When the talk turns to AI compute and the supply chain, the stars are usually hardware companies like Nvidia and TSMC. Perplexity looks far removed from all of that. It makes software and an answer engine, it buys no chips, and it builds no data centers. But here’s the interesting part: it still can’t escape that global compute supply chain.
This piece takes a lighter approach to show you where Perplexity’s compute comes from, and why it gets pulled around by the supply chain even though it owns no hardware. For how that whole hardware chain works, you can go straight to The AI Hardware Supply Chain, End to End. Here we focus only on Perplexity’s own situation. To get to know the company as a whole, start with What Kind of Company Is Perplexity.
Its compute is all rented
Perplexity’s compute strategy is simple: rent it.
It builds no data centers of its own. Its main cloud partner is Amazon AWS, and it has also signed a three-year, 750-million-dollar deal with Microsoft Azure, which means it stands on two big clouds at once. What’s more unusual is that the high-speed inference for its flagship Sonar model runs on chips from a company called Cerebras, the lifeblood of its speed experience. All in all, Perplexity holds no GPUs and no server rooms; it outsources nearly all of its computing.
This is the classic asset-light playbook. The upside is that it doesn’t have to pour a fortune into building its own compute, so it can concentrate its money and people on the product and move faster than asset-heavy rivals. For a fuller breakdown of its overall position as a middle layer, see Perplexity’s Middleman Problem.
Why it’s still tied to Nvidia and TSMC
Owning no hardware doesn’t mean you can decouple from the supply chain.
The logic is direct: the cloud Perplexity rents still runs on Nvidia GPUs underneath, and those GPUs depend on TSMC’s advanced process to manufacture. When any link in the chain tightens, the cost flows downstream. When GPU supply gets tight, when cloud providers pass rising costs on to customers, or when export controls tighten and make compute allocation harder in certain regions, all of these variables land on Perplexity indirectly, but very really, through the cloud bill it has to pay.
In other words, it looks like it’s just one layer away from the hardware. In truth, it has only tucked that exposure into the cloud bill.
Cerebras: a special and fragile thread
In this mostly rented landscape, Cerebras is one thread worth pulling out on its own.
That “answers seem to pour out almost instantly” speed of the Sonar model comes precisely from Cerebras’s chips. But the price of that speed is dependence. Cerebras is a chip startup that only went public in 2026, with a far weaker financial profile than the mature Nvidia, and about eighty-some percent of its revenue is concentrated in a single Middle Eastern customer group. Its specialized chips can only be made by TSMC and cannot easily be moved elsewhere. Should this supplier hit a bump in capacity, finances, or geopolitics, Perplexity’s speed advantage could wobble right along with it. For more on this layer of dependence, What Is Sonar has a finer-grained explanation.
Penchan’s take
Perplexity’s supply chain story is really a microcosm of its whole business model: by not owning and only integrating, it trades for speed and flexibility, but in doing so it hangs its own lifelines across several lines that belong to other people. The cloud, GPUs, specialized chips, none of these are in its own hands.
For the everyday reader, this brings a practical lens: when you look at any AI application company, don’t just look at how flashy its product is. Also ask, “Where does its compute come from, and whose moves shift its costs?” No matter how clever the answer engine, behind it runs that long hardware chain that stretches all the way from TSMC to the data center. To see that chain clearly, The AI Hardware Supply Chain, End to End is the best place to start.
Further reading: Perplexity’s Middleman Problem, What Is Sonar, The AI Hardware Supply Chain, End to End.