Bloomberg Tech

Krishna Rao on why the business is really about compute

Krishna Rao· CFO at Anthropic
·~18 min·English·Bloomberg
AI CompanyAI InfrastructureTrainingInferenceAI SafetyBusiness Strategy
TL;DR

Anthropic's CFO explains the company through a single lens, compute, arguing that superior efficiency, a diversified four-platform chip fleet, and disciplined, flexibility-first investment turn an $81B run-rate and a roughly $900B compute bill into a long-term bet, with safety as the enterprise moat.

01Core Mental Model

To understand the business, understand compute

You cannot read Anthropic's financials without reading its compute; the CFO treats compute as the one variable the whole business turns on.

to understand the business, you have to understand compute. Compute is incredibly complex.

— Krishna Rao, Bloomberg Tech
Key Insight
Framing compute as the master variable is itself a strategic signal: it tells investors that Anthropic's financial story is really a compute-allocation story, and that the CFO's edge is reading engineering, not just spreadsheets.

02Decision Framework

The cone of uncertainty

The further out you forecast the wider the uncertainty, so Anthropic sizes every compute bet to survive many futures rather than one fixed plan.

Basically, the further out you go, the wider that cone is.

— Krishna Rao, Bloomberg Tech
Key Insight
A cone-of-uncertainty mindset reframes a frightening $900B number as optionality: a deliberately conservative balance sheet lets Anthropic keep buying on price-performance as results come in, rather than committing to today's prices and today's guesses.

03The Efficiency Moat

Capacity is compute times efficiency

Effective capacity is physical compute multiplied by efficiency, not chips alone, and best-in-class efficiency lets 75% of the fleet go to R and D while demand is still served.

we think of our effective capacity as the product of our physical compute and our compute efficiency, and we have the best compute efficiency in the world, we believe.

— Krishna Rao, Bloomberg Tech
Key Insight
If capacity equals chips times efficiency, an efficiency lead compounds like a chip lead but is cheaper to defend. It also explains how 75% of the fleet can sit in R and D without starving paying customers: the efficiency multiplier effectively manufactures more chips.

04Compute Fleet

Four kinds of chips, one fleet

Four accelerator platforms behind one in-house orchestration layer make the fleet fungible, so Anthropic buys the best price-performance and routes each job to its best chip.

That's an orchestration layer that we've built internal to the company. That took a really long time to do and do well, but now we have this flexibility to use compute across these platforms, and we think that's a real advantage.

— Krishna Rao, Bloomberg Tech
Key Insight
The real asset is less the chips than the orchestration layer that makes them interchangeable, a software advantage built up over a long time. It is also what lets Anthropic treat each supplier, Nvidia included, as one option among four rather than a single point of dependence, even while calling Nvidia a close partner on both technology and capital.

05The $900B Question

$900 billion, spread over many years

The roughly $900B compute bill is spread over many years, not due at once; about half lands after 2030, and Krishna sets the roughly $70B due in the next year against both revenue and a large cash balance.

$900 billion of compute commitments, I mean, it's spread over many years.

— Krishna Rao, Bloomberg Tech
Key Insight
Quoting the commitment against revenue and cash, never in isolation, is the CFO's whole rhetorical move: a figure that reads as reckless in a headline reads as disciplined as a multi-year, half-after-2030 schedule financed by a fast-growing run-rate.

06Safety as Strategy

Not the revenue-maximizing move

Rather than release a powerful new model to everyone at once, Anthropic put it in cyber defenders' hands first, and frames safety and alignment as something its enterprise buyers demand alongside capability.

maybe the revenue maximizing thing to do in the moment was just to release it to the world and see what happens.

— Krishna Rao, Bloomberg Tech
Key Insight
For a lab whose customers are regulated enterprises, safety is not a tax on revenue but the feature that wins the contract. With 11,900 enterprise customers, up roughly eight times this year, the defenders-first rollout ships frontier capability without handing it to attackers on day one.

07The Opportunity

Still under 0.1% of the bill

Half the Fortune 500 already buy, yet by Krishna's own figures their cloud spend is under 0.1% of expenses, so he argues the market has barely started.

Their cloud spend represents less than 0.1% of their total expenses.

— Krishna Rao, Bloomberg Tech
Key Insight
The sub-0.1% figure is the bull case stated as arithmetic: even with half the Fortune 500 as customers, their cloud spend is still a tiny share of what they spend, which Krishna presents as evidence that most of the adoption is still ahead rather than behind.

08The Long Game

One stop on the train

The record IPO is one stop on the train, not the destination; the goal is the mission, and commercial success is how it gets funded.

I think of this IPO as, um, one stop on the train, right.

— Krishna Rao, Bloomberg Tech
Key Insight
Calling a record IPO one stop on the train is a deliberate expectations reset: it tells public-market investors not to price the stock on this quarter's margins, and signals that the mission, not the share price, will drive the big calls.