is not a single IQ score. "We’ve found that many benchmarks are already saturated. The real measure is what customers tell us—how models perform in real-world settings."
- Accelerated Enterprise Adoption: Anthropic’s net dollar retention rate exceeds 500% (annualized), with nine of the Fortune 10 as clients. Rao reveals: "On the Uber ride here, I signed two eight-figure commitments in just 20 minutes."
Scaling Laws Remain Effective, Model Efficiency Also Improves
Rao confirms that scaling laws are "still alive and well" within Anthropic, while model efficiency continues to improve, creating a "win-win" cycle.
- Scaling laws have not slowed down: Despite maintaining a "scientific method-style skepticism" within Anthropic, Rao states: "From what we have seen, scaling laws have not slowed down." Several of the company's founders are authors of the scaling laws paper.
- Efficiency improvements: From Opus 4 to 4.5, 4.6, and then 4.7, each new generation of models not only achieves capability leaps but also multiples the efficiency of processing tokens. Rao explains: "This is not just about serving customers; it also accelerates our internal reinforcement learning—because RL is essentially sandbox reasoning with a reward function."
- Recursive self-improvement: Over 90% of Anthropic's internal code is written by Claude Code, and Claude Code's own code is also written by Claude Code. Rao notes: "The model itself is helping us build the next generation of models."
Pricing Strategy: Staying Stable, Leveraging the Jevons Paradox
Rao explains why Anthropic chose to keep pricing stable rather than continuously raising prices—by lowering prices to trigger the Jevons Paradox, consumption grows far beyond expectations.
- Stable Pricing: Anthropic has kept pricing for the Haiku, Sonnet, and Opus series almost unchanged, with the only major adjustment being a reduction in the price of Opus 4.5. Rao explains: "We found that Opus-class models were underutilized relative to their capabilities. People were trying to cram Opus-level problems into Sonnet."
- Jevons Paradox: After lowering the price of Opus, consumption growth far exceeded expectations. Rao says: "We lowered the price, but consumption grew far beyond expectations. Because we found the sweet spot for customers, enabling them to use it more."
- Marginal Thinking: Rao emphasizes that Anthropic does not price based on the "variable cost" mindset of traditional software companies. "Compute supports all activities—inference supports today's revenue, while model development may unlock TAM six months down the line. We measure the return on the entire compute package."
Internal Application: Claude Reshapes the Finance Team Workflow
Rao shared how Anthropic's finance team became "power users" of Claude, compressing monthly financial reviews from hours to 30 minutes.
- Automated Financial Reporting: Statutory financial statements for all legal entities are generated by Claude (with human review). Rao stated: "Claude doesn't just report the weather; it helps think through the drivers—why the numbers are changing the way they are."
- Skill Library: The finance team has built over 70 Claude skills, accessible via a public repository. The Monthly Financial Review (MFR) skill produces 90%-95% of the first draft, shifting team discussions from "what happened" to "what should we do."
- Usage Patterns: The most senior team members are the largest token users. Rao noted: "Our tax director is the number one user, focused on automating the tax policy engine. If we aren't power users ourselves, how can we expect our customers to be?"
Capital Formation: From Linear to Exponential Thinking
Rao reviews Anthropic’s fundraising journey, noting that the hardest concept for investors to grasp is the "fungible use" of compute—a paradigm that does not exist in traditional enterprises.
- Fundraising journey: From Series D (2024, when the company only had frontier models) to Series E (end of 2024, with the first closing completed on the day of the DeepSeek news), and most recently signing compute commitments exceeding $100 billion with Google and Amazon. Since joining the company, Rao has raised $75 billion, with an additional $50 billion expected from the Amazon and Google deals.
- Investor misunderstanding: Rao believes the hardest concept to explain is that "compute is not a variable cost, but a fungible resource." He says: "In a traditional company, you can’t have the R&D team serve as COGS, or vice versa. Here, you really have that fungibility."
- Mindset shift: Rao acknowledges his own transition from linear to exponential thinking. "Dario has always been more accurate than me in predicting revenue. The first time I saw 10x growth, my mind was full of debates about the 'laws of physics' and the 'law of large numbers.' But when you see how the business works internally, see the adoption curve and exponential growth, you start to believe."
Culture: Collaborative, Transparent, Mission-Driven
Rao describes Anthropic’s unique culture—seven co-founders remain at the company, the vast majority of the first 20–30 employees are still there, and cultural interviews are a hard requirement.
- Cultural traits: Highly collaborative (no tolerance for "territorial behavior" or "taking credit"), humble (internal laptop stickers read "Our competitors are very strong, and success is far from guaranteed"), and transparent (Dario gives company-wide updates every two weeks and takes unfiltered questions).
- Talent retention: When companies like Meta poached with higher compensation, Anthropic lost only 2 people, while other labs lost dozens. Rao says: "People want to work in a truly collaborative place, not where they have to fight for something."
- "Talent density over talent quantity": Rao emphasizes this philosophy, arguing that the best models combined with the densest concentration of AI research talent and reasoning engineering talent form a "winning combination."
Risks and Outlook
Rao lists three risks that could push the company toward the lower end of the "cone of uncertainty," while expressing the most optimistic expectations for the biotechnology sector.
- Three risks: 1) Slower enterprise adoption diffusion; 2) Unexpected deceleration in scaling laws; 3) Failure to stay at the competitive frontier.
- Most optimistic area: Biotechnology and healthcare. "We may live in a world where you are diagnosed with an incurable disease, yet a cure is found within your lifetime. AI is a perfect fit in drug discovery—molecules and proteins are so complex that small changes have massive impacts. When laboratory throughput increases tenfold or a hundredfold, it can reshape the global healthcare landscape."
Mentioned Positions
| Position |
Analyst View |
Key Data |
| Amazon (Trainium) |
Bullish (deep partner) |
Signed up to 5GW Trainium computing capacity agreement; team collaborates deeply with Annapurna Labs |
| Google (TPUs) |
Bullish (deep partner) |
Signed 5GW TPU agreement (in partnership with Broadcom), deliveries starting in 2027 |
| NVIDIA (GPUs) |
Neutral (one of three platforms) |
Used as one of the three chip platforms |
| Broadcom |
Bullish (partner) |
Participates in TPU chip collaboration |
| xAI |
Neutral (short-term partnership) |
Collaborates using the Colossus facility in Memphis |
| Meta |
Mentioned (competitor) |
Previously poached talent with higher compensation; Anthropic lost only 2 employees |
| OpenAI |
Mentioned (competitor) |
Referenced as a frontier lab |
| DeepSeek |
Mentioned (market event) |
News about DeepSeek triggered market volatility on the day of Series E first closing |
Judgments Worth Remembering
1. "Compute is not a variable cost, but a fungible resource" (Krishna Rao) — Anthropic can switch the same chip from inference to model development within the same day. This flexibility, absent in traditional software companies or factories, is the core paradigm most difficult for investors to grasp.
2. "Returns on frontier intelligence continue to rise in the enterprise sector, with no slowdown" (Krishna Rao) — Q1 annualized revenue surged from $9 billion to $30 billion, net dollar retention exceeded 500% (annualized), and 9 of the Fortune 10 are customers, proving enterprise adoption has moved from "pilots" to "large-scale deployment."
3. "Scaling laws have not slowed down" (Krishna Rao) — Despite maintaining scientific skepticism internally, actual data continues to validate the effectiveness of scaling laws, while model efficiency is also improving in tandem (Opus 4→4.5→4.6→4.7, each generation delivers multiple-fold improvements in token processing efficiency).
4. "Lowering prices triggers Jevons paradox — consumption growth far exceeds expectations" (Krishna Rao) — After reducing the price of Opus 4.5, customer usage surged, proving that the core of pricing strategy is to deliver sufficient value to customers, not to maximize short-term profits.
5. "Over 90% of code is written by Claude Code, and Claude Code's own code is also written by Claude Code" (Krishna Rao) — Recursive self-improvement has already occurred in practice; models are helping build the next generation of models. This is the fundamental reason Anthropic sets a "minimum guarantee line for model development" in its internal compute allocation.
6. "Talent density beats talent quantity" (Krishna Rao) — When companies like Meta poached talent with higher compensation, Anthropic lost only 2 people, while other labs lost dozens. The core lies in a collaborative culture and mission-driven ethos.
7. "We may live in a world where you are diagnosed with an incurable disease, but a cure will be found within your lifetime" (Krishna Rao) — Biotechnology and healthcare are the areas Rao is most optimistic about. AI is a "perfect fit" in drug discovery, and a 10x or 100x increase in lab throughput could reshape the global healthcare landscape.
8. "Uncertainty cone" (a framework proposed by Krishna Rao) — In exponentially growing businesses, tiny changes in monthly or weekly growth rates lead to vastly different compounding outcomes. Anthropic addresses this through scenario analysis (rather than point estimates) and low-threshold prior updates, while maintaining flexibility in compute procurement to cover the upper end of the cone.