How AI will grow over the next 12 months

Imagine your grandmother has been baking the best cake in the neighbourhood for forty years. One day a robot walks into her kitchen and, in ten minutes, makes a batter identical to hers. Is grandma's business finished? My answer is no, as long as she understands what her neighbours are really buying from her.

That same thing is happening to the software world today. What used to take a team years of work, a machine does in an afternoon. This is my reading of what is coming over the next twelve months: who wins, who loses, and what is worth doing starting today.

The ink and the story

When the printing press arrived, ink and paper became dirt cheap. The scribes who spent their lives copying books by hand lost their trade almost overnight. The curious thing is that good writers, editors with a sharp eye and booksellers who knew their customers became more valuable than ever. There were many more books, and someone had to decide which ones were worth reading.

Today code, that is, the instructions that make an app work, is the ink. And artificial intelligence is making it cheaper at a speed we have never seen. To give you an idea of the pace: in just six weeks, between September and mid-October of this year, the big AI companies (OpenAI, Anthropic, Google and Meta) released at least seven new models. And each version does more and costs less than the one before.

There are already cases that make your head spin. One company said that one of these machines finished, in less than a day, a system migration that would have taken a team of engineers weeks. You have to take it with a grain of salt, because the maker itself is the one telling the story, but the direction is clear. An independent lab that measures how much work an AI can finish on its own has seen that amount double every few months, so much so that its measuring stick has already fallen short.

So, if ink costs almost nothing, what is still worth something? The story. Knowing which problem is worth solving, knowing the customer by name, having their trust, having data nobody else has, and things that cannot be copied with a click: a shop, a product you can touch, a government permit, a years-long relationship with a supplier. That is the most important idea in this piece. Everything else follows from it.

What is pure noise

Before looking at the future, let's turn off the noise. You will hear that AI has already replaced every programmer. It is not true, at least not yet. In the United States, job postings for programmers have been picking up for more than a year, although they are still well below where they were before the pandemic.

You will also hear that every tech layoff is AI's fault. It is true that many companies mention it when they cut staff. But the founder of the very site that counts those layoffs says there is little evidence that AI is doing the work of the people who left. Many companies, rather, are spending so much on AI that they cut in other areas to pay for it.

The ice cutters and the refrigerator

In the 19th century there was a huge business that sounds like a fairy tale today: cutting blocks of ice from frozen lakes and delivering them by cart to people's homes. The refrigerator did not break them all at once. First it was expensive and rare; then, normal. The ice cutters did not close the day the refrigerator was invented. They closed the day the refrigerator became boring.

With AI that writes code, we are right at that crossing: it is going from an expensive novelty to an everyday thing. That is the dangerous moment for anyone who makes a living cutting ice.

In the coming months

Before going on, an honest note. What I have told you so far has already happened, and I checked it. What comes next is my bet. I may be wrong about the dates, but I do not think I am wrong about the direction.

Between now and January, I expect each company's "middle" models to reach what was the best on the market a year ago, at a fraction of the price. Many mid-sized companies will cancel one or two simple digital tools they were paying for every month, like forms or dashboards, because now an assistant builds them their own version in a few days. It will not be an avalanche, but a steady trickle.

Agencies that charge by the hour for programming will hear something new from their clients: "Tell me what the result costs, not how many hours it takes you." Those that do not change the way they charge will start losing clients.

And mark a date: in December, Europe starts requiring that AI-made images and videos can be identified as such.

Towards spring

Around April, the normal thing on tech teams will be to leave an assistant working alone for hours, or all night, on tasks that used to keep several people busy for a week. It already happens at some companies; soon it will be routine.

In the investment world, companies that only sell software and raised money in the years of plenty will have to face an uncomfortable mirror. Investors no longer pay what they used to for them. Early this year, their shares fell hard within a matter of weeks; some people called it the "software apocalypse". Many recovered, but not all: some are still being punished even though they earn well, because people fear their entire business will become unnecessary.

A year from now

A year from now, any serious investor will ask one question before putting in money: "Could an assistant do this in an afternoon?" If the answer is yes, the conversation ends there.

By then the physical limit of all this will also be visible. AI lives in enormous data centres that use electricity like entire towns. Turbines, transformers and permits to connect to the power grid are already scarce today. My bet is that, for quite a while, what holds AI back will not be a lack of intelligence but a lack of plugs.

And after that

Beyond that year, I suspect programs will no longer be sold, but finished jobs: the books closed, the campaign launched, the insurance claim settled. The winners of 2030 will be those who gather customers, their own data and physical things today, not those who gather code.

The weavers and the looms

In England in 1811, a group of weavers, the Luddites, broke into factories at night to smash the mechanical looms. They were right about one thing: their trade of weaving fast by hand was losing value. But the amount of cloth being produced multiplied, and trades that did not exist before were born: loom mechanics, pattern designers, merchants, carriers. Those who only knew how to throw the shuttle suffered. Those who understood cloth, customers and machines prospered.

Something similar will happen with jobs. Anyone who only knows how to take a well-written instruction and turn it into code is now competing against a machine that charges pennies. Programmers who are just starting out feel it first: the recent rebound in jobs is almost all for experienced people. The entry rung is narrowing, and nobody has solved that problem.

On the other hand, people with judgement come out ahead. The person who looks at what the machine did and can say what is wrong and why. The accountant, the lawyer or the fashion buyer who knows how to ask AI for things properly and check what it delivers. Whoever sells, negotiates and looks after customers, because if making things became cheap, persuading became expensive. And the physical trades connected to all this: electricians, technicians, energy and robotics installers.

And the ones who took two years to build it?

This is the question that hurts the most. Think of a company that spent two years and millions of dollars building its app, and today watches an assistant do something similar in minutes. It is like having dug a well by hand, stone by stone, in the very year the electric pump arrived. The well still gives water, but it no longer stops the neighbour from having their own.

The good news is that those two years were not entirely lost. Code lost its value as a wall, but what was gained along the way is still there: customers who trust you, accumulated data, knowledge of the odd cases you only learn by stumbling, permits, connections to each customer's systems. Teams will get smaller and change shape: fewer people writing code and more people understanding the problem, reviewing and testing with real customers. And the agencies that survive will stop being hour factories and become workshops that deliver guaranteed results.

The photo after the phone

Remember when developing photos was a big business? When we all had a camera in our pocket, photos did not disappear: they multiplied by the millions. What sank was the film business. The winners were those who controlled the place where we share photos.

The same will happen with software. There will be much more of it, and that is why each loose piece will be worth less. A startup that only has a pretty app on top of public information can be copied in an afternoon. Here I am blunt: those companies will not survive as venture capital bets. Some will remain as small, profitable businesses, others will be bought, and many will close.

Mind you, the money has not gone away. This year there is more investment than ever. But it is extremely concentrated: in the United States, almost 8 of every 10 venture capital dollars went to artificial intelligence companies, and a good part of that to a handful of giants. For everyone else, the bar went up.

What are investors looking at now? The same thing your grandmother looks at before lending money to a nephew. Do your customers buy from you and not from the one across the street? Do you have information nobody else has? Do they trust you? Do you have the permits the law requires? Are you so embedded in your customer's day-to-day that taking you out would be a mess? Do you have something physical? And do you charge for the result you deliver?

And if a startup has none of that, what is it worth? I say this with affection: it is worth more or less what it has in the bank.

What I know and what I am guessing

I know AI is improving fast and getting cheaper and cheaper, because it is measured and it shows in every launch. I know money is more concentrated than ever and that the value of software companies has split in two. I know that, for now, most small businesses barely use it: in the United States, only one in five businesses had used it in the two weeks before the official survey.

What I am guessing are the exact dates, how many companies will close and how fast all of this will reach the corner shop. If I am wrong, I think it will be about the speed, not the direction.

The shovels in the gold rush

In the gold rush, a great many miners went broke. The ones who got rich were those who sold shovels, sturdy clothes and food, and the owners of the land. Today the "gold", which is the artificial intelligence models, is being fought over by a few giants with amounts of money that are hard to even imagine. Fabián has no reason to fight for that gold. It suits him better to own the land.

I am thinking, for example, of My Pet Model, which lets people try on clothes virtually before buying them. Its value is not in generating pretty images, because soon anyone will do that. It is in helping brands get fewer garments returned and charging them for exactly that, while gathering something no giant has: what it learns from each try-on about real sizes, fits and tastes. The first step fits in an afternoon: write to five fashion brands and propose a one-month trial in a single category, comparing how many returns they had before and how many after.

I am also thinking of a bilingual service for small businesses in Mexico and Canada. The owner of a shop or an office does not want to learn how to manage AI assistants. They want someone they trust, who speaks their language, to replace that expensive program they hate paying for every month with one made to measure, at a fixed price and with a guarantee. That gap is huge, because the giants sell tools; they do not sit down at the table with the owner. To start, it is enough to ask ten acquaintances which program hurts them to pay for, and offer to replace the most common one in a week.

And I am thinking of AlmaLibre. Nobody can copy an açaí with a click, nor its place on the shelf, nor its relationship with its suppliers. There AI is not a threat, it is a helper: it can anticipate how much will sell in each store and flag when to restock, so that a small brand operates like a big one. The test is simple: gather the sales from the last six months, ask an assistant to guess next month and compare with what really happens.

To close

The next time someone tells you AI is going to wipe out everything, remember grandma and her cake. The robot can copy the recipe in ten minutes. What it cannot copy is that the whole neighbourhood knows her, trusts her and comes back every Sunday.

Code now costs what ink costs. What is still worth gold is who you know, who trusts you and what you can touch with your hands.

---

References

1. "Time Horizon 1.1". METR, 29 de enero de 2026. https://metr.org/blog/2026-01-29-time-horizon-1-1/
2. "Task-Completion Time Horizons of Frontier AI Models". METR, actualizado al 8 de septiembre de 2026. https://metr.org/time-horizons/
3. "GPT-6 Astra: A new generation of intelligence". OpenAI, septiembre de 2026 (actualizaciones del 22 y 29 de septiembre). https://openai.com/index/gpt-6-astra/
4. "OpenAI begins rollout of new powerful AI model GPT-6 Astra". France 24 / AFP, 3 de septiembre de 2026. https://www.france24.com/en/live-news/20260903-openai-begins-rollout-of-new-powerful-ai-model-gpt-6-astra-1
5. "Introducing Claude Opus 5.5". Anthropic, 22 de septiembre de 2026. https://www.anthropic.com/claude-opus-5-5
6. "Introducing Claude Haiku 5.5". Anthropic, 7 de octubre de 2026. https://www.anthropic.com/claude-haiku-5-5
7. "Introducing Muse Wonder and Muse Wonder Full". Meta, 12 de octubre de 2026. https://about.fb.com/news/2026/10/introducing-muse-wonder-models/
8. "Introducing Muse Spark Open". Meta AI Research, 15 de octubre de 2026. https://research.meta.ai/blog/muse-spark-open
9. "Gemini 4 Argon is now available for everyone". Google, 13 de octubre de 2026. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/
10. "AI and Job Postings: From Destruction to Creation?". Indeed Hiring Lab, 8 de julio de 2026. https://hiringlab.indeed.com/2026/07/08/ai-and-job-postings-from-destruction-to-creation/
11. "Software Development Job Postings on Indeed in the United States". FRED, Banco de la Reserva Federal de St. Louis, dato del 16 de octubre de 2026. https://fred.stlouisfed.org/series/IHLIDXUSTPSOFTDEVE
12. "Tech Layoffs Outpace 2025 As Big Companies Shift Spending To AI". Crunchbase News, 25 de septiembre de 2026. https://news.crunchbase.com/layoffs/2026-layoff-numbers-rise-ai-shift-orcl-meta-amzn/
13. "PitchBook-NVCA Venture Monitor, Q3 2026". NVCA y PitchBook, octubre de 2026. https://nvca.org/pitchbook-nvca-venture-monitor/
14. "Crunchbase Data: Q3 2026 Posted A Record Count Of Billion-Dollar Rounds". Crunchbase News, 5 de octubre de 2026. https://news.crunchbase.com/venture/q3-2026-global-startup-funding-ai-billion-dollar-rounds-exits-data/
15. "Good news: AI Will Eat Application Software". Andreessen Horowitz (a16z), 2 de marzo de 2026. https://a16z.com/good-news-ai-will-eat-application-software/
16. "The SaaSpocalypse Is Officially Over...". SaaStr, aprox. mayo de 2026. https://www.saastr.com/the-saaspocalypse-is-officially-over-public-software-is-back-to-green-at-the-index-level/
17. "Large Firms With at Least 20 Employees Biggest AI Users". U.S. Census Bureau, 26 de mayo de 2026. https://www.census.gov/library/stories/2026/05/ai-use-businesses.html
18. "Data centre electricity use surged in 2025...". Agencia Internacional de Energía (AIE), 16 de abril de 2026. https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions
19. "AI Act | Shaping Europe's digital future". Comisión Europea, consultado el 25 de octubre de 2026. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
20. "Timeline – Artificial intelligence". Consejo de la Unión Europea, consultado el 25 de octubre de 2026. https://www.consilium.europa.eu/en/policies/artificial-intelligence-act/timeline-artificial-intelligence/

What I could not verify: I did not find a reliable source for the number of startups that closed in 2026, nor independent measurements of the newest models; that is why I do not include them.