But eras don’t end because demand disappears. They end because the bottleneck moves. GenAI might mark that shift for software. One question remains: what comes next?

The inflection point: Coding becomes optional

Early in 2024, Nvidia’s founder and CEO Jensen Huang made headlines by arguing, at a summit in Dubai, that programming should no longer be the primary skill children focus on learning.

“It is our job to create computing technology such that nobody has to program. And that the programming language is human. […] Everybody in the world is now a programmer. This is the miracle of artificial intelligence.” explained Jensen.

Critics pushed back. Some argued AI would not reach human coding level—or at least not “organizational” coding level. Others suggested AI would simply enable more people to code.

We are now two years later. A short amount of time, but enough for Cursor, Codex, Claude Code and others to rise. And despite the criticism, it increasingly looks like AI is good enough at coding to change incentives: not perfect, not magical, but capable of producing senior-level output in many contexts. The same is becoming true at the codebase level: not “solved” in every company, but increasingly viable in real organizations.

So the most interesting remaining question is not whether AI can code. It’s whether it empowered more people to code, or whether it changed the paradigm of building software.

A useful way to think about that is with an analogy: painting.

Painting has long been an art of the brush. You study it, practice it, correct it—then eventually your idea takes shape on a canvas. Then, in the 60s, came interactive drawing software. Anyone with a computer could now “paint” 1. Did this tool kill painting? Not fully. But it did change its economics—and, in many segments, its purpose.

Cartoons were once hand-drawn; now most are made through software. Background illustrations for movie sets, ads, and corporate visuals used to provide a living for skilled drawers and painters. Few people could do it well because it required years of training, and clients understood how long it took because it was a task of the physical world. Digital tools collapsed iteration time. Clients started expecting faster delivery with more edits. The work became more accessible, supply increased, competition intensified, and prices fell. Painting did not disappear—yet the “center of gravity” shifted. Today, its biggest economic use is arguably fine art, where taste, identity, and originality matter more than speed.

Interactive drawing software didn’t kill painting. It shifted the utility of the job: where the jobs are, what customers expect, and which skills get rewarded.

My belief is that the software world will undergo similar shifts in the coming years—maybe decades.

AI-powered tools have made access to software creation much easier. But it didn’t make more people software engineers. It gave more people the ability to produce software-like outcomes without having to learn to code—or even to interact with code. The same way you don’t need to know how to paint with a real brush to produce compelling visuals with modern tools, you increasingly don’t need to know how to code to build an app with platforms like Replit.

This nuance matters. It’s not “AI enables everyone to code.” It’s “AI makes coding optional for many outcomes.”

As with painting, this doesn’t mean the world will require no developers. It means expectations will change. Something that once required a team of ten trained workers over ten months can, in some cases, be prototyped by one person in a day. Developers will still be required—but the number of roles, the nature of the work, and the premium skills will likely shift.

The bottleneck: Adoption

To understand what might come next, it helps to name the main constraint facing AI’s expansion right now: technology adoption.

Model providers—Anthropic, OpenAI, and others—have shown that capability will continue to improve. It might not feel like it, because our brains compress timelines, but remember: four years ago most people had never heard of GPT. Two years ago it was “fun text that gives quick answers.” Today it can generate graduate-level analysis and turn it into a podcast or a website in minutes—work that used to take hours of engineering time.

Yet despite these capabilities—and despite the investment and attention—we still haven’t seen changes to society at the scale that the hype would suggest. If you are a white-collar worker, your personal workflow may have changed a lot. But did your company truly change? Was its business model redesigned? Did it triple revenue because of AI? In most cases, no.

The world is advancing at dual speed: models are getting faster, better, more extraordinary—while many industries are lagging behind.

A plausible reason is adoption. We see an increasing number of pilots, MVPs, and tests, but fewer systems that stick, scale, and reshape core workflows. Adoption is complex, and there are many reasons: corporate structure, the explore–exploit dilemma2, employee training, data readiness, risk and compliance, incentives. Fixing this will likely take years because organizations themselves need to change.

But adoption doesn’t fail only because companies are slow. It also fails because the gap between the people building the technology and the people operating complex industries is wide—and widening.

Incumbent industries are complex. Their processes and operations evolved over decades (sometimes centuries), making it nearly impossible for an outsider to grasp their real functioning quickly. You can learn the basics of metallurgy from a documentary, but you’ll still be extremely far from understanding how a company like ArcelorMittal actually runs day-to-day: its constraints, internal politics, regulatory boundaries, unions, supply contracts, safety culture, asset cycles, and the thousand “small” exceptions that are, in practice, the business.

Yet for AI to transform how such a firm operates, the people implementing AI must build that understanding first.

Most tech employees have scarce knowledge of how other industries work. Many studied STEM, then joined startups or big tech. They may be excellent engineers and still be unprepared for the operational reality of a complex incumbent.

Vice versa, many people in incumbents lack a practical understanding of modern AI systems. They don’t know what LLMs are good at, where they fail quietly, how to evaluate reliability, or what it means to deploy these tools inside regulated or safety-critical workflows.3

This dual knowledge gap is not new. The software world already struggled with it. But as technology becomes more powerful—and therefore more consequential—this gap becomes a harder blocker. It is now a central constraint on adoption.

The gap is complex, but it isn’t immutable.

The rise of the bridge-builders

This gap helped create entire industries: digital transformation consulting (Accenture, IBM, and others). It also created roles inside tech itself: solution architects—often senior technical professionals who design, develop, and implement technical solutions to solve business problems. In a nutshell: people bridging business needs and technical execution to enable adoption.

Until now, in the software-led economy, large tech companies often ran at something like one solution architect for ten software engineers. My bet is that ratio will change.

As code production becomes cheaper and more automated, fewer people will be required to produce the same volume of software. Meanwhile the adoption gap between tech “producers” (Google, OpenAI, Anthropic…) and tech “receivers” (incumbent industries) remains—and in some ways grows—because what matters shifts from “can we build it?” to “can we embed it safely, reliably, and profitably into how work actually happens?”

This creates demand for a broader family of bridge-builders: not just “solution architects” as a title, but people who combine domain knowledge, technical fluency, and organizational change ability. Call them architects if you want. The label matters less than the function.

To illustrate the shift: a company integrating AI into supply chain management may no longer need ten engineers to build everything from scratch. Tools will handle a growing share of generation, refactoring, and integration scaffolding. What the company will desperately need is someone who understands both (1) the real operational constraints of supply chains and (2) the practical capabilities and failure modes of AI systems—well enough to decide where automation creates value, where human judgment must remain, how exceptions get handled, how risk is governed, and how the transition is orchestrated without breaking the business.

That is the architect’s domain.

We’re already seeing early signals: “AI transformation” roles, product teams tilting from building to integration, and growing demand for people who can translate capabilities into operating reality.

Over time, the scarce resource won’t be code. It will be trustworthy change: deciding what to automate, proving it works under real-world exceptions, governing risk, and moving a human organization through the transition without breaking it.

In the software era, leverage came from writing code. In the AI era, leverage comes from shipping new behavior.

The new premium skill isn’t programming. It’s integration.

2025 was not just another fast year in AI. It marked a shift: from experimentation to scale, from promise to pressure, from models to markets. Reasoning systems replaced simple LLMs, price per compute became a competitive weapon, AI companies pivoted toward B2B survival, and capital flooded the ecosystem at levels never seen before.

This article is a synthesis of the most important shifts that defined AI in 2025 — and a set of informed bets on what is likely to matter in 2026.

The 2025 shifts: Toward reasoning models, toward price per compute, toward B2B, toward scale.

2025 was a fast-moving year with a lot happening in the AI ecosystem. Models evolved fast and switched from basic LLMs toward reasoning models. It was also the year of multimodality, and we started seeing image and video generation both amazing and terrifying. As a last point on the technical side, the entry of DeepSeek on the market reshaped expectations on compute cost and led to a strong focus not only on “smartness” but on pricing.

Over 2023 and 2024, model providers (e.g. Anthropic) gained unbelievable adoption on B2C; however, the business model does not seem sustainable, and they are now turning toward B2B as the solution to their financial hardship.

Finally, 2025 was the year AI was pushed to another scale. Thanks to massive investment from giants like Nvidia but also to some political play, we are now seeing the rise of massive projects like Stargate and unbelievable effects on the market (e.g. NVIDIA valued at $5T) leading most to fear a bubble. In a sentence: if you thought 2025 was quiet, think again.

Models evolution

The “LLMs have peaked” argument did not survive contact with 2025. We did not just see new versions. We saw new behavior.

2025 was the year we moved from “simple” LLMs toward reasoning models capable of working on their own for hours. Add to this new capability considerable improvement in the capacity of the base models, and major improvement in multimodality (especially when looking at image — Nano Banana — and video — Veo or Sora — generation).

Models have drastically improved in the span of a year, perhaps more than what we had seen over 2023 and 2024.

Price per token

Back in January 2025, the general belief was that more compute and more power would provide better results, then came DeepSeek. A Chinese PE company which introduced a model nearly as powerful as top players. But aside from the fact that no one would expect a PE company to produce a top notch model, the crazy improvement of DeepSeek was its cost efficiency.

Media claimed the company had trained the model for 5.6 million dollars — a massive improvement compared to the billions poured by OpenAI or Google. This news slowly created a shift in mindset and revised the expectation that smarter would mean more expensive.

(It soon came to many’s attention that the $5.6 million cost was highly misleading; still, the cost of training the model was considerably lower than most Silicon Valley companies had achieved and the impact was there.)

Fast forward to the end of 2025, competition among the top model providers is focusing increasingly on “price per 1M tokens”. Of course, this is also due to top players achieving similar performance in terms of capabilities, thus needing other areas to differentiate.

A new focus on B2B

There is no question that LLMs have made their way into the public’s hands; however, while the traction is massive, profits are not.

With subscriptions at $20 per month on average, model providers are far from generating enough cash to compensate for their monstrous spending. Sam Altman even said that their $200 Pro subscription was costing OpenAI money and OpenAI was expected to lose $5 billion in 2025.

Facing a challenge in B2C, tech companies are turning to B2B to generate more revenue. Within this shift, I observed 3 main plays:

1. Incumbent platforms embedding AI into existing channels: Google and Microsoft are both leveraging their cloud B2B business to push more and more AI into existing services (think Word, Teams, Docs, Gmail…). Each with different pros and cons. Microsoft is clearly profiting from its market monopoly but — lacking a strong enough in-house technology — needs to rely on OpenAI and Anthropic models. On the other hand, Google provides its Gemini model with fast and better (in my opinion) cross-tool deployment, but is far from having the scale of Microsoft when it comes to company access. This strategy comes, however, with a downside: it takes time, costs a lot of money to prospect companies, and doesn’t always work perfectly.

2. Frontier labs selling “Moonshots”: On the other hand, new tech players like Anthropic or OpenAI have no previous B2B services they can rely on. They need to create new business. Not able to compete with tech giants, their solution has been to focus on two axes: (1) they provide their models to peers who will then create corporate solutions (e.g. Copilot); (2) they sell themselves as top technology providers and promise fast iteration and deployment of “moonshots” - versus integration to existing services.

3. Tailoring like …: Among the AI tech companies, which one showed the strongest stock growth this year? Is it Nvidia? Google? It is Palantir with a whopping +150% YoY. The lesson? If you can’t match Microsoft’s distribution or Anthropic’s frontier models and branding, go vertical. Build tailored solutions for specific industries. European players like Mistral and Aleph Alpha are pivoting in this direction

Scaling laws are… broken?

Finally, perhaps the most noticeable aspect of 2025: markets went crazy.

Since the end of September, we have observed rising concern and debate about a potential AI bubble. There are multiple reasons behind the claims, mainly NVIDIA’s circular investment or the investment schemes used by model providers and startups in the ecosystem. But at the core of it is one undeniable thing: we are freaking out because we’ve never seen numbers like this.

Nvidia’s valuation exceeded $5T for a bit — more than the annual GDP of Germany, the 3rd strongest economy in the world. Palantir closed the year with 150% growth, AMD with 78%, Alphabet 65% and Nvidia 39%. We did also see some slower players too: Apple stock only grew about 9%, Meta 13.9%, (both lower than the S&P 500 which grew about 16%).

Investment also became crazy. The year started with the announcement at the White House of the Stargate project with a plan to invest $500B in infrastructure by the end of 2029; then for a moment, it was rare to live a day without seeing a new multi-million-dollar acquisition from a giant; thus it is not surprising that some estimated the total amount of AI CapEx in 2025 to have exceeded $400B.

So is this just the new normal? Or should we expect a massive correction akin to the year 2000? At the end of the day: no one knows. One thing is for sure, investors are placing heavy bets on the future of AI and its dreamy potential.

My bet for 2026: A future full of potential and dangers

As we are stepping into 2026, some key trends are emerging.

First and foremost, 2025 saw the emergence of “agents” — pre-prompted models with access to sets of tools. It is very likely that 2026 will see more and more agents and that they will be able to work together to provide better results. But current tech is still limited to computer ecosystems. So what comes next? Robots.

We should expect the rise of robotics companies producing AI-trained robots, smarter and more human-like than ever before. All those developments will for sure have an impact on the workforce and we should expect several developments there.

On another front, AI has proved itself to not only be a technology but a major political tool, and given the current geopolitical developments we should expect more AI political powerplay especially on the US-China front.

Finally, what’s the far (or perhaps not so far) future? Besides LLMs, other types of models have also improved a lot and we are seeing and should expect even more improvement in a couple of areas including physical simulation and new AI domains like World Models.

Development of multi-agent systems

Spring 2025 marked the beginning of the agents play. Slowly they became really useful and most tech companies started rolling out ways to build your own.

At the core of it, agents are LLMs pre-prompted to play a certain role and given access to a memory and a set of tools they might need.

To better understand the future of agents, I invite you to picture them as a specialized employee. Imagine you have one. Great, she is brilliant, does the work perfectly, but a company is not just one individual and she cannot change everything on her own. But what if you have five specialists together — perhaps each with a different and complementary set of skills? That gets you a very high performing team. Double it to 10 and add an orchestrator — a maestro to manage them all. You get a department. A hundred more? You get an unbelievably efficient company.

That’s the opportunity the realm of multi-agents creates.

Systems like Makers or N8N already allow users to create and make agents interact with each other. Thinking ahead, it is not hard to imagine a future where entire companies are agent-run.

Reaching the physical world

One of the shortcomings of AI as we’ve been talking about it for the last 2 to 3 years is its ability to reach the “real world”. Think of the infamous boomer’s comment “ChatGPT is smart but can it fold my laundry?”. Well, as a matter of fact, it might soon be able to.

A lot of expectations are building up on robotics, perhaps even more than we had with AI. We are seeing new entrants such as NEO or Figure 3 — working on creating humanoid robots acting as butlers, folding your laundry, watering your plants… But also some incumbent players such as Tesla with its Optimus robot.

Robotics and AI don’t stop at funny looking humanoids; AI technology is greatly helping improve robots everywhere, from robot vacuum cleaners, Waymo and Tesla autonomous taxis to Boston Dynamics’ Spot dog robot.

As AI models get faster and lighter, this gives engineers new opportunities for embedded systems capable of using them. Combined with increased productivity thanks to AI helping with design or coding, it is very likely to expect robotics companies to produce great feats in the next couple of years.

Impact on the workforce

Technology is nice but it is not all there is to this revolution. 2026 will probably mark history as the year we start observing real impact of AI on the workforce.

We can expect multiple levels of impact:

1. Job scope/task change: The basic level of change would be some task rescoping. If an agent is capable of performing part of your job, then you might not need to perform it yourself anymore. The good aspect of this is that it allows you to focus on more decisive tasks and on orchestration instead of operation. However, if poorly managed this can lead to ii.

2. Job redundancy: This is perhaps one of the biggest fears surrounding AI, and many like to say that it is not true, that we’ll find ways around it. I personally don’t believe so. The same way agriculture machinery reduced the number of people needed to work in the field, then computers reduced the number of people needed to work in certain areas, AI will for sure make some jobs redundant.

Of course, we have been talking about (1) and (2) for a while, so why would it not happen in 2025 and happen now? The context has changed.

2025 saw a big shift in tech companies going from building the technology and providing it through B2C to now having a working technology that they are pushing in B2B. On the other side, attracted by the potential efficiency increase, big corporates are investing massively in implementing those technologies. If implementation works, then we should definitely expect the occurrence of (1) and (2) to happen more often.

But there is a third impact that is less talked about. The (sad) impact on our brain.

GenAI gained massive adoption in the office over the last 3 years. However, the more people use AI tools without being fully trained on them, the more we can expect them to produce repeatable low quality work — or “slop” — and to progressively lose their own ability to reason. In its strategic prediction report for 2026, Gartner frames it as the “atrophy of critical-thinking skills, due to GenAI use”. Think of that one time you needed to brainstorm and perhaps your first reflex was to open ChatGPT.

There is a very high chance that to counter this, companies will need to actively fight the problem.

How? Given the current massive adoption context combined with the dreamy top line growth potential, it seems unrealistic that the way to safeguard companies from becoming brain dead because of AI is to fully lock AI. Instead, some companies are pushing their effort in the other direction: unlock access, let employees use as they see fit, but: train them.

Geo-politics: US, China and EU

This should come as a surprise to no one: 2025 was driven by strong unpredictability in geo-politics. We should not expect 2026 to be any different and AI is partly a tool, partly a driver of political discord.

To understand the implication AI has on our geopolitics, I would encourage any reader to review the AI 2027 report. While you might find that their predictions/ideas are somewhat far-fetched, the general dynamic they are painting is well thought through and good to have in mind. Allow me to summarize it.

GenAI is becoming a powerful tool: a booster. With AI, you can enhance your productivity, code faster, develop new products, accelerate growth, and ultimately strengthen your economy. Having the smartest AI isn’t just about having a fast Claude; it’s about having a better economic booster than everyone else. Therefore, it’s logical that all major powers would invest heavily in developing AI capabilities and ensuring that the smartest models are developed at home. It also gives you strong leverage in negotiations with others. However, AI relies on a complex production chain. Training AI requires massive data centers, which need graphics cards and energy, which in turn require minerals, oil, or uranium — all driven by different industries with strong ties to their operating locations. Think of Taiwan. Thus, securing AI requires geopolitical leverage.

This circular rationale shows how AI might lead us into more geopolitical instability driving global import and export regulation or country partnership/feud.

You can observe the intricate relationship between AI companies and political power by looking at the game their leaders are playing. Increasingly, tech CEOs gather with presidents or prime ministers, get invited to fancy dinners or appear as advisors to political figures. Silicon Valley is definitely done restricting itself to the west coast, and is developing close ties with Washington that both parties surely intend to leverage to accomplish their objectives.

At the forefront: research and world models

There is what is happening now in the public, and there is what is happening in the lab.

Too often we confuse LLMs with AI. However, large language models are but a tiny fraction of the research conducted on AI. According to some scientists like Yann LeCun — former chief scientist at Meta — LLMs would not even be the key to “super intelligence”. So what else is happening in the AI world that we should be paying attention to and expecting for 2026?

To answer this question easily I would recommend looking at one lab: DeepMind.

DeepMind has been at the forefront of AI Research for more than a decade. They are the reason Altman and Musk created OpenAI, they are the team behind LLMs, and they are the team at the frontline of Google’s AI effort. So what are they working on now?

Two main themes emerge:

1. Physical simulation models: AI models capable of simulating physical phenomena with unprecedented accuracy. Think of weather prediction, natural catastrophe modeling, or material science. These models could revolutionize how we understand and predict complex physical systems, from climate patterns to the behavior of new materials under stress.

2. World Models: Perhaps the most exciting development is the emergence of “World Models” - AI systems that can understand and predict how the world works by building internal representations of physical reality. Unlike traditional AI that learns patterns from data, world models aim to understand causality and physics, enabling them to reason about scenarios they’ve never encountered before. This could be the bridge between narrow AI and more general intelligence.

These developments represent a fundamental shift in how AI interacts with the physical world. While LLMs revolutionized how we process and generate language, physical simulation and world models could revolutionize how we understand and interact with reality itself. The implications for fields like engineering, manufacturing, urban planning, and scientific research are exciting.

So how should we interpret all of this?

The world is changing fast, and the control remote seems to have been lost. Humanity is most likely witnessing a new revolution cycle. The implications it will have on work, life, the economy, disparity, the world order, are unclear. Many predictions are dooming and many notorious voices are rising, asking urgently to slow the carriage. Can we? Most likely not.

Revolutions have always brought their share of uncertainty. They change society quickly and have impacts that ripple over centuries. Part of those impacts are negative — unemployment, displacement, exploitation. But they also unlock tremendous opportunities — longer life expectancy, higher quality of life, access to education and healthcare. The key question is not whether we can stop it, but whether we can shape it to maximize benefits and minimize harm. And that will only happen by paying attention and being actors of this revolution.

Subscribe now