Bloomberg Tech

Andrew Feldman on why fast tokens cost more than slow ones

Andrew Feldman· Co-founder and CEO at Cerebras Systems
·~7 min·English·Bloomberg
InferenceAI InfrastructureBusiness StrategyGPU
TL;DR

Cerebras says fast inference commands premium token pricing, while AWS, Oracle, and Microsoft give its compute service near-global enterprise reach.

01Core Mental Model

Fast Tokens Cost More Than Slow Ones

Cerebras can charge a premium for speed itself — the same answer delivered faster is worth more, so fast tokens price higher than slow ones.

We can charge more for fast tokens than you can for slow tokens.

— Andrew Feldman, Bloomberg Tech
Key Insight
If speed is the thing customers pay extra for, then latency, not raw model quality, is the lever Cerebras is pricing against — an edge that only holds while rivals cannot match its speed at the same cost.
02Spend Discipline

Spending to Spend Is Waste

Feldman draws a line between burning tokens to look busy and allocating them to genuinely hard problems, where getting answers faster is where the value sits.

allocating tokens to really hard problems and using AI for what it's really good at is enormously efficient.

— Andrew Feldman, Bloomberg Tech
Key Insight
The spend-wisely framing is also a sales argument: it reassures the CFOs doing token budgeting that buying more inference is disciplined investment rather than runaway cost, which lowers the barrier to buying more of it.
03Distribution

Three Clouds, Every Enterprise

Signing AWS, Oracle, and Microsoft hands Cerebras near-instant global distribution it could never have built alone, reaching almost every enterprise on earth.

between these three, you reach just about every enterprise worldwide.

— Andrew Feldman, Bloomberg Tech
Key Insight
Leaning on hyperscalers for reach is a tacit admission that distribution, not silicon, is the hard part of scaling — and it ties Cerebras's growth to partners who also sell competing compute.
04Competitive Stance

Superior in a Lane, Not the Whole Road

Feldman claims an edge only in specific segments like fast inference, not across the whole market, and expects customers to choose on the merits of the product.

our products are superior in segments of the market like fast inference

— Andrew Feldman, Bloomberg Tech
Key Insight
By conceding it leads only in segments, Feldman sidesteps a head-to-head framing against larger rivals and recasts the contest as many winners in an expanding market — a safer story to tell when you are the smaller player.
05On the Rival

He Won't Pile On a Rival

Pressed on turmoil around a rival, Feldman declines to claim a windfall, praises the incumbent's run, and insists customers still decide on merit.

Jensen had one of the great entrepreneurial runs in history.

— Andrew Feldman, Bloomberg Tech
Key Insight
A competitor refusing to claim a rival is weakened signals confidence that his own demand is structural rather than borrowed — and it avoids staking the pitch on another company's bad news, which can reverse.
06Where the Build Goes

The Build Is Leaving the U.S.

About three quarters of Cerebras's new data-center capacity will come online outside the U.S. by 2027, in Canada, France, Norway, and Finland.

about 75% of our data center capacity outside The US

— Andrew Feldman, Bloomberg Tech
Key Insight
Placing most new capacity abroad hedges against U.S. policy swings and power constraints at once, and quietly recasts an American chip company as a globally distributed infrastructure operator.