AI: It's a F*cking Warzone Out There
Nobody in the business calls it a war. They call it a race. A boom. A transition. On a bad day, a bubble.
We looked at 7 sectors, and at the numbers in each one. We found a fight in every single one of them.
One. Society. The war over what happens to you.
64% of American adults say AI will mean fewer jobs. Ask the people who build it, and 39% agree. Same question. Two different countries.
While they argue, the layoffs list recorded 48,716 people in cuts attributed to AI. In one quarter of this year.
And when Americans are asked whether they trust their own government to regulate it, 31% say yes. The lowest of 30 countries.
Two. Economics. The war of the cheque books.
70 companies filed $631 billion of capital spending for 2025. The year before, $432 billion. Spending up 46%. Revenue up 12%.
The gap gets covered with borrowed money. $349 billion of debt raised in fiscal 2025. Fiscal 2026 has already passed that, and the year is not over.
Nobody has blinked.
Three. Industry. The war for the chips.
Two thirds of all the AI computing power held at the end of 2025 sits in chips designed by one company. Nvidia. 67.4%.
The same company took 58.5% of the advanced packaging those chips cannot be built without.
And 4 companies, Google, Microsoft, Amazon and Meta, hold 63% of the computing power between them.
Want to build something? Get in line.
Four. Politics. The war by paperwork.
52 final export-control rules in 2024. 23 in 2025. 12 so far this year. Every one of them decides who is allowed to buy what.
Chips built for the Chinese market came to about 1 part in 21 of what was sold last year.
This is the one front where nobody pretends. Governments wrote it into law.
Five. Technology. The war of the labs.
The largest training run on record is 19 times the largest of 2023. That took 3 years.
Of 138 frontier models, OpenAI built 18. Google built 16, 21 with DeepMind. Nobody else has more than 4.
And of 107 notable models released in 2025, 30 came with open weights anyone can take. The rest come on somebody else's terms.
Six. Energy. The war for the socket.
Ohio hosts more AI data-centre power than any other state. 1,457 megawatts, across 7 sites.
An Ohio household paid 19.45 cents a kilowatt-hour this July. In July 2020 it paid 12.26.
We cannot tell you one caused the other. Prices rose in states with few data centres too.
Here is what we can tell you. They are on the same grid. And across the country, a household pays 18.31 cents. Industry pays 9.77.
Seven. Science. The war for the work itself.
Mathematics. The field where the whole job is proving it yourself.
In September 2024, 0.4% of mathematics preprints said they had used AI.
A year later, 1.7%.
This September: 38.1%. That is 3,012 papers out of 7,899. In one month.
And the best model with closed weights outscores the best open one by 9 points. The strongest tool in the room belongs to whoever owns it.
One more fight. Among the people building it.
“We are witnessing an extraordinary transfer of power away from governments to a handful of private corporations and individuals.” António Guterres, UN Secretary-General, to the General Assembly, 22 September 2026.
One more fight. It is not in the numbers. It is among the people building it, and they said so on the record, in their own words.
“If managed poorly, I even believe AI could be a risk to humanity as a whole.” Dario Amodei, Anthropic, to the UN Security Council, 23 September 2026.
“We could lose control of the future to AI.” Sam Altman, OpenAI, to the UN Security Council, 23 September 2026.
“Dario is right.” Elon Musk, on X, answering Amodei’s essay, September 2026.
“Dario’s essay points towards the right path forward.” Demis Hassabis, Google DeepMind, to Reuters, September 2026.
“2030 is not going to be the end of the world. There is 0% chance that’s going to be the end of the world.” Jensen Huang, Nvidia, to CBS News, September 2026.
Four say slow down. One says it will not end the world. These are their words as reported, not numbers we measured.
The close.
7 sectors. 7 fights.
We are not telling you who is right. We are not telling you it ends well, or badly.
We are telling you it stopped being a debate.
It's a f*cking warzone out there.
Token Runner. The Baseline. October 2026.
Sources
- Pew Research, 2025, as published in Stanford HAI, AI Index Report 2026, figure 9.2.4: 64% of US adults and 39% of AI experts say AI will lead to fewer jobs. Values as published.
- Layoffs.fyi, the AI layoffs list: 48,716 employees in layoffs attributed to AI in the second quarter of 2026. Attributions the source records, not a count Token Runner made.
- Ipsos, 2025, as published in Stanford HAI, AI Index Report 2026, figure 9.3.1: 31% in the United States trust their government to regulate AI, the lowest of the 30 countries shown. Values as published.
- SEC filings, annual reports of 70 companies with both figures in each year (of the 71 Token Runner tracks): capital spending $631.4 billion for 2025 and $431.8 billion for 2024; revenue $6,097.2 billion and $5,465.4 billion. The two percentages are arithmetic on the filed sums.
- SEC filings, cash-flow lines, 60 companies: debt raised $349.4 billion in fiscal 2025; $381.5 billion so far in fiscal 2026.
- Epoch AI, AI Chip Owners, median estimates at 31 December 2025: 67.4% of compute held is in chips Nvidia designed; Google 24.1%, Microsoft 16.3%, Amazon 11.7%, Meta 11.0% of compute held (63.1% together, a sum of those four).
- Epoch AI, AI Chip Components, median estimates, 2025: Nvidia 58.5% of advanced-packaging (CoWoS) wafers.
- Federal Register, Bureau of Industry and Security: 52 final export-control rules in 2024, 23 in 2025, 12 in 2026 to 24 September.
- Epoch AI, AI chip sales, median estimates, 2025: 646,000 H100-equivalents in chips marked as made primarily for the Chinese market against 12.97 million not so marked. The table marks a chip's intended market, not where it ended up.
- Epoch AI, Frontier AI Models: the largest training run on the list, 9.4e26 FLOP, against 5.0e25 for the largest published in 2023; 18 of 138 models from OpenAI, 16 from Google and 5 more from DeepMind.
- Epoch AI, Notable AI Models, 2025: 30 of 107 notable models released with unrestricted open weights; the other 77 through an API, hosted, restricted or other forms.
- Epoch AI, the data centres file: Ohio 1,457 megawatts of AI data-centre power across 7 sites, the most of any US state in the file.
- US Energy Information Administration, average residential price of electricity: Ohio 19.45 cents a kilowatt-hour in July 2026 and 12.26 in July 2020; US average in July 2026, households 18.31 cents and industry 9.77.
- Epoch AI, AI acknowledgment in mathematics preprints: 0.4% in September 2024, 1.7% in September 2025, 38.1% in September 2026 (3,012 of 7,899).
- Epoch AI, Epoch Capabilities Index: best closed-weights model 166.82, best open-weights model 157.42.
- Associated Press (Seth Borenstein), "AI firm leaders tell UN Security Council that it could be a risk to all humanity", 23 September 2026, as carried by PBS NewsHour: the words of Dario Amodei and Sam Altman at the UN Security Council.
- Reuters, "Tech leaders, governments split over AI doom fears", 21 September 2026, as carried by BNN Bloomberg: the words of Elon Musk, Demis Hassabis and Jensen Huang (Huang to CBS News).
- António Guterres, address to the opening of the General Debate of the 81st session of the UN General Assembly, 22 September 2026 (the words as reported by Tech Policy Press, 27 September 2026; to be confirmed on the UN text before publishing).