The AI Index Report 2026 is over four hundred pages and hundreds of charts, and out of all that material the Spanish press always picks the same three: the growing market, the company adopting, the public that's afraid. The ones left in the drawer are the ones that really make you uncomfortable. Geographic concentration, dependence on a handful of compute providers, public data that runs out, a population using a tool it doesn't trust. I'm going for those. I read them; the tech-section reporter with a seven o'clock deadline didn't.
The European Desert Isn't a Metaphor
Let's start with how the money is split, which is where you see everything else before anyone explains it. Private investment in AI during 2025 draws a map with two big blots and a void. The United States takes the lion's share; China, a fraction of that; and the rest of the planet, the European Union included, ends up lumped into generic categories because its figures don't even make up a column of their own. Spain, of course, doesn't appear broken out in the main chart.
The market's growth is real, nobody disputes that, and the report documents it in detail. What the press reports is precisely that, the growth. What it doesn't report is the geometry of the split, which is the interesting part: the money doesn't rise equally everywhere, it rises concentrated, and the distance between whoever has a model factory and whoever doesn't widens with every passing year.
Here a nuance is in order that's barely mentioned in Spain. The figure for Chinese private investment is systematically underestimated relative to the country's real effort, because much of China's AI spending doesn't run through the private-capital lane the AI Index measures, but through government-guidance funds that venture-capital databases don't capture. The report itself recognizes that this state-directed investment falls outside its private tallies. If you add the two lanes, the private and the public, the picture of "American dominance beyond dispute" becomes less emphatic. They're still two different funding models competing, and the Spanish press narrates the race as if only one of the two columns existed.
The Technical Gap Has Almost Entirely Closed
There's a cliché that holds up poorly against the 2026 data: that of American technical superiority as an abyss. The report measures the performance distance between the best US model and the best Chinese model and puts it, at the start of 2026, at around a couple of percentage points. To get a sense of the speed of the narrowing, at the start of 2025 a Chinese model momentarily matched the US leader in the public measures of user preference.
In the top spots of those preference tables there's no longer a single flag. Among the best-placed models, US labs and Chinese labs coexist, and the technical frontier, far from being one country's property, is split between two. The rest of the world watches the race from the stands.
The United States still publishes most of the relevant models of the year; China, not far behind. Europe doesn't show up with a model of its own in the front line. And then things get interesting for us, because with stunted investment growth and no model at the frontier, you have to ask what exactly Europe is doing in this matter. What it's doing is regulating. It decided that its playing field wasn't to build the models, but to set rules for the models others build, and the European AI regulation approved in 2024 is the materialization of that bet. As a legal piece and a political gesture, the regulation does its job. As an industrial engine it doesn't start, for the simple reason that it was never designed to start anything.
The Spanish-language debate on AI goes, almost entirely, in that regulatory direction. What to ban, what to control, what risk to watch. The other debate, the one about who puts up the money, who builds the infrastructure, with what data and under what architectures, either arrives translated from English months late or doesn't arrive.
The Cost Curve Switches Off the European Conversation
The figure that really closes off the fantasy of a European frontier model has nothing to do with talent or regulation. It has to do with the arithmetic of training. The cost of training a frontier model has been growing at a rate that roughly doubles it each year, and that turns the barrier to entry into a wall that rises on its own.
Where it used to be hundreds of millions per model, it's now thousands. And looking a bit further out, Epoch AI has modeled how frontier training clusters scale toward the end of the decade: in its analysis, the compute needed for a cutting-edge training run toward 2030 could be on the order of tens of billions of dollars, with aggressive scenarios topping a hundred billion. It's not a closed-model bill, it's the scale of the compute apparatus you have to build to play in that league, and it's best read as a lower bound as much as an upper one.
At those prices, no European public institution gets into the room. It's not a matter of political will or wounded national pride, but of an inequality you can settle with a calculator.
Spain funds the public program Generación D, known as "the AI Generation," with a hundred and twenty million euros spread over four years among the CSIC, the Barcelona Supercomputing Center, the CNIC and the CNIO, according to the government's official note. It's real money to train hundreds of researchers. Compared to the cost of a single frontier training run, however, that entire budget barely grazes the scale of a cutting-edge compute bill. If the declared goal were to compete at the frontier, the numbers didn't add up before and add up less now. It's worth saying it without anesthesia, because there are institutional speeches that hint at the contrary.
Which doesn't condemn the idea of funding public AI. The condemnation it deserves is another: any program sold to the citizen as "the European alternative to OpenAI" is lying to them about what its budget allows. An honest conversation would be possible —what kind of public AI is in fact viable with Spanish money: vertical, specialized, built on open models and fine-tuned with proprietary data— but that conversation isn't being had.
Public Data Runs Out, and That Divides the Future
Epoch AI has spent years estimating when the high-quality public text available to train models will run out. The range they handle across their various revisions of the study is wide, placed roughly between 2026 and 2032, with a median that tends to fall toward 2028. The margin depends mostly on how much you "overtrain," that is, on how many times you reuse the same data to squeeze more performance out of it.
If the models are trained at the ratio considered optimal, the stock lasts longer; if the data is reused aggressively, as happens in practice, the horizon comes closer. The exact date matters less than the direction of the movement, which is unequivocal: the well is drying up.
The consequence is the part the Spanish press leaves out. When quality public data grows scarce, the advantage goes to whoever has proprietary data nobody else can touch. The companies that already hold that asset —search engines with their index of the web, social networks with years of conversations, professional and office platforms with the work trail of half the world— are going to keep it like gold. The ones that don't have it are left with two ways out: manufacture synthetic data or buy exclusive access to someone else's.
And that's already happening. No need to imagine it. OpenAI signed content-licensing deals with first-rank media groups —News Corp, Axel Springer, the Financial Times, among others— to feed and back its models with their archives, as the sector's specialist press reported. Paying for access to a corpus someone else controls stopped being an oddity and became one more supply route.
Now, not all labs play the same game. Anthropic, for one, hasn't closed licensing deals of that kind with publishers and faces a Reddit lawsuit precisely for having used the platform's content with no prior agreement; its famous payout of one and a half billion dollars was a court settlement with authors over the use of pirated books, according to Bloomberg Law, not a data license. Lawsuits on one side, checkbooks on the other: the economics of training is closing with both instruments at once, before most Europeans have even understood what training a model means.
Fear Doesn't Slow Adoption, It Accompanies It
Here comes the most disturbing chart in the report, the one the general press barely grazes because it doesn't fit in a clean headline. The adoption of generative AI has reached a little over half the world's population in about three years, faster than the personal computer, faster than the internet, faster than the smartphone. It's one of the fastest diffusion curves on record.
And it isn't uniform. Some small, highly connected countries are well ahead of the average, while certain large economies adopt below what their wealth would lead you to expect. Spain isn't broken out in the main chart; the available indications place it in the European mid-range, with no more reliable detail to offer here.
What really throws you is what happens with trust while usage soars. The report documents an enormous gap between what AI experts expect and what the general public expects about this technology's impact: the specialists are consistently far more optimistic than the citizenry on jobs, on the economy and on health, with distances that in some sections comfortably exceed twenty points. The population uses the tool without trusting it, integrates it into its daily life without having fully decided that it suits them, adopts it without understanding it.
That pattern breaks the comfortable binary of "good AI versus bad AI" that sustains much of the coverage. The reality the data draws is messier: a technology people take up with growing distrust. That sentence doesn't fit in a headline, and that's why almost nobody writes it.
The Phantom Chart Is the One That Isn't There
There's a chart worth looking at precisely because it doesn't exist. The AI Index 2026 doesn't devote a chapter to the cognitive impact of the prolonged use of generative assistants. There are scattered mentions, there are references to recent work on what leaning on these tools does to attention and memory, but there's no chapter with that title. And the absence is eloquent.
It says, I suppose, that the conversation about what AI does to the brain isn't yet measurable enough to enter a report that presents itself as a compendium of verified data. Studies exist, but there's no methodological consensus, and the AI Index prefers to wait until there is. It's an honest decision. It's also revealing of how much of this territory remains in the dark.
The other absence that weighs is that of the concrete labor impact, sector by sector and country by country. The report brings aggregate data on productivity and on which tasks are exposed, but not an industrial breakdown with national figures, probably because that data isn't yet comparable across countries. The practical effect is that the public conversation about "AI and my job" keeps feeding on projections from consultancies and investment banks instead of an independent academic report. Two honest gaps, and at the same time the mirror in which you see what's left to measure.
What the Spanish Gap Means
Spain doesn't appear broken out in the main charts, and nobody at Stanford did it out of contempt. It's simply that the country doesn't generate the sectoral data that would fit the report's categories. There's no AI Index Spain to cross-check anything against, there's no breakdown by Spanish frontier model because the model doesn't exist either, and the private-investment line of its own is missing because the figure doesn't come close to the threshold a globally scoped chart demands.
Being left out of a global dataset isn't a grievance to defend against, but a symptom worth reading slowly. What that symptom says is that the conversation about AI in Spain is imported, translated and discussed a couple of years late, because the country itself still has no apparatus capable of producing the data that would seat it at the table.
Meanwhile, other countries that aren't powers —some quite small— do appear with a line of their own, because at some point they decided to manufacture their statistics rather than wait for someone else to report them. That's the difference the report leaves on view without needing to underline it: there are those who produce their data and those who receive it in translation, and Spain, for now, remains in the second group.
Definitions
AI Index Report is the annual report published by Stanford HAI on the state of artificial intelligence; its 2026 edition runs over four hundred pages spread across several chapters.
Private investment in AI is the capital that funds and companies pour into startups and projects in the sector. It doesn't include public spending or the big corporations' internal investment in infrastructure, which explains why the effort of countries with a strong state component ends up underrepresented.
Frontier model is the one that sits at the technical edge of the sector at a given moment, with the highest capabilities available.
Government-guidance funds are investment vehicles directed or backed by the State, common in China, that channel capital toward strategic sectors and that private-capital databases don't tally.
Overtraining is reusing the same data more times than the optimal ratio would advise, to get more performance out of it at the expense of compute efficiency.
Generación D (publicized as "the AI Generation") is a Spanish public program under the Ministry of Science, Innovation and Universities, endowed with a hundred and twenty million euros, that funds AI research contracts over four years through the CSIC, the Barcelona Supercomputing Center, the CNIC and the CNIO.
References
Stanford HAI — AI Index Report 2026 (full report in PDF). The main source of the figures on investment, model performance, adoption and public opinion cited throughout the article; the concrete magnitudes should be read against the official document.
Stanford HAI — Inside the AI Index: Takeaways from the 2026 Report, the official summary of the report's conclusions.
Epoch AI — Can AI Scaling Continue Through 2030?, the basis for the estimates on the cost and scale of frontier training compute toward 2030. https://epoch.ai/blog/can-ai-scaling-continue-through-2030
Epoch AI — Will We Run Out of Data? Limits of LLM Scaling Based on Human-Generated Data, the basis for the range on the exhaustion of high-quality public data. https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data
The Next Web — "Stanford AI Index 2026: China narrows US lead while spending far less," coverage of the narrowing of the US-China technical gap. https://thenextweb.com/news/stanford-ai-index-2026-china-us-performance-gap
Government of Spain, Ministry of Science, Innovation and Universities — "El Gobierno financia más de 370 contratos de investigación en centros públicos para formar la Generación IA" (2025), the official note with the 120-million-euro endowment and the beneficiary entities (CSIC, BSC, CNIC, CNIO) of the Generación D program. https://www.lamoncloa.gob.es/serviciosdeprensa/notasprensa/ciencia-innovacion-universidades/paginas/2025/070425-contratos-generacion-ia.aspx
Press Gazette — "News generative AI deals revealed: Who is suing, who is signing?", a tally of OpenAI's licensing deals with News Corp, Axel Springer, the Financial Times and other publishers. https://pressgazette.co.uk/platforms/news-publisher-ai-deals-lawsuits-openai-google/
European Union — Regulation (EU) 2024/1689 on artificial intelligence, approved in 2024, the European regulatory framework alluded to in the article. https://eur-lex.europa.eu/eli/reg/2024/1689/oj
Bloomberg Law — information on Anthropic's one-and-a-half-billion-dollar court settlement with authors over the use of copyrighted books. https://news.bloomberglaw.com/ip-law/anthropic-to-pay-1-5-billion-to-settle-author-copyright-claims
Media and the Machine — an analysis of Reddit's lawsuit against Anthropic over the use of content with no licensing agreement. https://mediaandthemachine.substack.com/p/reddit-v-anthropic-the-lawsuit-that
También te interesa
- How the Press Is Using the AI Index 2026 Figures
- Why Stanford HAI Rules and Spanish Universities Are Nowhere
- Epoch AI Says the Data Runs Out in 2028
- The Real Cost of Training GPT-5
No comments yet
No comments yet. Be the first.
Leave a comment