BUSINESS── ARTICLE 009── Oct 5, 2026── 12 MIN READ

An AI factory next to a power plant: what Chiba's ¥2.3 trillion data center is really buying is time

Japan's largest power producer, JERA, has teamed up with Dell and RHAELM to plan an AI data center of up to 400 MW, worth more than $15 billion, inside the grounds of its Chiba power plant near Tokyo. How big is 400 MW, and how many GPUs could it run? Why build at a power station instead of a normal site? We look at the eight-to-ten-year wait for grid power in Tokyo, Japan's shifting demand forecasts, the US precedents and the CO2 math.

DATA WORLD Editorial Team Feature · 12 min

A bay at night where glowing data center buildings stand right next to an LNG power plant's chimneys and gas tanks, linked by power lines. DATA WORLD editor-in-chief Kaito Ogasawara looks on from the edge of the frame. Large gold text reads "BUYING TIME", with a smaller line "Chiba's ¥2.3T AI factory"

Japan is about to get what its backers call the country's largest AI data center, and they want to build it inside a power station.

On October 1, 2026, three companies announced the plan: JERA, Japan's biggest power producer; Dell Technologies, the US tech giant with more than $100 billion in annual revenue; and RHAELM, a London-based AI infrastructure developer. The site is JERA's gas-fired power plant on Tokyo Bay in Chiba, just east of Tokyo. The facility would draw up to 400 MW, and the whole project is budgeted at more than $15 billion (about ¥2.3 trillion). Operations are slated to start around 2028.

WHAT IS

JERA

The company that generates about a third of Japan's electricity, and buys the fuel for it from around the world

2015.04Founded by two utilitiesTokyo Electric Power and Chubu Electric Power set it up 50:50, starting by pooling their fuel buying
2019.04Takes over the power plantsThe two utilities' thermal power stations move into JERA. Fuel and generation now sit in one company
2020.10Pledges net-zero CO2 by 2050Sets a goal of zero emissions across its businesses in Japan and abroad
2025.10Invites data centers inSigns an MOU with Yokohama to explore data centers inside a power plant; agrees to talks with AWS in February 2026
NOWAn AI data center in ChibaOctober 2026: plans what it calls Japan's largest AI infrastructure with Dell and RHAELM

Most people outside Japan have never heard of JERA. Many inside Japan haven't either. Yet in fiscal 2025 it generated 228.6 billion kWh, about 30% of all the electricity produced by Japan's power companies. Its 55 GW of generating capacity is the largest in the country, and its plants ring Tokyo Bay, from Chiba and Futtsu to Kawasaki, Yokohama and Shinagawa. When the lights come on in greater Tokyo, a good share of that power is JERA's.

What makes the company unusual is everything that happens before the power plant. JERA buys liquefied natural gas (LNG) around the world, ships it on a fleet of 22 vessels, stores it at 11 terminals in Japan and burns it to make electricity. It handles roughly 40 million tonnes of LNG a year, one of the largest volumes in the world by its own account, and its revenue in fiscal 2025 was about ¥3.05 trillion (around $20 billion). In other words, JERA controls both the front door and the back door of Japan's energy supply.

In June 2026, JERA's management named power for data centers as one of its main growth bets. AI runs on huge, uninterrupted amounts of electricity, day and night. A company that makes electricity is now moving closer to the business of running AI, and Chiba is its first big move.

¥2.3 trillion and 400 MW are hard numbers to picture. And the location raises an obvious question: why a power plant?

The more we dug, the clearer the answer became. What this project is really paying for is not just electricity. It's the years it won't have to spend waiting for it.

SHORT ANSWERS

Four questions about Chiba's AI data center, answered up front

The ¥2.3 trillion buys less of a power supply than a head start.

Q1What was actually decided?

The three companies signed a memorandum of understanding to study the project. The ¥2.3 trillion covers everything from land to AI computers, and JERA isn't paying it alone.

→ PART 01

Q2How big is 400 MW?

Running at 80% on average, it would use as much electricity in a year as about 720,000 Japanese homes. In GPUs, that's well over 100,000 of today's top chips.

→ PART 02

Q3Why next to a power plant?

One analysis puts the wait for grid power in Tokyo at 8 to 10 years. Taking power directly from the plant, the partners say, can cut that wait by several years.

→ PART 03

Q4What's still open?

The fuel is LNG. At 80% load, burning it would emit roughly a million tonnes of CO2 a year. The power price and water use haven't been disclosed.

→ PART 04

What was signed: an MOU, and a ¥2.3 trillion price tag that covers everything

Start with what has, and hasn't, been agreed.

On October 1 the three companies signed a memorandum of understanding on a basic concept and a feasibility study. No final investment decision or completed financing has been announced. According to RHAELM, it has also signed heads of terms with JERA setting out the main conditions.

Here is the plan as announced:

Site: inside and next to JERA's Chiba Thermal Power Station in Chiba City
Size: up to 400 MW
Cost: more than $15 billion (about ¥2.3 trillion) across all phases, covering land, power equipment, buildings and the AI computers
Timing: operations to start around 2028. A JERA official told Reuters the partners expect full operation in 2029 and a 15- to 25-year power supply contract
Power: supplied behind the meter, straight from the plant's operating generators

JERA brings the power and the land, Dell the computing platform, and RHAELM the development and financing. US investment firm Apollo is expected to join as RHAELM's investment and financing partner. How much each party will put in hasn't been disclosed, so this is not JERA spending ¥2.3 trillion.

RHAELM itself is young. UK company records show it was incorporated in January 2026 and took its current name in August. Its leadership lists past experience, but we couldn't find a data center it already operates in public records.

"Japan's largest" is also the companies' own description: the largest single-site AI infrastructure in Japan, and among the largest in Asia outside China.

400 MW: 720,000 homes' worth of power, or about 180,000 GPUs

How big is 400 MW? A few yardsticks help.

The nearest one is the power plant next door. Chiba Thermal Power Station has 4.38 GW of capacity across 11 LNG-fired generating units: eight of 360 MW and three newer ones of 500 MW, added in 2014. The data center would equal about 9% of the plant, or 80% of one of its largest units.

The Chiba power plant's 11 generating units drawn as towers sized by output, with 80% of one large unit lit in gold. 400 MW is about 9% of the plant's 4.38 GW — The Chiba plant has eight 360 MW units and three 500 MW units. 400 MW equals 80% of one large unit, or about 9% of the whole plant. (DATA WORLD, based on JERA data)

Now compare it with homes. Data centers run around the clock. If this one ran at 80% of its maximum on average, it would use about 2.8 billion kWh a year. The average Japanese household uses 3,911 kWh a year (fiscal 2023). Divide one by the other and you get about 720,000 households.

That is also about 8% of Chiba Prefecture's entire annual electricity demand (fiscal 2024), from a single site. These comparisons match annual volumes only. They don't mean homes in Chiba would go short.

About 300 MW of the 400 would reach the GPUs

So how much AI hardware could this power?

Assume a PUE of 1.2, and about 333 MW of the 400 goes to computers; the rest runs cooling and power systems. If 90% of the computing power goes to racks of GPUs, the GPU racks get about 300 MW.

A grid of 2,500 gold squares, one per GPU rack, showing that 400 MW could run about 180,000 GPUs. A side bar splits the 400 MW into about 67 MW for cooling and power, 33 MW for other IT equipment and 300 MW for GPU racks — A DATA WORLD estimate assuming a PUE of 1.2 and 90% of IT power going to GPU racks, filled with NVIDIA GB200 NVL72-class racks (about 120 kW and 72 GPUs each). The chips and server counts for Chiba have not been disclosed. (DATA WORLD, based on NVIDIA, HPE and TrendForce data)

An NVIDIA GB200 NVL72 rack holds 72 GPUs and draws about 120 kW. At 300 MW, that's 2,500 racks and about 180,000 GPUs. With the next-generation GB300 the count would be about 164,000. For the Rubin generation after that, one research firm estimates about 225 kW per rack, which works out to roughly 96,000 GPUs.

Newer generations fit fewer GPUs into the same power. But each chip does far more work, so a lower count doesn't mean less computing.

Why a power plant: skipping the wait for the grid

Normally a data center connects to the utility's grid. Around Tokyo, getting that connection has turned into a long queue.

TEPCO Power Grid, which runs the grid in the Tokyo region, says it had received connection requests from data centers and similar users totaling about 17.9 GW by the end of March 2026. That figure includes projects slated for 2037 and beyond, and not all will be built. Still, it is roughly 45 Chiba projects' worth.

Building the grid takes time, too. In Inzai, a Chiba city packed with data centers, Tokyo Electric Power built a new substation and cut a job that normally takes eight years to four years and nine months, which it reported as an achievement. Even the fast version took almost five years.

The real estate firm JLL estimates the wait for a data center to secure power at 8 to 10 years in Tokyo and 3 to 5 years in Osaka.

A timeline comparing waits for power: 8 to 10 years in Tokyo and 3 to 5 years in Osaka, the Inzai substation built in four years and nine months instead of the usual eight, and the Chiba plan going from announcement in October 2026 to operation around 2028 — Tokyo and Osaka are JLL's estimates of the time to secure power, Inzai is a substation construction period, and Chiba shows the planned schedule. They measure different things. (DATA WORLD, based on JLL, TEPCO and JERA data)

That is where the Chiba plan makes sense. Inside the plant's grounds, a data center can take power without joining the grid queue. JERA says this approach can get it running several years sooner than a standard grid connection.

In AI, a year is a hardware generation. Wait eight years for power, and the computers you planned for are long obsolete. That's why we think what this ¥2.3 trillion is really buying is time.

The government says "spread out." Why build near Tokyo anyway?

There's some tension with national policy here.

Since 2025, Japan's government has run a public-private council on what it calls "Watt-Bit collaboration", planning power and data networks together. Its interim summary pointed toward spreading data centers to regions like Hokkaido and Kyushu, where power is more plentiful and renewables are available. According to JLL, about 90% of Japan's data centers are now in Tokyo and Osaka.

Chiba won anyway because a big power plant, transmission lines and usable land were already there, next to Tokyo. The partners themselves cite the existing generators and adjacent land as the reason they can start early. No customer has been named, so we can't say whether being close to users was also a factor.

Demand forecasts keep rising, and keep slipping

How much power will Japan's data centers need overall?

The latest forecast from OCCTO, the national grid coordinator, puts the added peak demand from new and expanded data centers at 6.61 GW by fiscal 2035, or about 16 Chiba projects.

But the forecast moved a lot in a year. For fiscal 2030, the previous edition expected 4.4 GW. The new one says 3.28 GW, about 25% lower.

Look further out, though, and the picture changes. For fiscal 2034, the two editions are almost identical: 6.16 GW versus 6.15 GW.

Line chart comparing OCCTO's 2025 and 2026 forecasts of added peak demand from new data centers. Fiscal 2030 was revised from 4.4 GW to 3.28 GW, about 25% lower, while fiscal 2034 is almost unchanged at about 6.15 GW — The fiscal 2030 estimate fell about 25%, but fiscal 2034 barely changed. The growth was pushed later, not cancelled. Both lines are forecasts and cover only new and expanded data centers. (DATA WORLD, based on OCCTO demand forecasts for FY2025 and FY2026)

In other words, the destination stayed the same and the arrival got later. OCCTO reviews each project's likelihood and timing before it counts the demand. The most natural reading: the demand is real, but power and construction can't keep up, so it keeps slipping. That is exactly what makes a "start now, next to the plant" model valuable.

The US hasn't settled on one answer either

The US got to the power-plant-adjacent model first, with mixed results.

Talen Energy and Amazon had a data center running next to a nuclear plant. In 2025 they said they would switch to a grid-based supply deal of up to 1.92 GW. One trigger was a 2024 decision by FERC, the US power regulator, rejecting a contract change that would have expanded supply to the co-located site, amid concerns about costs for other customers and grid reliability. FERC has since asked the grid operator to write new rules.

Microsoft's Three Mile Island deal is often described as a direct nuclear hookup. It is actually a 20-year agreement to buy power from a restarted reactor that feeds the grid. xAI took a third route, installing its own gas turbines and batteries to get power fast, and drew criticism from environmental groups over air pollution.

Every country, and every project, is still figuring out how to get power quickly. In Japan, Chiba looks set to be the first big test.

Still unanswered: CO2 and the price of power

Finally, the questions that remain open.

The first is carbon. The Chiba plant burns LNG, which emits less than coal but far from nothing. Using reference emission factors for this type of plant (0.341 to 0.362 kg of CO2 per kWh), running 400 MW at 80% would emit roughly 960,000 to 1.02 million tonnes of CO2 a year from combustion.

JERA has pledged net-zero emissions by 2050 and plans to use hydrogen at its LNG plants. But when the data center switches on around 2028, its power will still come mostly from LNG.

The second is price. At 400 MW running all year, a difference of just ¥1 per kWh adds up to about ¥3.5 billion a year. The rate fixed in a 15- to 25-year contract will make or break the economics, and it hasn't been disclosed.

Cooling and water use, local jobs, and backup power for when the plant's units are down are also still unknown. Whether the project is good for the area can only be judged once those details are public.

EDITOR'S TAKE: How Ogasawara reads it

The AI race moves from chips to power, and from power to time

Three things we took away from digging into the Chiba plan.

The cost of AI is starting to be set at power plants

The AI we use every day runs on electricity in some data center. Converting 400 MW into homes and GPUs made it obvious how directly the amount of AI available maps onto the amount of power. The global scramble for power that AI has set off is now visible on the shore of Tokyo Bay.

Waiting costs more than money

Eight to ten years to get power in Tokyo is several generations of AI hardware. Starting in 2028 is worth more than the ¥2.3 trillion itself. To us, this deal shows that in AI infrastructure, the thing that sells for the most is a start date.

Power companies are becoming AI companies

JERA has stepped beyond selling electricity toward offering power, land and computing as a package. If its plan for several gigawatts in the 2030s comes through, Japan's map of AI computing will start to look a lot like its map of power plants. How it answers the questions on carbon and price is what we'll be watching next.

SOURCES ── References

#BUSINESS#Data centers#JERA#Dell#RHAELM#AI and power#Chiba#Japan energy