HPC · National Compute
Step off the podium,
build the used machine.
RIKEN, Fujitsu, and NVIDIA's next-generation supercomputer "Fugaku NEXT" stepped away from the goal of topping the TOP500 list and toward a machine that industry and research can actually use up. What this shift means for Japan's national compute — a read on the design intent.
What Changed
Inside the "no more #1" stance
RIKEN Center for Computational Science (R-CCS), Fujitsu, and NVIDIA — the trio building Fugaku NEXT — declared that "topping the TOP500" is no longer the primary design goal. That's an explicit break from the current Fugaku (in operation since 2020), which took the #1 slot on the Linpack benchmark for multiple cycles.
What replaces it: real-application performance and the fraction of time the machine is actually productive. Instead of a system that stands on the podium at peak FLOPS, the next machine is being positioned as one that keeps producing results in sustained operation, balancing power, cooling, and software fit.
By The Numbers
The comparison axis is rewritten
Why It Matters Now
Why the pivot lands now
TOP500 rank was the proof of "we built it." In the AI era, "we kept it running" is rarer and harder.
HPC has shifted in five years. Generative AI training and inference, materials, weather, drug discovery — most demand is bursts of large parallel jobs, and the practical fight is whether rack-scale GPU aggregates can be kept at high utilization. A world-leading peak that under-delivers against real applications wastes both power and budget.
The broader context: budget debates around Japan's Ministry of Education (MEXT) have grown sharply less patient with HPC investments that lean on "world-leading" as the pitch. A design that promises "outcomes over rankings" upfront threads both the fiscal debate and domestic user demand at once — a pragmatic policy fit. NVIDIA's role as design partner, likely leaning on Grace-Hopper and next-gen CPU-GPU nodes, adds credibility on the AI-effective-performance side.
Who Should Do What
How this hits, by role
Framed as "Japan's supercomputer," this is really about whether to use national compute at all. It lands differently by role.
Research engineers: portability math changes
Code optimized for the current Fugaku's ARM (A64FX) will likely need to run on NVIDIA GPU + general-purpose CPU in the next machine. Reoptimization cost and the upside from GPU acceleration should be re-estimated per project. Effective throughput on CUDA / PyTorch stacks can vary by multiples depending on how well the port lands.
PIs and PMs: allocation criteria change
Allocation policies that favored large Linpack-style MPI jobs are being re-drawn toward AI-training-friendly resource splits. Prepare proposals in the language of per-job effective TFLOPS and wait time, not benchmark headlines.
Industrial users: expect a wider industry lane
The "used" framing rhymes with policy signals to broaden industrial access. This is a good moment to re-evaluate national compute as a cloud GPU alternative on TCO. Materials, drug discovery, weather, and other mixed HPC + training workloads are prime candidates for redoing the go/no-go analysis.
What Comes Next
Short-term outlook and moves
Three near-term expectations. (a) Application-focused benchmarks (weather, molecular, AI training) are published as primary, with TOP500 numbers demoted to "reference." (b) Migration roadmaps from A64FX-optimized assets to GPU emerge, with re-porting support scaled up for industry users. (c) With NVIDIA co-design as the axis, and compatibility with the DGX software stack secured, differentiation lands in the domestic scheduling / job-planning layer. Two recommended actions: start pricing GPU-migration cost for your current Fugaku workloads early; and rewrite your cloud-GPU-vs-national-compute TCO in "application effective" numbers, not "Linpack-equivalent" ones.
Design not for the moment on the podium,
but for the everyday the machine gets used.
Counterpoint
Limits and things easy to miss
It's premature to cheer without caveats. First, walking away from "world's fastest" as a headline costs a real amount of international recruiting and domestic budget-defense leverage. TOP500 is legible, and it remains a strong external marketing asset. Whether "application effective" as a new KPI can carry the same political weight is untested.
Second, tightening co-design with NVIDIA means national compute becomes more exposed to U.S. export-control winds. If U.S.–China AI geopolitics keeps its current tension, availability of key parts and top-bin models will move on political calendars. Without a "purely domestic" contingency alongside, that exposure partly cancels the pivot's upside. Third, application-benches are harder to compare cleanly across vendors than Linpack, which risks reading as "nicely constructed self-report" — an operational care point. Two things to watch over the next six months: third-party validation of application-effective numbers, and the depth of the migration guide for existing Fugaku users.