AI is running short of power and text. Its next meal: your company's records

The smarter AI gets, the hungrier it gets.

Summarizing text, making images, analyzing data, writing code: the list of jobs you can hand to generative AI grows almost monthly. Some AI can now look things up, pick tools and pull the results together on its own.

On screen, AI keeps getting faster and smarter. Pull the camera back beyond the screen, though, and a different picture emerges.

Huge data centers wired into power plants and the grid. Chips redesigned to curb power use and heat. And a dwindling stock of human writing for AI to learn from.

AI is devouring electricity and text at a staggering pace.

So what does it eat next? Our answer: meaning, or context. Follow AI's three hungers (electricity, text and meaning) and you can see where the next AI race will be fought.

At the heart of generative AI such as ChatGPT sits a model. A text model splits huge amounts of writing into small units called tokens, then trains over and over on predicting the next word.

The key is balancing the size of the brain against how much it reads. In the Chinchilla study Google DeepMind published in 2022, researchers compared two AIs on the same computing budget. "Gopher" had 280 billion parameters (the adjustable values inside a model, a rough measure of size). "Chinchilla" was a quarter of that size but read about four times as much: 1.3 trillion tokens. The smaller Chinchilla won on almost every task measured.

On the same computing budget, the smaller AI that read more came out ahead. Chart by DATA WORLD based on Google DeepMind.

A smart AI needs enough reading to match the size of its brain. And today, computing doesn't stop at training. It runs every time AI is asked to think longer before answering or to use outside tools.

In other words, AI eats when it learns, and eats again when it works. What keeps those meals coming is electricity.

According to the International Energy Agency (IEA), the world's data centers used 485TWh in 2025, up 17% in a single year. By 2030, they're expected to use about 950TWh, roughly 3% of global demand (IEA).

How big is 950TWh? Japan generated about 991TWh in fiscal 2024 (April 2024 to March 2025) (Agency for Natural Resources and Energy). By 2030, data centers alone will use nearly as much electricity as Japan generates in a year.

The IEA's regional outlook from April 2025. Solid lines are actuals through 2024, dashed lines are projections; the vertical axis is annual consumption (TWh). Source: IEA, “Energy and AI”, CC BY 4.0.

The same is starting to happen in Japan. The Organization for Cross-regional Coordination of Transmission Operators (OCCTO), which coordinates Japan's regional grids, expects extra demand from new and expanded data centers to rise from 4.8 billion kWh in fiscal 2026 to 49.4 billion kWh in fiscal 2035, about tenfold (OCCTO).

The world isn't sitting still. Its response is running on three clocks, each at a different speed.

With enough money and time, you can add electricity, up to a point. Text is another matter.

Epoch AI, a research group that tracks AI progress through data, estimates the stock of human-written public text, adjusted for quality and duplication, at about 300 trillion tokens (central estimate). If training data keeps growing at today's pace, leading models will reach that between 2026 and 2032 (Epoch AI).

The effective stock of human-written public text is about 300 trillion tokens at the central estimate (range: 100 trillion to 1,000 trillion). Chart by DATA WORLD based on Epoch AI, “Will we run out of data?”.

Picture the web as a giant library that AI has been reading end to end: news, papers, blogs, forums, code. People add books daily, but AI reads far faster than we write.

Sooner or later, AI will have read every shelf in the library. Rereading the same books teaches it nothing new. That's the 300-trillion-token wall.

But is all the world's writing and data really in that library?

A 2025 survey by NEDO (the New Energy and Industrial Technology Development Organization, a Japanese government research funder) put the data created worldwide at 175ZB (NEDO). One zettabyte is 1 trillion GB, so that's 175 trillion GB. The survey assumes 17.5ZB of that is unique after removing duplicates, and says about 16ZB of it is hard to use for AI training today.

Most of the world's data isn't yet in a form AI can read.

Think factory sensor readings, experiment results, maintenance logs, customer inquiries, and daily reports where veteran workers noted why they made a call. None of it is on the web. These records come from the front lines of actual work. We call them "real data."

Real data often isn't much use if you just hand it to AI as is.

Say a factory log reads "motor temperature 82°C." Is that a problem?

You can't tell. Which motor? What's its usual temperature? When was it measured? Was a part replaced three days ago? Is vibration up? Only with all that can you say, "this might be an early sign of failure."

82°C becomes actionable once it's linked to the usual value, time of measurement, maintenance history and change in vibration. Chart by DATA WORLD based on the example in this article.

That surrounding information is context: what the data represents, where and under what conditions it was created, and who may use it and how. The number 82 only means something in context.

The trouble is, context is scattered across the company. The temperature sits in the equipment database, the replacement history in the maintenance contractor's system, a worker's sense that "something seems off" in a daily report. A U.S. National Academies report calls research data that can no longer be found or reused for lack of description "dark data". It's dark not because it's secret, but because the people who need it can't see it.

Many Japanese companies are at exactly this stage.

AI adoption doubled in a year, yet only 7.0% of companies have their own data ready to feed it, according to the Information-technology Promotion Agency, Japan (IPA), a government IT agency. In the same survey, just 3.9% said AI had "increased sales or profits." For now, AI mostly speeds up individual tasks such as summarizing and translating documents (IPA).

Picture a fridge where nothing has a name or an expiration date on it. That's where many companies are.

Japan's national answer to this shortage of meaning is Open Data Spaces (ODS).

WHAT IS

Open Data SpacesODS

A Japanese technical approach for connecting data across companies and organizations while keeping its meaning and terms of use intact

“The era of Data Scarcity. AI knows the world, but it has yet to understand the domain context.”── IPA Open Data Spaces official website

1. 2025.10Becomes a common specIPA, the University of Tokyo and others agree to make it Japan's common technical spec for data sharing
2. 2026.04Deliverables releasedDesign concepts, protocols, trial software and adoption guides published
3. 2026.04Going globalPresented at Hannover Messe, the industrial trade fair in Germany
4. NOWToward a context layerPromoted as the layer supplying the ‘business context’ AI uses

ODS, led by IPA, is a technical approach for managing data while it stays spread across its owners (IPA). Rather than pooling data in one place, each company keeps its own and shares it with trusted partners, only for agreed purposes.

It aims to connect three things: where data is, what it means, and who can use it, for what and how far. In April 2026, design documents, communication protocols, prototyping software and adoption guides were released (IPA).

The ODS website's language is telling. The first video in its intro series is titled "Data is Eating the World," and ODS calls itself a "context layer" for AI. The idea: AI's next food is real data, and context decides how it gets eaten. That's exactly the path this article has traced.

Companies worldwide now use nearly the same AI models. So where does the difference come from? From how much material about your company's own work you can hand to AI.

Only 7.0% of Japanese companies have their own data ready for AI. That's a weakness, but also an opportunity. Companies that start writing down when, where and why each entry in their daily reports and maintenance logs was made should pull ahead of the other 93%. They'll be the first to see AI's answers match how their work actually runs.

One honest caveat: in the same IPA survey, the top challenge in preparing data was "securing talent" (51.2%) (IPA). The people who add context are the people who know the front lines. AI's next race is over data, and just as much a race to put front-line know-how into words.

More than 10GW for one site in Ohio. A national strategy to cluster data centers in nine regions. The electricity race is fought by nations and giant corporations, not an arena most of us can enter.

The meaning race is different. You don't need a power plant to write "usually 65°C" next to 82°C. A small factory, a shop or a local government can start tomorrow. Of all the AI race's arenas, it's the closest to home and the easiest to join.

Finally, a few words about DATA WORLD.

Give data meaning, and the way you see the world changes, for people as much as for AI. The moment you learn who collected data, why and under what conditions, plain numbers start telling a story.

I love that moment. Ocean observations reveal a changing planet. Local statistics reveal how people live. Industry figures turn up seeds of new work. Each unfamiliar dataset lays another map over familiar scenery.

As the technology for handing context to AI advances, DATA WORLD wants to hand context to people: finding buried data, checking its sources and background, and editing it into a form that reaches even first-time readers. That's why I wrote this article.