TWIN-TECH – Twin Transition for VET Resilience
Project number: 2025-1-SI01-KA220-VET-000360870
Program: Erasmus+ KA220-VET
Project website: twin-tech.eu
Do we need a Bugatti for buying groceries?
The data center building boom in the United States is worth watching. The country is building huge facilities to support its AI ambitions. These facilities require not only new buildings but also substations, cooling plants, high voltage lines and power plants.
It also means resistance.
Whereas so far in the US power demand, water consumption, noise, land use and the financial and technical effort to build the infrastructure are frequently considered to be a barrier for data centers, communities across the country are already resisting these enormous projects today. Some of these have already been delayed or even been put on hold, in some cases even completely stopped.
Again, it seems that just building more power and more cooling would solve these problems.
Is it necessary to use this amount of computing in the first place?
Imagine someone going grocery shopping in a Bugatti. The car is a Bugatti – it’s a really, really fast, exotic[RB1] , fantastic car to drive. The owner of this car will arrive at a supermarket in this car in a rather spectacular manner. He’s got to park it and so on and then he can go in and buy some milk and some bread and some potatoes and so on for the week, and then he’s got to walk back out to this car.
There is nowhere to put half of it. Next to him pulls up in a Renault. There is a small crowd, but no one pays attention to the handsome automobile. The owner of the car opens the boot of the car, puts in the groceries and off he drives.
For going to the supermarket and buying groceries for example, a Renault is much more suitable than a Bugatti.
That is roughly where we may be with AI.
Today’s GPUs are huge technical wonders and can be used for just about everything: training a language model, generating images from text, running simulations of complex systems, and much more. Yet, once a model has been trained, most of the work that needs to be done with that model is, in fact, repetitive.
After a large compute resource has been set up to train up a model, the model is typically run over and over again to serve up queries. This is very different from general computing where programs execute in vastly different ways every time they are run. Therefore, using an extremely flexible resource to execute a lot of repetitive work is not efficient.
This is where Taalas becomes interesting. A Canadian firm from Toronto that has taken a different approach. Taalas.com[RB2]
We[RB3] build a model into a chip, the model is the computer as Taalas so eloquently puts it.
A GPU is like a super-talented chef who has the best kitchen in the world. He can make you just about any dish you desire provided you give him a recipe and the required ingredients to complete it. Taalas is like building a specialized cooking machine for a particular recipe – while it cannot make you all the dishes on the menu it completes the one it was designed to prepare very fast without needing to get up to retrieve more ingredients from the cabinets.
The HC1 system from Taalas has been used to demonstrate this new approach, with the HC1 system running Llama 3.1 8B. This achieves roughly 17,000 tokens per second per user, and in designing around the memory constraints of conventional high-end AI accelerators, provides an interesting new direction for high performance computing. In August 2026 AMD announced that it would acquire the company Taalas and use its AI chips in conjunction with its existing line of AMD Instinct GPUs for deep learning.
That last part may point toward the more interesting future. Not Taalas instead of GPUs. Both.
A mixed datacenter with general-purpose machines (GPGMs) for rare tasks like training and special processors for very common work (like a pattern recognition in a factory).
The set-up of a materials research institute could look like this: established models like to be analyzed for molecular structures of chemicals are run on dedicated hardware, a hospital works with specialized systems for validated models that analyze CT scans, a pharmaceutical company can make do with stable protein-analysis models and a factory with dedicated hardware for its machine vision, quality control and predictive maintenance models.
The Bugatti needs to be driven fast the first time to reach that fantastic shopping result, but after that, driving that same route in a Renault is a lot more practical.
When that same task needs to be performed 10 million times, it should be handed off to the Renault.
This could also change where AI runs.
Gorenje[RB4] does not necessarily need a giant American-style data center to use AI for production, maintenance or quality control. Turna[RB5] might need it for a narrow set of manufacturing tasks. Premogovnik Velenje [RB6] could use local systems for industrial monitoring, equipment data or technical documentation.
If specialized hardware makes useful AI smaller, cheaper and less power-hungry, some computing could move out of giant centralized facilities and into companies, hospitals, universities and regional institutions.
The AI would be where the data is.
Sensitive production information could remain inside a company. Medical data could stay within hospital infrastructure. Researchers could keep control of valuable datasets. Organizations would not need to send everything to a distant cloud provider simply to use AI.
It could also spread access. Instead of serious AI capacity being concentrated in a few giant campuses, useful systems could be distributed across thousands of smaller locations.
There are limitations.
Specialized hardware is specialized. [RB7] A GPU can be repurposed tomorrow. A chip designed around one model cannot simply become something completely different after a software update. Models change quickly, while manufacturing silicon takes time and money. Taalas is also still proving its approach, and its headline figures come from its own technology demonstrator.
Not every workload should be specialized. Research, new model development and unpredictable tasks will continue to need flexible hardware.
But that is not an argument against specialization. It is an argument for using the right machine for the job.
The benefits could be substantial: data centers that need less electricity and cooling, lower costs for repetitive AI work, smaller local deployments, greater control over data, less dependence on a handful of cloud companies, and better resilience when an outside service or internet connection fails. It could also reduce the habit of solving every new AI problem by adding another rack of expensive general-purpose hardware.
The American data center boom is showing us what happens when enormous computing demand meets finite electricity grids, water systems, land and communities. The lesson does not have to be that AI itself is unsustainable. It may be that we are still using an expensive, resource-hungry way of delivering it.
Sometimes you really do need a Bugatti. You need the speed, flexibility and engineering excess.
But if the job is buying milk and potatoes, perhaps the future of AI does not need another hypercar.
Perhaps it needs a Renault with a decent boot.
Do we need a Bugatti for buying groceries?
Big data centers in the US have become a kind of poster child for AI, but the question of infrastructure is also relevant in Slovenia. SAŠA region is facing industry’s transition to low-carbon economy and companies are seeking to modernize their production in various ways. For companies from manufacturing, energy and other technically oriented sectors the question is not so much if they use AI, but rather what kind of computing infrastructure they need for it.
By building new data-centres for AI more infrastructure is created to ease some of the constraints, but other constraints emerge, such as higher demand on the electricity grid, water supply, land and surrounding communities. Most AI applications however don’t require such powerful general-purpose computing. Grocery shopping for example in a Bugatti is very fast, but not very practical for shopping for milk, bread and potatoes for a small family.
Comparing going grocery shopping for milk, bread and potatoes in a Bugatti to using a small family car for daily errands is a good analogy for evaluating the use of infrastructure for repetitive tasks. A modest family may do the everyday job better and, in some cases, faster.
That is roughly where we are with AI hardware.
Today’s GPUs are to some extent highly flexible. But once a model has been trained to recognize language, generate images or perform scientific calculations, running the model again and again to answer new queries or performing analyses of a similar nature can become a rather monotonous task. Using highly flexible, expensive hardware for every repeated task may therefore be inefficient.
This is what makes Taalas, a Toronto-based chip company, interesting. Its approach is to build a model directly into specialised hardware, in effect, making the model part of the computer itself. Its HC1 system, demonstrated with Llama 3.1 8B, is designed to avoid some of the memory bottlenecks of conventional AI accelerators. In August 2026, AMD announced plans to acquire Taalas and combine its technology with AMD Instinct GPUs.
It seems that in the future, we shall have two types of hardware: highly flexible GPUs for training as well as for unusual uses, and dedicated processors for work that is repeated millions of times.
That could matter for regions such as SAŠA. Gorenje, the Slovenian home-appliance manufacturer, may not need a giant American-style data centre to apply AI in production, maintenance or quality control. Turna, a local manufacturer of components and systems for household appliances, could use specialised computing for narrowly defined manufacturing tasks. Premogovnik Velenje, the Velenje Coal Mine, could potentially use local AI systems for industrial monitoring, equipment data or technical documentation.
Specially designed hardware will make useful AI small, cheap and energy efficient. Therefore, in addition to working in giant data centers, AI could be running in companies, hospitals, schools and other institutions in the country and in the SAŠA region.
There are trade-offs. Specialised hardware is less flexible: a GPU can be reassigned to a different workload, while a chip designed around one model cannot be transformed as easily when models change. Manufacturing silicon also takes time and money, and Taalas’ performance claims are based on its own demonstrator. Flexible hardware will therefore remain essential for research, model development and unpredictable tasks.
The case for specialisation is not that every AI workload should move to a dedicated chip. It is that the hardware should match the task.
For Slovenia and regions such as SAŠA, that could mean using AI without copying the scale of the American data-centre model. Some tasks may need the Bugatti. But when the same route must be driven ten million times, a Renault with a decent boot may be the more practical choice.