Predictions about the future of technology usually fail in one of two ways: they assume today’s constraints disappear, or they assume nothing changes at all. The useful approach is to look at what is already working, what is expensive, and which bottleneck has to break before the next step becomes normal.
AI moves onto the device
The clearest shift is from “send everything to a data centre” to “run what you can locally”. Phones, laptops and wearables already ship with neural processors capable of transcription, translation, photo editing and small assistants without a round trip to the cloud. The direction of travel is obvious: smaller models, better compression, and hardware tuned for inference.
What this changes: latency, privacy and cost. What it does not change: the largest models will still live in data centres, because frontier training and long-context reasoning need memory and power that a pocket device cannot supply. Expect a split — local models for fast, private, everyday tasks; cloud models for heavy reasoning.
Computing becomes ambient
Today you open an app. In five years more interactions will start with a sentence, a glance or a sensor event. Glasses, earbuds, watches and home devices are converging on the same idea: the computer is present, but not the centre of attention. The blocker is not the model. It is battery life, social acceptability, and the fact that voice interfaces are still poor at correction and privacy in public spaces.
Agents that do rather than answer
The current wave of AI assistants answers questions. The next wave acts: booking, filing, comparing, monitoring and escalating. That requires something harder than a better model — reliable permissions, audit trails, rollback, and a clear answer to “who is responsible when it gets it wrong”. Expect agents to arrive first in narrow, well-instrumented workflows such as support triage, software testing and back-office processing, not as a general-purpose digital employee.
Robots leave the demo floor
Robotics has quietly improved in warehouses, farms and factories. The humanoid robot demonstrations are impressive but still expensive and fragile outside controlled environments. The realistic path is boring: more autonomy in structured spaces, better manipulation for specific tasks, and remote supervision for the rest. Homes will get useful single-purpose robots before general-purpose ones.
Quantum, and the security deadline it creates
Quantum computers will not replace laptops. They are specialised tools for problems such as molecular simulation and some optimisation. The practical story for most organisations is not “buy a quantum computer” but “migrate to post-quantum cryptography before stored data becomes readable”. Standards are already published, and migration is slow. That work belongs on a 2026 roadmap, not a 2035 one.
The constraints that decide the timeline
- Energy. Data centres are now constrained by power and grid connections, not just chips. Efficiency gains matter more than benchmark scores.
- Memory bandwidth. Many AI workloads are limited by moving data, not by arithmetic.
- Semiconductor capacity. Leading-edge fabrication is concentrated in a few places and takes years to expand.
- Trust. Adoption stalls when people cannot tell what a system did or why.
- Regulation. Rules such as the EU AI Act shape what ships in which market, and when.
What probably will not happen by 2030
- A general-purpose robot doing all household chores reliably.
- Fully autonomous vehicles everywhere, in every weather and city.
- Quantum computers breaking everyday encryption on demand.
- AI that needs no human review for high-stakes decisions.
Each of those may arrive eventually. The mistake is treating a lab result as a shipping product, and a shipping product as a solved social problem.
How to prepare, practically
- Keep your data portable. The tools will change; your documents, code and customer records should not be trapped.
- Learn to verify AI output. The skill that matters is checking, not prompting.
- Treat security as identity. Passkeys, multi-factor authentication and short-lived access age well.
- Watch the unit economics. A demo that costs £50 per task is not a product yet.
- Prefer boring integration. The future usually arrives as a feature inside software you already use.
The most likely future is not a single dramatic breakthrough. It is a decade of quiet defaults: AI in the background, computers that fade into objects, and security work that decides whether any of it can be trusted.
