Patents, models, investment

Every row on this page has a different date, and the distance between those dates is the first thing to read. Granted patents stop at 2024. Scientific publications at 2024. Models at 2025. Quality scores at 2 October 2026. Anyone who puts these numbers in the same sentence is describing a world that has never existed at any single moment.

The second thing to read is the perimeter. The same training run costs 218 million dollars or 850 depending on the method used to count it, and both numbers are correct. Next to every figure below it says what it measures.

Where patents are filed

AI patents granted worldwide in 2024, the last year for which the figure exists. These are the ones that made it through examination, so far fewer than the applications filed. The count is by family, meaning one invention counts as one even if it was filed in ten countries, and the attribution is to the office that grants, not to the nationality of the patent owner.

AI Index Report 2026, PATSTAT data from the European Patent Office. The absolute values are read off the figures and rounded: the report writes 15,290 US patents in the text and marks about 15,900 in the chart, and 12.1% of 131,121 makes 15,813. I use the percentage, which is consistent in both places.

The American share has fallen from 42.8% in 2015 to 12.1% in 2024. The world total grows by 8.2% in a year and has multiplied by thirty-four since 2010, when AI patents granted numbered 3,866.

Volume says who files, and on the value of those filings there is a second indicator. The United States collects over half of all citations received by AI patents with 12.1% of the volume, and only 19% of American patents are left with no citation at all, against 32-44% for the other areas.

AI patents granted per 100,000 inhabitants, 2024: South Korea 14.31, Luxembourg 12.25, China 6.95, United States 4.68, Japan 4.30, Singapore 1.31, Germany 1.30, Sweden 0.70, Finland 0.67, France 0.62, United Kingdom 0.60, Australia 0.45. Italy does not appear in the ranking, which stops at fifteen countries.

On published applications, which is the opposite perimeter and arrives earlier, the 2025 picture comes from WIPO: over 435,000 AI patent families published in one year, of which 37,808 in generative AI, against 18,862 in 2024. China passes 43,000 generative families over the two years 2024-2025; the United States goes from 1,185 to over 4,380 in two years; Japan from 398 to 3,835 and is third; Germany is sixth and first in Europe. The largest holder worldwide by cumulative portfolio is SoftBank with 2,985 families, then Tencent, Ping An, Baidu. OpenAI, at the end of 2025, has 35 filings in the whole world: those who build the most used models patent very little.

Where models are born

"Notable" models by country of the organisation that built them.

Columns 2024 and 2025: AI Index, which photographs the Epoch AI database on 17 March 2025 and 22 April 2026. Column 2026: Epoch AI database re-read on 3 October 2026, year in progress, counting a model once for each distinct country present among the affiliations. Three models with an Italian affiliation appear in the database, all academic: Ceramic-MLP from Sapienza University of Rome (1994), the graph neural networks of the University of Siena (2008) and Selective Search from Trento with Amsterdam (2013). After 2013, none.

Four things about the way this count is made, because without them you cannot tell what it says.

The list is compiled by hand by Epoch AI, according to three declared criteria: advancing the state of the art, historical importance, high citation rate. The report says so explicitly: it is not a census of every model released.

Attribution to a country is triggered if even a single author is affiliated with an institution in that country. A model with American and British authors counts for both, and the sum of the columns exceeds the total number of models.

The database is alive, backwards too. For 2025 the United States is worth 59 in the April photograph and 64 on re-reading the database on 3 October 2026: models added after the fact. Comparing two editions of the AI Index without saying so produces differences that nobody has ever observed in reality.

The 2026 column shows it better still. Between 14 August and 19 September China goes from 19 to 27 and the United States from 34 to 42, but the models actually released in those five weeks are twelve: nine American, two Chinese and one with no country declared. The rest is back-cataloguing. Re-reading the same database today and stopping at 14 August gives 33 American and 25 Chinese, not 34 and 19: the past changed afterwards. And it has kept changing: re-read on 3 October, the database counts 29 Chinese models published by 19 September, not 27. In the following two weeks seven models came out, all American: Grok 4.7, Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, Claude Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon. A number read twice a month apart does not measure only what happened in the month.

The European models of 2025 are two, one French and one British: Mistral OCR and SPAR3D by Stability AI. European 2024 too, in the updated database, includes a German model by DeepL and a British one by Stability. Anyone repeating that Europe produces only Mistral models is quoting a photograph from March 2025.

In 2025 new releases fall in every area. Building them is almost entirely industry: 93 models against 2 from academia, 91.2% of the total. By organisation: OpenAI 20, Google 14, Alibaba 11.

The model database, updated daily →

The gap between the United States and China

The number that circulates as "the benchmark gap" comes from a single ranking, the Arena, where people vote on which answer they prefer out of two they receive without knowing whose they are. How that mechanism works you will find on the market page. What counts here are the values and how little they are stable.

The same month, February 2025, is worth 1.7% in the edition of the report published that year and 0.4% in the one from the year after. Arena scores are recalculated as new votes arrive, so the historical series moves under your feet: a gap in the past can change without anything having changed in the past.

The percentage, then, divides a difference in Elo points by an Elo score, and the zero of that scale is arbitrary. Quote the points: 37 on a scale where the leader is worth 1,525. Across the rows above that carry points, the distance runs from 5 to 39.

The last row comes from a direct reading of the ranking, taken on 3 October 2026 on the update of 2 October; the others come from the AI Index. That is why the two rows do not subtract: between March and October the scores were recalculated, and the model that was worth 1,503 in March, Claude Opus 4.6, still reads 1,505 in the "high" variant and 1,497 in the base one on 2 October. On 30 September Google presented Gemini 4 Argon, which enters straight at the top with 1,525 but with fewer than five thousand votes and the label "preliminary". The Chinese model closest to the top is still Kimi K3 Max by Moonshot, 1,488, in sixteenth place.

In the same ranking, in March 2026 the best closed-weights model (Claude Opus 4.6) leads the best open-weights model (GLM-5, 1,454) by 49 points, 3.4%. In August 2024 the distance was 7 points: it had almost closed and it has reopened. On 2 October 2026, measured entirely on the same ranking on the same day, it is 45 points, 3.0%: Gemini 4 Argon high at 1,525, preliminary, against MiMo-V2.6-Pro by Xiaomi at 1,480, MIT licence, weights published on 21 September. Without the preliminary model the distance is 25 points, 1.7%, from Claude Opus 4.6 high at 1,505. GLM-5.3 Max by Z.ai, the leading open-weights model on 19 September, drops to 1,478; GLM-5 still reads 1,458. At 1,488 sits Kimi K3 Max, which publishes its weights under a bespoke, non-standard licence: here I count the first one with a verifiable licence.

What it costs to train a model

No laboratory publishes what it spends on training. Every figure you read is an estimate, and it changes fourfold depending on what you put inside it. Grok 3, trained on 80,000 H100 chips for ninety days, is the best documented example.

Epoch AI, 2023 dollars. For the full development cost Epoch publishes the composition but not the series: hardware 47-67%, staff 29-49%, energy 2-6%. The same table carries two higher columns, 2.45 and 4.45 billion, whose method is not documented in the field guide: I do not use them.

On the first method, the one Epoch documents and uses for its historical series, the highest estimates ever calculated are these.

The list contains no GPT-5, no GPT-6 Astra, no Gemini 3, no Claude Opus 4.x and no Grok 5. For no frontier model of 2026 does a cost estimate exist, because the technical data needed to calculate one have stopped coming out. The last estimate for a frontier model concerns a model from July 2025; since then Epoch has estimated only models it does not classify as frontier, such as A.X K1 and A.X K2 by SK Telecom: 9.6 million dollars for the first, December 2025, and 12.4 for the second, July 2026. Re-read on 3 October 2026, the database has the cost field empty for GPT-6 Astra and for the frontier models that came after it: GPT-6 Sol, Claude Opus 5.5, Gemini 4 Argon, Grok 4.7.

The growth rate measured by Epoch is 2.4 times a year on the amortised method and 2.6 on the rental one. The forecast published in 2024 says the largest training runs will pass a billion dollars "by 2027". As of 3 October 2026 it has not happened yet on any documented perimeter: the highest figure ever calculated is 850 million, and it concerns a model from February 2025.

One thing has changed in the meantime, and it concerns the sense of the question itself. Epoch estimates that GPT-5 consumed less compute than GPT-4.5 to be trained. With reasoning techniques, a 2-million-dollar post-training run yields on a 20-million pre-training run what previously took 200. The spending has shifted: what a model costs no longer sits entirely in the final run.

On the DeepSeek case the number to quote is 5.576 million dollars, not "about six": 2.788 million H800 GPU hours, valued at 2 dollars an hour. That price is a rental list price applied to cards that DeepSeek owned, and the paper excludes in an explicit note the preliminary research, the architecture experiments and the data. On the following model, R1, the company declared in a peer-reviewed article 294,000 dollars for the reinforcement learning phase alone, a figure that excludes the cost of the base model it rests on.

What the laboratories spend in a year

Collected by Epoch AI and taken up by the AI Index. This is rented compute capacity, not amortisation of owned data centres. Mind the nature of these figures: neither company publishes accounts. The entries come from press reports on documents intended for investors, and Epoch itself marks them as "likely". I report them because they are the best existing reconstruction, with that limit declared.

The row that counts for anyone designing a product is the third. In 2024 OpenAI spent 1.8 billion to run the models against 4.0 to build them. In 2025 it spends 8.0 against 8.3. The item growing fastest is the cost of each single answer, and that one multiplies by the number of users.

The hundred-billion-and-more figures you read in the headlines are another thing. Stargate, announced by OpenAI with SoftBank and Oracle in January 2025, is a 500-billion commitment over several years; the Oracle deal of July 2025 adds over 300 more across five. These are purchase promises, not spending incurred: between the 500 billion announced and the 16.3 billion spent in a year there is a factor of thirty. On capital actually invested the certified figure exists only for listed companies: Alphabet spent 91.4 billion dollars in 2025 and states for 2026 a forecast between 195 and 205 billion.

How much those promises weigh is now also readable in published accounts, from the side of the party collecting. Oracle, first quarter fiscal 2027 results, 10 September 2026: contractual obligations still to be performed rise to 664 billion dollars, 209 more in a year, with over 30 billion of AI cloud contracts signed in that quarter alone. These are contracts signed and not yet performed, the same material as the 500 billion of Stargate. The Oracle release names neither OpenAI nor Stargate: anyone attributing those 664 billion to a single customer is quoting the press, not the company.

On the physical side the distance between announcement and electricity is measured in years. On 17 August 2026 OpenAI announced roughly 8 gigawatts in Pike County, Ohio, with the first 800 megawatts expected in 2028. In that release too the word Stargate does not appear.

Who puts up the money

Private investment in AI companies in 2025, billions of dollars.

AI Index 2026, Quid data. Italy does not make the ranking of the top fifteen countries, which closes with Sweden at 0.97 billion: the Italian value sits below that threshold and is not published. In 2024 it was 0.9 billion.

The United States is worth 23.1 times China and 48.5 times the United Kingdom. On generative AI alone, 163.64 billion against the 4.7 of China and Europe put together. Since 2013 the American cumulative total is 757.3 billion against China's 131.8.

Before using these ratios in an argument, four things the series does not contain: mergers and acquisitions, which are a separate item and in 2025 grew by 132.6%; public offerings; the capital the large listed companies put into their own data centres, and that is Alphabet's 91.4 billion, more than three times all Chinese and British venture capital combined; public funds. On this last point the report is explicit: Chinese state guidance funds deployed about 912 billion between 2000 and 2023, of which about 184 towards AI companies, and they sit outside the count. The ratio of twenty-three to one holds inside that perimeter and not outside it.

For 2026 the figure on this perimeter does not exist and will not exist before April 2027. On another count, which keeps every sector inside, in the first half of 2026 global venture capital passed 510 billion against the 440 of all of 2025, and OpenAI and Anthropic alone absorbed 217 of them, 43% of the world total.

In Europe, inside the window, a single round weighs as much as a country. On 8 September 2026 Mistral raised 3 billion euros at a post-money valuation of over 21 billion, with Samsung Electronics leading the round, and the company calls it the largest equity fundraising ever completed by a European technology company. At the ECB reference rate of 18 September, 1.1460 dollars to the euro, that is about 3.4 billion dollars: almost as much as everything French AI companies raised in the whole of 2025, which in the table above is worth 4.36 billion.

Has the university left the game?

It depends on what you look at, and the three available indicators give three different answers.

On models academia is no longer there: 2 out of 95 in 2025. Building a frontier model takes a machine park no department owns.

On publications academia dominates: 68.1% of AI articles in 2024, against 11.5% from industry and 12.4% from public bodies. In Europe the academic share rises to 55.3%, in the United States industry reaches 24.6%, in China public bodies 25.1%. AI articles are by now 40.9% of all indexed computer science, 257,890 in 2024.

On doctorates the direction has reversed. The share of new AI PhDs going into industry has fallen from the 77% peak of 2022 to 65% in 2024, and the share staying in academia has almost doubled. New doctorates have grown by 22% in two years, and all the growth has gone to the university. The figure covers the United States and Canada.

What has changed

3 October 2026 · revision

A model the ranking still marks as preliminary has entered at the top of the Arena. On 30 September Google presented Gemini 4 Argon, and in the update of 2 October its high variant leads with 1,525, on 4,932 votes and with the label "preliminary". Counting it, the leading Chinese model, Kimi K3 Max at 1,488, is 37 points behind; counting only consolidated scores, from Claude Opus 4.6 high at 1,505, it is 17.

The leading open-weights model with a verifiable licence has changed. MiMo-V2.6-Pro by Xiaomi, weights published on 21 September under the MIT licence, reads 1,480 on 2 October; GLM-5.3 Max, the leader on 19 September, drops to 1,478. Two points apart, with a declared margin of plus or minus 9.

Seven new models in two weeks, all American. The 2026 column moves to 49 American models and 29 Chinese ones. The seven American models came out between 21 and 30 September; the two extra Chinese ones are models published before 19 September and catalogued by Epoch afterwards.

A correction on Italy. I had written that no Italian model appears in the Epoch database. Three do, all academic: Ceramic-MLP from Sapienza University of Rome (1994), the graph neural networks of the University of Siena (2008) and Selective Search from Trento with Amsterdam (2013). After 2013, none.

What was not updated, and why. Patents: WIPO has published nothing since the report of 14 July. Scientific publications and doctorates: no new count. Laboratory compute spending: the OpenAI and Anthropic figures for 2024 and 2025 are unchanged. Oracle's obligations, the Mistral round and the OpenAI Ohio announcement: re-read, unchanged. Training cost estimates: none for the frontier models released after July 2025, and the maximum is still 850 million.

19 September 2026 · revision

The model count grew more than production did. The 2026 column goes from 34 to 42 American models and from 19 to 27 Chinese ones, but the models actually published between 14 August and 19 September are twelve. Epoch added the rest backwards, onto months already closed.

At the top of the Arena the leading Chinese model is twenty-one points behind. On 19 September Claude Fable 5 high leads with 1,506 and Kimi K3 Max follows with 1,485; the leading open-weights model is GLM-5.3 Max, 1,483, twenty-five points above GLM-5, which held that place in March. The March row from the AI Index does not subtract from this one, because the scores are recalculated: Claude Opus 4.6, worth 1,503 then, reads 1,505 and 1,497 in its two variants on 19 September.

Purchase promises now have published accounts that count them. Oracle takes contractual obligations still to be performed to 664 billion dollars, 209 more in a year. And in Europe Mistral raises 3 billion euros on 8 September, about 3.4 billion dollars: almost as much as every French AI company in the whole of 2025.

What was not updated, and why. Patents: WIPO has published nothing new since the report of 14 July, and the 2024 figures from the AI Index remain the last data that exist. Scientific publications and doctorates: no new count inside the window. Training cost estimates: the Epoch database has added none, and no training run has passed a billion. Laboratory compute spending: the OpenAI and Anthropic figures are the same as in August. The rows on this page I did not re-measure keep their old date.

14 August 2026 · first publication

The laboratories have stopped saying how their models are made. Code, number of parameters, data size and training duration are no longer published for the most expensive systems, those of OpenAI, Anthropic and Google. The average score of the foundation model transparency index has fallen from 58 to 40. Anyone working with this notices when looking for a figure: every number on 2026 is less verifiable than one on 2024.

Running the models has come to cost as much as building them. At OpenAI the ratio between the two items was 1.8 against 4.0 billion in 2024; now it is 8.0 against 8.3. Anyone designing a service has to factor in the marginal cost of each answer, which before could be ignored.

On models China has almost closed the gap; on basic research it had closed it long ago. In March 2026 the leading American model is 39 points ahead of the leading Chinese one on 1,503. At the same moment China has 74.2% of AI patents granted, 17.8% of publications, 20.6% of citations and 41 of the hundred most cited articles of 2024. Two leads remain to the United States: the number of notable models, 59 against 35, and the citations to their patents, over half the world total despite having filed little more than one in eight.

Six figures that circulate wrong

You find them everywhere, including some pages of this book in previous editions.

"Gemini Ultra cost 191 million to train." That is the cloud rental price applied to the hours consumed. On the amortised method the same model costs 31. The number is worth six times what it seems.

"GPT-5 cost between 500 million and 2.5 billion." Epoch AI, which is the source usually cited for this range, argues the opposite, and for GPT-4.5 estimates 366 million. On GPT-5 no estimate exists.

"Each generation of models costs twenty-eight times the previous one." The growth rates published by Epoch are 2.4 and 2.6 times a year, depending on the method. I have not found the origin of the twenty-eight, and I have not found it either for the figures of 3 and 170 million attributed to Llama 1 and Llama 3.1: the published estimates for those two models are 0.58 and 53 million.

"The United States-China gap in benchmarks." That number comes out of a single ranking of human preferences, recalculated continuously, and no automatic test enters it.

"DeepSeek actually spent 1.3 billion to train V3." That figure derives from an estimate of the machine park of the company and of the fund that controls it, accumulated over several years and across all projects. The two quantities measure different things and do not contradict each other.

"Google spent more than 150 billion in capital expenditure in 2025." Alphabet's accounts say 91.4. The figure above 150 concerns 2026 and is a company forecast, not recorded spending. The AI Index too reports it in the wrong place, on a third-party chart.

The AI Index Report 2026, in full →