The important qualification is hardware: NVIDIA specifies a GPU with at least 24GB of VRAM. This is not a new capability for every PC with an RTX badge. For Windows users with qualifying graphics hardware, the launch could offer a useful local-first option for document, data, development, and research workflows. But the product’s practical value depends on more than the word “local.” Cloud escalation remains available, its sandbox and connector claims come from the vendors rather than independent testing, and the precise Windows model lineup is not fully clear.
What NVIDIA and Perplexity say has launched
Portable Computer is Perplexity’s local version of Perplexity Computer. NVIDIA describes it as a system for multistep work rather than a conventional one-question-at-a-time assistant. According to NVIDIA, local models can be used to analyze data, combine material across files, and address recurring tasks. Perplexity similarly positions Portable Computer as a way to run its Computer experience on-device.
The Windows timing matters. Perplexity introduced Portable Computer on August 25 for NVIDIA DGX Spark systems running Linux, while saying Windows support was coming soon. NVIDIA’s September 14 announcement changes that status: Portable Computer is now described as available for supported Windows systems with GeForce RTX or RTX PRO graphics.
NVIDIA says the Windows version includes Computer’s built-in browser and Perplexity’s proprietary SPACE sandbox. Perplexity says code and tool execution occur in isolated sandbox environments with controlled access to files and connected applications. In theory, that architecture lets the system do more than summarize a file while imposing limits on the environment in which tools and code run.
Those are meaningful vendor-described safeguards, but they are not a guarantee of security in every use case. The available launch material does not supply independent security testing, detailed evidence about permission granularity, testing against malicious files or websites, or reliability data for long sequences of AI-directed actions. Organizations considering sensitive accounts or confidential data should inspect the actual access requests, file scope, and available controls before treating the sandbox as a complete security boundary.
Local-first does not mean offline-only
For eligible work, the on-device design is the release’s most consequential feature. Perplexity says locally completed work does not consume Perplexity Computer credits. For people repeatedly working with documents, data, or coding tasks, that could reduce dependence on a metered cloud workflow. For organizations, routine local handling may also be appealing when internal material should not automatically be uploaded for each request.
But NVIDIA and Perplexity describe Portable Computer as a hybrid design, not a permanently disconnected one. NVIDIA says cloud support is available for advanced research and reasoning, while Perplexity says users can authorize escalation to the cloud. Both companies say the product asks for permission before information is sent off-device.
That produces a more accurate privacy picture than either extreme. A task completed locally may keep the relevant content on the Windows PC. A task escalated to cloud assistance can involve off-device transfer after authorization. Portable Computer should therefore be understood as local-first, rather than as an unconditional promise that every prompt, file, output, and workflow will remain offline.
For Windows users, that distinction has practical consequences:
- Read cloud-escalation prompts carefully. Customer records, source code, financial information, health data, and internal plans may carry privacy, contractual, or compliance implications if they are authorized for off-device processing.
- Treat connectors as access decisions. A connection to mail, storage, collaboration, or source-control services can expand the information available to a workflow. That can be useful, but accounts and folders should be limited to what a specific task genuinely needs.
- Keep data location separate from answer quality. Processing material locally does not establish that an analysis, summary, recommendation, or automated step is correct. Locality is a data-handling characteristic, not a substitute for human review.
- Test policy-sensitive workflows before deployment. Teams that need a strict no-cloud process should verify what occurs when advanced reasoning is requested. A local model setting alone should not be assumed to meet a policy without examining the product’s actual behavior.
Permission before sending information off-device is a meaningful stated control. Its usefulness in practice will depend on whether users can understand what data is about to be shared, why cloud assistance is needed, and whether a local alternative is available. The vendor launch material does not establish how detailed those disclosures are in the Windows app.
The 24GB VRAM requirement is the real compatibility test
NVIDIA’s stated Windows requirement is a GeForce RTX or RTX PRO GPU with 24GB or more of VRAM. That is the central constraint for prospective users and IT buyers.
It rules out treating Portable Computer as a standard AI feature for the broad base of consumer RTX laptops and lower-memory desktop GPUs. Compatibility depends on available graphics memory, not just the presence of RTX branding. Before installing or purchasing anything for this feature, users should check the VRAM capacity of the specific GPU installed in their PC.
The 24GB threshold also means the first Windows audience is likely to skew toward high-end desktop owners, professional creators, developers, technical teams, and workstation users. That is an inference from NVIDIA’s minimum specification rather than a market classification published by NVIDIA or Perplexity, but it is a useful correction to the assumption that the launch reaches all RTX PCs.
For an organization, VRAM is only the first deployment question. The available launch material does not settle subscription eligibility, pricing, regional availability, enterprise administration, user-management controls, logging, or support commitments. A local model can change where some work occurs, but it does not by itself resolve questions around identities, connector permissions, approvals, retention, and auditing.
NVIDIA says support for DGX Station is expected soon, but it has not provided a date or technical requirements in the available announcement. That future support should not be confused with the earlier DGX Spark release, and it should not be treated as an available feature until NVIDIA supplies details.
Connectors and workflows are promises to test, not proven outcomes
NVIDIA lists connectors for Microsoft Outlook, OneDrive, Word, Google Drive, Gmail, Slack, and GitHub. If these integrations perform as NVIDIA describes, Portable Computer could work across many places where teams already store documents, discussions, email, and code. The Outlook, OneDrive, and Word connectors are especially notable for Windows-centered workplaces built around Microsoft productivity tools.
NVIDIA also gives examples involving pull requests, financial documents, and startup funnels. These examples describe the category of work NVIDIA and Perplexity intend Portable Computer to support: reviewing development activity, synthesizing business records, and examining operational information across different sources.
They should not be mistaken for independently validated results. The available evidence does not provide task-completion rates, error rates, execution times, handling of ambiguous instructions, resilience across real-world folder structures, or proof of consistent behavior with every listed connector. It also does not show how often a workflow needs cloud escalation or human correction.
That limitation matters most in consequential uses. A financial-document summary may omit a qualification; pull-request analysis can misunderstand project context; and an agent connected to email or shared content can encounter misleading or adversarial instructions. The sensible first deployment is bounded and reviewable work: preparing a draft, organizing a research set, identifying possible changes, or summarizing information that a qualified person checks.
It is less prudent to begin with unsupervised decisions, irreversible account actions, or sensitive communications. A controlled trial using non-sensitive, reversible tasks is a more realistic way to evaluate NVIDIA and Perplexity’s claims on a particular PC and within a particular organization.
Check the current Windows model before buying hardware
The named local model deserves special caution. NVIDIA’s Windows announcement identifies Qwen 3.8 27B as an example of a local model, describing it as post-trained for Perplexity Computer and optimized for NVIDIA RTX GPUs. However, reporting based on Perplexity’s Windows product page says Windows systems use PPLX 27B and that stock Qwen 3.8 27B is not available on Windows RTX PCs.
That leaves the exact Windows model roster unresolved in the available material. Prospective users should not buy a 24GB GPU or set a deployment requirement around a named model until they confirm the model shown in the current Perplexity app or in current official documentation.
This is more than a naming detail. The local model used can affect capability, performance, memory requirements, update behavior, and the reproducibility of results. Until the Windows lineup is clearly documented, hardware planning should focus on NVIDIA’s explicit 24GB VRAM requirement and on the model actually offered to the user at setup.
A high-end Windows capability, not a universal AI-PC feature
Portable Computer is a significant Windows development because NVIDIA and Perplexity are offering a route for multistep, tool-using work to run locally on sufficiently capable RTX hardware. The vendor-described benefits are concrete: local execution for eligible work, no Perplexity Computer credit consumption for that on-device work, permission-based cloud escalation, a sandboxed execution environment, and connectors aimed at common productivity and development services.
The restrictions are just as concrete. NVIDIA’s 24GB VRAM minimum sharply narrows the compatible PC population. Cloud escalation remains part of the design, so privacy depends on workflow choices and user permissions. The Windows model roster needs verification. And the claimed security protections, connector behavior, performance, and workflow outcomes have not been independently validated in the material available here.
For owners of qualifying systems, the practical next step is not blind trust or an immediate high-stakes rollout. Confirm GPU memory, inspect connector scopes, observe when cloud permission is requested, and measure the review burden on real but non-sensitive tasks. For most other Windows users, Portable Computer is best viewed as a local-agent capability for high-end RTX PCs and workstations today, rather than a feature that has arrived across the entire RTX PC market.