Bloomberg Technology

Rodrigo Liang on why inference now comes down to economics

Rodrigo Liang· Co-founder and CEO at SambaNova Systems
·~9 min·English·Bloomberg
InferenceAI InfrastructureBusiness Strategy
TL;DR

SambaNova's CEO argues that as inference scales the winner is decided by economics — payback speed, energy efficiency, and reused data centers — not by who owns the newest or most chips.

01Core Thesis

Economics is paramount

At scale, Liang says inference is won on how fast a rack pays for itself — about six months on SambaNova versus eighteen to twenty-four for a typical build.

we're able to give you six month returns, where, otherwise, it might take you eighteen, twenty four months to do it.

— Rodrigo Liang, Bloomberg Technology
Key Insight
By making payback the yardstick, Liang shifts the contest from peak performance to return on capital, where a lower-power challenger can compete. It implies buyers now judge an inference rack like any capital asset — on how fast it pays back, not on raw specs.
02The Signal

Strategic backers, not just financial

The Series F did not just add money — its extended close brought in strategic corporate backers, big tech companies partnering with a challenger rather than only funding it.

These are huge tech companies who are coming in and partnering with us to actually help someone someone else grow in the market.

— Rodrigo Liang, Bloomberg Technology
Key Insight
Strategic backers, not just financial ones, means incumbents are buying optionality: a seat beside a challenger that undercuts the dominant GPU economics. Their money reads less like a bet and more like a hedge against being locked to a single supplier.
03Market Structure

Hybrid is the end state

Liang bets the end state is hybrid: data centers that intermix NVIDIA, AMD, and SambaNova chips and mix frontier with private models, each workload routed to whatever gives the best economics.

people are looking at heterogeneous environments using chips from NVIDIA and AMD and SambaNova to intermix

— Rodrigo Liang, Bloomberg Technology
Key Insight
If the end state is multi-vendor and multi-model, then "who wins AI chips" is the wrong question; the prize is being one accepted tile in every rack. That reframes SambaNova's goal from displacing NVIDIA to coexisting with it — a claim JPMorgan's use of a wide array of chips, cited on its earnings call, already supports.
04The Real Constraint

Financing is the bottleneck

Scaling inference is capital-intensive and slow — chips take nine to twelve months to reach data centers, so the capital sits in process while it turns into deployed value, and financing that gap is the real limit.

it's a capital intensive business. And as you scale, you're having to put all these chips into the factory and taking nine to twelve months

— Rodrigo Liang, Bloomberg Technology
Key Insight
Liang quietly moves the scarce resource from GPUs to capital plus time: if a chip takes nine to twelve months to earn a return, whoever can finance that gap, or shorten it, sets the pace. It implies the moat is a balance sheet and payback speed, not only silicon.
05The Build-Out Play

Brownfield, not greenfield

Rather than fight the power and permitting wall that stalls new data centers, SambaNova reuses brownfield sites because its ten-kilowatt racks fit the power already there.

brownfield data centers that already have the space, already have the power, already have the infrastructure developed, and actually reusing it and reenergizing that

— Rodrigo Liang, Bloomberg Technology
Key Insight
Reusing existing data centers turns SambaNova's low power draw from a cost story into a go-to-market one: it unlocks sites, like the ones its partner Vista is reenergizing, that greenfield builders cannot touch. The efficiency becomes a permitting-and-speed advantage, not only a lower energy bill.
06The Macro Answer

Early innings, not a bubble

Answering the bubble question, Liang says we are still in early innings — most enterprises are barely started, and the market is only now entering its efficiency phase.

I think we're in early stages still despite all this build out. You do look at most enterprises are just getting started with delivering applications on AI, and so I think we're just at the early days. Now we're entering the phase where it is about efficiency.

— Rodrigo Liang, Bloomberg Technology
Key Insight
Framing now as the start of the efficiency phase lets Liang answer the bubble question without denying the overbuild: spending may be ahead of demand, but the demand curve has barely begun, so the fix he offers is efficiency, not retreat.