Perplexity's new Portable Computer is a serious bet that some AI agent work belongs on your own Nvidia hardware, not on a rented cloud meter.
Perplexity released Portable Computer on August 25. The hook is simple: if you've got the right Nvidia machine, the agent can run its work locally before it ever asks the cloud for help. That changes the cost conversation. It also raises a harder question for anyone buying into AI agents: do you trust a smaller local model enough to let it handle real work?
According to Perplexity's launch post, Portable Computer is a local version of its Computer agent platform built first for Nvidia's DGX Spark, the 128GB unified-memory desktop system based on the Grace Blackwell GB10 platform. The company says it will also support Nvidia RTX GPU PCs. Other reports on the launch put the practical floor at a Linux machine with an RTX card carrying at least 24GB of VRAM, roughly an RTX 3090-class setup or newer.
That isn't a small ask.
The reason this product matters is not that it makes AI free. It doesn't. You still need a paid Perplexity plan, a Linux setup for now, and hardware most people don't have sitting under the desk. But once the local system is running, Perplexity says local work doesn't consume credits. The cloud only enters when a task needs web access, connected apps, or a stronger advisor model. You decide what leaves the device.
OpenAI Says Its First Chip Jalapeño Beats Nvidia's GB300 on Inference OpenAI published independent benchmark results on August 25 showing its debut Jalapeño inference chip beats Nvidia's GB300 by up to 1.9x in performance per watt and up to 3.6x in latency. The figures are self-reported and unaudited on a per-precision basis, and volume deployment isn't expected until 2027, but they mark the first outside-validated... - openai jalapeño chip beats nvidia gb300 - openai custom inference chip energy efficiency benchmark
For anyone running agents over private files, code, client notes, or financial documents, that permission step is the product. Perplexity says the local orchestrator, planner, tool router, scheduler, durable task queue, and local search index all run on the machine. That's the whole stack, on your desk. Before an advisor call goes out, the system selects the relevant context and flags sensitive information with a PII classifier. It shows you what would be sent. You approve it, or you don't.
The small model has to do real work #
The local models are not pretending to be the biggest systems in the market. Perplexity ships Portable Computer with Qwen 3.8 27B and PPLX 27B, its post-trained version of the same base model. Nvidia's Nemotron 3.5 Lightning, a 30B open model, is listed as coming soon.
Perplexity's own research blog gives the useful numbers, with the obvious caveat that these are vendor-run tests. On its 53-task Local Knowledge Work Bench, Computer running Qwen 3.8 27B on DGX Spark scored 82.6%, compared with 77.6% for Pi and 74.0% for Hermes using the same model. PPLX 27B lifted the score to 85.4%. On BrowseComp, a 1,266-task web research benchmark, Computer reached 66.7%, ahead of Pi at 50.2% and Hermes at 43.9%.
The benchmark detail worth keeping is not just the score. Perplexity says Computer averaged 402.1 seconds and 852,000 tokens per BrowseComp task, compared with 826.0 seconds and 2.82 million tokens for Pi. That is the sort of number that gets an operations team to pay attention. If your agent runs all day, token waste stops being an abstract annoyance and starts looking like a line item.
There is still a ceiling. On Terminal Bench 2.1, Perplexity says the fully local Qwen 3.8 27B setup scored 59.6%. Let it escalate to Claude Opus 5 as an advisor, and the score rose to 73.0% at an estimated API cost of $0.415 per rollout. Claude Opus 5 alone scored 82.4% at $0.65. So no, the local model doesn't beat the frontier model. It gives you a cheaper middle step, and for many jobs that may be enough.
Nvidia wins either way #
This launch lands while Nvidia and Perplexity are already being pulled closer together. The Information reported on August 23 that Nvidia has discussed investing in Perplexity at a valuation above $30 billion, up more than 50% from the company's roughly $20 billion valuation in its last financing. The same report said Perplexity's annualized revenue has risen to more than $750 million from less than $250 million at the start of the year, helped in part by the cloud-based Perplexity Computer product used by professional customers.
That relationship only looks strange if you think Nvidia cares where the workload runs. It doesn't. Sell chips to a cloud provider and the agents burn compute in a data center. Sell a DGX Spark to a developer instead, and the compute burns on a desk instead of a rack - same electricity bill, different address. Nvidia gets paid either way.
Sam Altman Fears a Handful of Companies Will Control AI OpenAI CEO Sam Altman told podcaster David Senra that his biggest fear is a small number of companies gatekeeping AI, framing it as a choice between "AI authoritarianism or liberty." The remarks land awkwardly given OpenAI's own dependence on Microsoft's cloud and Nvidia's chips, and its pending push for a $1 trillion IPO valuation. - sam altman fears ai companies controlling market access - how many companies will dominate artificial intelligence industry
The upfront price is the catch. Nvidia announced in February that it raised DGX Spark Founders Edition pricing from $3,999 to $4,699 because of memory supply constraints, and its marketplace still lists the machine at $4,699. Local doesn't mean cheap to start. It means the meter changes after you've bought the box.
Portable Computer is available now for Perplexity Pro and Max subscribers on DGX Spark, with Linux first and Windows support coming soon. Apple silicon is not in the launch plan Perplexity has described. That leaves a narrow first audience: developers, analysts, founders, and teams already willing to buy Nvidia hardware to keep more work on premises.
Frankly, that's the real test. The benchmark charts help, but customers will judge this by whether a 27 billion parameter local agent can finish useful work without constant supervision. If it can, Perplexity has a sharper argument than another cloud agent with a nicer interface. If it can't, the token bill was never the main problem.
Also read: The Cautionary Startup Tale of Leopold Aschenbrenner: Inexperience Means Inexperience • Bitcoin Breaks Above $80,000 Again, But Traders Want Proof It Holds • SpaceX and Nvidia Unveil Starmind AI1, an Orbiting Data Center Satellite