Krishna Rao - Anthropic's CFO on Compute, Scaling to $30B ARR, and the Returns to Frontier Intelligence - [Invest Like the Best, EP.472]

Invest Like the Best with Patrick O'Shaughnessy · May 13, 2026

5 Selected Lessons:

case · 2 framework · 1 heuristic · 1 reasoning_pattern · 1

Cone of uncertainty for exponential compute planning

framework based on: lived experience Krishna Rao

Source

Invest Like the Best with Patrick O'Shaughnessy — Krishna Rao - Anthropic's CFO on Compute, Scaling to $30B ARR, and the Returns to Frontier Intelligence - [Invest Like the Best, EP.472] (May 13, 2026) · at 2:27 ▸ · at 7:25 ▸

When this applies

when a business is growing exponentially and must commit to large, hard-to-reverse resource purchases years before demand is known

The takeaway

When committing to scarce, lead-time-heavy resources under exponential uncertainty, model a range of future scenarios and work backward from the distribution rather than a point forecast, and bias toward flexibility to cover the gap between forecast and reality.

Why it works

Small differences in weekly/monthly growth rates compound into vastly different outcomes over 1-2 years, so a single point forecast is almost certainly wrong; planning for a range of scenarios and working backward lets you size commitments to survive most of that range.

One level up

Under genuine exponential uncertainty, the discipline is in the process (scenario range + flexibility), not in the precision of any single forecast.

Where it stops holding

This stops working when growth is actually linear/predictable or when the resource being procured has short lead times and can be adjusted cheaply and quickly, making heavy scenario planning unnecessary overhead.

Detail

procedure:

Model bottoms-up what demand will likely be
Estimate compute needed to stay at the frontier
Look at a range of scenarios across a 1-2 year cone rather than a single forecast
Work backward from desired position in that cone to today's purchasing decisions
Build flexibility into deal terms and into how compute is used across workloads
Bias toward the top end of outcomes while planning for a wide range

benchmarks: {"time_horizon":"1 to 2 years","time_allocation":"Rao spends 30-40% of his time on compute"}

rationale: Buy too much compute and you go out of business; buy too little and you can't serve customers or stay at the frontier — both are existential, so disciplined scenario planning plus flexibility is the only way to survive the exponential cone.

In their words

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.

spoken at 2:27 ▸

Interpretability/safety research as a dual-use investment

reasoning_pattern based on: lived experience Krishna Rao

Source

When this applies

when a company must decide whether to fund a mission-driven investment that has no obvious near-term commercial payoff

The takeaway

A mission-driven investment made 'for its own sake' can create unplanned commercial leverage (e.g. trust, product quality) that becomes a competitive advantage — look for and recognize these downstream linkages rather than assuming mission and commercial goals trade off.

Why it works

Investing in understanding a product more deeply (interpretability) both serves a stated mission and improves product development quality; it also builds the trust required for risk-averse enterprise buyers to adopt the product for sensitive workloads, creating a commercial flywheel that wasn't the original motivation.

One level up

Investments justified on values grounds should still be scrutinized for hidden commercial upside, because the two are not always zero-sum.

Where it stops holding

This stops working when the mission-driven investment has no plausible link to product quality or customer trust, in which case it remains a pure cost rather than a flywheel.

Detail

approach: Fund interpretability/alignment research for mission reasons, but recognize and communicate its downstream effects on both model-building quality and customer trust for sensitive enterprise workloads.

signals:

ability to look inside the model (interpretability) improves model-building itself
enterprise customers entrusting sensitive data require demonstrable trustworthiness
selling to 9 of Fortune 10 requires trust beyond raw capability

rationale: Investors initially saw safety and scale as being 'at odds'; over time the business proved the opposite — safety investment supported both better models and enterprise trust needed to close large deals.

In their words

Those things are important for our mission, and that's why we did them. They had these downstream effects where it turns out if you can look inside the model, you're better at building them.

spoken at 41:30 ▸

Jevons paradox pricing: lower price to unlock volume

heuristic based on: lived experience Krishna Rao

Source

When this applies

when a premium product is underutilized relative to its capability because of price friction

The takeaway

If a premium capability is underused because of price, cutting price to the point where customers fully adopt it can raise total consumption/revenue far more than the discount, i.e. expect a Jevons-paradox effect.

Why it works

Lowering price below the threshold where customers previously avoided the premium tier causes consumption to expand far more than the price cut, because usage was being artificially suppressed, not because demand was inherently limited.

One level up

Pricing decisions should target unlocking latent demand, not just preserving margin on existing usage.

Where it stops holding

This stops working when demand is inelastic or already saturated, so a price cut simply reduces revenue without materially increasing volume.

Detail

rule: When a premium-tier product is underutilized relative to its capability because customers are trying to fit a bigger problem into a cheaper tier, cut the premium price to unlock adoption.

threshold: Price change made once (Opus 4.5 launch); subsequent model upgrade (Opus 4.6) shipped at same price, showing pricing stability is itself valued

rationale: Consumption went up 'way, way more than what you would have expected' after the price cut, confirming the Jevons paradox effect and justifying price stability afterward.

In their words

Changing the pricing for Opus, you see this Jevons paradox. We lowered the price of it, but the consumption went up way, way more than what you would have expected.

spoken at 30:42 ▸

Series D fundraising amid skepticism and the FTX liquidation

case based on: lived experience Krishna Rao

Source

When this applies

when raising capital for an unprecedented business model during a period of external turmoil and investor skepticism about the category itself

The takeaway

When raising capital for a category-defining business, expect investors to misapply familiar frameworks from adjacent industries, and expect each funding round to be stress-tested by external shocks that force a full re-underwriting — persistence through this and the business proving out its own growth pattern is what builds investor conviction over time.

Why it works

Investors tried to fit a novel business into prior frameworks (enterprise SaaS sales models, 'mission vs scale' tradeoffs) and were wrong because the business's actual growth mechanics (model-led growth) didn't match those templates; each subsequent funding round the thesis got re-tested by new shocks (DeepSeek news) and kept being validated by continued growth.

One level up

Categories without precedent are judged against borrowed mental models first, and are only re-rated once the business's own track record supplies a new one.

Where it stops holding

This stops working when the business's growth is in fact well-modeled by an existing template (e.g. ordinary SaaS), in which case investor skepticism about novelty is simply correct.

Detail

situation: Series D closing occurred during the FTX share liquidation and with only a nascent frontier model; investors doubted the mission-vs-scale tradeoff and small sales force. Series E closed the same day DeepSeek news broke, prompting re-underwriting fears about AI broadly.

decision: Anthropic continued raising and let the growing business (model-led growth, enterprise adoption) serve as the counter-argument rather than trying to force-fit prior frameworks.

outcome: Both rounds closed successfully; subsequent growth (run-rate revenue from ~$250M to $30B over time) continued to validate the thesis that returns to frontier intelligence are high, despite persistent early skepticism.

lesson: Novel business models get judged by borrowed frameworks first; sustained execution and growth are what eventually re-rate investor understanding.

In their words

The business continued to prove out the thesis that the return to frontier intelligence is really high.

spoken at 41:30 ▸

Phased release of Mythos due to elevated cyber-capability risk

case based on: lived experience Krishna Rao

Source

When this applies

when launching a powerful new capability/product that has a specific, unusually high dual-use risk in one dimension

The takeaway

When a new capability has an unusually elevated risk concentrated in one specific dimension, release it to an expanding, controlled group rather than either withholding it entirely or releasing it broadly immediately.

Why it works

Rather than withholding or fully releasing the capability, restricting initial access to an expanding group lets the organization observe and manage defensive-vs-offensive use while still capturing the model's broader positive value, buying time to build mitigations before full-scale exposure.

One level up

Manage asymmetric risk by controlling exposure surface and pace, not by binary go/no-go decisions.

Where it stops holding

This stops working when the risk is diffuse rather than concentrated in a specific, identifiable capability, making a targeted phased release impossible to design around.

Detail

situation: Mythos was found to be far more capable than prior models at finding security vulnerabilities (250 vs. 22 for a prior model on the same open-source codebase), raising concern about offensive cyber use.

decision: Anthropic released it via a phased approach to an expanding group, focused on demonstrating defensive (patching) use rather than halting release entirely.

outcome: Described as a template the company believes could be used for future releases with similarly concentrated risk profiles.

lesson: A capability spike in one specific risk dimension warrants a phased, controlled release rather than a binary release/withhold decision.

In their words

We have this phased approach to it because we think that when a model is this capable, and again, cyber is the thing that people focus on, but there are other things as well.

spoken at 56:17 ▸