Image generated by Deeptech Times using Google Gemini
The AI industry likes to frame its future around increasingly powerful models and ever-more-capable AI agents. However, beneath the headlines about reasoning breakthroughs and autonomous agents lies a less glamorous reality: enterprises are struggling to operationalise AI because the bottlenecks are no longer merely algorithms or compute.
According to WEKA’s chief AI officer Val Bercovici, the industry’s biggest constraint may be something even more fundamental.
“It’s really energy,” he told Deeptech Times during his visit to Singapore for SuperAI 2026.
The observation is both simple and profound. Enterprises today can access some of the world’s most advanced AI capabilities only through hyperscale cloud providers because few organisations possess the power capacity required to host frontier models themselves. Running state-of-the-art AI systems increasingly demands tens of megawatts of data centre capacity, or levels of power that traditional enterprise infrastructure was never designed to accommodate.
The result is a growing disconnect between AI ambition and operational reality. For many organisations, AI adoption has begun with individual productivity tools and knowledge assistants. Yet the next phase of replacing business processes with autonomous agents that continuously interact with enterprise systems and data presents a vastly different challenge.
The problem is not simply where data resides. It is whether organisations can generate, store, process and govern intelligence at scale.
The new infrastructure war
For much of the past decade, enterprise technology revolved around cloud-native transformation. Applications were rearchitected to become distributed, scalable and accessible from anywhere.
AI-native transformation is considerably more disruptive. Bercovici likens the current moment to the transition from horses to automobiles.
Early cars initially operated on the same dirt roads. Only later did entirely new systems of highways, fuel stations and logistics emerge around them. AI appears to be following a similar trajectory.
Today, many organisations are simply inserting AI into existing workflows. Tomorrow’s enterprises, however, may fundamentally redesign business processes around teams of autonomous AI agents, or what Bercovici calls “agent swarms”.
These systems do not sleep, do not require breaks and can execute complex tasks simultaneously at machine speed. Rather than eliminating work altogether, Bercovici believes this shift will likely create entirely new markets and revenue opportunities.
History suggests he may be right. The internet created businesses that previously could not exist. Smartphones generated entirely new industries. AI-native operating models may prove equally transformative. But they also demand entirely new infrastructure assumptions.
Welcome to the age of agent traffic
One of the least appreciated developments in AI is that machine-to-machine interactions are rapidly overtaking human prompts.
According to Bercovici, the overwhelming majority of token traffic today is already agentic in nature.
Thousands or even millions of autonomous agents simultaneously accessing data, generating responses and initiating actions create unprecedented infrastructure demands. Every interaction generates memory requirements, processing loads and governance considerations.
At the centre of this challenge lies a seemingly obscure concept: memory efficiency.
Much of AI’s cost and scalability problem stems from redundant re-computation. The ability to retrieve and reuse previous context efficiently has become one of the most important determinants of cost and performance.
This is why AI infrastructure discussions are increasingly shifting from compute acceleration towards memory optimisation and data architecture. Ironically, making systems more efficient often creates even greater demand.
Economists refer to this phenomenon as Jevons Paradox: improvements in efficiency lower costs and encourage increased consumption. AI appears to be following precisely this pattern.
As models become cheaper and more efficient, organisations simply want to do more with them. The race therefore is no longer merely about building bigger models but creating infrastructure capable of supporting exponentially growing intelligence demands.
Sovereignty enters the AI conversation
The infrastructure challenge is also geopolitical.
Governments and enterprises increasingly recognise that intelligence itself is becoming a strategic capability. Data sovereignty concerns have accelerated initiatives aimed at ensuring that AI capabilities, sensitive information and digital infrastructure remain within national jurisdictions.
AI has also dramatically changed the threat landscape. Autonomous systems can now generate attacks at scales previously unimaginable. The emergence of agentic cybersecurity means future conflicts may increasingly resemble battles between defensive and offensive agent swarms.
In such environments, infrastructure capacity becomes a strategic advantage. Whoever can process more intelligence loops faster gains an operational edge.
This reality suggests that future national competitiveness may depend not only on gross domestic product but also on what Bercovici describes as “gross token product” – the ability of economies to generate, process and apply intelligence at scale.
Scaling intelligence is not enough
The AI industry remains obsessed with scaling. Notwithstanding, as organisations distribute workloads across multiple models and increasingly autonomous systems, governance becomes exponentially more complicated.
Which models should be permitted to process sensitive information? How should organisations audit decisions made by agents? How can enterprises distinguish between human and machine identities? These questions are no longer theoretical.
Research increasingly shows that AI systems can behave opportunistically, manipulate information and optimise for outcomes in ways humans did not explicitly anticipate.
As Bercovici notes, intelligent systems can become remarkably Machiavellian if incentives are poorly designed. The answer is not less AI. It is better guardrails.
Future enterprise architectures will require rigorous policy frameworks, robust auditing mechanisms and increasingly sophisticated governance models capable of monitoring machine actions continuously.
In many situations, AI systems may increasingly supervise other AI systems. Software development is already moving in this direction. AI-generated code is now frequently reviewed by additional models before reaching human oversight.
The future of governance may therefore become less about humans performing every task directly and more about humans establishing policies and supervising systems of autonomous intelligence.
The misconception of 2026
If there is one idea that may look naïve by 2030, it is today’s tendency to equate AI with chatbots, according to Bercovici.
That assumption already feels outdated. The future of AI is unlikely to be defined solely by language models. It will increasingly encompass systems that understand physical environments, process multimodal information and reason about the world with unprecedented contextual awareness.
The real transformation, then, may not be bigger models or smarter agents. It may be recognising that intelligence itself has become infrastructure.
And the organisations that thrive in the next decade may not necessarily be those with access to the most powerful algorithms. They may simply be the ones that understand that AI’s most important building blocks are energy, memory, trust and the ability to govern intelligence responsibly.
