AI on a Card

AI on a Card

Written on 11 September 2026.

What if you could buy an AI model as a computer card? You put it into an expansion slot, install the software, and have a capable AI running locally. You would not have to keep paying for access to a remote service. The intelligence would come with the hardware.

This is the possibility that makes local AI much easier for me to imagine. A fixed model might sound like a disadvantage, because you cannot simply replace it with the next model. But if it is already good enough, why would that make it useless? Something that can help you read, write and reason today could still be useful five years from now.

There is real technology behind this idea. But there is also a considerable distance between demonstrating a model in specialized silicon and selling an affordable card containing the kind of reasoning people associate with a powerful cloud assistant.

Putting the model into the hardware

An AI model contains learned numerical values, usually called weights. On a conventional system, those values are stored in memory and used by processing hardware to calculate the response. A discrete GPU normally uses its own fast memory. Other systems use shared or ordinary system memory.

The alternative being pursued by Taalas is to build hardware around a particular model, combining its storage and computation. The company says this avoids HBM, the expensive high-bandwidth memory used in many large AI accelerators.

So “burned into silicon” is a useful informal description. It does not mean that a factory records every future answer. The machine still calculates answers from the input. What becomes fixed is much of the machinery and numerical information used to do that.

Nor does it mean that somebody must burn billions of future operations individually into every card. Semiconductor manufacturing creates many circuit features through patterned layers. The complexity has to be designed and manufactured, but the finished object can then perform calculations repeatedly.

A fixed model also does not mean a fixed conversation. You could give it a new document, ask a new question or provide facts that were not in its training. Whether it handles those facts correctly depends on the model, just as it does with a software model.

What has actually been demonstrated?

Taalas lists its HC1 demonstrator as running Llama 3.1 8B, with an 815-square-millimetre chip containing 53 billion transistors. Its product page specifies a 2.5 kW server and claims approximately 17,000 tokens per second per user. Those are manufacturer figures, not independent measurements.

That matters. A server demonstration is not evidence that an equivalent device already fits into a small, low-power consumer card. Lower energy per answer also does not automatically mean low total power consumption.

And speed is not intelligence. Making an eight-billion-parameter model answer extremely quickly does not establish that it reasons as well as a much more capable model.

Taalas also acknowledges quality loss from its initial weight compression. It says the largely fixed model supports low-rank adapters and configurable context, so “fixed” does not mean absolutely nothing can be adjusted.

On 6 August 2026, AMD announced an agreement to acquire Taalas, subject to closing conditions and regulatory approvals. AMD described plans to combine the technology with Instinct GPUs in system-level solutions. That announcement did not offer a consumer AI card, a retail price or an offline Astra product.

Are laptop and phone AI chips the same thing?

Generally, no. The NPU in an AI laptop is a neural processing unit: hardware specialized for the calculations used by neural networks. It normally runs models supplied through software. Microsoft explicitly describes bringing and optimizing different models for these devices.

The hardware is specialized, but one particular assistant is not permanently manufactured into it. That distinction is easy to miss when both devices are called AI chips.

Phone AI also does not necessarily require a separate graphics card with dedicated VRAM. Mobile processors can combine CPUs, GPUs and neural accelerators, drawing on the device's memory. Apple's description of its on-device foundation model, for example, discusses compressed model weights and software adapters, alongside separate server models.

So there are more than two choices. We do not have to choose between an expensive desktop GPU and a model permanently embedded in silicon. Programmable neural accelerators already provide another route. And an AI feature on a phone may run locally, remotely or through a combination of both; the presence of an AI chip does not establish that everything is offline.

Local AI already exists. The question here is whether a much higher level of capability could become affordable and convenient for an ordinary person to own.

Could there be an Astra card?

I use “an Astra card” here to describe the hoped-for experience: a local assistant with the level of reasoning and writing ability that makes a powerful cloud model useful. It is a hypothetical product, not an announced specification.

Putting such a device on a PCI Express card is a plausible engineering arrangement. PCIe is the modern expansion interface people often mean when they say a PCI slot. The card would need suitable drivers, application software, electrical power and cooling. A small M.2 module would impose tighter physical and thermal limits than a larger desktop card.

But the slot is not the main unanswered question. The question is whether the desired model can be implemented within an affordable amount of silicon, memory and power.

Hardwiring does not make an arbitrarily large model fit into one chip. The learned information still needs a physical representation. A larger design might require several chips, and those chips still have to communicate. Without a disclosed implementation, there is no sound way to calculate how many would be needed for the particular capability being imagined.

There is also working memory. A language model must handle the text being processed and intermediate results. Transformer systems commonly retain a cache associated with the conversation, and that cache can become substantial as the context grows. Embedding the weights does not eliminate this changing information.

Finally, a commercial cloud assistant can include tools and services beyond its language model. A local card would need local software to read files, search an archive or run code. Live internet information would still require internet access. Reproducing useful reasoning is one target; reproducing an entire online service is another.

When, and at what price?

There is no verified retail price or release date for the card being imagined here. Naming a year and a price as though they had been announced would turn a possibility into a promise.

Taalas has described a planned frontier-model implementation on its second-generation HC2 platform, targeting winter in its published roadmap. That is a development plan, not proof of delivery or a consumer-card launch.

My expectation is that commercial server deployments are a more likely first market than inexpensive cards for home computers. That is an inference from the announced direction and the economics: a business running the same model continuously has a strong reason to pay for lower operating costs.

Could consumer versions follow in the late 2020s or early 2030s? Possibly. I would treat that as a period in which to watch for products, not a forecast that one must arrive. A useful fixed assistant could appear before a card matching the much higher capability being imagined. The latter might take longer, or this particular approach might remain mainly a server technology.

A price such as €500–€1,500 is an attractive hypothetical target for a mass-market product. It is not an estimate supported by an announced design. An early specialist product could cost several thousand euros or more, and there is no basis yet for putting a reliable price on Astra-level capability in this format.

The cost would depend on chip size, manufacturing yield, packaging, memory, the circuit board, cooling, software support, model licensing and the number sold. Large production runs can spread development costs over many buyers. They cannot make silicon area or manufacturing defects disappear.

There is a business problem as well. A company has to decide that enough people will want the same model for long enough to justify manufacturing it. If buyers insist on changing models every few months, programmable hardware becomes more attractive.

But that is exactly where I think the ordinary user's perspective can differ. I would not necessarily need the latest model. I would need one that was good enough.

Good enough five years from now

If a model can explain a difficult passage, help organize an article or examine a document today, the existence of a better model tomorrow does not automatically remove those abilities. A fixed assistant could continue doing useful work while the cloud moves ahead.

Its stored knowledge would age. Its mistakes would remain unless the product allowed some means of correction. New software and file formats could create compatibility problems. But new documents could still supply information in context, and the surrounding software could potentially be updated without replacing the fixed core.

For lasting independence, the product would also need to work without continuing authorization from a server. Owning a card would not, by itself, guarantee freedom from subscriptions or activation requirements. That would depend on how it was sold and supported.

I think this could be an important breakthrough for local AI if the engineering and price work out. The basic idea is convincing: manufacture a sufficiently useful capability into a device that a person can keep using. Specialized chips are not new, but applying this degree of specialization to a capable language model opens an interesting possibility.

It may also be that smaller models and better programmable accelerators reach the same practical result first. Either way, the point that interests me remains. Local AI does not have to stay equal to the best remote system. It has to become good enough that the intelligence you can own is worth having.

References

• Opening Terminator's Skull, taking out Chip | Terminator 2: Judgment Day | , https://youtu.be/IMapohyDPXg

AI Disclosure: Parts of this page may have been created, edited, or assisted by artificial intelligence tools (such as ChatGPT or other language models). All AI-assisted content is reviewed by a human before publication. For questions, contact the site administrator.