Table of Contents
In 2015, a car that drove itself was supposed to be in your driveway by 2020. It wasn’t. That year’s most exciting gadget was a $300 drone, and voice assistants still stumbled over the word “call.” Meanwhile the paper that actually reset the decade, Attention Is All You Need, spent its first year as a niche citation in machine translation circles.
That gap between what we expect and what lands is why predictions are worth making carefully. Not as prophecy, but as bets you can check. Each of the seven below comes with a reason to believe it and a specific observation that would prove it wrong.
1. AI stops being something you talk to and starts doing the work
The demos are already impressive. The reliability isn’t. Ask an agent to book a flight, file an expense report and reconcile two spreadsheets, and it might chain forty steps together. If each step succeeds 95% of the time, which is generous for current models on open-ended software tasks, the whole run finishes roughly 36% of the time. Agents fail less because they’re stupid than because mistakes compound.
Fix that and everything downstream changes. By the early 2030s I expect agents to run multi-hour workflows inside narrow, well-instrumented domains: insurance claims triage, tier-one support, code maintenance, regulatory filing prep. The general-purpose digital employee that handles anything you throw at it is a much later problem, if it’s a problem at all.
The most interesting version of this isn’t office work. It’s the possibility that machine reasoning accelerates research itself, which is the argument behind how AI could reboot science and revive long-term economic growth. Slow, expensive fields like materials discovery and drug formulation are where a 20% speedup is worth billions.
2. The compute crunch decides who gets to build
Frontier training runs now consume tens of megawatts, and the bottleneck has quietly moved away from chip design toward three grittier constraints:
- Electricity. Grid interconnection queues in the US stretch four to seven years in some regions.
- Memory bandwidth. High-bandwidth memory supply is sold out well ahead of demand.
- Advanced packaging. The CoWoS capacity that stitches accelerators together has been the real ceiling, not wafer starts.
That pushes the industry in two directions at once. Giant models get concentrated in a handful of well-capitalised labs, while small distilled models get good enough to run locally on a phone or a laptop. I’d bet that by 2032, most consumer AI inference happens on-device, not in a data centre, purely because latency and privacy are worth more than a few extra benchmark points.
3. Robots leave the demo video, but not the ones you’ve seen
Humanoid robot footage is the most misleading genre of video on the internet right now. Plenty of those clips are teleoperated, the hardware costs more than a house, and battery life is measured in a couple of hours. The pattern is familiar from the technologies most likely to transform the future and what’s actually real behind the hype: the arrival is later and narrower than the keynote suggests.
Meanwhile the boring robots are already working. Warehouse fleets move millions of totes a day. Laser weeders cut herbicide use by 80% on some vegetable farms. Automated fry stations and pizza lines quietly staff late-night kitchens. Expect the next decade to be dominated by these specialised machines, with a home humanoid landing somewhere around 2033 as a $30,000 appliance for people who really hate laundry.
4. Energy becomes a product feature
Solar module costs have fallen roughly 90% since 2010 and grid-scale batteries have followed a similar curve. What’s changed recently is who’s buying. Microsoft signed a deal in 2024 to restart a reactor at Three Mile Island. Google ordered small modular reactors from Kairos Power the same year. Data centre operators are now the most aggressive energy investors on earth.
My prediction: by the mid-2030s, “powered by” becomes a marketing label the way “organic” did for food. Cheaper electricity also unlocks things that were previously uneconomical, from direct air capture to desalination to aluminium smelting, and it makes electrifying heating and transport a straightforward financial decision rather than a moral one.
5. Biology gets cheap enough to be boring
Sequencing a human genome cost about $3 billion in 2003. It’s now a few hundred dollars, and the $100 genome is close enough to argue about. CRISPR-based therapy moved from lab curiosity to approved product when Casgevy was cleared for sickle cell disease in 2023. AI protein design tools are generating candidate binders that would have taken a PhD student a year to build by hand.
The consumer side deserves scepticism. As experts unpacking consumer genetic tests have pointed out, some of these products have accuracy close to zero, and a risk score is not a diagnosis. Clinical sequencing with genetic counsellor support is a different animal from a spit kit with a slick dashboard.
6. Provenance gets built into the pipes
Watermarking individual files was never going to work. What’s arriving instead is provenance at the capture point: cameras and editing tools that sign content cryptographically under the C2PA standard, so a photo carries a verifiable record of where it came from and what was done to it. Passkeys are doing something similar for identity, replacing the password by tying login to a hardware key on your device.
Regulation is moving in the same direction. Expect major markets to require labelling of synthetic media in advertising and political content before 2030. The result won’t be a world without deepfakes. It’ll be a world where unverified content starts to carry a quiet reputational penalty, the way an unsigned contract does now.
7. Quantum computing gets real in a narrow, unglamorous way
Google’s Willow chip demonstrated error correction that improves as the system scales, which is the milestone that matters. It does not mean encryption is about to collapse. Breaking RSA-2048 needs millions of physical qubits and a level of fault tolerance nobody has shown, and I’d be surprised to see it before 2040.
The commercial wins will come from chemistry and materials simulation, where quantum systems are naturally good at modelling other quantum systems. Fertiliser catalysis, battery electrolytes, new superconductors. My bet is the first genuinely useful commercial quantum result in one of those areas lands before 2032, and almost nobody outside the field notices when it does.
What to watch instead of the headlines
Predictions are only useful if you know what would falsify them. Track these five numbers rather than the launch events:
- Agent reliability on long tasks. When success rates on 30-step workflows cross about 80%, the labour market feels it within two years.
- Signed power purchase agreements. Gigawatt-scale contracts tell you where compute actually goes, regardless of what any roadmap says.
- Robot cost per operating hour. The day a machine undercuts a forklift driver’s fully loaded hourly cost, adoption stops being a pilot programme.
- Genome price and clinical reimbursement. Both must fall together for sequencing to become routine.
- Provenance adoption rates. Watch how many cameras and publishing tools sign content by default, not how many press releases announce a standard.
None of this requires a leap of faith, and none of it arrives in a single dramatic moment. The next decade will be shaped less by one breakthrough than by the slow, unglamorous work of making seven half-working things dependable enough to trust with money, health and reputation. That part never makes the keynote, and it’s the part that decides everything.
