Table of Contents
In 2019, a research lab could train a frontier AI model for a few million dollars. By 2025, the leading efforts were spending well over $100 million per training run, on clusters of hundreds of thousands of specialised chips drawing power on the scale of a small town. That single number, the cost and wattage of computation, explains more about where technology trends and innovation are heading than any keynote.
When compute gets that expensive, it stops being a neutral utility. It becomes an asset to hoard, a bottleneck to engineer around, and a reason some countries build products while others only consume them. Most of the shifts below trace back to that pressure.
AI has stopped being a product and started being plumbing
Chatbots were the headline act of 2023 and 2024. The interesting work now is far less photogenic. AI is being folded into ordinary workflows: claims triage at insurers, transaction screening at banks, first-pass code review at software companies, yield prediction on farms.
You can see the change in how it gets sold. Enterprise software vendors bundle AI features into existing seats instead of charging separately for them. Prices for capable mid-tier models have fallen by more than 90% in two years. When a capability becomes cheap and invisible, it has graduated from trend to infrastructure.
Inference bills, not benchmarks, drive decisions
A model that scores two points higher on a reasoning benchmark but costs five times as much per query will lose most procurement fights. Engineering teams now optimise for tokens per dollar, caching aggressively, routing easy requests to small models and hard ones to expensive ones. That discipline is new, and it changes which startups survive.
The uneven distribution of compute is its own story. A handful of regions control most advanced chip capacity, and access increasingly determines who can build at the frontier, as this analysis of how AI computing power is splitting the world into haves and have-nots sets out. Smaller players respond by specialising: fine-tuning open-weight models on narrow, high-value data rather than competing on general capability.
Cloud computing is the ground everything stands on
None of the above runs on a laptop under someone’s desk. It runs in data centres, which means the cloud layer sets the ceiling on what any team can attempt. The past two years made that dependency more complicated rather than less.
Three things changed in cloud economics
- Egress fees fell. Regulatory pressure in Europe pushed providers to waive or cap the cost of moving data out, making multi-cloud architectures far more practical than they were.
- GPU capacity became a reserved commodity. Long-term commitments for accelerated instances now look like commercial lease agreements, complete with brokers and renegotiation windows.
- FinOps grew up. Cost accountability moved from a quarterly spreadsheet to a daily engineering concern, with teams tagging and budgeting their own workloads.
For anyone choosing a platform today, the layers matter more than the marketing suggests. Understanding what infrastructure, platform and software services each take off your plate is worth doing before signing anything, which is why a plain-language guide to IaaS, PaaS and SaaS beats a feature-by-feature comparison.
The edge got quietly clever
Away from the data centre arms race, something more modest happened: models got small enough to run in your pocket. Phone chips now ship with neural accelerators capable of running multi-billion parameter models locally. Hearing aids perform real-time speech separation. Cars handle lane reasoning without a round trip to a server.
The reasons are practical rather than ideological. Latency measured in milliseconds, privacy that doesn’t depend on a policy document, and no per-query bill. Expect the split to harden: heavy reasoning in the cloud, routine perception and personalisation on the device.
Quantum computing moved from physics papers to procurement
Error correction crossed a threshold. Several teams have now demonstrated logical qubits that stay coherent long enough to be useful, which shifts the conversation from “is this possible” to “when do we need to be ready.”
The near-term impact isn’t breaking encryption next Tuesday. It’s the migration work already underway. Standards bodies have published post-quantum cryptography algorithms, and organisations holding long-lived sensitive data are being told to inventory and re-key now, because traffic intercepted today can be decrypted later.
Biology is becoming an engineering discipline
Protein structure prediction, generative design of new molecules, and cheaper DNA synthesis have turned parts of biology into something closer to software development. Clinical timelines are still measured in years, but the discovery phase has compressed from decades to months in several areas.
The counter-current is just as real. Medicine depends on biology’s stubborn problems, and antimicrobial resistance is among the hardest. Bacteria are evolving around our last-line drugs faster than replacements reach the market, and the fixes are as much about prescribing habits, diagnostics and agriculture as about novel compounds. The practical steps in this piece on antibiotic resistance as a silent pandemic show how much of the solution is operational rather than technical.
Trust is now a technical problem
Generative tools made it cheap to produce convincing fiction at scale: synthetic voices, cloned faces, fabricated studies. The damage lands hardest where people are already uncertain, in health and medicine especially. Content that looks authoritative and cites nothing spreads faster than corrections, a dynamic traced in the rise of health misinformation from snake oil to wellness influencers.
The response is partly technical. Content provenance standards like C2PA attach signed metadata to images and video, and watermarking tools keep improving, though detection remains imperfect. Regulation is arriving in patches, with the EU AI Act phasing in obligations over several years while other jurisdictions take a lighter approach. None of it replaces the boring fix: verification habits, better platform incentives, and readers who check a claim before sharing it.
How to tell a real trend from a loud one
Trade press produces a new paradigm every fortnight. A few filters help.
- Follow the capex. Announced spending on fabs, chips and data centres is harder to fake than a press release.
- Look for boring adoption. When a technology disappears into payroll systems and insurance back offices, it’s staying.
- Check the cost curve. Falling unit costs create markets; flat ones create demonstrations.
- Watch who’s excluded. Any trend that concentrates capability in few hands generates a political response, and that response becomes the next trend.
Quarterly roundups are useful precisely because they track what practitioners build rather than what analysts promise. Digests such as the Radar Trends to Watch: May 2026 review tend to flag shifts months before they surface in a board deck.
Where the next eighteen months likely go
Expect consolidation rather than explosion. A smaller number of well-capitalised labs push frontier capability, while most organisations extract value from open-weight models, retrieval systems and careful data work. Energy becomes a constraint nobody can ignore: grid connections, permitting and cooling design are now bottlenecks as real as chip supply.
Agentic software will keep improving and keep disappointing in equal measure. Reliability at 95% per step still fails most ten-step tasks, and that gap is where the useful work sits: verification, rollback, human checkpoints, deliberately narrow scopes.
The habit worth building is unglamorous. Ask what a new capability costs to run, who controls it, and what it quietly replaces. That question cut through the last decade of hype, and it will cut through the next one just as cleanly.
