Artificial intelligence is ceasing to be just a tool for conversation and is becoming a technology to realistically do our jobs and close business processes. We discuss how to invest wisely in AI infrastructure, cut costs and count the real return on deployments in an interview with Wojciech Janusz, EMEA Data Science & AI Horizontal Lead at Dell Technologies.
Klaudia Ciesielska, Brandsit: Over the past year, the market has been wowed by Generative AI, and now Dell is starting to talk about Agentic AI – autonomous agents performing tasks. However, many Polish companies are still at the stage of testing simple chatbots. Are you running away from technology too fast to move forward? Why do you think this is the moment to invest in infrastructure for autonomous agents, when companies often have yet to see a return on investment in simpler GenAI models?
Wojciech Janusz, Dell Technologies: Agentic AI is not a new technology. Rather, it is a natural transition from chatbots to agents that can perform specific actions for us. While large language models allow us to unleash the knowledge we have in the company, the real revolution starts when we turn knowledge into skills and actions.
I get the feeling that we’re all a little over-saturated with chatbots. At the end of the day, we don’t want to read the advice of a wise assistant who will tell us what to do, but someone who will do that thing for us – or at least relieve us of most of the task.
“Agentic AI is not a new technology. Rather, it’s a natural transition from chatbots to agents that can perform specific actions for us.”
We always consider the implementation of new technology in three aspects: people, processes and technology. Unfortunately, in the last two years we have too often focused on technology instead of the first two categories. AI agents are a way to bring it all together. It’s about integration with processes, it’s about human-machine collaboration and leveraging existing technology.
To answer the question: this is a very good moment, because only when AI starts to perform specific tasks for us will we be able to determine the actual yield, count the efficiency and better plan the next steps and implementations.
K.C.: There is a lot of talk about Sovereign AI and local models, but where does the point of profitability lie? At what scale of operations does it realistically pay for a Polish company to withdraw data from a hyperscaler and invest in its own AI Factory? Is this a solution only for corporates or a viable financial alternative for the SME sector?
W.J.: The break-even point lies much lower than we think. Few people realise how big a technological leap we have made in the last two years. That goes for both the hardware and the AI models themselves.
Firstly, the AI market has split. ‘Open’ models have emerged, giving us the opportunity to download and run them on our hardware in a secure controlled environment, but also to further customise them to fit our use case even better.
Simply downloading a model and running it won’t do much if it doesn’t meet expectations – and here too we have a big leap. Open models are catching up with the best closed cloud models in terms of capability and correctness. Of course, a model 1000 times smaller will not have the same capabilities as one in the cloud. But that is not the purpose and applications – instead of universal models that can speak all the languages of the world and solve every problem regardless of the domain, we can use specialised but smaller models and focus on specific activities. This gives us more flexibility and control over what is happening.
Instead of a single ‘universal genius’, we choose a team of expert AIs working together in a controlled and efficient manner. Appointing the necessary resources when required to solve a specific problem.
Such models with a developed ability to reason and break down problems into smaller tasks form the basis of AI agents.
Through high computing power and energy costs, models optimised to run on simpler hardware have also emerged. Here, the biggest contributions come from new architectures – such as Mixture of Experts (MoE), new training methods – including the use of Reinforcement Learning, and advanced ways of optimising the model itself.
The final element is the development of hardware platforms. Here too, new developments are emerging. We have a whole new category of hardware designed to use AI rather than train it.
It is estimated that the cost of running the model per token cost is decreasing by a factor of 10 each year, and so far, since the advent of GPT 3.5, this trend is managing to continue.
Tasks that only 2-3 years ago required powerful servers are today easily performed on an AI PC, for example, the Dell Pro Max with GB10 allows you to successfully work with models up to 200 billion parameters.
Of course, the appetite is growing and the list of tasks we want to do with AI is growing too, but it is becoming increasingly clear that the technology is no longer blocking us. The main question now is what we actually want to do with AI, not how to run it on our infrastructure.
K.C.: Poland has some of the highest energy prices in Europe and AI servers are extremely power hungry. Does the implementation of efficient AI solutions in Polish conditions force companies to overhaul their server rooms and switch to liquid cooling? Is it not the case that the main barrier to AI adoption in our region will not be the price of the server itself, but precisely the cost of electricity and the need to upgrade the cooling infrastructure?
W.J.: It will depend on the scale. Earlier we talked about changes in AI technology itself. We have new models, new uses, but also new architectures to enable AI to run on even modest resources.
“On a large scale, there will be no escaping energy costs and changes to the Data Centre infrastructure, but I am optimistic.”
My impression is that quite a few companies assume a massive cost of entry. Meanwhile, we can start AI projects with single applications. In any case, this is a very sensible and recommended approach: to limit ourselves to a few use-cases, well grounded in the realities of the company, with a clear budget and projected profit, and, most importantly, lying close together in terms of the technology and integration needed. This approach means that we can start modestly, with single devices, such as the Dell Pro Max with GB10 just now, without a huge revolution in our DC. Of course, when we are successful, these examples will be the basis for further steps while providing a solid foundation.
Start Small, Think Big, scale fast. This is the basis of our AI strategy.
Of course, on a large scale there will be no escaping energy costs and changes to the Data Centre infrastructure, but I am optimistic. I think for most companies it will be a gradual evolution rather than a revolution requiring drastic changes.
Investments can also yield very rewarding results. One new Dell PowerEdge server can replace up to seven older servers, and this translates into a reduction in energy costs of up to 65-80 per cent. Dell customer Wirth Research has reduced energy consumption in HPC environments by up to 70 per cent at its Verne Global data centres thanks to liquid-cooled PowerEdge servers.
K.C.: The great hardware replacement is underway, but does it make economic sense to buy PCs with NPUs (AI PCs) today when there are still few business applications that make real use of this chip? Aren’t companies today paying a ‘novelty tax’ for hardware whose potential will only be realised in 2-3 years, i.e. at the end of its life cycle?
W.J.: We are seeing a lot of interest in AI PC among business customers, with organisations looking to enhance their AI capabilities used locally.
Every computer we presented at CES 2026 is a computer with an AI processor and an NPU. This chip is not just for new applications yet to be developed – it is actively used during the user’s day-to-day work, providing benefits such as extended battery life – up to 27 hours of video streaming in the case of the XPS 14.
K.C.: Finally, a request for an honest forecast. Looking to 2026 and your experience of working with companies: in which area will Polish business (regardless of the industry) “overspend” with investments – spend too much money without a quick return, and which area will they drastically underestimate, which may negatively affect the performance of companies?
W.J.: I think in 2026 companies will be preparing more thoroughly for AI projects. We no longer want to have AI for the sake of having an AI project. There will be more cost-effectiveness analysis and looking for those applications that actually bring real benefits. We will also focus on the efficiency of using AI, not just the cost of purchase.
We have new metrics and tools to better choose the right approach to AI.
“A model that achieves 80% in a test having 8 billion parameters is considered much more impressive (and effective) than one that achieves 82% but requires 70 billion parameters.”
Until recently, we have only focused on maximum quality and speed.
Currently, we are increasingly looking for a reasonable compromise between quality and efficiency. An example might be the Frontier Pareto methodology: instead of looking only at the top of the scoreboard, we look for models on the ‘Pareto front’, i.e. those that offer the best ratio of quality (e.g. MMLU score) to model size (number of parameters) or inference cost. A model that achieves 80% in a test with 8 billion parameters is considered much more impressive (and efficient) than one that achieves 82% but requires 70 billion parameters.
Another example is a metric showing the real cost of an AI decision or action – Tokens per Decision/ Tokens per Action – A more efficient model will make an accurate decision using a few hundred reasoning tokens, while a weaker one may need several times as many.Choosing the former significantly reduces TCO and allows for a faster return on investment.
A final but very effective way of showing which way we are heading is the Cost per Resolved Task (or Cost per Resolution) metric: how much it realistically costs us to perform a specific activity using AI or, more commonly, an AI Agent.
In my opinion, 2026 will be the year of prudent AI project building, well-founded, grounded in reality and backed up by numbers.
