An AI feature is often sold as a model and a prompt. The feature only runs if someone has data centres, chips, networks, power and people who can operate them. McKinsey's 2026 technology outlook puts that infrastructure in the foreground. For a UK firm, the lesson is not to build a data centre. It is to treat hosting, data access, security and cost as part of the AI decision, before a pilot is declared a success and pointed at the whole customer base.
Infrastructure has become a business question because the bill, the latency and the privacy risk show up in operations, not in a research slide. A demo on a vendor's tenant can hide all three.
What the 2026 outlook is pointing at
McKinsey's Technology Trends Outlook 2026 says spending on AI infrastructure doubled in a single year, and that energy technologies attracted nearly $200 billion of investment in 2025. The report examines 14 trends. Among them are AI infrastructure and model architectures, agentic software development, and AI for scientific discovery and engineering. McKinsey also describes AI moving into robotics, scientific research, cybersecurity and semiconductor design, beyond the usual software products.
Those are global investment figures, not a forecast of a UK company's cloud invoice. They explain why capacity, power and chips are now part of the technology story rather than a footnote. McKinsey's State of AI and the Global Tech Agenda 2026 are the companion reads if you want the adoption picture next to the infrastructure one. The direction is consistent: more AI work means more demand on the systems underneath it.
What a UK firm actually has to decide
Performance, cost and availability depend on where the workload runs. A team should know where the data is stored, whether an AI service is hosted by a vendor or in an environment the firm controls, and what happens when usage grows. Security controls matter at the point information is sent out. Customer records, staff files and unpublished commercial material do not become harmless because they were pasted into a chat box. UK GDPR still applies to personal data, including data sent to a model.
Energy use and computing cost belong in the same plan. A pilot that is cheap at ten users can be a different product at a thousand, especially if every request calls a large model and keeps a long context. Agentic software, which takes more steps and calls more tools, multiplies that cost. Google's Gemini agent is an example of multi-step work that can keep running after the user has closed the laptop. Someone pays for those steps. GitHub's storage rebuild shows the same pressure on development infrastructure when agents increase activity.
- Name the business problem before you name a model.
- Compare a hosted service with an option you operate, on privacy, performance and cost.
- Use a small, defined dataset for the first pilot, not a dump of the customer database.
- Write down what may be sent to an external AI service, and what must stay inside.
- Estimate the cost of a busy month, not only the cost of the demo.
You do not have to own the data centre
Most UK businesses should not respond to an infrastructure trend by commissioning hardware. They can use hosted models and cloud capacity, provided they have asked the questions above. A small pilot on a limited dataset is how a team learns the real requirements: which documents the system needs, how slow or fast it is, what it gets wrong, and what the invoice looks like. That evidence is what makes a later commitment informed. A slide that says the market spent heavily on AI infrastructure is not evidence that this firm should.
Skills sit next to capacity. Someone has to judge output, watch access, and turn the pilot off if it is wrong. McKinsey's list includes people who can build and operate the systems. In a smaller firm that person may be the same developer who looks after the website and the integrations. The role still has to exist. An unused API key is not an AI strategy. Neither is a pilot that cannot name which records it used, or what a wrong answer would cost the business if it reached a customer.
Strategy includes the plumbing
AI strategy and infrastructure strategy are the same conversation. Data access, security, computing capacity and the skills to supervise the result decide which experiments are ready to become daily operations. Firms that skip that conversation discover the constraint in production: the data was not available, the vendor terms did not fit, or the cost only made sense as a demo.
Agentic development, which McKinsey lists among the trends, raises the stakes because the software does more than answer. It takes steps. Each step is a hosted call, a tool, and a possible mistake. The infrastructure check and the permission check belong together.
Start from the problem and the constraint
Cloud solutions are how most UK firms will host this work, with a clear line around what leaves their environment. AI development should start from a defined task and a dataset the business is willing to use, not from a commitment to match global infrastructure spend.
Web Works Rise is a UK-led team in Birmingham, with a development hub in Tunisia for delivery. If you are deciding whether a hosted assistant or a tighter integration fits the job, talk to the team about the data, the volume and the outcome you need. The outlook describes the industry. The pilot should describe your business.