Cisco AI Summit: After 2025's AI breakthrough comes 2026's

Contents

At the second Cisco AI Summit, the tech industry assured each other that 2026 will finally be the year of production-ready AI. Amidst remarkable insights and plenty of wishful thinking, one thing stands out: the battle for AI infrastructure has begun.

Cisco is clearly positioning itself as an infrastructure provider for the AI era, and a six-hour stage show with Jensen Huang, Sam Altman, and Fei-Fei Li is not a bad backdrop for that. Nevertheless, a closer look is worthwhile: behind the usual superlatives, there are quite interesting shifts in the debate.

From Chatbots to AI Factories

The most striking narrative of the day came from Nvidia's Jensen Huang. AI is not a feature, but a reinvention of the entire computing stack. Instead of writing code, developers will define intentions in the future – the transition from explicit to implicit programming. What was new was less this thesis than the consequence Huang draws from it: companies don't need individual AI tools, but so-called "AI Factories" – integrated systems of computing power, network, and security that produce intelligence industrially.

Huang's advice to executives not to primarily ask for classic ROI in the early stages of AI initiatives, but to proceed exploratorily and find out where AI can have the greatest strategic leverage, was also remarkable. Return metrics like those of a classic ERP rollout cannot be applied here.

The Agent Promise

Sam Altman delivered the expected escalation: AI will evolve from a tool to a team member that independently operates computers, writes software, and completes complex tasks end-to-end. He predicted that by the end of 2026, the range of problems that AI systems can meaningfully address will expand massively. Aaron Levie from Box added to this perspective with the thesis that companies could deploy many times more AI agents in the future compared to the number of their employees.

Such statements sound impressive, but remain in need of explanation as long as it remains unclear how progress is concretely measured. Altman relativized the vision and admitted that the biggest bottlenecks are currently not with the models, but with energy, infrastructure, and the sluggish adoption of AI in organizations.

Infrastructure as the Real Battlefield

Almost all of the prominent speakers ultimately landed on infrastructure. Google's Amin Vahdat put it particularly clearly: it is not primarily the models that determine success in the AI competition, but computing power, networks, and energy supply. The departure from general standard architectures enables significant efficiency gains – on the order of a factor of ten – but requires hardware cycles of currently around three years to be significantly shortened. Vahdat even brought space-based data centers into play as a long-term thought experiment to overcome physical scaling limits.

AWS CEO Matt Garman remained more down-to-earth. Many AI projects fail less due to technology than because companies do not define clear success criteria in advance. Progress is not made through individual clever experiments, but through systematically built context – for example, in the form of data, processes, and integrated expertise. A sober insight that was almost lost in the noise of grand visions.

China as an Uncomfortable Benchmark

The most geopolitically explosive passage of the summit was provided by Intel CEO Lip-Bu Tan. China has used its restricted access to high-end GPUs to build its own CPU and GPU ecosystems and systematically reduce technological dependencies. The Western lead still exists, but could shrink – also through targeted personnel recruitment, for example at Huawei.

Tan's real point went further: the differences in AI progress are less of a technological than a regulatory nature. While the expansion of energy infrastructure in Western democracies is hampered by lengthy approval procedures, China implements political decisions into construction projects much faster. Anne Neuberger and Brett McGurk added to this analysis from a security policy perspective: if democratic states significantly slow down their own AI development while geopolitical rivals scale faster, this could lead to a real strategic disadvantage.

How Software Development is Changing

Beyond the grand narratives, there were two areas where the changes are already tangible. In software development, several speakers reported significantly increasing proportions of AI-assisted code creation. Microsoft CTO Kevin Scott soberly stated that the bottleneck has shifted – away from pure code creation towards evaluation, quality assurance, and whether software solves the "right" problem.

Mike Krieger of Anthropic described how the human role is shifting more towards product vision and architecture, while Figma CEO Dylan Field predicted that designers could prospectively influence productive codebases directly via design interfaces.

Where is the Reliable Data?

What was largely absent from the summit was an honest assessment of the past year. 2025 was also announced as a breakthrough year. How many of the promises made then were actually kept? Where was the reliable data on productivity or ROI effects, the collection of which even proponents like Jensen Huang consider premature?

HUMAIN CEO Tareq Amin openly stated that the productivity gains achieved so far are limited in many places because AI is often only added on top of existing legacy platforms. His approach of building an entirely new, AI-centric operating system from scratch is radical – whether it is practical remains to be seen.

Also noticeable: the term "hallucination" was hardly mentioned. Issues of trust were discussed under the label of security, but primarily in the context of cyberattacks and geopolitics – less regarding AI systems that simply produce incorrect or misleading results.

The Cisco AI Summit 2026 was more insightful than many comparable industry events – less because of individual visions than because of the lines that emerge between them. The debate is shifting from the question "What can AI do?" to "Who has the infrastructure to run AI at scale?". For companies defining AI strategies today, the realization remains: the biggest problem is not the models, but energy, data, integration – and the courage to fundamentally rebuild existing processes. Those who want to watch the interviews from the Cisco AI Summit can find them thematically prepared individually here.

(nie)