Artificial intelligence (AI) can be used in computers in various ways. Services such as chatbots and assistants like ChatGPT, Gemini, and Copilot can operate in the cloud, while PCs with AI are beginning to incorporate specialized hardware to perform specific tasks directly on the device.
However, these two approaches do not necessarily have to compete with each other. According to Carlos Buarque, Marketing Director at Intel Brasil, the trend is that AI applications will combine local and cloud processing depending on the characteristics of each task. He stated in an interview with Olhar Digital: 'We believe that all AI applications will become hybrid applications.'
One reason for moving AI processing to the computer is the desire to avoid sending certain data to the cloud. Buarque points to privacy and the cost of using AI services as factors promoting local execution. In his words, language models installed on the PC itself are capable of performing operations such as text proofreading, translation, and data analysis.
Performing part of the workflow locally can reduce costs because some tasks no longer depend on cloud processing. In this scenario, the model is loaded onto the machine and performs inferences locally, without requiring an internet connection for that specific task.
Privacy is especially important when AI works with personal or confidential information. Buarque shared that he used AI to analyze data from his tax return, including information about stock transactions. He noted that he would not want to send such content to a public model without guarantees of data confidentiality.
The director also provided an example from healthcare where local processing makes sense. For instance, during a consultation, transcription could be performed directly on the computer before being included in the doctor's medical record. He emphasized: 'I transcribe the consultation. This is critical, confidential information of my patient. It will be performed locally and transcribed on my computer to go into my patient's file on my computer. I will not run this in the cloud, you understand? Because there is a confidential relationship between the doctor and the patient.'
Furthermore, there are situations where connectivity is a decisive factor. At a factory, Buarque gives the example of using computer vision to detect defects on a production line. In this case, the analysis can happen locally so that the work does not depend on the network. He asserts: 'You will not run this in the cloud; you run it locally due to performance.'
A similar principle applies to security systems. Buarque mentions cameras capable of identifying certain situations and making decisions locally, such as activating an alert or alarm.
Local processing does not negate the need for cloud services. Buarque himself mentions chatbots as applications that continue to function well in the cloud. He says: 'I run my Copilot, Gemini, ChatGPT in the cloud.'
The cloud remains an option when part of the work requires processing that should not happen on the device. Instead of a final choice between architectures, Buarque believes that applications can distribute tasks between the computer and external servers. He concludes: 'There are different types of applications that are very sensible to run locally, and there are others for which I would say the meaning is hybrid. There will be part that you run locally, and another that you run externally due to performance.'
This model can also be applied to separating different types of information within one application. Buarque gives examples where more sensitive data remains on the device, while publicly available information or tasks requiring external searches are processed in the cloud.
The discussion of local processing is not limited to placing a chatbot on a laptop. Buarque argues that new applications can use the ability to run models directly on the device to create an experience different from the current one. In games, for example, he mentions non-player characters (NPCs). In the expert's opinion, locally executed AI models can make these characters smarter, make them behave more naturally, and react better to the player's style.
With AI, these non-player characters can become much smarter and act naturally. The game can adapt much better to your playstyle and become much more personalized for you.
Another example appears in video conferences. Buarque notes that platforms like Microsoft Teams use CPU, GPU, and NPU to perform functions such as noise cancellation and background blurring. In this case, the advantage is not only in the visible or audible result but also in energy efficiency: this task could be done by the CPU, but it would require more energy.
Thus, the combination of local and cloud processing should not be determined by the complete replacement of one technology with another. For Buarque, the place of execution for each part of the AI will depend on the characteristics of the application, including factors such as privacy, cost, performance, and the need to work offline.
During the interview, the expert emphasized that the choice between local and cloud processing must be based on the specifics of each application. In his opinion, different tasks can be divided between the computer and external servers depending on factors such as performance, cost, and privacy.
