The Compute Bill Sitting Under J-MID

The Japan-Medical Image Database paper records, in a sentence written like an inventory line, that by April 2024 J-MID had collected 534 million CT and MR images across 1.65 million cases from ten university hospitals, pushed over the SINET research network into a central cloud store that migrated off-premises in 2023. The number lands quietly. Ten hospitals, one wire, one bucket, half a billion images ready to be fed into whatever radiology model a Japanese research group decides to train next. What the paper does not put a figure to is the second bill that arrives when the models start to run.

That second bill is the point of this piece. Japan and Korea are the two countries whose hospitals moved from AI-in-radiology as a pilot to AI-in-radiology as a shift-by-shift dependency during 2025 and 2026, and both are running the same experiment on the electricity side of the ledger without having priced it. The Green Radiology review the Japanese Journal of Radiology published in 2026 is the first paper I have read in either country that names it out loud: the operational phase of imaging systems, particularly the electricity they draw, is the dominant contributor to radiology-related emissions, and the non-productive idle window between scans is where most of the loss sits. That has been the working suspicion in facilities-management circles for a decade. Seeing it in the radiological literature, from a Japanese author group writing to a Japanese readership, changes what a hospital director can plausibly say she did not know.

What the Green Radiology paper actually establishes

The Asian-Oceanian Society of Radiology survey that precedes and feeds the review is worth reading on its own. Japanese institutions ranked at or near the top of the region on the specific practices that trim the idle draw: energy-conserving lighting in reading rooms at 79 percent of surveyed sites, workstations powered down between sessions at 73 percent, and CT scanners with vendor-provided auto-shutdown modes enabled at the highest prevalence in the sample. That is real, and unusual internationally, and it is where a Japanese chief radiologist can point to a decade of quiet engineering discipline the rest of the region has not matched.

None of it prices the AI stack. The idle load the survey characterises is the equipment on the scan side of the room. What has landed on the other side of the room in the last eighteen months is a set of inference workloads that do not sleep between patients: report drafting on chest films, brain-tumour MRI classification, ultrasound segmentation running at the bedside, admission triage lifted out of an emergency-department queue. Every one of those is a continuous compute load layered on top of an intermittent scanner load. The Green Radiology framing has not caught up to it yet, and neither, in any published document I can find, has the facilities plan of the ten J-MID hospitals feeding the corpus.

The number the hospital's power meter now has to answer for

Take the Fujitsu-NVIDIA hospital admissions programme, reported by Nikkei this year, where inference sits behind every step of the patient journey from intake to post-op. The compute is placed inside the hospital's own facility in the Fujitsu reference architecture, on GPUs whose power draw the hospital's chief engineer now owns. Take the RIKEN and National Cancer Center Japan MoU revised on 18 February 2026, which folds pharmaceutical R&D into a joint programme resting on supercomputers and, in the fuller press language, quantum computers, based at RIKEN's own site. That is not the hospital's meter, it is the research institute's. The two energy budgets are shifting toward each other regardless. A radiopharmaceutical pipeline that inherits a training run against genomic and imaging data from a partner hospital is a compute footprint that no one accounting practice today assigns cleanly to either the treatment ledger or the research ledger.

The counter-move, and the more interesting one to my ear, is the Korean track. Deepnoid, one of the country's larger medical AI vendors, rebuilt its radiology-report inference stack around chips made by FuriosaAI, the Seoul startup whose RNGD accelerator runs LLM inference at roughly 2.25x the tokens-per-watt of an H100 at LG Research's benchmark, on a 180-watt part rather than a 700-watt one. Furiosa began mass production of the RNGD in January 2026. Both are Korean vendor claims and the usual pinch of salt applies until a hospital's facilities director publishes an actual after-versus-before draw on the same scan volume. The direction, though, is unambiguous. A hospital can either put an H100 rack in its basement to run report inference at scale, or it can put a Furiosa rack there and cut the electricity line by a factor between two and three at the same throughput. If a Korean tertiary hospital is running a hundred thousand chest reports a month through an LLM assistant, that factor is the difference between one substation feeder and two.

Samsung Medical Center picks a side, quietly

The move that says the most about where a serious Korean hospital thinks the energy question sits is at Samsung Medical Center in Seoul, which in June 2026 began developing an on-device real-time ultrasound AI for emergency and critical care. On-device is the operative word. The alternative architecture, inference over a cloud link to a co-located GPU pool, is the one every major vendor is nudging the hospital toward. Samsung chose to put the model on the machine at the bedside. That is a bandwidth call in part and a latency call in part. It is also, in the medium term, a power-and-cooling call. The scanner draws what it draws. The bedside inference chip draws what it draws. Nobody has to keep a warm eight-GPU node in a chilled room to serve the query.

The Korean Ministry of Health and Welfare's medical AI training programme, running through Samsung Medical Center, Seoul National University Hospital, and two peers since 2025, is honest about the clinical-adoption gap. What no version of that programme's public documentation does, and what no ministry document in either country has yet done, is publish the electricity assumption underneath the model deployment plan.

The disconfirming reading, from the radiologists themselves

The most careful pushback on the whole build-out is coming from the radiologists in the room. The Korean Society of Radiology, in a position captured by Korea Biomedical Review, flagged an objection to broad reimbursement of AI-based diagnostic tools on the grounds that the evidence base for clinical improvement over an existing radiologist workflow is thinner than the deployment pace implies. Read on the merits, this is a serious argument. If the marginal AI read produces at best a small, uncertain uplift on the final report, the incremental energy cost of running the model on every scan, everywhere, is a real cost against a small, uncertain benefit. That is a question a hospital chief medical officer should be answering with data, not deflecting with a slide deck.

Where I part from that reading is on where the AI stack has already earned its place. Report drafting on chest X-rays and prioritisation on brain MRI are two workflows where the throughput gain is large enough that even a lossy energy trade clears the bar. The National Cancer Center Japan's August 2026 press stream includes work on models that predict IDH mutation status from a brain-tumour MRI at a level the reporting oncologist can act on. That is a scan where the model earns its watt. Every hospital rolling AI onto every scan on the same tariff has not made the discrimination.

The question the plan has not answered

Japan's five-year commitment to stand up ten AI-powered hospitals, on the numbers the trade press is quoting the government at, is one of the largest single healthcare compute build-outs any advanced economy has publicly committed to this decade. Korea's small-hospital rollout, model by model, is the same programme in a different register. The energy line under both is a number I have not seen published by either ministry, either national radiological society, or either of the flagship university-hospital consortia that will run the racks. It is not that the calculation is hard. It is that nobody has been made to do it in public.

What is the operating electricity budget, per hospital, for the inference load Japan and Korea are about to switch on across every teaching hospital in both countries by the end of 2027? I do not know. I have not found the number in any of the primary documents I read for this piece. I have not found it in either country's press briefings for the last twelve months.

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan for Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.