A $109 Crowdfunded Hub Would Pool Spare RAM for Bigger Home AI Models
TechRadar reports a crowdfunded hub that would combine memory from old computers and phones to run larger AI models, built on free open-source software.
TechRadar reports that TradeMAV, a small company, is crowdfunding a device called RAMDeck on Indiegogo. The hub is a small, low-power box, and early backers can get it for $109. Its pitch is that people running AI models at home do not need to buy more memory. Instead, they can borrow the RAM already sitting in their older PCs, Macs and phones.
According to TechRadar, the software behind the idea is the RPC backend of llama.cpp, an open-source project that anyone can already run for free. The hardware, in other words, would act as a way to connect existing machines rather than as a new technology for running models.
The outlet says the strongest demonstrations pair the setup with an Nvidia RTX 3060 graphics card and mixture-of-experts models, a type of model that activates only part of its network for each word it produces. Denser models reportedly manage around two tokens per second, a pace TechRadar presents as a real limit on the approach.
The report links the pitch to market conditions. TechRadar says DDR5 memory now costs roughly five times what it did in mid-2025, which makes adding RAM an expensive option for many home users and gives the borrowing idea more appeal.
Details beyond the pitch, including shipping timing, final specifications and how the hub performs with a wide range of everyday hardware, are not settled in the reporting. Readers should treat the early demonstrations as a starting point rather than a verdict.
Background
Running a large language model at home depends heavily on memory. A model's weights have to sit somewhere the processor can reach quickly, and graphics cards carry their own pool of fast memory that is separate from the system RAM in a computer. When a model is too large for the graphics memory, software can spill part of it into system RAM, which is far slower, and performance drops sharply. That is why hardware enthusiasts often focus on memory capacity as much as raw processing power.
Open-source projects have made local inference practical for many people. llama.cpp is one of the best known. It runs quantized models, which store weights at lower precision to shrink their size, and it can run on ordinary processors as well as on graphics hardware. Its RPC feature allows one instance of the software to offload work to another machine over a network, so a model can in principle be split across several computers.
The catch with splitting work across machines is the network. Every exchange of data between devices adds delay, and that delay compounds with each step needed to produce text. A link that is fast enough for simple file transfers can still be a bottleneck for generating words one at a time. Speed measured in tokens per second is the common yardstick for how responsive a model feels.
Mixture-of-experts models are designed to reduce this burden. Rather than using every part of the network for every token, they route each token to a subset of specialized sub-networks. This lowers the computing work per token while the total model can remain large, which is why such designs are popular for local setups.
Memory prices are also cyclical. Demand from data centers, consumer electronics and AI hardware can push costs up quickly, and shortages can last for months. Crowdfunding, meanwhile, is a common route for small hardware makers, but backers often wait for months and carry risk if a campaign fails to deliver.
Why it matters
The idea behind RAMDeck is appealing because it treats memory as something people already own. Many households have a drawer of older machines with capable processors and sizable RAM that rarely get used. If software can reliably combine them, local AI becomes more accessible without new purchases. That fits a broader push to keep AI tools on personal hardware, where users control their own data and avoid subscription fees for cloud services.
The story also highlights a tension in the market. A hardware company is asking people to pay for a device when the core software is free. That is not unusual, since many successful products bundle open tools with convenient packaging. But it raises the question of how much value a dedicated box adds beyond setup help and a tidy design.
The speed figures are the most important detail for judging the claim. A setup that runs well with specific models on specific hardware does not automatically extend to every combination of devices. Readers who want to experiment will need to check which models suit their machines and whether a network link is good enough.
For the wider technology landscape, the story is a reminder that expensive components shape what people try. When memory costs rise, people look for workarounds, and some of those workarounds turn into products.
What to watch
The first thing to watch is whether the campaign delivers on schedule and at the stated price. Crowdfunded hardware has a mixed record, and early backer pricing can change once a product moves to retail. Clear communication about shipping and support will matter.
Independent testing will be more informative than demonstrations. Reviewers who run a range of models across mixed hardware, and report real token rates and latency, will show whether the approach holds up outside controlled examples.
Watch also how the llama.cpp project and its contributors respond to commercial packaging of their work. Open-source licenses generally allow this, but community expectations around attribution and support can shape a product's reputation.
Finally, keep an eye on memory prices. If DDR5 costs stay high, interest in borrowing memory across devices may grow. If prices fall, the appeal of a dedicated pooling box could weaken.
Viewpoints
Commentary from our columns: three angles on this story.
The Skeptic open questions and weak points
The most appealing part of this pitch is the least complicated: a box that connects machines you already own. The harder part is the one that gets less attention. When a model is split across devices, every piece of information has to travel over a link between them, and that link is usually far slower than the memory inside a single computer. The speed figures described in the coverage, around two tokens per second for denser models, hint at how much that gap matters.
None of this makes the idea wrong. It makes the claim conditional. Mixture-of-experts models are a good fit because they activate less of the network for each word, and that is exactly the case where demonstrations look best. Readers should ask what happens with the models they actually want to run, on the network they actually have, in a house with older Wi-Fi and mismatched machines.
Then there is the money question. The core software is free and well documented. A campaign asking for early payment is asking backers to trust a small company to finish a product, ship it and keep it working. That trust may be deserved, but it should be earned with independent testing, not with a launch-day pitch. Healthy skepticism here is not hostility. It is the normal way to judge a hardware promise.
The Ledger the business angle
Look at the business shape of this offer before the technology. The software at the center is free to anyone. What a company can sell is convenience: a dedicated device, a setup that works out of the box, and perhaps support when things break. That can be a real business, but it is a narrow one, and it depends on how many people will pay for convenience they could assemble themselves with some patience.
Crowdfunding adds its own economics. Early backer pricing is designed to fund production upfront, which means the company carries risk that buyers do not see until the product ships. Campaigns sometimes grow more expensive to fulfill than planned, and the price at retail is often higher than the price that first attracted buyers. Anyone weighing the offer should compare the total cost with what the same money could buy in new memory.
The memory-price backdrop is the strongest selling point. If DDR5 costs remain far above their level from a year earlier, a product that avoids that purchase has a clear pitch. But that is a timing advantage, not a permanent one. Memory prices move in cycles, and a product built around a shortage needs a plan for the day the shortage eases.
The User what changes for people who use it
For most people, the question is simple: will this make my old computer useful again? The idea of putting an old laptop, a desktop and a phone to work together is appealing, but it sounds easier than it may be. Someone who wants to run a local assistant will need to know which models fit, how fast they respond and whether the setup stays stable over time.
The speed figures matter in everyday terms. A reply that arrives at a couple of tokens per second feels slow, like watching text appear one word at a time. Faster results depend on demanding hardware such as a capable graphics card, which many households do not have. Someone without that card should expect a different experience from the demonstrations.
There is also the question of privacy and control. Running a model at home means your prompts stay on your own devices rather than going to a company's servers. That is a real benefit for people who worry about data. Still, that benefit only holds if the setup is simple enough to maintain. For many users, the best next step is to test free software on the machines they already have before deciding whether a dedicated box is worth backing.