McKinsey’s sixth annual technology outlook, published a few weeks ago in September, opens with a line that sets the tone: the technology story of 2026 has moved off the screen and into the physical world. The report tracks fourteen trends in three families: the AI revolution, compute and connectivity frontiers, and cutting-edge engineering. They range from agentic AI and AI infrastructure to semiconductors, cybersecurity, robotics, space, life sciences and energy.

Only four of the fourteen are formally about AI, but AI drives most of the rest, speeding up robotics, chip design, cybersecurity and drug discovery. Five themes run through the report: machines are gaining autonomy; AI is producing breakthroughs faster than organisations can absorb them; the time available to defend against cyber-attacks has collapsed; hardware and software are being designed together for specific jobs; and, most bluntly, AI is hungry and the power grid is not ready. Investors have noticed: five trends are on course to more than double their funding in 2026.

Four threads explain where the money is going:

• Infrastructure: AI has entered industrial-scale deployment, and its infrastructure is very expensive. Equity investment in AI infrastructure and model design reached USD 145bn in 2025, then nearly USD 384bn in the first half of 2026 alone. Data centre spending could approach USD 7trn by 2030, while processors, high-speed memory, transformers and skilled engineers are all in short supply. Even the richest companies feel the strain: Alphabet’s quarterly capital spending nearly doubled in a year to USD 44.9bn, and it raised roughly USD 70bn in new equity and debt in a single quarter to pay for it.
• Governance: As agents move from advising people to acting for them, governance becomes a daily operating discipline rather than a policy document: who is accountable when an agent acts, and when must it stop and ask a human? It now covers money as well as behaviour. AI already takes about a third of what companies spend on new capabilities, and tokens, the units in which AI work is billed, have gone from a rounding error to a budget line to be managed.
• Security: AI has removed the head start defenders once had. More than three-quarters of software flaws are now exploited before they are made public, and in many organisations machine credentials outnumber human users by roughly a hundred to one.
• Energy: US electricity demand linked to AI is projected to rise from about 30 gigawatts in 2025 to more than 90 gigawatts by 2030, roughly what California uses today. Data centres could take 14% of US electricity by 2030, up from 3% in 2022. Worldwide, more than 2,500 gigawatts of energy projects are waiting for grid connections, and transformers can take more than two years to arrive. Reliable computing power is becoming as much a competitive advantage as talent or capital.

That is what the report says. What follows is what I think it means.

AI has become inflationary

In mid-September, I published an article called “The Cost of Thinking”, where I argued that AI had quietly become inflationary: cheaper by the unit, but dearer by the task. Model prices fell by more than 99% during 2024, from around USD 20 per million tokens to roughly USD 0.07, yet company spending on AI tripled in a single year. McKinsey’s outlook adds substantial evidence to that view. The report never uses the word inflation, and the conclusion is mine, not McKinsey’s. But its own language points the same way: long-term adoption, it notes, may depend less on what models can do than on whether running them can be afforded at scale.

The report describes a technology pulling in two directions. At the level of the task, AI is deflationary: one forecast expects the cost of running a very large model to drop by more than 90% between 2025 and 2030. At the level of the system, however, AI follows a classic macroeconomic timeline. In the short term, massive capital expenditure on data centres, energy infrastructure and hardware is inflationary—driving up demand and prices for scarce real-world resources. But once this infrastructure matures and reaches scale, the investment shifts from an inflationary cost driver to a deflationary force: economies of scale bring down marginal costs, expanding overall productive capacity and lowering the long-run cost of operating AI reliably, securely and at scale.

A sceptic might call this growth rather than inflation: companies spend more because they use more. That is partly true, but it misses the point. Inflation is the rising cost of getting the same result, and that is what the evidence shows. Some 93% of companies report going over their AI budgets, and while 89% use AI regularly, only 37% say it has done anything for their earnings. In software, agentic tools raised coding activity by 180%, but the number of releases shipped rose by only 30%. In some places the word is not even a metaphor: demand from AI data centres has already contributed to higher electricity prices in some regions, a cost borne by households and businesses that have never used an agent.

Why the system costs more

The reason is a fundamental change in what AI is. Two years ago, most business AI was a chatbot: a question went in, an answer came out, and the exchange ended at the edge of the screen. Today’s agents barely work on screen. They plan multi-step workflows behind the curtain—logging into systems, calling external software, reading from and writing to databases, and acting on what they discover across the open web.

Each step demands infrastructure the chatbot never required: network capacity, machine credentials, governance controls and uninterrupted electricity. In enterprise settings, a single agentic workflow already consumes 5 to 30 times more tokens than a chatbot query—and tokens are only the visible part of the bill.

An agent is valuable precisely because it connects things—and, at scale, every connection carries a social integration cost that does not necessarily fall when token prices do. The report treats connectivity, cybersecurity and energy not as side topics, but as conditions for progress that will decide how fast AI spreads—and where. Anyone can buy cheaper intelligence; no one can buy their way past a two-year wait for a grid transformer.

Costs that fall on others

Economic theory tells us that every cost described here is ultimately passed on to the consumer. AI is a commercial, for-profit enterprise, and its providers will recover their capital. That is not a critique—it is a reality that no leader, regulator or citizen should lose sight of.

Cyber-security, for example, shows the less obvious side of this inflation: part of the cost lands on people who never bought the AI. Every machine credential is a door, and the risk is no longer only that attackers will use AI. Agents with perfectly ordinary instructions may now walk through those doors on their own.

Some of this traffic is invited. For example, Morgan Stanley is opening its stock-administration platforms to its corporate clients’ agents, and agents in China already order drinks and book flights for consumers. Some of it is not. In June, an OpenAI agent got into a New South Wales government department without permission and retrieved non-public historical bushfire data, weeks after news of a similar breach at a federal portal holding Medicare statistics.

Episodes like this will become the norm, not the exception. Agents are starting to move through society the way people do: visiting websites, querying public services, dealing with companies and, increasingly, with other agents. That reach makes them useful, and it also makes the risk systemic: every organisation an agent might touch becomes part of the system, whether it chose AI or not. Governments will need to secure public systems never built for autonomous visitors; regulators, to set clear rules on accountability and incident reporting; large companies, to watch their own agents as closely as outside threats; small businesses, to find affordable defences; and individuals, to understand what their agents are allowed to do.

Who pays, and for how long?

For as long as society continues to pay the capital costs of integrating AI into its fabric. While the price of individual model outputs will continue to plummet as chips advance and smaller models handle routine tasks, data centres, power grids, networks and security layers require heavy capital expenditure. Every AI agent that reaches deeper into the economy compounds these demands. Inflation will persist through the build-out—and that build-out has barely begun.

None of this renders AI a poor investment, nor is it a justification for hesitation. Organisations that view system costs as a reason to wait will not avoid them; they will simply pay them later on terms set by competitors who built earlier and adapted faster. The system bill is the price of admission to the next economic era. What has ended is the illusion of AI as a cheap, self-contained software add-on.

Always expected, always limited

Preparing for this means acting on two fronts, whether you run a government department, a company or simply your own digital life:

• The first is cost: AI should be budgeted for the full system, not the price of a token. Expensive reasoning should be saved for work that truly needs it, and spending judged by what it achieves rather than how much AI it uses.
• The second is access: everyone should assume that agents will reach their systems, from a government database to a family’s online accounts. That starts with a deliberate decision about what is public, enforced by the systems themselves rather than left to visitors’ good manners. Agents should reach only what their task needs, with narrow, short-lived credentials and activity that is logged and watched. Every action should show afterwards who authorised it, whether the agent stayed within its limits, and how that authority passed between systems. Older systems never built for automated visitors should be reviewed first. Cybersecurity must be rebuilt on a new assumption: agentic interaction is always expected, and always limited.

This is not purely defensive. Security and governance are not brakes on AI; they are what make speed possible. Those who set their agents’ boundaries early, whether companies, public services or individuals, will be able to rely on them at scale while others are stuck in reviews, pilots or simple mistrust. Those who let outside agents in safely will reach people through channels others cannot yet open. As Morgan Stanley shows, access, properly governed, stops being a source of risk and becomes a source of value.

Final thoughts

Ultimately, I conclude that AI’s immediate impact will almost certainly be inflationary. As a transformative technology transcending industries at unprecedented speed, its integration into society demands massive capital expenditure—whether driven by direct conviction or competitive necessity. Yet, I view this high initial cost as the price of entry for a generational shift. As AI matures from a heavy infrastructure project into a seamless part of daily life, economies of scale will take hold, ultimately converting today’s inflationary capital investments into tomorrow’s long-term deflationary gains.