When people talk about AI compute, the stars are usually hardware companies like Nvidia and TSMC. Perplexity makes an answer engine. It buys no chips and builds no data centers, so it looks far from hardware. Yet it still can’t escape the global compute supply chain.
For how the hardware chain works, see The AI Hardware Supply Chain, End to End. To get to know the company, start with What Kind of Company Is Perplexity.
Its compute is all rented
Perplexity has no GPUs and no server rooms. It outsources almost all of its computing to the cloud:
| Cloud | Details |
|---|---|
| Amazon AWS | Main cloud provider (a Perplexity spokesperson told Bloomberg in January 2026) |
| Microsoft Azure | Three-year, $750 million deal (January 2026, according to Bloomberg) |
| Crusoe | Multi-year partnership (September 2026, amount not disclosed), renting Nvidia GB300 clusters to train and deploy its own models |
A newer path is moving some computing onto users’ own computers. Starting in August 2026, Perplexity has been letting Portable Computer run fully locally on PCs with Nvidia or AMD chips. The models include PPLX 27B, which Perplexity post-trained itself. Users pay for that hardware.
This is the classic asset-light playbook. It doesn’t have to spend heavily on its own compute, so it can put its money and people into the product. For 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 breaking free of the supply chain.
The cloud Perplexity rents still runs on Nvidia GPUs; the Crusoe clusters are Nvidia GB300. Those GPUs depend on TSMC’s advanced process to be made. Tight GPU supply, cloud providers passing costs on to customers, or stricter export controls that make compute harder to allocate in some regions: all of these land on Perplexity through its cloud bill.
In other words, it is only one layer away from the hardware. The exposure is hidden in the cloud bill.
Cerebras: a past source of speed
The new Sonar launched in 2025 ran on specialized chips from a company called Cerebras. That “answers pour out almost instantly” speed came from those chips.
In September 2026, Sonar moved into the developer Agent API. Perplexity has not confirmed since then that Cerebras still powers its inference, so treat this link as 2025 information. Renting lets you switch suppliers; the cost is that none of those lines are in your own hands. For more on Sonar, see What Is Sonar.
Penchan’s summary
Perplexity’s compute story is a small picture of its business model: it owns nothing and only integrates. That buys speed and flexibility, but its lifelines hang on other people’s lines: the cloud, GPUs, and specialized chips.
When you look at any AI app company, don’t just look at how flashy the product is. Also ask: “Where does its compute come from, and who moves its costs?” However smart an answer engine is, it sits at the end of a long hardware chain that runs from TSMC all the way to the data center.