790 Slurm Profiles, 2 Say 'SRE.' Where GPU Cluster Engineers Really Are

If you're trying to hire GPU cluster engineers by typing "SRE" or "data center engineer" into LinkedIn, you're searching for job titles that almost none of them have. Demand for them keeps climbing anyway. Indeed Hiring Lab's October 22, 2026 snapshot puts data center job postings at 247 on its index (June 2024 = 100), while tech postings outside data centers sit at 99. Kelly's 2026 Data Center Salary Guide lists GPU cluster operations as a shortage skill, with 84% of employers reporting high demand.

Why GPU cluster operations is now a shortage role

GPU cluster operations is a shortage role because data center hiring has pulled away from the rest of tech, and the skill sits on Kelly's 2026 list of the most in-demand data center skills. Indeed's index shows data center postings climbing slowly through 2024 and then speeding up sharply from mid-2025. They finished about 2.5x ahead of the rest of tech.

Kelly's guide ranks the scarcest skills this way:

• Liquid-cooling commissioning: 92% of employers report high demand
• Medium-voltage electrical expertise: 88%
• GPU cluster operations: 84%

The same guide says 90% of data center operators see workforce gaps as a critical risk. Pay points the same way. Indeed shows advertised wages for electrical engineering up 5.6% year over year and IT systems and solutions up 4.7%. Overall posted wage growth is 2.4%, and CPI inflation is 3.7%.

Why a search for "SRE" returns the wrong people

A search for "SRE" fails because almost nobody who runs Slurm uses that title. In Refolk's index of professional profiles, 790 US profiles list Slurm as a skill, and only 2 of them have the title "Site Reliability Engineer": one at SpaceX and one at Synopsys.

0.25%

Slurm is a workload manager. It queues jobs and assigns them to nodes across a cluster, and SchedMD says it runs on about 65% of TOP500 supercomputers. Site reliability engineering is a different job. SRE is the practice of keeping web services available through automation, error budgets, and on-call rotations. The two jobs sound alike, but day to day they look nothing alike.

The gap exists because GPU scheduling is something people pick up while doing another job. A research software engineer learns Slurm because her lab's jobs keep getting stuck in the queue. A systems architect learns it because the ML team needed a fair-share policy. A postdoc learns it because his thesis ran on a shared campus cluster. None of them changes their title, so a title search can't find them.

One caveat: skills in the index are self-reported, and the HPC count matched exact job titles, so 199 is a minimum.

What GPU cluster engineers actually call themselves

Most GPU cluster engineers carry generic engineering or research titles such as Software Engineer or Postdoctoral Research Scholar, and a much smaller group uses explicit HPC titles such as HPC Engineer. In a 25-profile sample of the Slurm-skilled pool (treat it as directional), the titles included:

• Software Engineer
• Distinguished Engineer
• Senior Systems Architect
• Computational Toxicologist
• Postdoctoral Research Scholar

The most common employers in that sample were NVIDIA, SambaNova, and Meta (two profiles each), followed by Google, Qualcomm, Snowflake, Pfizer, and Tempus AI.

The explicit-title group is much smaller. Among the 199 US profiles with HPC titles, the most common were:

1. HPC Engineer
2. HPC Systems Engineer
3. HPC Systems Administrator

So the skill-based pool is about four times the size of the title-based pool, and even the title-based pool is more useful than "SRE." This is the practical answer to how to recruit HPC engineers: write your search as a description of the work and ignore the title field.

Where to source: the pools that actually run clusters

The people who have run schedulers at scale sit in three places: the Slurm and Slinky projects, national labs and university HPC centers, and industrial companies with on-prem clusters. Converted bitcoin miners are a fourth pool, but for a different hire, which the next section covers.

Slurm and Slinky contributors

NVIDIA announced on December 15, 2025 that it was buying SchedMD, the company behind Slurm. SchedMD's second project, Slinky, runs Slurm on top of Kubernetes, which makes it the bridge between the HPC talent pool and the Kubernetes one. People who contribute to or file issues on Slurm and Slinky have shown both skill sets in public. Look at who files the detailed bug reports, not only who merges code. The people writing careful reports about stuck GPU nodes are often the operators you want.

National labs and university HPC centers

Slurm was co-developed at Lawrence Livermore National Laboratory, and Sandia National Laboratories appears in the HPC-title sample. With Slurm on about 65% of TOP500 systems, the deepest group of people who have run schedulers at real scale is at national labs and university research computing centers. Now that the Slurm vendor sits inside NVIDIA, and NVIDIA is already a top employer in the Slurm sample, expect more competition for these people. Reach them early, and lead your pitch with the scale of the problem they would work on.

Industry HPC benches far from the coasts

Ford Motor Company is the top employer in the HPC-title sample, followed by CoreWeave, then single profiles at Sandia, AMD, KLA, Shell, and GDIT. Lansing, MI and Houston are among the top regions. Automakers, energy companies, and chip-equipment makers like Shell and KLA have run on-prem clusters for years, and they sit well outside the default AI hiring map.

The coastal picture runs the other way. Indeed's local job postings indexes put Seattle at 67 and San Francisco at 72, among the lowest in the country. Weak local demand can make people there easier to approach, but "hire in SF" shouldn't be your default.

Ex-miners solve your facilities shortage, not your scheduler shortage

Hire ex-miner operations staff for the liquid-cooling and medium-voltage electrical gaps (92% and 88% in Kelly's guide), not for Slurm. Five major public miners have shifted toward AI and HPC and together produced 47% fewer bitcoin this year:

• IREN
• Core Scientific
• Bitfarms, now Keel, which stopped mining bitcoin entirely in June 2026
• TeraWulf
• Cipher

Most of their AI deals are colocation. In colocation, the miner provides power, cooling, and the physical building, and a tenant such as Fluidstack runs the GPUs and the software on top.

The capacity is real. On August 17, 2026, IREN delivered Horizon 1, the first 50 MW of the 200 MW it is building for Microsoft. TeraWulf had 102 MW of revenue-generating critical IT capacity at Lake Mariner, and Cipher's Barber Lake site belongs on the same list. The people who brought those sites online have done power, cooling, and physical infrastructure work at scale, which maps directly onto the two biggest shortages in Kelly's guide.

IREN is the main exception. It runs its own GPU cloud, so its teams are more likely to include people who work on the GPU layer itself. At the other converted miners, assume the GPU layer belongs to the tenant until a profile shows otherwise.

For data center engineer recruiting, this changes how you build the search:

• For facilities roles: search miner site operations, electrical, and commissioning staff at IREN, TeraWulf, Cipher, Core Scientific, and Keel.
• For scheduler roles: search the tenant side (CoreWeave and similar GPU clouds), plus the Slurm, lab, and industry pools above.
• For hybrid roles: start with IREN, then check each candidate's evidence individually.

How to check that someone has really run a GPU cluster

Ask for operational detail that people only learn by getting paged, then compare it with their public footprint. A profile that lists Slurm proves very little on its own, since skills in Refolk's index are self-reported.

A screening sequence that works:

1. Ask about a hung job. "Walk me through the last time a multi-node training job hung." Real operators name a cause: a bad link, a misconfigured partition, a GPU dropping off the bus, a node that should have been drained. Résumé padders talk about "monitoring."
2. Ask who owned the scheduler config. Did they write the partition layout, the fair-share or quality-of-service policy, and the GPU resource definitions, or did they only submit jobs? Many people in the 790 are users, not administrators.
3. Ask about node health. How did a bad node get detected, drained, and returned to service? Anyone who has run a large cluster has an opinion here.
4. Ask how they moved between Slurm and Kubernetes. With Slinky bridging the two, ask which one they would pick for a given workload and why.
5. Check the artifacts. Look for issues or patches on Slurm or Slinky, talks at HPC user groups, or co-authorship on lab technical reports. Ten minutes of reading is enough.

Step 5 goes fastest when the evidence is already linked to the person. When I return a candidate in Refolk, the repos, papers, and profiles behind the match come with it, so you check the work and skip rebuilding the trail by hand.

Pay and poaching: why your first offer has to be your best

Your first offer has to be your best because Kelly reports that 25% of data center personnel are hired away by competing operators, so assume you'll face a counteroffer. Pay also varies by 46 percentage points across the US, from Silicon Valley at 34.2% above the national average to Omaha at 11.8% below.

25%

A 25% poaching rate tells you two things. First, the realistic source for experienced operators is your competitors' ops teams: CoreWeave, the converted miners, and other GPU clouds. Second, anyone you hire is also on someone else's list. The 46-point geographic spread works for you if you're willing to hire away from Silicon Valley. Lansing and Houston already rank among the top HPC regions, and an offer below Silicon Valley rates goes further there.

Facilities offers need to move first. Electrical engineering wages are rising more than twice as fast as posted wages overall (5.6% vs. 2.4%), so a facilities offer that was competitive in the spring may already be behind. The usual result: a requisition for a liquid-cooling commissioning lead or a cluster operations engineer stays open for months, the agency shortlist is full of web SREs, and the one strong finalist accepts a counteroffer from the operator next door.

The fix for that cycle is to start from the work, not the title. Search the skill pool, which is about four times the size of the title pool, keep the miner pool for facilities, and put your strongest offer on the table first.

FAQ

What is a Slurm administrator, and how is it different from an SRE?

A Slurm administrator configures and runs the Slurm workload manager: partitions, scheduling policies, GPU resource definitions, and node health on a shared cluster. An SRE keeps web services available through automation and on-call practices. Some people do both, but in Refolk's index only 2 of 790 US Slurm-skilled profiles have the SRE title. That's why searches for Slurm administrator jobs and SRE jobs bring back almost completely different people.

Can former bitcoin miner staff run AI training clusters?

Usually not the GPU layer. Most miner AI deals are colocation: the miner provides power, cooling, and the building, and a tenant such as Fluidstack runs the GPUs. Ex-miner staff are a strong pool for liquid-cooling commissioning and medium-voltage electrical roles. IREN, which runs its own GPU cloud, is the main exception.

Does NVIDIA owning SchedMD change how I should source Slurm talent?

Yes. NVIDIA announced the SchedMD acquisition on December 15, 2025, and it's already tied for top employer in a sample of the Slurm-skilled pool. Expect more competition for experienced operators at national labs and university HPC centers. Contributors to SchedMD's Slinky project, which runs Slurm on Kubernetes, are a pool to watch because they've worked on both sides.