Discussions in Davos showed that AI is entering a phase of operational maturity. Jeetu Patel, President and Chief Product Officer at Cisco, said that after a year of experimentation with chatbots and AI agents in 2025, the coming year will see these systems move into production environments. In his view, “2026 will be the year of agentic AI in production,” alongside the first practical applications of so-called physical AI and large world models. At the same time, Patel pointed to three major constraints on further growth: infrastructure (shortages of computing power, energy, and network bandwidth), trust in AI systems, and access to data, including machine-generated data.
In the short term, energy is expected to be the most limiting factor. Varun Sivaram, founder and CEO of Emerald AI, warned that as early as 2026, a lack of grid connection capacity could begin to slow the expansion of data centers in the United States and India. He noted that while investments of tens of gigawatts in new data-center capacity are planned, only a portion of that power can be connected to the grid quickly. By contrast, China could have as much as 400 GW of reserve capacity for AI infrastructure by 2030, according to his estimates. Emerald AI is developing solutions for real-time, flexible management of data-center power consumption, enabling faster grid connections without increasing the burden on local users.
At the same time, AI is becoming increasingly important as a tool for business decision-making amid volatile supply chains and rising geopolitical uncertainty. Chakri Gottemukkala, co-founder and CEO of o9 Solutions, argued that large language models alone are not sufficient for operational use in enterprises. He said the key will be to combine them with structured corporate knowledge through so-called neuro-symbolic approaches, allowing advanced analytics to move from small specialist teams to operational staff and frontline managers.
Security is also emerging as a major challenge. Jonathan Zanger, CTO of Check Point Software, emphasized that many AI solutions were not designed with resilience to attacks in mind, creating new threat vectors. He noted that boards and executive teams are significantly increasing budgets for securing AI systems, treating this as a strategic priority. Patel added that as models scale and infrastructure becomes more distributed — now spanning clusters of hundreds of thousands of GPUs linked into virtual “superclusters” across multiple data centers — architectures must be designed to ensure both predictable performance and resilience to failures and cyberattacks.
The conclusions from Davos are consistent: in 2026, the key to further AI expansion will no longer be the capabilities of algorithms themselves, but the ability of energy systems, networks, and security frameworks to support industrial-scale infrastructure, and the readiness of enterprises to embed AI into critical decision-making processes. How quickly and safely the shift from pilots to widespread deployment occurs will depend on resolving these bottlenecks.
