Perplexity Just Priced the Agent Stack at Zero on a Consumer RTX. Qwen Is the Substrate.

The Nvidia blog dropped the sentence on Sunday, September 14, 2026: Perplexity Portable Computer is now shipping inside the Windows app on any RTX or RTX PRO GPU with at least 24GB of VRAM. The entire agent stack runs on device (orchestrator, planner, tool router, scheduler, durable task queue, local search index) with two launch models to choose from at setup, Qwen 3.8 27B and PPLX 27B (Perplexity's post-trained Qwen), with Nvidia's Nemotron 3.5 Lightning listed as the next arrival. The pricing sentence sits two paragraphs below the spec sheet: local work consumes no Perplexity Computer credits.

Headline: an agent lab just put the token line on the entire harness at zero, provided the buyer supplies the silicon.

What Shipped

Portable Computer is the local twin of Perplexity Computer, the browser-based multistep agent that has been metered on a credit balance since launch. Same product surface, same task language, same tool loop, one important switch: on the local build, the model does not phone home for inference. The stack escalates to the cloud only when the task needs current information, a hosted browser session, a connected app, or frontier reasoning past what the local model can carry, and it asks permission before sending content off the device.

Portable Computer first shipped in August 2026 on Nvidia's $4,699 DGX Spark and on Linux workstations with RTX cards. That was a developer product with a nine hundred dollar taste, priced at a spec no ordinary Pro subscriber owned. The Sunday drop puts the same stack on the RTX card a paying gamer already has sitting under a desk. That is the only piece of the announcement that moves a market number, and it is the only piece the headline should keep.

The Fifth Harness Receipt in Twelve Days

Line up the last twelve days of harness pricing on one page. September 5, Anthropic published the Prove2Me Fermat receipt, a research artifact with no SKU sitting on general-availability Claude. September 8, OpenAI published the Navier-Stokes swarm run, a research post on an unshipped post-Astra model with no SKU. September 10, OpenAI shipped the Agents API in public beta with the harness fee explicitly set at zero and the token line on the underlying model paying the bill. September 11, Sakana priced Fugu Max at $2 input and $6 output per million tokens on a recursive orchestrator over open-weight sub-models. September 14, Perplexity and Nvidia shipped Portable Computer for Windows with the entire token line at zero for anyone who supplied a 24GB RTX.

Five receipts, five pricing postures, three theories of where the margin lives. The Anthropic and OpenAI research posts keep the margin on the model layer. The Sakana receipt puts it on the orchestrator itself. Perplexity puts it on the subscription line alone, and hands the compute cost of every local step to the customer's power bill. Perplexity Pro at $20 per month is now the flat rate for as much local agent time as the RTX can render, which is a very different product than a metered Comet session or an Agents API call priced on Astra tokens. On the same shelf the three postures decompose the agent bill into three lines (model tokens, orchestration policy, host compute) and each receipt this month picked which one carries the invoice.

The Substrate Is Chinese Open Weights

Both launch models on the Portable Computer setup screen are Qwen shapes. Qwen 3.8 27B is Alibaba's dense open-weight release from earlier this quarter, and PPLX 27B is the same base weights post-trained by Perplexity on its own agent harness. Neither is a proprietary American frontier model, and neither is a Nvidia model (Nemotron 3.5 Lightning is listed as forthcoming, but not on the launch card). An American AI company just made the default local substrate under its consumer agent a Chinese open-weight base. That is the second concrete data point this year that the coordination layer at the top of the stack has decoupled from the identity of the weights doing the arithmetic underneath, after the Sakana Fugu launch dispatched to the same open-weight pool at commodity rates. Two labs pointing at the same model layer this month, one Japanese and one American, both routing around the frontier margin at the local step.

The export-control read on this is not that Qwen escaped a rule, it is that the rules were never written for the case of an American company post-training an Alibaba base model into its own product SKU and shipping the result inside a Microsoft Store install. PPLX 27B is Qwen with a Perplexity lift on top and a paid subscription in front. The policy conversation for the next quarter is going to have to answer what that SKU counts as, and the answer is going to matter for anyone who wants to price a competing local agent against it without touching an Alibaba release note.

What the 24GB Floor Does to the Buyer List

Twenty-four gigabytes of VRAM is not a laptop spec, it is a desktop or workstation spec. The RTX 5090 lists at $1,999, the RTX 4090 at $1,599 where inventory is available, the RTX PRO workstation lineup climbs from there. Portable Computer for Windows is a product that runs on a machine that costs at least twice what an M4 MacBook Pro base does. The floor names the audience: a paying Pro or Max subscriber who already owns a top-tier Nvidia card for gaming, rendering, or model tinkering, and who is willing to trade a couple of hundred watts of desk power for a zero-token agent loop. That is a real audience and it is not a mass-market one, and Perplexity is not pretending otherwise. The Windows number matters because it is the first time a hosted-agent vendor drew the line at consumer Nvidia rather than at a data-center accelerator.

The bigger read is what a zero-token subscription line does to the pricing pages up the shelf. Anthropic's Claude Code just cut weekly limits 17 percent on the seat product because the seat could not absorb the load. OpenAI's Agents API uncapped the harness and let the Astra token line pay the bill. Sakana priced the orchestrator against the mid-tier shelf. Perplexity Portable Computer priced the whole stack at zero for anyone with the GPU. Four different answers to the same question, and the Perplexity answer is the one that concedes the least revenue on the token line and the most cost on the customer's side of the ledger. The buyer supplying the compute is the pricing model, and the RTX card is the receipt.

Caveats Worth Naming

Three things the launch does not do. It does not replace the cloud Perplexity Computer for tasks that need a hosted browser session, a connected SaaS app, or a frontier model the local 27B cannot carry, so heavy research runs still hit the credit line. It does not guarantee Qwen 3.8 27B and PPLX 27B stay free of licensing wrinkles as the US-China rule cycle continues (Alibaba tightening the Qwen license, or a US delegated act reclassifying a post-trained Chinese base, would land squarely on the launch models). And it does not solve the mobile-agent story: the 24GB floor keeps this off phones, laptops without workstation GPUs, and every macOS device Apple has shipped since Rosetta.

Our Take

The interesting decision this week is not the model choice and it is not the hardware spec, it is the credit line. Perplexity has decided that anyone willing to buy the silicon can run the agent stack for free at the token layer, and it has priced that decision at zero on the Pro and Max subscription sheet. Read against the harness receipts from the last two weeks, this is the fifth pricing artifact in twelve days and it is the first one that puts the entire agent stack (orchestrator, planner, router, scheduler, queue, index, model) on the customer's side of the compute ledger. Anthropic sells the coordination policy on the seat. OpenAI sells it on the token. Sakana sells it as a priced model of its own. Perplexity gave it away on the local build. Four defensible answers, one of them is the one that eventually rewrites the pricing floor if the RTX install base gets large enough to matter.

Practical implication for builders. If your product needs a local agent on a Windows machine with a 24GB RTX, Portable Computer is the first hosted-vendor build on the market that treats the token line as a solved problem. If your product needs to run on a laptop, on a phone, on a Mac, or on a machine with a smaller GPU, nothing about the Sunday drop changes what you can ship today, and the pricing floor at the frontier is still the pricing floor at the frontier. The local zero-token line is a genuine change to the ledger, but only for the specific hardware footprint the announcement names. Keep the read narrow, keep the RTX 24GB caveat attached to every quote, and watch the next 60 days for either a laptop-tier version of the same product or a competing announcement from a lab that has decided to route around Nvidia consumer silicon entirely.

Three signposts for the next 60 days. Whether Anthropic or OpenAI ship a comparable local agent build with a zero-token line, on any hardware footprint, before the end of Q4 (the direct test of whether the harness-priced-at-zero pattern generalizes past Perplexity or stays a Perplexity plus Nvidia one-off). Whether Perplexity extends Portable Computer to Apple silicon or to AMD Ryzen AI workstations inside the same window (the direct test of whether the substrate is genuinely open-weight-agnostic or effectively Nvidia-locked at the plumbing layer). Whether an EU AI Act delegated act, a CAISI advisory, or a US delegated rulemaking cites Qwen 3.8 or PPLX 27B by name in the next 90 days (the direct test of whether the Chinese-substrate question forces a policy artifact in the same quarter it forced a product artifact). Two of the three fire and the harness pricing conversation for 2027 gets written from the local side of the bill, not the frontier side. We are tracking Portable Computer credit consumption and the Windows install base on our Perplexity provider page, and the RTX edge-agent buildout on our earlier RTX Spark writeup. The next number to watch: whether Nemotron 3.5 Lightning arrives inside 30 days and whether Nvidia prices it at zero on Portable Computer or holds a token line.

Marcus Chen, September 16, 2026. TensorFeed tracks AI model releases, provider status, and the pricing layer underneath them. Sources for this piece: the September 14 Nvidia developer blog post announcing Portable Computer for Windows, Perplexity's launch post inside the Windows app, VentureBeat and Tom's Hardware coverage of September 14 and 15, and the Anthropic, OpenAI, and Sakana harness receipts referenced above.