Summary

  • San Diego Supercomputer Center publishes user and allocation counts, but every figure is bundled across several systems at once, so the flagship machine named in the news release never appears on its own.
  • Its operational pages carry thirty days of queue telemetry for the older V100 GPU partitions and an N/A for the H100 AI partitions, while disagreeing with themselves over how many nodes exist at all.

Picture the meeting. A programme officer at a public funder has to decide whether the AI hardware bolted onto a national supercomputer in 2024 did what it was bought for. A researcher writing an allocation request wants to know whether her job will queue for a day or a fortnight. Both go looking for the same thing: what the machine delivered, at what scale, over which period. Both come back with a number that covers three systems at once and a chart whose own denominators contradict each other two sentences apart.

That is the state of the public record at San Diego Supercomputer Center (SDSC) as its flagship system enters its last chapter. On 15 September 2026 the centre announced a $10M NSF award for Expanse2, which "will replace the NSF-funded Expanse supercomputer", in service since late 2020. Production is expected in 2027. The announcement does not say when Expanse stops accepting work — replacement is not a retirement date, and no document I inspected supplies one.

The bundle hides the machine

The most recent annual report inspected, carrying a 2025 copyright and a quantitative panel labelled fiscal year 2023/24, gives 4,015 UC users and 970 UC active allocations for Expanse, Voyager and the NRP combined. Two things are worth pausing on. The allocation figure counts allocations active at the period, not awards granted during it. And the scope is three systems, so no reader can subtract the two others and recover Expanse alone.

The previous annual report's fiscal 2022/23 panel gives 2,339 UC San Diego users and 651 UC San Diego active allocations for Expanse, Voyager, the NRP and Comet combined. Those are larger machines' worth of context packed into one line: four systems, one of them since retired from the bundle. Read the two panels side by side and you are not looking at growth or decline. You are looking at two different bundles and calling it a series, which is precisely what the documents invite and what a careful reader cannot do.

The only throughput figure in the set is the most general of all. A User Services statement in the July 2021 – June 2023 summary report gives nearly two billion core-hours and more than 12 million accelerator-hours per year across all SDSC supercomputers, and about 100,000 users "in recent years". Per year is a rate, not an audited fiscal window; "in recent years" is not a period at all. Neither figure is broken down by Expanse, by partition, by project or by allocation award.

So the flagship machine has no published consumed-hours total of its own, no award count of its own, and no utilisation percentage against installed capacity. What exists is a three-system rate and a four-system headcount separated by one fiscal year and one changed bundle.

The denominator will not hold still

Suppose you decide to do the arithmetic yourself. The operator's own pages stop you. The Expanse CPU documentation, as of the 10 November 2026 metrics update, describes a 728-node cluster in its narrative. The queue table on the same page shows 734 nodes across the compute and shared rows. No reconciliation appears between them.

The GPU page repeats the defect on the half of the machine that matters most to AI buyers. Prose says 52 GPU nodes. The V100 gpu and gpu-shared queue rows total 56. A utilisation percentage needs a denominator; this publisher cannot fix its own within a single document, so any figure an outsider computed would be an artefact of which sentence they trusted.

Those pages do carry real operational telemetry: rolling thirty-day wait-time trends, wall-time trends and jobs-run counts for the general-purpose resources. That answers a queueing question. It does not answer a delivery question, because neither page has a utilisation field or a consumed-hours field.

The AI column is blank, not zero

The operator does draw the line AI policy cares about. The GPU page states that GPU nodes are allocated as a separate resource, and it distinguishes the V100 gpu and gpu-shared queues from the H100 nairr-gpu and nairr-gpu-shared queues — partitions named for the national AI research resource, sitting inside the same machine but administered apart from the general-purpose allocation.

For the 10 November 2026 metrics update, the V100 queues show thirty-day job and timing metrics. The H100 NAIRR rows show N/A. That is a blank column. It is not a zero, and nobody should read it as one: an absent metric proves an absent measurement, not an idle accelerator. But it does mean the one partition a reader would consult to judge the AI expansion is the one partition with nothing in it.

The blank is a choice rather than an engineering impossibility, and the ACCESS programme's own vocabulary shows why. The XDMoD data catalog defines period-aggregated metrics for unique accounts with jobs, CPU core-hours allocated and used, active allocations, active PIs, active projects, and XD SUs allocated and used, with raw data reachable through the Data Analytic Framework or the Data Export tab. The instrument exists. The definitions are published. What is missing, for this resource in this period, is the reading.

One honest limit: that catalog page supplies definitions, and the linked dashboard returned a JavaScript shell during retrieval. No numeric ACCESS-level value is asserted here, and none should be inferred from the definitions.

What the record actually supports

A programme officer can establish from these documents that SDSC runs a large, real, publicly funded estate with a documented AI-specific allocation route; that its reports count users and active allocations; that its operational pages expose queue behaviour for general-purpose resources; and that a $10M successor system is funded and expected in 2027. She cannot establish, from the same documents, what fraction of Expanse's installed capacity was delivered in any fiscal year, how much of the AI partition's capacity was consumed, or what relation any of it bore to the money spent.

That is the finding, and it is narrow enough to be worth stating precisely: the publishing is not absent, it is aggregated past the point of accountability for the specific machine. The door is documented in detail. The meter reads blank.

Evidence boundaries

The two annual reports and the summary report do not print publication dates in the files inspected; they are cited here by the fiscal-year panels and the July 2021 – June 2023 coverage they carry, not by a stated issue date. The absence of an Expanse-only consumed-hours figure is an observation about these documents on 12 November 2026, not proof that no such figure exists anywhere. The node-count discrepancies are recorded exactly as the pages state them, with no attempt to adjudicate which count is right. N/A entries are reported as blank.

The 2027 production target for Expanse2 is not treated as an Expanse retirement date, and the "tens of thousands of researchers" in that announcement is a cumulative framing, not an annual user count.