AI seen shifting to local edge models under stricter rules

Artificial intelligence developers expect a shift towards smaller, local models, stricter energy rules and greater government control over data in 2026, according to forecasts from quantum and AI software firm Multiverse Computing.

Roman Orús, Co-founder and Chief Scientific Officer at Multiverse Computing, said advances in model compression, Edge AI and regulation will drive changes in how AI is designed and deployed across consumer, defence, healthcare and public sector environments.

He said local AI systems on devices will challenge cloud-based services, as concerns over privacy, resilience and energy use rise.

Local models

Multiverse expects large language models running on laptops, phones and other consumer electronics to become more common.

"The industry is entering a period where local LLMs (Large Language Models) will become genuine competitors to cloud-based services. This fundamental shift is driven by ultra-compressed AI models that are enabling a transition from the data centre to the device, making Edge AI applications both practical and essential," said Orús.

He said manufacturers are embedding compressed models in hardware.

"This decentralisation offers two primary advantages. First, it drives hyper-personalisation on the device, as laptop manufacturers and consumer electronics providers are increasingly embedding compressed AI models directly into the hardware. This allows for the creation of an ultra-personalised digital assistant, everyone's very own ChatGPT, that runs locally, dramatically improving response speed and enabling users to work offline," said Orús.

He said users are focusing on privacy and data control.

"Second, privacy is the primary driver for consumers, offering the core benefit of data ownership. By processing sensitive information, such as conversations, documentation, and personal planning on a local machine, users eliminate the risk of data leaks and retain full control, bypassing privacy concerns associated with third-party cloud providers. Future architectures will also employ smart orchestration, using a small "Router AI" to intelligently balance power, deciding whether to process a request instantly on the local, compressed model or route it to the cloud for more complex, intensive tasks," said Orús.

Orús linked this shift with debates over ethics and bias in AI.

"While the urgency to adopt localised AI is inextricably linked to the growing global demand for robust ethical governance and the necessity to address inherent model challenges, such as algorithmic bias - areas where the industry must continue investing to make AI systems safer, more transparent, and less prone to unwanted filtering or over-restriction," said Orús.

Edge AI in defence and health

Orús expects Edge AI to gain ground in sectors that operate in low-connectivity settings.

"The technical race is now focused on deploying highly capable, compressed Large Language Models (LLMs) directly onto local devices, moving sophisticated intelligence out of centralised cloud environments and onto the network's edge. This leap is made possible by advanced compression techniques, which drastically reduce the size and computational requirements of AI models, while preserving strong operational accuracy," said Orús.

He said the impact will be strong in defence.

"While this enables unprecedented hyper-personalisation for consumers, its impact is most transformative in sectors, such as defence where this shift provides a tactical edge. Edge AI is essential because traditional military operations often take place in austere environments where network connectivity is limited or contested, making reliance on distant cloud processing impossible. By embedding AI directly into platforms like drones, ground vehicles, naval ships, and wearable soldier systems, military forces gain operational resilience and speed. Applications include immediate situational awareness, where AI detects threats and anomalies from real-time sensor feeds locally, and autonomous navigation for unmanned vehicles in hostile environments, reducing dependence on GPS," he said.

Healthcare organisations are exploring similar approaches, according to Orús,

"In healthcare, ultra-compressed AI models are becoming essential as they enable complex AI-driven diagnostics and patient history summarisation to run locally on devices like hospital workstations and secure private clouds.

"This localisation ensures that highly sensitive patient data remains within the organization's firewall, upholding strict data privacy and ethical requirements. By dramatically reducing the need for high-end hardware, these efficient systems make advanced AI assistance viable and cost-effective. This ultimately allows medical staff to achieve quicker, more secure diagnoses without risking external patient data exposure, meeting both operational and regulatory demands," said Orús.

Digital sovereignty

Governments are also looking at compressed and self-hosted models, according to Multiverse.

"The public sector is adopting compressed and self-hosted LLMs to meet mandates for efficiency, speed, and strict security compliance. AI tools are being used to automate time-consuming administrative tasks, document summarisation and the analysis of large policy consultation responses, enabling policymakers to quickly derive insights and ensure decisions are evidence-based," said Orús.

He said security and data control are key reasons for self-hosting.

"A major driver for deploying these models locally is the requirement for data security and digital sovereignty. Government departments routinely handle highly confidential data, and to mitigate the risks associated with third-party data management and potential data leakage, agencies are taking the important step of self-hosting LLMs within secure, internal networks. This approach ensures data is contained within the organisation's network, isolated from the public internet, and restricts access only to authorised users," said Orús.

Energy scrutiny

Orús said growing attention on AI power use will affect model design and regulation.

"The staggering energy footprint of AI is now a global concern, pushing model efficiency from an engineering challenge into a regulatory and environmental mandate. Driven by the immense power required for large-scale training and inference, Goldman Sachs estimates that AI will drive a 165% increase in data centre power demand by 2030," said Orús.

He pointed to emerging European rules over AI and data centre energy consumption.

"In response, regulators are building accountability into the system and the conversation is likely to intensify in 2026. The EU AI Act already requires General Purpose AI (GPAI) providers to document their models' energy consumption, while the 2023 EU Energy Efficiency Directive requires data centres to report on their energy consumption, water usage and use of renewable energy, and a 2024 EU scheme for rating the sustainability of data centres requires them to report on key performance indicators on energy and sustainability."

"In addition, the European Parliament plans to balance ambitions on competitiveness and concerns over decarbonisation. The Cloud and AI Development Act will aim to triple EU data centre capacity by the early 2030s, while the "strategic roadmap for digitalisation and AI for the energy sector" and the "data centre energy efficiency package", planned for early 2026, will address the energy impacts," said Orús.

He said these trends will shape which models succeed.

"This development forces enterprises and governments alike to prioritise AI models that deliver maximum performance with minimum power draw. By focusing on compression and efficiency at the edge, organisations are not only lowering operating costs and latency but are actively aligning with intensifying global ESG and climate reporting standards. This current and future regulatory landscape ensures that the models that succeed in the coming years will be those that are not merely powerful, but are also demonstrably energy-efficient and responsibly governed, a shift that places efficiency at the centre of AI competitiveness," said Orús.