14/09/2026
Why we never farmed lions, and the energy arithmetic of the AI race
Every mammal domesticated for meat lives on plants. Cattle, sheep, pigs and goats. No predator ever made the list, not the lion, not the leopard and not the wolf, and the reason is not the danger of raising one. The reason is a simple energy calculation, and the same calculation is running today in the data centres of artificial intelligence.
A cow eating grass converts about a tenth of the plant's energy into meat. To raise a lion for food you would have to raise a herd of cattle for it, and whole fields of feed for that herd, and each step loses most of the energy again. The carnivore is simply too expensive, before anyone even gets to what it might do to the farmer 🦁
Jared Diamond, in Guns, Germs, and Steel, puts that conversion at around ten percent, and a study published in Science Advances in July 2026 found even less. In modern farming it comes to about twenty five kilograms of feed for one kilogram of edible beef, which is why we raise cows.
Of 148 species of big mammals that were candidates for domestication only fourteen passed every condition, and carnivores never made the list at all. They are disqualified by the first of Diamond's conditions, diet, before the question of their temperament is ever reached.
When the energy cost of a way of doing things rises, the outcome is decided by conversion efficiency, by how much use is extracted from each unit of energy, and size stops being the advantage. That is what is happening now in artificial intelligence.
To improve a model by a given amount, the compute required grows faster than the rate of improvement, so each further step costs more than the one before it. This is not peculiar to machines: a human brain is about two percent of body weight and consumes about a fifth of the energy the body burns at rest. Intelligence is expensive in whatever substrate it runs.
A frontier model is the newest and largest model a company runs at a given time, the highest version of GPT, of Claude and of Gemini, rather than the smaller versions alongside it. It handles the hardest tasks any model can take on today, and it is also the most expensive to operate, because every question sent to it uses more electricity than the same question sent to a small model. When each further degree of capability costs more than the last, running it on a routine question pays the full price for capability that is not needed, rather like sending a lorry to fetch milk from the corner shop. The result shows up in the electricity market, where technology companies are already buying the output of entire nuclear reactors: Microsoft 835 MW from Unit 1 at Three Mile Island, Amazon 1,920 MW from the Susquehanna plant, and Meta 2,609 MW over twenty years.
A signed contract is not electricity flowing. The reactors ordered come online between 2030 and 2035, lead times for large power transformers run two to four years, grid connection is set by the queue, and water for cooling by a local planning proceeding. No amount of money speeds those up.
When it became clear that lions could not be farmed, people did not stop eating meat; they moved to animals that convert cheap feed efficiently. The same response is visible now. Parameters are the internal numbers in which what a model learned is held, and every use of it draws electricity in proportion to how many of them are actually activated. DeepSeek-V3 holds 671 billion parameters and activates only 37 billion at a time, because the model is divided into parts and each piece of text activates only a few of them, so that a very large model works each time with a small part of itself. The result is the knowledge of a large model at the running cost of a far smaller one, and in June 2026 Apple presented a model that runs on the device itself and sends nothing to the cloud. That saves electricity, and it also changes the status of the material: text that never leaves the device is not transferred to a third party, is not stored on a provider's server and is not subject to that provider's terms of use.
What is interesting is that the constraint itself produced the solution. DeepSeek built that design under restrictions on the export of advanced chips to it. When chips are plentiful, an idea that makes a model more efficient reduces a bill you are paying anyway. When they are not, that same idea is what lets you run a model that would not otherwise run at all, so a team that meets the constraint puts more time and more people into that idea, and it arrives there first. There are two ways to grow, to add electricity or to get more capability out of each unit of it. The first depends on transformers, grid queues and planning permits, and its timetables are measured in years. The second depends on an idea, and whoever improves conversion efficiency grows without adding a single megawatt.
Full article, with the tables and the sources, in the first comment.
