AI Data Centre Co-location: The Chiba Test

AI data centre co-location sounds like a shortcut: put the machines beside a power station and skip the queue. JERA, Dell and RHAELM have now put that idea on paper in Japan. Their proposed Chiba campus would draw power behind the meter from an operating thermal plant. It is a serious design choice, but the announcement is a memorandum of understanding, not a working 400 MW data centre. That gap is where the interesting work sits.

The 1 October announcement describes power capacity of up to 400 MW, a target to begin operations around 2028 and expected capital deployment above US$15 billion across land, electrical infrastructure, buildings and AI compute. Apollo intends to be a strategic investment and financing partner for RHAELM. These are plans and stated intentions. None is a commissioning certificate or a signed customer workload.

Behind the meter, not beyond the system

Chiba Thermal Power Station is an operating plant, with maximum output of 4,380 MW. A data centre next to generation can be wired to receive power directly rather than wait for the conventional connection path the partners say they want to avoid. That is the advantage JERA is selling. You can design the electrical path and the compute campus together, on land near an existing source of power.

But "behind the meter" describes a commercial and physical arrangement. It does not tell us that 400 MW is uncommitted power, that the campus will be electrically islanded, or that the rest of the grid will never feel it. The press release does not publish the operating agreement or the dispatch rules. It does not explain what happens to power otherwise sold into the market when the data centre starts drawing it. Those are different questions from whether an onsite connection can be built.

This matters whenever someone presents a plant nameplate as a ready-made compute allocation. A generator's rated output is a ceiling under specified operating conditions. A data centre needs a delivered profile through maintenance, fuel disruptions and peak demand. The useful number is not the plant's headline capacity. It is the capacity the project can actually take, at what hours, under what contract, and with what backup when the arrangement fails.

I have argued before that AI infrastructure power procurement now belongs at the front of the project plan. Chiba pushes the thought further. The buyer isn't merely securing a power purchase agreement. The generator, developer and hardware supplier are proposing to design the supply chain as one project. It may remove one queue while creating a much more intimate dependence on the plant.

The bottleneck moves into the contract

JERA says the collaboration would combine its LNG procurement, shipping, import, regasification and generation capabilities with Dell's rack-scale AI equipment and RHAELM's development work. The partners want a repeatable design so they do not have to integrate power, cooling and compute from scratch at every site. On paper, that is sharper than buying chips first and asking a utility for megawatts later.

The price of this speed is coupling. If the power station is the supply plan, the campus inherits questions about fuel procurement and generator outages. If the compute hardware is standardised, the developer has to decide how that standard stays useful across successive chip generations. If the electrical and cooling design is fixed early, the team needs room to change the mix of training and inference loads without ripping the site apart. The memorandum tells us what the partners intend to integrate, not that they have settled those operating terms.

A lender should want to see who bears the cost when the plant is unavailable, what minimum energy is guaranteed, how replacement power is priced and whether the data centre can curtail less urgent work. A customer should ask what the contracted compute capacity means during a plant outage. An electrical one-line diagram and a fuel agreement will say more than a glossy rendering. These are questions, not allegations about Chiba's design. The public announcement does not answer them.

Power redundancy cannot be hand-waved away because a generator sits over the fence. There may be a grid backstop, onsite storage or another generation source in a final design. We do not have those details yet. The distinction is important: an architecture can be credible before every cable is specified, but a proposed power route should not be counted as delivered compute.

A repeatable AI factory needs local proof

The strongest part of the announcement is its ambition to make the design portable. JERA supplies the energy and site, Dell supplies a pre-integrated compute layer, and RHAELM leads delivery. One template for the electrical interface, cooling and racks could spare each new project from rediscovering the same engineering conflicts. In that sense, Chiba is a prototype for a way of building, not merely a big building.

A repeatable template still has to survive the local site. One station may have available land and a practical route for new electrical works. The next may have different export commitments, thermal limits, cooling water constraints or fibre paths. Standardising the rack does not standardise the permits. Nor does a successful connection at one station establish that the next site has spare firm capacity. JERA and RHAELM say they intend to explore other JERA locations. That wording matters: no fleet of finished campuses has been announced.

This is where energy-elastic AI infrastructure becomes more than an efficiency talking point. If some compute can move or pause, a campus can be designed around the power it can secure rather than an imaginary flat draw at full load. But a training job that can be shifted is not the same as an inference service with tight latency commitments. Any claimed flexibility needs workload-level evidence: which jobs, how often, how long and at what cost to the customer.

Cooling deserves the same discipline. The partners explicitly include it in the proposed standard model, which is sensible. Dense racks move the bottleneck from a megawatt figure to heat removal and water or air handling at the site. Without published specifications, nobody can tell whether the eventual Chiba design has a particular cooling method or water footprint. That uncertainty is a reason to ask for the design, not to invent one.

What would count as progress

Over the next two years, watch for commitments rather than larger numbers. A final project agreement would narrow the distance between MoU and build. A disclosed power delivery structure would show what "behind the meter" means in practice. Construction and electrical milestones would tell us whether the schedule is holding. First energisation and customer workloads would finally let us judge whether the integrated model delivered capacity faster than a conventional development path. JERA's target is to begin operations around 2028; it has not claimed the whole proposed capacity will be live on day one.

There is also an energy-accounting question worth keeping in view. The Chiba project is tied to an existing thermal power station, and JERA's own release says reliable gas-fired generation and the LNG chain are central to its model. That may be a workable route to early compute capacity. It is not a renewable-power claim. Anyone comparing the project with a proposed solar-backed campus should compare actual delivered hours, fuel exposure and emissions, not marketing labels attached to either one.

The industry loves a scale figure because it travels well in a headline. US$15 billion across all phases tells us the intended scope is large. It does not tell us the capital is committed, the site has customers or the plant has handed over 400 MW. This deal is worth following precisely because the companies are trying to solve a real sequencing problem: power, building and compute have to arrive together. If they can show the contracts and then the operating plant-to-rack path, Chiba will be a stronger case than another campus announcement. Until then, the disciplined word is "proposed."