In 2025, generative AI had a packed year. Plenty of announcements, plenty of promises... but above all a few genuine paradigm shifts that quietly moved the playing field. We didn't just get better models: we changed the way we work with them.
What struck me is this constant contrast: AI is, at the same time, far smarter than I expected... and far dumber on simple tasks.
You've probably lived it: a stunning result at 9am, then an absurd mistake at 9:05. And despite these quirks, it's already extremely useful, and more and more teams are seeing concrete gains in time and quality.
In this edition, I'm offering a simple format: the 5 shifts of 2025, then my 7 predictions for 2026, and a January checklist to get ahead.
TLDR : 2025 confirmed an AI that's very useful but still uneven. In 2026, the difference will come down to moving into real-world use (deployment, reliability, ROI).
Looking back at 2025
Here are the 5 shifts I take away from 2025, the ones that genuinely change how we use AI at work (and that explain why 2026 could be even more interesting).
1️⃣ No sign of slowing down + the year of reasoning
We spent the year telling ourselves “ok, this is going to calm down now”... and then no. Every 2 to 3 months, a new milestone falls: more quality, more speed, more options, more use cases. Bottom line: if you feel behind, that's normal, the bar keeps moving.
But beware the trap of benchmarks, which are increasingly criticized.
Why? Because they're becoming hackable: you can optimize a model to score very high... without it translating into reliability in our real cases (internal docs, business exceptions, long instructions, compliance constraints). In short, the leaderboard can be reassuring, but it doesn't tell you whether it works in real life.
And 2025 was above all the year of reasoning. The turning point is the emergence of explicit reasoning with “reasoned” models, and a strong signal like DeepSeek R1: we saw a shift toward reinforcement learning with rewards you can verify (RLVR), rather than “classic” RLHF (where a human judges the output). Consequence → the key lever is no longer just “more parameters,” but better training after pre-training (post-training), with measurable objectives.
Another point that surprised a lot of people: the training cost to obtain these reasoning behaviors is lower than expected, and behind that, all the big labs released reasoning-oriented variants.
That's the real 2025 signal: we learned to make models more “useful at work” not just by making them bigger, but by training them more intelligently at the end.
2️⃣ Unevenness + deployment: the paradox (and ROI anyway)
The real barrier in companies isn't AI's intelligence... it's its unevenness. It can be brilliant, then get a basic detail wrong, so you lose a bit of certainty, and that forces you to set guardrails (validation, sources, steps, rules). And yet, despite this, we're already at the start of “serious” deployments and the ROI feedback is starting to come in: time saved, better quality on first drafts, less mental load on repetitive work.
The twist is that this leverage should amplify with the next wave: AI agents, able to chain several steps (search, decide, execute, verify) and therefore maximize ROI where AI (in conversational mode) or more rigid AI automation workflows hit a ceiling.
Leverage > certainty
3️⃣ GenAI has become an industry... and the bubble question is back
Money clearly poured in. The tech giants invested amounts that are enormous in data centers and infrastructure: over the last 3 months, we're talking about a cumulative total close to 100 billion dollars spent by just 4 companies.
And the comparison that makes you think is the one with the great historical waves of innovation: how far will this spending (CapEx) climb as a % of GDP, to the point of nearing the peaks of the railways in the 1880s?
At the same time, we saw giant funding rounds and contracts (OpenAI, Claude/Anthropic, etc.), with an ambition summed up bluntly: to meet demand “at unprecedented levels.”
And where it gets a bit worrying is the “circular investment” angle: OpenAI buys Nvidia chips, Nvidia invests in OpenAI, Oracle rents Nvidia chips to OpenAI, OpenAI takes stakes in AMD... It makes for a very frothy market, where a single bad surprise (typically a bad quarter on Nvidia's side) could trigger a correction.
The situation summed up by a Bloomberg chart:
Even if the music stops for a few moments in 2026, it wouldn't necessarily be a disaster. A little air coming out of the sails is sometimes healthy. But with announced commitments running up to 1,000 billion dollars for some players, things could move...
4️⃣ The market rebalanced... and developers changed jobs
By the end of 2025, the market balanced out: Google clearly got back in the race (with visible progress on Gemini and a build mode that's very convincing in Google AI Studio), Anthropic established itself on the enterprise side via the API, and OpenAI remains ultra-dominant on the consumer side.
Put differently: you no longer buy “a model,” you choose an ecosystem depending on the use (professional, consumer, integrations, etc.).
And to me the strongest signal comes from devs. We saw an explosion of tools like Claude Code, Codex and Antigravity: AI is no longer just a web page where you paste text, it's an assistant that lives on your machine (in the command line), with your context, your files, your habits.
This has a huge impact on developers, who no longer ask for code, they steer a mini-colleague that chains the steps (understand, modify, test, fix).
What's interesting is that this wave exceeded what many expected on paper: not just scores, but a real change of workflow.
5️⃣ Revenues take off and the job market is being “reconfigured” (especially for juniors)
The real 2025 shift is the revenue of AI solution vendors. OpenAI announced an annualized run-rate of $10B as early as June 2025, then ~$12B at the end of July. And on Anthropic's side, the trajectory is even crazier in the enterprise segment: ~$1B run-rate in early 2025, then >$5B in August 2025.
The result: models are becoming products with clear revenue lines (subscriptions + API), which fuels even more investment... and even more competition.
AI companies' revenues have overtaken SaaS revenues.
But there's a very concrete flip side: capital market vs labor market. On one side, billions keep coming in. On the other, in certain roles, especially junior positions, we're seeing real pressure: falling salaries/demand in companies exposed to AI (e.g. -4.5% on starting salaries, with a stronger impact on juniors).
Here's the chart comparing the S&P 500 curve with US hiring.
And several public signals point the same way: entering a career is getting harder, because part of the learning tasks (the ones we used to give juniors) is now automatable.
Late 2025: Google comes back strong... and visuals explode
In late 2025, Google hit the accelerator hard again. Between Gemini 3 (and its variants) and the Build mode of Google AI Studio, you sense a simple intent: make AI more of a product, more accessible, more integrated into daily work, not just a model in a lab.
And above all, we saw a shift in usage: text is no longer the only reigning format. With Nano Banana (and Nano Banana Pro), Google is pushing an AI that generates/edits images “natively,” and that helped explode visual formats: images, mockups, infographics... and even increasingly simple video workflows. I talked about it here:
AI image generation: concrete examples and tips
And to put a marker down on a 2026 trend right now (we'll talk predictions just after), Google has a strong chance of outpacing the others.
Simply because it plays on every front at once (models, products, research, visual formats, deployment), while the others will be pushed to specialize.
2026: my predictions (the ones that will really matter)
After a very tech-heavy 2025, 2026 looks more like a “usage + organization” year. Fewer debates about the best model, more very concrete questions: who automates what, with what level of confidence, and what ROI behind it?
1️⃣ Companies will pay more for AI agents than for some humans (on targeted tasks)
Not because AI is magic, but because once you count recruitment, onboarding, management and turnover, a reliable agent can cost more... and still be profitable.
2️⃣ Agents will work longer, without supervision, to the point of running for more than 24h.
We'll go from occasional help to complete workflows: preparation → execution → verification → report. And there, the leverage explodes.
3️⃣ Memory becomes a real feature... and it opens up cascading use cases
When AI remembers your preferences, your constraints, your projects and your context, you stop “re-explaining” every time = much more useful personal/team assistants
4️⃣ The web shifts to “agent-first”
More and more decisions (purchases, providers, comparisons) will be prepared by agents that read 50 sources and synthesize them. So docs, product pages and content will be written to be understood by robots... before being pretty for humans.
5️⃣ Knowledge bases and search within your content become central again
The more AI becomes multimodal (text + images + videos), the more you need a solid architecture to find internal information quickly. The companies that win will be the ones that cleanly connect models ↔ data ↔ business rules.
6️⃣ Data centers will keep swallowing budgets... with possible tremors
Infrastructure will remain the crux in 2026. And since we're in a very “massive investment” phase, it only takes a shift in risk perception for the market to shake.
In no particular order, other predictions that smell of a changing era
- AI-augmented micro-businesses : smaller, more numerous, more niche, but viable.
- The return of the closed web (invitation, reputation, communities) to escape mass-generated content (polluted by AI)
- Human support becomes a premium option : talking to a real person = rarer, so more expensive.
- Apps become disposable : you generate a mini-app, use it for 72h, then move on to something else.
- Horizontal SaaS gets nibbled away by ultra-specialized vertical solutions.
- Local / on-device makes a strong comeback as soon as there's a big security incident: “my data doesn't leave”
- Blue-collar trades become safe havens : electricians, plumbers, AC repair technicians ... their wages rise while automation accelerates on office tasks.
- The “historical” web gets archived : the pre-2023 internet becomes an ultra-precious dataset, because it's one of the last great reservoirs of 100% human content.
- The job “breaks apart” : fewer 40-hour permanent contracts, more outcome-based missions, and AI-driven micro-businesses.
- Therapy shifts to AI-first : before reaching a human, part of the journey first goes through AI-guided protocols (and a lot of people prefer it... because it's available at 3am and it judges no one).
