TALAMANA · THE AI LITERACY MAP FOR ARCHITECTURE AND DESIGN · Foundations · AGE 16—18 · FACTUAL · EVOLVING
The compute story
Capability tracks data, chips and money, and whoever owns the compute shapes it.
The idea
Capability in this field has tracked three things together: data, chips and money. Researchers who measured it found that the computing power used to train the largest models doubled roughly every six months after 2010 — far faster than the older pace of about twenty months — with a separate era of very large training runs from around 2015–16. Compute is one large part of why capability jumped when it did — data, methods, engineering and money are the others — and it is why capability sits where the chips are. Compute costs money and burns electricity, so whoever owns compute shapes capability.
Why it matters
A tool's power does not come from nowhere. It was bought, with chips, data centres and capital. Knowing that lets you read AI news as industrial news, which is mostly what it is.
See it in the studio
The render you get in twelve seconds ran on hardware you will never see, in a building you could one day be asked to design. The data centre is now a building type, and a client.
Watch for this
Treating doubling curves as laws of nature. They are measurements of what companies chose to spend. Spending can slow, chips can be rationed, and efficiency gains can bend the curve. That is why this card is marked EVOLVING.
Try it
Look up the reported training compute or cost of one well-known model. Then look up the electricity demand of data centres in the IEA report. Write down which numbers you found, who reported them, and which you could not find. The gaps are part of the story.
Prove it
Explain, in plain terms, why capability rose when it did, and name the three inputs it has tracked.
How it works
Training compute is counted in floating-point operations; the doubling times above come from Sevilla and colleagues' 2022 survey of landmark models, which found a pre-2010 era tracking Moore's law at roughly twenty months, a deep-learning era at roughly six, and a large-scale era from late 2015 using ten to a hundred times more compute than the trend. On the energy side, the International Energy Agency's 2025 report Energy and AI gives the measured and projected electricity demand of data centres. Quote the report, not a headline. EVOLVING because the curve, the chips and the policy around them are all moving this year.
What this idea builds on
What this idea opens up
Sources
Open this idea on the map · The complete map · Logika · RBDS AI Lab, India · revised every edition.