Bradd Lewis describes his company's founding problem in one phrase: stranded power capacity. On 1 October, RHAELM signed a memorandum of understanding in Tokyo with Japan's largest power generator and with Dell Technologies. The site is JERA's Chiba Thermal Power Station. The commitment runs to 400 MW of AI compute.
More than US$15bn — 2.3 trillion yen — goes into land, power infrastructure, the building and the compute inside it. Operations are targeted for around 2028. The host plant lists a maximum output of 4,380 MW. So the data centre load is roughly a tenth of the generator standing next to it.
Whether that is the fastest route to compute or the most expensive way to tie compute to one fuel for a generation is the argument the announcement started. It has not settled it.
Behind-the-meter power deletes the interconnection queue from the critical path — and that queue is the reason the structure exists at all.
The unresolved parts are fuel and financing, not megawatts: Apollo has said it intends to serve, and the supply agreement runs for fifteen to twenty-five years.
What the MoU commits, and what it leaves open
Three companies signed a document about repeating a process.
JERA brings the generation, the LNG chain from procurement through shipping and regasification, and the site. Dell supplies what it calls the AI Factory, standardising the compute layer with pre-integrated rack-scale infrastructure built for rapid deployment. RHAELM, headquartered in London, leads development and delivery.
The MoU itself asks for a standardised, repeatable and scalable model, built first in Japan and applied in other markets later. Chiba is the reference case, not the product. The agreement carries a Heads of Terms between JERA and RHAELM for the investment, with Apollo Global Management intending to serve as strategic investment and financing partner. That is the whole financing structure, for now. Capital deployment is expected to exceed US$15bn across all phases, covering land, power infrastructure, facility construction and compute.
$15bn is the Chiba project across every phase. The $140bn figure reported by the Financial Times is Japan's wider programme with Dell and JERA. They are not the same money, and the gap between them is the reason standardisation matters.
Jiji Press, via bne IntelliNews on 3 October, stated the engineering claim plainly. The facility draws electricity directly from the plant that hosts it. That removes the need for large-scale transmission. Jiji also puts first operations at around 2028, and JERA's release calls the campus the largest single-site AI infrastructure deployment in Japan.
The case for stranding nothing
The queue explains the whole structure. RMI, a US research institute tracking generator queues, put the average stretch from an initial interconnection request to commercial operation at nearly five years in 2024. It had been under two years before. A queue that long does not care how quickly your racks arrive.
Behind-the-meter siting removes it from the critical path. What JERA already holds at Chiba is land, switchgear and a fuel chain. Site area is about 0.76 million square metres. The host asset lists 4,380 MW of maximum output. The queue is the part it does not have.
The second argument is standardisation, and it is the one the MoU is actually about. Every AI campus designed from scratch is a bespoke engineering project. Its own substation, its own cooling loop, its own schedule. Dell AI Factory fixes the compute layer to a pre-integrated rack-scale standard instead. JERA and RHAELM intend to explore the same pattern at other JERA sites. Their stated ambition is multi-gigawatt AI capacity across Japan in the 2030s, and Chiba either proves that pattern or kills it. That makes it a high-value first site.
The third argument is demand, and it is the weakest of the three because it rests on somebody else's forecast. Kuniyoshi Suzuki, senior director of SoftBank's cloud AI service division, has projected that Japan will need 320 times more AI computing power by 2030 than it used in 2020. SoftBank sells compute, so treat that as a pitch. Even discounted, a claim that size sets a floor under demand for anything that reaches 2028.
Land, switchgear and firm around-the-clock output already exist on the site, so the delivery schedule depends on construction rather than on a queue position.
In September we wrote about TAR's $120M Series A for off-grid power to AI data centres. Same thesis, different balance sheet. TAR had to build generation from scratch. JERA already owns the asset, and can sell the output under a supply agreement reported at fifteen to twenty-five years. That structure matters more than the headline gigawatt figure.
Eiji Ueda, Apollo's partner and head of Asia Pacific, named two of the firm's conviction themes. The first is financing the industrial renaissance, AI and digital infrastructure included. The second is serving as a long-term capital partner to Japan. Both themes are live here. A generator with firm load, a developer with delivery capability and an infrastructure investor with a mandate is an unusually complete stack. Few projects collect all three on a single signature page.
The case against one 400 MW bet
The document is thinner than the headline. What exists is a memorandum of understanding, a Heads of Terms, and a statement that Apollo intends to serve as financing partner. Intends is doing real work in that sentence. No final investment decision, construction start date, environmental permit or anchor tenant is disclosed.
By co-locating at JERA's Chiba station and deploying the Dell AI Factory as our compute standard, we can deliver a 400MW facility years ahead of a conventional grid-connected timeline.— Bradd Lewis, CEO & Chief Revenue Officer, RHAELM
Ahead of which baseline, and at what price, is not stated anywhere in the release.
The second objection is the one the announcement walks past. Four hundred megawatts of firm load behind the meter is a long claim on thermal generation, which the Financial Times reports as a fifteen-to-twenty-five-year supply agreement. That is a generation contract, not a short lease. JERA's own corporate materials describe a growth strategy to 2035 and a zero-CO2 2050 target. JERA runs 26 thermal power stations with about 59 GW of combined capacity. It also operates 11 LNG terminals. Domestic assets total roughly 60 GW, close to 30% of Japan's electricity.
Co-location solves the queue by relocating the constraint, from interconnection to fuel. Whether that is a hedge or a trap depends on gas prices and on carbon prices in Japan, and the release prices neither. Neither number is public here.
The third objection is concentration. One site, one host asset, one cooling loop, one rack-scale standard. A 400 MW cluster inherits every single-site failure mode that a distributed build would have spread across several campuses.
JERA and the Financial Times also disagree about the date. The release targets operations around 2028. The FT puts full 400 MW capacity in 2029. Somebody will have to explain which year is real.
Yesterday we argued that the fastest route to a grid connection runs through giving up control over when a data centre consumes power. A behind-the-meter firm load bets the other way. It makes the campus a fixed industrial customer of a generator rather than a dispatchable participant in a market. Both bets can be right. They are not the same bet.
Where the two readings split
The disagreement is narrower than the framing suggests, and it lands on specific line items.
| Line item | The reuse case | The sceptic's case |
|---|---|---|
| Power path | Behind-the-meter from an operating 4,380 MW asset | Firm fifteen-to-twenty-five-year claim on thermal output |
| Interconnection | Queue leaves the critical path | Transmission scope deferred, not avoided |
| Timeline | Operations targeted around 2028 | About two years from MoU to first load |
| Financing | Apollo intends to serve as financing partner | Intention rather than a closed facility |
| Replication | Multi-gigawatt across JERA sites in the 2030s | Chiba may stay the only site |
Sources: JERA press release, 1 October 2026; Financial Times, 1 October 2026; bne IntelliNews citing Jiji Press, 3 October 2026.
Two of those rows get settled by filings nobody has made yet.
Chiba AI campus power commitment
The campus draws behind the meter of a plant JERA lists at 4,380 MW of maximum output. JERA, 2026
Put plainly, the campus asks for less than a tenth of the host plant's nameplate. Whether that is efficient co-location or a rounding error depends on what the spare megawatts are worth. That is a question for the capacity market, not for this press release.
What would settle it before 2028
Four disclosures would move a framework into the category of a project.
✔ Arguments for co-location winning
A named anchor tenant or compute offtaker for the first phases, so the demand behind 400 MW has a name.
Apollo's facility closed with a disclosed structure, and at least one further JERA site committed.
Confirmation criteria: a second site in the 2030s pipeline, cooling architecture published, first load before 2029.
✗ Arguments against it working
Delivered capacity phased well below 400 MW, leaving the standardisation case unproven at the only site that matters.
Falsification criteria: capacity below 200 MW at first load, or a second JERA site that never appears.
None of that requires a new law, a subsidy or a technology that does not exist. It requires a filing — and a filing is exactly what this announcement is not.