New Year, new predictions. As with previous years, I’ve collected my thoughts and the progress we’ve made in 2025 to see how we did on our past predictions and how that can inform this year’s prognostications.

We’ll cover technology, markets, and products, leaving no stone unturned, from the first big AI IPO coming in 2026, my personal bets for companies that will do great in 2026, the next big trade, what’s next after LLMs, my fears of what might go wrong with AI, and finally, things that won’t happen in 2026.

Let’s dive in.

AI has changed how consumers shop by speeding up research. But one thing hasn’t changed: shoppers still trust people more than AI.

Levanta’s new Affiliate 3.0 Consumer Report reveals a major shift in how shoppers blend AI tools with human influence. Consumers use AI to explore options, but when it comes time to buy, they still turn to creators, communities, and real experiences to validate their decisions.

The data shows:

The most effective brands are combining AI discovery with authentic human influence to drive measurable conversions.

Affiliate marketing isn’t being replaced by AI, it’s being amplified by it.

Download the full report to see what this means for your brand.

Let’s start by taking a brief look at how I did in my ‘predictions for 2025’ piece back on December 30th, 2024. A humbling way to see what real progress we’ve made this year.

My first prediction was that future reasoning models would use search algorithms to search the solution space. I’ll give me half of the points on this one. Not because I was wrong about the essence, but I was wrong about the structure.

Human intelligence is a combination of three things: intuition, search, and feedback. Our experiences and knowledge build intuition, which is essentially pattern recognition that suggests how to approach a problem.

But intuition only takes us so far. With complex problems, we need time to think. This introduces the idea of intuition-guided search. Using our intuition of what might be good ways to solve a problem, we search for the optimal solution.

But human intelligence has one last feature: the ability to adapt to feedback, something AIs do not yet have, because it’s a very problematic procedure that usually leads to performance decreases (known as catastrophic forgetting).

Humans continuously experience the consequences of their actions, shaping their beliefs and thus priming them for better predictions in the future. AIs do not encode or hold beliefs (at least not explicitly), nor do they adapt to new data in real-time (for now).

Knowing all this, the prediction was that, in 2025, we would see AI conquering steps 1 and 2; an LLM serving as the intuition machine, using explicit search over possible solutions at runtime.

At that time, most models behaved like an option second-to-the-left, just a straight line of thoughts. Instead, I predicted models would evolve into the thing on the right, where they would explicitly reevaluate their thoughts and decide which reasoning chains to pursue and which to “kill”, creating a “tree of thoughts”.

And do top models look like today? Well, not one, not the other, but the one in between, self-consistency models.

Today, most models behave the same way they did one year ago, but with two big chances:

Put another way, they are still a single chain of thought, but a more powerful one that imitates the tree-of-thoughts implicitly, not by discarding reasoning chains and opening new ones, but self-correcting in the same chain.

Therefore, the only reason I don’t give myself full points here is that, even though the most advanced models use verifiers and other stuff very reminiscent of a tree of thoughts, Labs explicitly rejected the use of MCTS-style search (Monte Carlo Tree Search), despite my intuitions.

I first predicted that 2025 should see the first AI-led discoveries, even some worthy of Nobel Prizes. This prediction was accurate… sort of.

In 2025, AI did, in fact, present itself as a great tool for discovery, but not in the flamboyant, spectacular way I predicted. However, it did:

And others.

Thus, I’ll give myself half points on this one. Not sure if they are “Nobel-worthy” today, but in my view, there’s no doubt that people assisted by AIs will win most Nobel Prizes in the future.

Even if AIs aren’t the ones that actually came across the solution, they can help you think, which has always been my primary use case for this technology and for me and many others.

This one’s fast. Full scores, because Factory AI released a $2,000/month subscription a few months ago, called the Ultra subscription.

But will they get more expensive? We’ll answer that in our 2026 predictions section.

I believe this was spot on. In 2025, all AI Labs were pressured to focus not just on brute-force intelligence but on intelligence-per-cost.

Probably nothing exemplifies this better than the cost comparison between o3-preview, from around a year ago, and Gemini 3 Flash, released just a few weeks ago, which has similar nominal scores in the ARC-AGI 1 benchmark, yet costs just $0.17/task. At the same time, the former required $4.5k/task for similar results, meaning Gemini 3 Flash represents a 26,000x reduction in the “cost of intelligence.”

2024 ended with o3-preview “telling the world” that progress was becoming too expensive. Thus, in 2025, we saw a clear effort by AI Labs to be more frugal and improve intelligence-per-cost.

Full scores, although I don’t feel like a genius predicting the S&P would go up, seeing its performance over the last decade.

Instead, I’m much more proud of the fact that:

The truth is, the only reason we are performing so well is that AI has absorbed all attention, and AI is precisely what I’m good at. As they say, everyone is a genius in a bull market, so even though we’ve been choosing big winners consistently, let’s keep ourselves humble.

Full scores. The first gigawatt data center came from xAI, with Colossus 2. Now, this site is planning to grow much larger, and they just raised another $20 billion, which signals to me that they are probably going to be the first to reach 2 GW deployed on a single site.

But why such large data centers? As Jensen briefly hinted at during CES, NVIDIA’s Rubin platform is expected to support training models with up to ten trillion parameters, the size he expects the next frontier to lie in (and thus preparing its hardware for).

Training such a model requires an equally impressive training budget, likely exceeding 10²⁸ FLOPs, about ten times the size of Grok 4’s training. But how much is that? This is a very coarse estimate, but Transformers require around {6xN} FLOPs per training prediction, where ‘N’ is the number of parameters.

This means that if the training budget is that size, we can write 10²⁸ = 6×10×10¹²*D, where ‘D’ is the number of training tokens. Assuming a 10 trillion-parameter model, that gives us a size for D of 1.67×10¹⁴, or a dataset with 167 trillion tokens.

This sounds like a lot, because it is. And that is why you have to make models bigger. For instance, if we use a model that is ten times smaller, like China’s largest models today (Kimi K2 has 1 trillion parameters), that would still require a dataset ten times larger than the one above to achieve the desired training budget.

Naturally, this is a very coarse estimate that doesn’t account for the fact that models are mixture-of-experts (reducing activated parameter count), use tools during training, and, more importantly, it doesn’t consider Reinforcement Learning (RL) at all, a type of training that skyrockets traning budgets without having to make training datasets larger.

Consequently, the actual dataset is likely much smaller than the number above, as Labs increasingly rely on the RL part of the training process. Rather, the takeaway is that training budgets are still huge and thus require many concurrent accelerators to process the dataset in a manageable time.

Overall, I was correct on this prediction, but I won’t give myself full credit because I wasn't specific enough about what ‘traditional’ meant.

When I shared my very bearish sentiment regarding SaaS, I was thinking of companies like Salesforce or Adobe, which indeed had a terrible year (see below), but to get full scores, I should have mentioned that some SaaS companies, especially those more linked with AI, would still shine.

Nonetheless, pure AI plays like Palantir were one of the best-performing stocks of all industries and sectors. Cloudflare’s omnipresence on the web and its role as a gatekeeper of data from agents ensured it had a great year too; Shopify's unequivocal embrace of AI is paying off; and data platforms for AI, like MongoDB or Snowflake, or mission-critical businesses like CrowdStrike, did, unsurprisingly, absolutely fine.

The idea was correct, but saying “SaaS will die in 2025” was an exaggeration.

Wrong. It’s something I do not discard could happen in 2026, but it nevertheless did not occur in 2025.

Correct, as demonstrated by complex world-model systems like Waymo, which leverage video and text models to drive cars. More on this later.

Wrong. I predicted that 2025 would see a shift in how we “build software” toward declarative software, one where the entire logic of the application is controlled by LLMs. This is still not the case today, even though I still believe it is inevitable.

Overall, a not too shabby score of 6.5 out of 10 (considering 0.5 for partial scores). And now, it's time for the 2026 predictions, which, as you’ll see, hold several key insights to watch out for.

We are off to a strong start here with the rise of state capitalism in the West and its strong impact on AI.

National security has become the center point of all decisions since Trump's explicit embrace of the Monroe Doctrine, proclaimed by US President James Monroe in 1823, a foundational US foreign policy that warned European powers against further colonization or interference in the Western Hemisphere.

However, the exception is that the newer version, the playfully referred to as “Donroe” Doctrine, is mainly aimed at China and Russia, not at the European countries whose best days are gone unless they execute profound reform, and which have completely lost the AI train.

And what does this have to do with AI? Well, everything.

Trump, initially much less interested in foreign interventionism and instead being a more “profits-focused” president, seems to have veered back into what the US does best: make decisions based on national interest, trying to secure the supply chains of the technologies and assets considered a priority.

In one way, one could say the US is starting to behave as if it were at war, similar to how it behaved for much of the 20th century. This may not sound particularly transformative, but let’s not forget that John Kerry, Obama’s Secretary of State, said in 2013 that “the era of the Monroe Doctrine is over.”

More specifically, with respect to AI, this means we now have confirmation that AI is under the flag of state capitalism, a technology in which the government’s role will be huge.

Interestingly, among all the things that divide US citizens, hawkishness toward China is bipartisan, so that even if Trump were not a larger-than-life figure who draws all the attention, this decision would have happened either way.

And guess what technology falls neatly in all this? Semiconductors.

Amongst all the many battles that will be “fought” in the US-China war for supremacy, semiconductor supply chains are, without out doubt, one of the main ones. If Operation Barbarossa represented the beginning of the end for the Axis powers in 1941 during the Second World War, the US's failure to secure the semiconductor supply chain could be its particular Barbarossa failure.

The problem is that one of China’s biggest levers over the US today, besides its tight grip on rare Earth refining, is the threat it poses to the global semiconductor chain, which is extremely bottlenecked in chip manufacturing and packaging in Taiwan.

For the US, a country that wants to maintain its supremacy, this is intolerable, and they urgently need to diversify out of Taiwan (while still securing the island for years to come, because this dependency is not ending anytime soon).

Put mildly, the world’s dependence on Taiwanese semiconductor production is absolute, even though Taiwan’s national chip champion, TSMC, has opened fabs in the US. But let me be clear, the relief that the US might feel on this is an illusion. At the end of the day, it’s a Taiwanese company defending Taiwan's interests, which, in this case, are not aligned with US interests.

For all these reasons, nothing matters more than being absolutely “AI and chip independent”, so I predict interventionism in AI markets will become the norm, a term known in geopolitical circles as ‘state capitalism’, best represented by China, be that via direct investments or subsidies (the former is much more likely, given Trump’s well-known dislike of subsidies). I believe this support will be mostly based on providing liquidity at any cost, which could be very inflationary, but that’s another story.

Moreover, the US could also start looking at/pressuring other countries for turbines and access to power in general, potentially offshoring large-scale data center projects and keeping MW-sized data centers inside the US, with the Middle East being a clear-cut candidate (especially considering data centers aren’t popular infrastructure investments because they generate few long-lasting jobs).

Likely candidates for intervention? The AI Labs, especially OpenAI. And talking about AI Labs…

I predict we will see two big AI Lab IPOs in 2026 because liquidity isn’t eternal. The best proof of this is China, which is well underway with its AI Lab IPO craze, with the upcoming IPOs of Zhipu and Minimax.

And while liquidity in US private markets remains strong, seeing the very recent xAI $20 billion Series E, the capital requirements are so enormous that these Labs can’t raise the funds needed to fund their ambitions. For instance, OpenAI alone has $1.4 trillion in commitments for +40GW.

So, who is going to IPO? Well, funnily enough, it’s probably not OpenAI, but two other much more likely candidates you might not expect. And the answers are Anthropic and Mistral.

For the latter, their bet to concentrate efforts on coding and agents, meaning models are developed almost exclusively for those use cases, is paying off.

Not only are the models the best in those areas, giving them pricing power and better revenue per user (Anthropic’s models are way more expensive than the rest, despite, ironically, being worth it), but, crucially, it also allows them to concentrate and minimize R&D compute.

This last point is critical. As we suspected from OpenAI’s cost structure leaked by The Information (shown below), and as we have also learned from the IPO filings coming from China, R&D compute, the costs of running experiments, trying out new stuff to see what works, are one of the largest expenses.

For Anthropic, this is not as big a problem because they can concentrate efforts and teams in two simple directions: making their AI models code and call tools better.

All this is to say that Anthropic is likely to be way more profitable (or, to be more specific, less unprofitable) than OpenAI. In fact, their respective internal cost projections reveal this huge disparity:

More visually, this just means Anthropic is expected to be much more efficient than OpenAI over the long run, measured by the percentage of revenue server spending represents:

These better overall business numbers, coupled with the fact that Anthropic doesn’t have nearly as much access to liquidity in private primary and secondary markets as OpenAI (i.e., their cost of capital is higher), make their IPO much more likely.

On the other hand, we have Mistral. Let me be clear, this prediction is totally based on vibes and zero evidence.

I believe they will also IPO in 2026, but for very different reasons: Their revenue numbers are likely to be horrifically bad relative to rivals, and they desperately need access to huge liquidity to grow (the AI Lab market is very much grow-or-die). On the margin level, their numbers are probably not as bad and thus won’t look terrible on an IPO; their issues are likely mostly about top-line growth.