Every year I get sent the same slide. Gartner's trend list, Deloitte's tech predictions, a dozen posts summarising both, and a client asking whether we should be doing something about quantum computing. The honest answer for most businesses is no, not this year — but that answer is less useful than explaining how to tell the difference between a trend that will affect your roadmap and one that will not.
So rather than ranking the 2026 lists, here is the filter I actually use, applied to the technologies people keep asking me about.
The Filter: Three Questions
Before I let a trend influence a plan, I ask three things.
Can I buy it today from someone who will still exist next year? Not "does a demo exist" or "has a research lab shown it." Is there a product, with pricing, support and customers who are not in the press release.
Does it change a cost curve, or just a capability? The technologies that reshape industries usually make something dramatically cheaper rather than newly possible. Cloud did not invent servers; it made capacity cost nothing when you were not using it. That is the shape to look for.
What breaks if I ignore it for two years? For most trends, the answer is nothing — you adopt later, slightly behind, at lower cost and lower risk. For a small number, waiting two years means a competitor restructures their cost base while you are still evaluating. Those are the only ones that deserve budget now.
Applied honestly, that filter kills most of the list. Here is what happens to the ones people ask about most.
AI Infrastructure: The Trend Under All the Other Trends
This is the one I would put at the top, and it is the least exciting to talk about. Not models — the boring layer underneath. Vector storage, retrieval, evaluation tooling, observability for model calls, cost attribution, prompt and context management, the gateway that sits between your services and three providers.
The reason it matters is that this is where teams are currently losing money and shipping unreliable features. Everyone can call an API. Very few organisations can tell you which of their AI features is costing what, whether quality dropped after last week's prompt change, or what happens when their provider has an outage.
Passes all three questions: buyable today, changes the cost curve of running AI in production, and ignoring it for two years means accumulating a pile of AI features nobody can operate. If you have anything using a language model in production and you do not have per-feature cost tracking and a saved set of evaluation cases, that is a more valuable project than anything else on this page.
Physical AI and Robotics: Real, But Capital Intensive
Physical AI — models controlling machines that move in the world — has moved faster in the last two years than I expected. Warehouse robotics, autonomous inspection, agricultural equipment, quality control on production lines. The progress is genuine, driven by better perception models and cheaper sensors rather than by any single breakthrough.
But the buyers are specific. If you run a warehouse, a factory, a farm or a fleet, this is on your roadmap now and you probably already know it. If you build software, this is a market you might sell into, not a technology you will adopt. The capital requirements and the safety engineering keep the barrier high, and the gap between an impressive demonstration video and a machine that runs a full shift without a human intervening is still substantial.
The part that is relevant to more people: the software patterns are converging with agentic AI. Perception, planning, tool use, a control loop with a safety envelope, a human able to intervene. If you build agent systems, you will recognise the architecture — and the robotics people have thought harder about failure modes than most software teams have, because their failures have physical consequences.
Quantum: Right to Ignore, Except for One Thing
Quantum computing will matter enormously and almost certainly not to your application in the next few years. The machines that exist are remarkable physics and are not solving your business problems. Nobody is optimising their delivery routes on one in production.
The exception, and it is a real one, is cryptography. The concern is not that someone will break your encryption today. It is that encrypted data captured today could be decrypted later, once capable machines exist. For most businesses, whose data has a short useful life, this is not urgent. If you hold data that must stay confidential for ten or twenty years — health records, state secrets, long-term financial or legal records — then the migration to post-quantum algorithms is a planning item now, because the standards exist and migrations of that kind take years.
What I would actually do this year: know where your organisation uses cryptography and what algorithms are involved. That inventory is useful regardless, and it is the prerequisite for any future migration. Everything else about quantum can wait for the vendors to make it a product.
IoT: The Quiet Comeback
IoT had its hype cycle around 2016, disappointed everyone, and then kept growing anyway while nobody was watching. What changed is unglamorous: connectivity got cheaper and more reliable, edge hardware got capable enough to run real models locally, and the platform side matured so that you no longer have to build the entire ingestion and device management layer yourself.
The current version is more sensible than the earlier one. Fewer internet-connected kettles, more industrial monitoring, predictive maintenance, cold chain tracking, energy management. These are boring problems with clear financial returns, which is exactly why they work.
If you are building here, the hard parts have not changed: device fleet management, over-the-air updates, security on hardware you cannot physically reach, and handling data from devices that are offline half the time. The failure mode is always the same — a pilot with fifty devices that works beautifully, and an architecture that collapses at five thousand.
AR and VR: Still Waiting
I have been wrong about the timeline here twice, so I hold this view loosely. But the pattern keeps repeating: impressive hardware, genuine enthusiasm, and no consumer use case compelling enough to justify wearing something on your face for hours.
Where it does work is narrow and industrial. Training simulations for expensive or dangerous procedures. Remote assistance where a technician sees what a field worker sees. Design review of physical products before they are manufactured. Real value, real budgets, small market.
The thing that would change my mind is not better resolution. It is a device people would wear anyway for another reason. Until that exists, this stays a specialist tool rather than a platform.
What I Would Actually Do About 2026
If I were setting technology priorities for a normal business this year, the trend list would come third, after two more boring things.
First, get your data in order. Almost every advanced capability on every trend list assumes you can find, trust and access your own data. Most organisations cannot. The company with clean, well-modelled, accessible data will adopt any of these technologies faster than the company with a better strategy deck.
Second, build the operational muscle for the AI you already have. Cost visibility, evaluation, human review paths, an owner for each AI feature. This is what turns experiments into systems.
Then, and only then, pick one emerging technology that plausibly touches your industry and run a small, honest pilot with a defined kill criterion. Not a strategy. A pilot, with a date on which you decide whether it worked.
The trend lists are useful as a map of what other people are excited about. They are a poor substitute for knowing which of your own costs and constraints a technology would actually change. That question you have to answer yourself, and the answer is usually smaller and more specific than the slide suggests.
