Invest Like the Best · May 13, 2026

Krishna RaoThe Cone of Uncertainty

Most CFOs manage costs. Krishna manages the canvas. In a rare act of candor, Anthropic's CFO pulls back the curtain on the decisions that actually drive one of the fastest-growing businesses in history — how they buy and allocate compute, why the old financial frameworks break completely at the frontier, and what it felt like to watch the business go from $9B to $30B run-rate in a single quarter. He also shares the walk with his chief compute officer that made him go home and tell his wife: "If even 10% of this is true, it will bend all paradigms." One of the most honest windows into frontier AI economics you'll find anywhere.

$9B→$30BRun-rate ARR, one quarter
500%+Net dollar retention
9 of 10Fortune 10 customers
$125B+Committed compute
90%+Code written by Claude Code
01

Key Frameworks & Mental Models

The Cone of Uncertainty
Don't build a single forecast — build a range of exponential scenarios across a 1–2 year horizon. Small movements in weekly growth rates compound into wildly different outcomes. Work backwards from each point in the cone to determine what resources you need to stay solvent and at the frontier regardless of where reality lands. The goal isn't precision; it's survivability across the full distribution of futures.
So what: Any exponentially growing business planning to a point estimate is almost certainly either over- or under-resourced. The right question isn't "what will happen?" — it's "what position do we need to occupy for any plausible outcome?"
Planning · Forecasting · Capital
Return on Compute — Not Gross Margin
Compute is not a cost to minimize. It's the single asset generating return across multiple time horizons simultaneously: inference revenue today, model capability in 6 months, internal productivity now. The same chip runs customer inference in the morning and model development in the afternoon. Traditional COGS/R&D separation completely misframes AI lab economics. The right metric is return on the full compute envelope.
So what: When evaluating AI labs, skip gross margin — it's the wrong frame. Ask: what's the ROC across all workloads, how is compute efficiency trending generation-over-generation, and is the gap between compute spend and revenue output widening or narrowing?
Investor Lens · Unit Economics
Jevons Paradox — Applied Twice
Lowering the price of a resource increases its consumption far beyond what you'd expect, expanding TAM rather than compressing margin. Krishna applies this in two places: (1) Cutting Opus pricing unlocked a new consumption band — usage exploded so much that net revenue impact was positive. (2) AI-assisted employees don't do less work — they surface more high-value work, because throughput multiplies what's possible for the whole team.
So what: In exponential-growth categories, pricing for margin capture is usually the wrong move. Find the price point where consumption explodes — the resulting usage depth, customer stickiness, and TAM expansion typically create more enterprise value than margin preserved at a higher price.
Pricing · Labor · TAM
Intelligence is Multidimensional, Not a Score
Benchmark leaderboards are saturated and easily gamed. Real model intelligence has distinct axes: long-horizon task completion, tool use, agentic speed, domain accuracy, reliability under load. Two models with equivalent "IQ" but different task completion speeds deliver radically different economic value — the faster one can be 7x more productive on the same assignment. Anthropic measures model value by what customers report, not leaderboard position.
So what: When comparing models for enterprise deployment, don't ask "which scored higher?" Ask: which dimensions matter for your workload, and which model optimizes those? Speed and reliability often matter more than raw intelligence for production use cases.
Product · Benchmarking · Enterprise
Compute Fungibility as Structural Moat
Most frontier labs are locked into one chip vendor. Anthropic runs Trainium, TPUs, and GPUs interchangeably — shifting workloads across all three intraday based on price-performance and workload type. This took years of compiler work and deep co-engineering with chip teams to build. The result: a dollar of compute goes further inside Anthropic than at any other frontier lab, by their own assessment. Flexibility that took years to build is not easily replicated overnight.
So what: Infrastructure flexibility is a moat that doesn't show up in any standard financial metric but shows up everywhere in the P&L over time. The companies that invest in it early create compounding cost advantages that become nearly impossible to close later.
Moat · Infrastructure · Strategy
Model-Led Growth (MLG)
Revenue acceleration at Anthropic is caused by model capability leaps, not by adding salespeople. Each new model generation unlocks new use cases, which pulls enterprise adoption forward. The GTM team accelerates the cycle but doesn't initiate it. The Sonnet 3.5/3.6 coding leap is the clearest example: capability threshold crossed, adoption followed, revenue followed. Anthropic now uses coding as the "analog" to predict what comes next in other domains.
So what: For AI companies, the most important leading indicator of revenue is model capability unlock, not pipeline or sales headcount. The right investor question: what new use cases does the latest model make economically viable that weren't before? That's the next revenue cohort.
GTM · Growth · Enterprise
Talent Density Beats Talent Mass
A tight team of exceptional people, empowered with the best tools, consistently outcompetes a larger team with average tooling. Headcount growth should be a consequence of productivity gains — not a prerequisite. Anthropic deliberately forgoes billions in potential inference revenue to accelerate employees with internal compute, treating workforce productivity as a compound investment. The best people want to work with other best people — density creates its own gravity.
So what: Before adding headcount, ask: have we made every existing person maximally productive? The leverage on a new hire is far higher in a high-productivity environment. Hiring into a low-productivity org to solve output problems typically makes the problem worse.
Org Design · Culture · Leverage
Platform Horizontal, Vertical Only Where You Have Unique Insight
Anthropic's default is platform — enabling the ecosystem to build. They go vertical in two cases only: (1) to demonstrate a model capability the market doesn't yet believe exists (Claude Code proving what an AI-native coding env looks like), or (2) to show how to compose the platform in a new way. The goal is never to compete with customers — it's to prove the platform's potential and let customers capture most of the resulting value.
So what: The platform vs. app tension at AI labs is real but often misframed. The question isn't "will they compete with me?" — it's "do they have unique insight that makes the platform more valuable if they demonstrate it first?" Fear of the former causes builders to miss the upside of the latter.
Product Strategy · Ecosystem
02

How Anthropic Allocates Compute

Model
Development
→
Better
Models
→
Efficient
Inference
→
More
Revenue
→
More
Compute
→
Internal
Productivity
↩
The self-reinforcing compute flywheel — every node feeds every other node, across different time horizons simultaneously
Patrick's hypothetical
"If I airdropped 2x, 5x, 10x the compute on you tomorrow — how fast would you consume it?" Krishna's answer: a year or two ago, absorbing heterogeneous compute quickly would have been very hard — different chip platforms have real idiosyncrasies. Today, Anthropic would deploy almost any type of compute nearly immediately across all three workload buckets. That shift — from months-to-deploy to near-instant absorption — is itself a measure of how much their infrastructure capability has matured.
Bucket 01
Model Development
Has a non-negotiable floor — protected even if it means serving customers less. The logic: returns to frontier intelligence are so high and compounding that short-term revenue sacrifice is always worth a long-term capability lead. This is where recursive self-improvement happens. 90%+ of Anthropic's own code is now written by Claude Code.
Bucket 02
Internal Use & Productivity
Employees using Claude accelerates model development itself — they find efficiency multipliers, ship products faster, and generate feedback that becomes training signal. Forgoing billions in potential inference revenue for this bucket is a deliberate long-term bet. The head of tax, not a junior engineer, is the top token consumer on the finance team.
Bucket 03
Customer Inference
The direct revenue driver today. Dynamically allocated against the other two buckets based on demand signals. Each model generation improves inference efficiency, meaning more revenue can be served per dollar of compute — compounding this bucket's return without proportional cost increases.
Chip Platforms & Deals
  • Amazon Trainium 2 & 3 — up to 5 GW committed, starting 2027
  • Google TPUs v5e/v6/v7 + Broadcom — 5 GW deal signed last month
  • NVIDIA GPUs — multiple generations used concurrently
  • SpaceX Colossus (Memphis) — near-term burst capacity for consumer/prosumer
  • Total committed: $125B+. Workloads shift across all three intraday.
Procurement Philosophy
  • Flexibility built into every contract — not just the strategy
  • Near-term spot deals layered on long-dated commitments ("layer cake")
  • Every deal evaluated on: price-performance, duration, location, workload fit
  • Deep co-engineering with Annapurna Labs to influence chip roadmaps directly
  • Capital raised because of the cone of uncertainty, not to fund operating losses
03

Business & Fundraising Timeline

March 2023
First dollar of revenue
The company had been operating for roughly two years as a pure research lab. Three years later, during this episode, this date is referenced as a reminder of how compressed the entire commercial arc has been.
Early 2024 — Series D
~$250M run-rate ARR · Krishna joins
Not a clean fundraise. FTX was liquidating Anthropic shares mid-process. Core investor objections: Can you build a real business alongside an AI safety mission? Where's the enterprise sales force? Krishna, still thinking linearly, asked Dario "in what year?" when he heard the $1B ARR target.
End of 2024 — Series E
~$1B run-rate ARR · DeepSeek drops on close day
Extreme fundraising volatility. DeepSeek news hit the same day as first close. Investors still skeptical — pointing to how long cloud and enterprise software adoption cycles take. The model-led growth thesis was still being proven in real time.
January 2025
$9B run-rate ARR · 30 product releases in January alone
Revenue base already beyond most multi-year investor projections. Pace of product releases accelerating — enabled by AI-assisted internal development. The flywheel between model capability, internal tooling, and product output is now clearly visible.
End of Q1 2025
$30B+ run-rate ARR — 3x in a single quarter
Driven by model capability leaps across the Opus 4 family and Claude Code enterprise adoption. Net dollar retention exceeds 500% annualized. Nine of the Fortune 10 are now customers. On the day of this recording, Krishna was signing double-digit million-dollar commits in a 20-minute Uber ride.
April–May 2025
$125B+ in committed compute · SpaceX Colossus announced
Amazon (Trainium) and Google/Broadcom (TPU) deals totaling over $100B committed. SpaceX Colossus in Memphis added for near-term consumer capacity. $75B raised since Krishna joined; another $50B incoming. Capital raised primarily because of the cone of uncertainty — the returns on compute today are already robust.
04

Key Lessons & Takeaways

1
Breaking linear instincts is a two-year project, not an insight
Krishna is a sophisticated finance professional who spent two years consciously rewiring how he forecasts. The shift isn't intellectual — it's behavioral. You have to catch yourself anchoring to last quarter's number and ask instead: what does this mean for TAM if capability continues to compound? Dario has consistently been a better revenue predictor than Krishna because Dario never stopped thinking exponentially.
Operator + Investor
2
Safety investment and commercial moat are the same bet
Interpretability research — the "MRI for models" — made Anthropic better at building models, not just safer ones. Alignment science made enterprises comfortable handing over their most sensitive workflows. The mission-driven investment came first; the commercial differentiation was downstream. Most investors initially didn't believe the linkage. Nine of the Fortune 10 are the answer to that skepticism.
Investor lens — competitive differentiation in enterprise
3
Use your own product harder than any customer — starting with leadership
90%+ of Anthropic's code is written by Claude Code. The finance team produces statutory financials and monthly financial reviews with Claude. The head of tax — not the 22-year-old engineer — is the top token consumer. If you're not a super-user internally, you can't credibly sell externally, can't find real product edges, and can't give accurate feedback on where the model falls short.
Operator lens — internal AI adoption
4
Culture is the highest-leverage retention tool — and it compounds
When Meta deployed massive packages to poach AI talent, Anthropic lost two people. Other labs lost dozens. All seven co-founders are still at the company. The vast majority of the first 30 employees are still there. The mechanism: rigorous culture interviews that can veto otherwise-perfect candidates, intellectual honesty as a real operating norm, zero tolerance for fiefdoms, and radical transparency — Dario takes live, unplanted questions from the whole company every two weeks.
Operator lens — retention & org design
5
Hire partners, not reports — especially when surface area is expanding faster than you can track
Krishna explicitly tells candidates: "I'm not hiring you as a direct report, I'm hiring you as a partner — which means I expect you to push back." In a business where the surface area is expanding faster than any one person can track, the quality of your team's independent judgment is the only scalable resource. Hiring for deference is a delayed failure mode that becomes visible at the worst possible time.
Operator lens — leadership at scale
6
Have a very low bar for updating your priors
In a normal business, quarterly forecasts and annual plans are reasonable anchors. In an exponential business, something true last month can simply stop being true. The operating posture Krishna describes: constant scenario re-evaluation, extreme willingness to discard last month's model, and pattern-matching within your own business history to predict what comes next. Anchoring to prior beliefs is the fastest path to being caught flat-footed at the frontier.
Operator + Investor
7
The three questions to ask any AI lab
Speaking from the CFO seat, Krishna names the questions that actually matter: (1) What is your return on compute, all up? (2) How are your customers actually experiencing ROI — are these real deployments or pilots? (3) Where does your compute come from over the next 3 years, and what's your flexibility if the market shifts? If a company can't answer these clearly and specifically, that's the signal.
Investor lens — AI lab due diligence
8
The $40T product is a virtual collaborator, not a better chatbot
The end-state Krishna is building toward is an agent with full organizational context, persistent memory, access to every tool in your stack, and the ability to work on a multi-month idea — not just a task. $40 trillion of annual knowledge work is the addressable market. Coding proved the pattern first. Cowork is already growing faster than Claude Code did at the same stage. The form factor that gets this right captures the decade.
Operator + Investor — product vision & TAM
05

Memorable Quotes

"If you buy too much compute, you go out of business. If you buy too little compute, you can't serve your customers — and you're not at the frontier. Same thing."
On the fundamental tension in compute planning
"Humans mostly think linearly. They think incrementally. I've been at the company for two years. That's a paradigm I've had to break for myself — to stop thinking linearly and think on this exponential."
On rewiring how you forecast
"We started the year with about $9 billion of run-rate revenue, and we ended the quarter with north of $30 billion. That kind of change is enabled by these model intelligence leaps — and then the products that we build around them."
On model-led growth in practice
"Our net dollar retention rate is over 500% on an annualized basis. Nine out of the Fortune 10. These aren't pilots anymore."
On the depth of enterprise commitment
"We have this sticker on our laptops: 'Our competitors are incredibly capable, and success is far from guaranteed.' If something good happens, there's not confetti on the floor. It's: what's next?"
On Anthropic's cultural posture
"On the way here, I was in an Uber, and I signed two double-digit million-dollar commits in the car ride, which was 20 minutes."
On the velocity of enterprise AI adoption today
06

What Krishna Didn't Say

Analytically honest gaps — read alongside the episode, not against it
These are questions worth bringing to the next conversation — not criticisms of the episode, which was unusually candid by any standard.
07

Discussion Questions

1
Krishna argues that traditional financial metrics — gross margin, COGS, R&D — fundamentally misframe AI lab economics because compute is simultaneously generating current revenue, future capability, and internal productivity. If you were building a new framework to evaluate frontier AI labs, what would your three most important metrics be, and how would you get the data to calculate them?
Investor Framework · AI Lab Economics · Due Diligence
2
Anthropic's ability to run three chip platforms fungibly took years of compiler work and deep co-engineering to build. Krishna sees this as a structural cost moat. Do you agree? What would it cost a competitor to replicate it today — and is that gap widening or closing as the ecosystem matures around multi-vendor AI infrastructure?
Competitive Moat · Infrastructure · Defensibility
3
The Jevons Paradox shows up twice — in token pricing and in labor economics. Where in your own business or portfolio is this dynamic playing out right now? And where might it be actively suppressed — by regulation, org inertia, or buyer psychology — even though the underlying economics would suggest it should be appearing?
Pricing Strategy · Labor Economics · TAM Expansion
4
Anthropic's safety investment became its most credible enterprise differentiator — Fortune 10 companies trust them with sensitive workflows partly because of it. Is this a moat that was only available to the company that invested first, before it became expected table stakes? Or can it be meaningfully replicated by labs that start investing now?
Competitive Strategy · Trust · Enterprise GTM
5
Krishna names three scenarios that could push Anthropic to the low end of its cone: enterprise diffusion hitting a wall, scaling laws plateauing, and losing the frontier position. If you were running a pre-mortem on Anthropic's next 18 months, which would you weight most heavily — and what leading indicator would tell you it was happening before it showed up in the revenue numbers?
Risk Assessment · Pre-Mortem · Leading Indicators
6
The virtual collaborator thesis — full org context, persistent memory, every tool in your stack, long-horizon task capability — is framed as a $40T TAM. Cowork is already growing faster than Claude Code did at the same stage. What do you think the biggest remaining barrier is: model capability, enterprise trust, integration complexity, or something else entirely? And which breaks first?
Product Vision · Agentic AI · TAM · Enterprise Readiness
7
Patrick asks whether there's something dystopian about an employee who "just does what the AI tells me." Krishna reframes it as amplification, not replacement. Where do you draw the line between AI as leverage and AI as decision-maker? How would you design an organization to preserve human judgment at the right levels while still capturing maximum AI leverage everywhere else?
Future of Work · Human + AI Collaboration · Org Design
8
Krishna's closing answer — his brother quietly chose to stay in-state for college six years before it would matter, so Krishna could have full financial freedom in his own choice — is a remarkable act of long-horizon generosity with no guarantee of payoff. What does it tell you about how Krishna thinks about investment horizons? And how does it mirror the way Anthropic is willing to forgo near-term inference revenue to make long-term capability bets?
Long-Term Thinking · Values · Character as Signal
08

Investor's Cheat Sheet

Screenshot this for your next IC meeting
3 Questions to Ask Any AI Lab
  • What is your return on compute, all up — across training, inference, and internal use?
  • How are customers actually experiencing ROI? Real deployments or pilots?
  • Where does your compute come from over the next 3 years, and what flexibility do you have if the market shifts?
5 Numbers That Matter
  • $9B → $30B run-rate ARR in a single quarter
  • 500%+ net dollar retention (annualized)
  • 9 of 10 Fortune 10 companies as customers
  • $125B+ in committed compute (Amazon + Google deals)
  • 90%+ of Anthropic's code written by Claude Code
What's Not in the Standard Framework
  • Compute is an asset, not a cost — evaluate ROC, not gross margin
  • Model capability leaps drive revenue, not sales headcount
  • Safety investment and enterprise trust are the same bet
  • Fungibility across chip platforms is a structural cost moat
"The returns to frontier intelligence are not slowing down — especially in enterprise." — Krishna Rao, CFO of Anthropic