A graphics card launched in 2020 just became the minimum recommended hardware for something that sounded like a data center job a year ago.
Perplexity launched Portable Computer on August 25. It is a local version of its Computer agent that can keep the model, agent harness, conversation, files, planner, tool router, scheduler, local search index and sandbox on the user’s own machine. Perplexity says work completed locally does not consume Computer credits. If a task needs a stronger cloud model or outside information, the system can ask permission before sending that part away.
Then came the hardware requirement that made me stop.
Perplexity says an NVIDIA GPU with at least 24GB of VRAM is the practical floor for the experience it wants to deliver. In normal PC language, that means an RTX 3090 qualifies.
That card will be six years old in September.
The RTX 3090 launched as NVIDIA’s ridiculous 8K gaming card. It has 24GB of GDDR6X memory, 10,496 CUDA cores and a 350W graphics power rating. In 2026 it is no longer the fastest GeForce card, no longer the shiny object in gaming PCs, and definitely not the hardware anybody would have chosen to headline a new AI platform.
Yet its 24GB memory pool is suddenly enough to host an entire modern agent stack.
That says more about the direction of local AI than another new GPU launch.
Access without medium partner: RTX 3090 Is Enough for a Local AI Agent
This is not just a chatbot running offline
The phrase “local AI” has become so broad that it barely means anything.
Running a 7B model in Ollama is local AI. Generating an image with a model on your GPU is local AI. Transcribing an audio file without uploading it is local AI.
Portable Computer is trying to move the line much further.
Perplexity says the local system includes the orchestrator, planner, tool router, scheduler, durable task queue and local search index. The model can read files, search documents, work through multi step tasks, execute tools inside an operating system sandbox and keep jobs running.
That distinction matters.
The difficult part of a useful agent is not merely producing text. It is keeping context, deciding what to do next, calling tools, handling failures, managing permissions and continuing long enough to finish something useful.
Until recently, most polished agent products solved that by putting the hard parts in the cloud.
Portable Computer is an attempt to move the operating system of the agent onto your PC.
And apparently a 24GB Ampere card is enough to enter.
Why the RTX 3090 refuses to die
The 3090 has had a strange second life.
Gamers moved on to newer cards. Local AI users kept looking at the same old specification: 24GB.
That amount of VRAM remains awkwardly useful.
The RTX 4090 also has 24GB. The RTX 5090 finally moves consumer GeForce to 32GB, but it is far more expensive. Plenty of cheaper modern cards still sit well below 24GB. For local models, capacity can matter before raw compute does.
A model cannot run entirely in GPU memory if its weights and runtime state do not fit there.
Perplexity’s launch models explain the choice. Portable Computer can run Qwen 3.8 27B or PPLX 27B, Perplexity’s post trained version of that model. NVIDIA Nemotron 3.5 Lightning is planned as another option.
Twenty seven billion parameters sounds large, but modern quantization makes that class of model realistic on a 24GB GPU. The model still needs room for context and runtime allocations, which is why 24GB is a sensible floor rather than an arbitrary marketing number.
Perplexity did not say every 24GB GPU will behave identically. It also did not claim a 3090 matches a DGX Spark. Its launch demos used NVIDIA’s DGX Spark, which has a much larger memory pool. The RTX requirement means the card clears the compatibility target for the product.
That is still a big shift.
Five years ago, the 3090 was marketed around 8K gaming and creator workloads. In 2026, the same memory capacity is enough to host a private knowledge worker that reads documents, reasons over them, calls tools and keeps tasks running.
Old hardware did not become new.
The software around it finally caught up.
The interesting part is the harness, not the model
Perplexity’s research around Portable Computer contains a detail that deserves more attention than the hardware requirement.
The company found that putting a smaller local model inside a harness designed for giant frontier models does not work particularly well.
That sounds obvious after you hear it.
Frontier models can swallow enormous prompts, inspect huge tool descriptions and recover from messy long tasks. A 27B local model has much less room for mistakes. Perplexity says Qwen 3.8 27B advertises a context window around 260K tokens, but its researchers found performance beginning to degrade beyond roughly 100K in agent work.
So Perplexity changed the agent rather than simply demanding a bigger model.
The core prompt is kept small. Tools are limited. Extra capabilities are loaded as skills only when needed. Old context gets compacted. Common connectors that would normally arrive as large MCP tool descriptions are converted into smaller command line tools. The agent is also encouraged to verify its own work.
This may be the most important lesson in the whole launch.
Local agents do not necessarily need frontier sized models if the software surrounding the model stops wasting its limited attention.
That is why a 3090 can suddenly look relevant again.
The card did not gain VRAM.
The agent got less wasteful.
Perplexity’s own numbers are promising, but they are still Perplexity’s numbers
Perplexity tested its local harness against Pi and Hermes using the same Qwen 3.8 27B model on DGX Spark.
On the company’s internal Local Knowledge Work Bench, Computer scored 82.6 percent. Pi scored 77.6 percent and Hermes 74.0 percent. Perplexity’s post trained PPLX 27B model pushed its own system to 85.4 percent.
That is encouraging.
It is not independent proof that Portable Computer is now the best local agent for everyone.
The benchmark belongs to Perplexity. The company says it plans to publish more technical detail and open source the evaluation. Until that happens and outside users reproduce the results, I would treat the numbers as vendor evidence, not a final verdict.
The public benchmark results are more interesting because they show both the progress and the limit.
On Terminal Bench 2.1, which tests coding agents, the fully local Qwen setup scored 59.6 percent in Perplexity’s research.
Give the local model access to Claude Opus 5 as an advisor and the score rises to 73.0 percent.
Run Claude Opus 5 as the main model and it reaches 82.4 percent.
There is the entire local versus cloud argument in three numbers.
Local is useful.
Hybrid is better.
Frontier cloud is still stronger on the hardest work.
So no, the cloud is not dead
The title of this article says the cloud lost one excuse.
Not its job.
For years, one of the strongest arguments for cloud AI was that serious agent systems required hardware and infrastructure ordinary users did not own.
That argument just got weaker.
If a 24GB RTX GPU can host a complete agent harness, local search, tools, model inference and sandbox, “you need our servers because your PC cannot do this” becomes much harder to say with a straight face.
But the cloud still has obvious advantages.
Frontier models remain better at difficult reasoning. Cloud services can provide huge models without asking you to buy a GPU. They can scale across many simultaneous users. They are easier for somebody who does not want to maintain drivers, Linux, model files and a machine that can dump hundreds of watts of heat into a room.
Portable Computer itself admits the point by supporting escalation.
When a local model gets stuck, it can ask a stronger remote model for advice. Perplexity says the user can decide whether to allow that call and can see what context is leaving the machine.
That is not the death of cloud AI.
It is a demotion.
The cloud stops being the default place where everything must happen and becomes the expensive specialist the local system calls when it needs help.
That is a much more interesting architecture.
“Zero token cost” does not mean free
There is another phrase that needs cleaning up.
Perplexity says local work does not consume credits and local inference has no per token fee.
Correct.
Your electricity company will remain disappointingly interested in the experiment.
An RTX 3090 Founders Edition is rated at 350W. NVIDIA originally recommended a 750W system power supply around the card. Actual power use varies with the workload and power limits, but running a local agent for hours is not zero cost.
The GPU itself is not free either.
The used market is messy in August 2026. Recent sold RTX 3090 examples on eBay have ranged from below $900 for some OEM cards to around $1,200 or $1,300 for several branded and Founders Edition examples. Condition matters enormously.
Then there is the rest of the PC.
You need a compatible system, enough system RAM, storage for model files, a decent power supply and cooling. A 3090 is physically huge and hot compared with modern compact AI boxes.
Portable Computer also requires an eligible Perplexity subscription. Perplexity Pro currently costs $20 per month and Max costs $200 per month. Enterprise plans cost more.
So “no token bill” is the accurate claim.
“Free AI agent” is not.
Privacy gets much better, but local does not mean disconnected
This is where Portable Computer makes its strongest case.
Sensitive documents can stay on the machine while the local model reads and processes them. Perplexity says the model, conversation and trajectory remain local by default. Tool execution happens in an operating system level sandbox. If the sandbox is unavailable, Perplexity says the harness disables tool execution instead of silently running commands with the user’s full permissions.
That is a meaningful design choice.
A lawyer reviewing contracts, a developer working on private source code or a finance team examining internal spreadsheets may prefer that model over uploading everything to a remote service.
But “local” needs another asterisk.
Web search is not local.
Sending a message to Slack is not local.
Reading Gmail through a connector is not local.
As soon as the agent intentionally uses an external service, some network traffic leaves the machine because that is how external services work.
Perplexity’s architecture tries to keep private model inference and private files local while allowing specific outside capabilities when approved.
That is better described as local first than offline.
The distinction matters.
The Linux requirement will stop plenty of people
Portable Computer launched on Linux.
Windows support is scheduled for September.
That alone prevents this from being a mainstream consumer product today.
A huge number of RTX 3090 cards still live inside Windows gaming PCs. Their owners technically possess the right class of GPU but cannot simply install Portable Computer on the operating system they already use at launch.
That should improve quickly if Perplexity hits its September target.
Apple Silicon is a bigger absence.
Mac users have become one of the most active local AI groups because unified memory makes large models unusually practical on Apple hardware. Perplexity has said it is focused on NVIDIA hardware for now.
That decision makes engineering sense. CUDA is mature and NVIDIA worked directly with Perplexity on the launch.
It also limits the audience.
“Runs on hardware you already own” is true for a specific slice of people.
A fairly wealthy slice with 24GB NVIDIA GPUs.
A 3090 is powerful, old and inefficient at the same time
I would not rush out and buy an RTX 3090 just because Perplexity supports it.
If you already own one, the story is excellent. You suddenly have a new class of workload for hardware that many people were considering replacing.
If you are buying from scratch, the decision gets harder.
A 3090 offers 24GB at a used price that can still be attractive compared with newer high memory NVIDIA cards. CUDA support is mature. Local AI tools know the architecture well.
But it is a 350W, three slot card from 2020.
Used cards may have years of thermal cycling. GDDR6X memory can run hot. Fan condition matters. Power supplies matter. Some systems simply do not have the space or airflow.
The 3090 has become a budget AI card in the strangest possible sense.
It is budget hardware only because newer options can cost so much more.
The bigger story is that agents are becoming PC software
For most of the chatbot era, the user’s computer was basically a terminal.
You typed.
The cloud thought.
The cloud stored the model.
The cloud ran the tools.
The cloud charged for the work.
Portable Computer flips a surprising amount of that relationship.
The local machine can hold the model. It can own the memory. It can keep a search index. It can run the agent loop. It can schedule work. It can execute tools in a sandbox. It can keep private files beside the model rather than shipping them to it.
The cloud becomes optional for certain steps.
That change has consequences far beyond Perplexity.
It gives NVIDIA another reason to sell high memory consumer GPUs.
It makes 24GB, 32GB, 64GB and 128GB memory tiers much more important than generic “AI PC” TOPS numbers.
It gives old workstation and gaming hardware a new job.
And it changes the economics of agents because a system that runs all afternoon no longer has to meter every local thought against an API balance.
The best part is that we are only at 27B models here.
Local models will improve.
Hardware will get more memory.
Agent harnesses will get more efficient.
The floor will move downward.
The RTX 3090 just became a very strange benchmark for the future
In September 2020, NVIDIA sold the RTX 3090 as a $1,499 monster for creators, researchers and people determined to play games at 8K.
In August 2026, Perplexity is using its 24GB memory capacity as the rough entry point for a full local AI agent.
I like that symmetry.
The card was always excessive for normal gaming.
Now AI has finally found a way to make 24GB feel modest.
Portable Computer does not make cloud AI irrelevant. Perplexity’s own results show why. A compact local model still falls behind a frontier model on hard coding tasks, and a hybrid setup can improve results by asking the cloud for help.
But one old assumption has broken.
A serious agent no longer has to live entirely in somebody else’s data center.
If your six year old GPU can hold the model, run the tools, search your files and keep the agent working on your desk, the conversation changes from “Can this be local?” to a much more useful question:
Why should this particular task leave my computer at all?