In just four years, artificial intelligence has emerged onto the world stage and transformed it. In the autumn of 2022, ChatGPT went viral: it made amusing mistakes and yet still gave the impression that something momentous was happening. Machines could now converse and understand human instructions. Their limits remain to be seen, but their impact on the economy and on everyday life is already clear. The following charts trace that rapid rise.
Models keep getting better
Chatbots from OpenAI, Anthropic and Google improve with every new release. To measure their progress, the industry has developed a series of competency tests known as benchmarks, and the results point to steady gains over time.
In 2023, the leading model of the day, GPT-4, scored 36% on GPQA, a benchmark designed to test doctoral-level scientific knowledge, well below the roughly 65% that PhD holders achieve in their own fields. Today, the best models are approaching 95%, pushing the benchmark to its limits as a measure of progress. The same pattern has been seen in coding and mathematics benchmarks. That is why new evaluations are constantly being developed, including variants of ARC-AGI, which are designed to pose problems that are easy for humans but difficult for machines. And, time and again, AI systems eventually learn to solve them.
Adoption has been massive: hundreds of millions of people use AI
Generative AI is likely the fastest-adopted mass technology in history. Personal computers took more than a decade to reach half of Americans. The internet took six years; the smartphone, five. Chatbots like ChatGPT or Gemini did it in just over two. In Spain, 55% of internet users say they already use generative AI tools, according to the Spanish National Observatory for Telecommunications and the Information Society (ONTSI).
This speed also reflects an increasingly digitalized world, and therefore a faster-moving one. Getting online once required installing a modem, and getting a cellphone meant going to a store. AI models, by contrast, are just software: a single click is enough to bring them into your home. Many of the technologies that follow are likely to spread just as quickly.
Equally, if not more important, is adoption by businesses. According to data from Ramp, which processes payments for tens of thousands of U.S. companies, more than half of businesses already pay for AI services from at least one provider. Forty-three percent pay for Anthropic’s Claude, which overtook OpenAI’s ChatGPT (40%) in professional settings a few months ago. In Spain, the trend is similar: according to ONTSI, 20% of medium-sized companies and 58% of large companies were using AI in 2025.
AI companies have seen their expectations and their market valuations soar
The world’s five most valuable companies are all technology firms: Nvidia, Apple, Alphabet, Microsoft and Amazon, and each has a major stake in artificial intelligence. Since the launch of ChatGPT in late 2022, Microsoft’s market value has doubled, Amazon’s has tripled, and Alphabet’s has nearly quadrupled.
Above all stands Nvidia, which went from making graphics cards for video games to powering the computation behind these models with hardware and software that are close to monopolistic. Since late 2022, its market value has increased fifteenfold. Its rise is hard to fathom: if you had invested $1,000 in the company in January 2010, that stake would be worth more than $600,000 today. It is little surprise that thousands of Nvidia employees have become millionaires.
The rest of the semiconductor supply chain has also enjoyed spectacular gains. Broadcom, which designs custom chips, is worth six times more than it was at the end of 2021. SK Hynix and Micron, which manufacture the memory used in AI chips, are worth 10 times more. And Dutch company ASML, whose machines are essential to the production of advanced semiconductors, has seen its value double.
Tech giants are investing like never before
This stock market boom has naturally raised concerns that a bubble may be forming. Even when a technology proves transformative, the companies betting on it can still fail if they arrive too early or back the wrong approach. That is one of the key lessons of the dot-com bubble and other speculative manias before it. The debate remains unresolved, but tech giants have shown no sign of slowing their spending: combined capital expenditures by Amazon, Microsoft, Alphabet and Meta have quadrupled since 2023.
The boom is showing up in job postings
Many of AI’s effects on the labor market are still unknown. Will jobs be automated? How much will occupations change, and in what ways? Will the impact fall most heavily on young workers or on knowledge workers? These questions are still the subject of intense debate. But one consequence is already clear, and hardly surprising: job postings that mention AI have multiplied across countries such as the United States, Germany and the United Kingdom.
The cloud is more physical than it seems
Training and running AI models requires an extraordinary amount of computing power: vast data centers packed with millions of processors, first to train the models and then to handle the requests of individual users. Forecasts suggest that computing demand will continue to grow rapidly, even as models become more efficient.
And that growth carries an increasingly tangible physical cost. Land is needed to build data centers, silicon to make their chips, factories to produce their components, and electricity to keep the machines running around the clock. According to the International Energy Agency, electricity consumption by data centers is expected to more than double by 2030, rising from 415 to 945 terawatt-hours and approaching 3% of global electricity demand. That would be more electricity than Japan consumes in an entire year.
The United States still leads, but China remains only months behind
The most advanced AI models are still being developed by U.S. labs such as OpenAI and Anthropic. But the gap with China has remained remarkably narrow, with Chinese companies typically matching leading U.S. models within a matter of months. According to the capabilities index compiled by Epoch AI, the top U.S. model scores 162 points (Claude Fable 5), while the leading Chinese model, Moonshot, has already reached 157.
In other words, China’s best model is now operating at roughly the same level as the leading U.S. model was earlier this year. Europe lags somewhat further behind: the top model from France’s Mistral scores 143 points. The race also highlights another divide. Many Chinese models are open weight, meaning they can be downloaded and run by anyone with a computer.
Using AI is getting cheaper
As models have improved, costs have fallen very quickly. Using the model behind the original version of ChatGPT cost around $20 per million tokens. Two years later, matching that level of performance with Llama, Meta’s open-weight model, cost just 10 cents, a 200-fold reduction.
According to estimates by Epoch AI, the cost of achieving a given level of capability falls by a factor of between ten and a thousand with each passing year. That is another reason for AI’s rapid rise: what is expensive and exclusive today is likely to become cheap and ubiquitous tomorrow.
What do we think about AI? Somewhere between enthusiasm and anxiety
The arrival of AI is likely to have social consequences across many areas of life. Its effects on education, the attention economy, artistic creation and cybersecurity are all being debated. This spring, Ipsos surveyed thousands of people in around 30 countries, asking whether AI-powered products “excite” them and whether they “make them nervous.” Spaniards feel both: 47% say they are excited by them, while 52% say they make them nervous.
The survey also points to a cultural divide. Although excitement and anxiety coexist in every country, often within the same person, anxiety tends to dominate in the West while excitement is more prevalent in Asia. Eighty-three percent of Chinese respondents and 79% of Indians say they are excited about AI, compared with 26% of Canadians and 33% of Americans. It will be interesting to revisit that question in three, five or 25 years.
Sign up for our weekly newsletter to get more English-language news coverage from EL PAÍS USA Edition
