Firmus Raised $2B. Australia's Immersion Pool Is 2 Engineers.

On August 7, 2026, Firmus closed a $2B Series G at a $10.5B post-money from Nvidia, Coatue, Blackstone and Jane Street, on the back of an Nvidia partnership covering 170,000 accelerators through 2028. The plan is $73.3B of build, 1.6 GW of capacity across five Australian sites, and a first Tasmanian campus that will eventually house 36,000 GB300s. The problem nobody in the term sheet is talking about: the number of engineers in Australia who have actually shipped immersion-cooled dense GPU deployments is two.

Why Firmus hiring is a talent problem, not a capital problem

Firmus has more capital than any Australian data-center operator in history and a smaller in-country immersion-cooling talent pool than most Series A startups. The Series G lifts post-money above $10.5B, nearly double April's $5.5B mark, and follows a $10B Blackstone-led debt facility from February 2026, one of the largest private credit transactions in Australian history. On top of that sits a three-year Aurora Energy deal for 104 MW of Hydro Tasmania supply and a build pipeline that would make Firmus Tasmania's single largest electricity user.

Now the counterweight. In Refolk's index of professional profiles, the phrase "immersion cooling data center" returns exactly two people in Australia: a Global Sales Director at FogHashing and a Data Center Liquid Cooling Consultant at Innofabrics, both in the Sydney area. Broaden to "liquid cooling GPU data center" and the Australian pool collapses to one profile. Neither query returns a single hit in Tasmania, which is where Project Southgate is being built.

18,000

What Project Southgate actually needs

Project Southgate needs immersion-cooling operators, dense-GPU critical-facilities engineers, and AEMO grid-integration specialists, in a state where none of them currently work. The public plan is a $4.5B campus in Launceston, Tasmania with a 90 MW first stage in 2026, expanding to Melbourne, Sydney, Canberra and Perth for 1.6 GW by 2028 at a projected $73.3B build cost. The Nvidia leg adds 170,000 accelerators between Q1 2027 and early 2028, mostly through Batam in Indonesia, with up to $30B in first-six-year revenue based on customer commitments.

Firmus's own marketing is the tell on how specialized the labor is. The company claims its modular liquid-cooled HyperCube design uses up to 60% less energy and costs up to half as much to build as a traditional data center. That premium is not free. It means an air-cooled NABERS-rated colo operator is not a drop-in hire. The mechanical stack, the coolant chemistry, the pump and CDU integration, and the failure modes are all different.

The three roles that gate the build

• Immersion-cooling operators. Dielectric fluid handling, two-phase versus single-phase tradeoffs, CDU and manifold design, seal and gasket compatibility with GB300 cold plates.
• Dense-GPU critical-facilities engineers. 130 kW+ per rack power delivery, busway sizing, redundancy topologies that assume liquid, not CRAC.
• AEMO grid-integration engineers. FCAS market participation, NEM dispatch response, load-shaping software that ties into Nvidia's DSX architecture.

The third role is the one that will actually decide whether Southgate hits its 2028 date, and it is the one that basically does not exist as a titled job on LinkedIn.

The Refolk index: HyperCube engineers by country

Here is the pool data straight from Refolk's index. The Australian numbers are not "small," they are effectively a rounding error against the build plan.

The 14 US profiles cluster around Green Revolution Cooling (Austin), Pentair, Baltimore Aircoil, and Denvr Dataworks, spread across Houston, the Bay Area, Boston and NYC. That is the realistic visa-sponsorship pipeline for Firmus's Melbourne and Sydney sites. There is no comparable cluster in APAC outside of Singapore, where STT GDC already runs Firmus HyperCubes.

Firmus is building 400 MW of GPU capacity in a talent desert and will have to fly engineers in or convert Bell Bay industrial workers.

Nvidia is already competing with Firmus for the same two people

One of the only two Australian data-center power-engineering profiles in Refolk's index already works at Nvidia, which is also a Firmus investor and the sole GPU supplier for the entire 170,000-accelerator plan. That is not a coincidence, and it is not fixable with money. Nvidia's DSX program pulls exactly the kind of engineer Firmus needs to run a grid-aware AI factory, and Nvidia pulls them first, because DSX is the reference architecture Firmus is deploying against.

For a recruiter, this reframes the outbound entirely. Every "obvious" candidate in Sydney has already had a Nvidia conversation. The real move is upstream and sideways: the ex-crypto immersion operators nobody is looking at, and the US and Singapore transplants who need visa sponsorship. Which is the exact gap Refolk closes: you describe the person in plain English ("Australian, ran two-phase immersion at a crypto miner, now doing anything") and get a ranked shortlist that a Boolean string on LinkedIn will not surface, because none of these people self-title as "immersion cooling engineer."

The transferable-adjacent pool is bitcoin miners, not colo operators

The densest source of HyperCube-competent engineers in Australia is ex-crypto mining staff, not traditional data-center hires. Firmus itself was founded in 2019 by Oliver Curtis, Tim Rosenfield and Jonathan Levee, initially targeting Bitcoin mining and HPC on Tasmanian hydropower. The founding team's own network is the country's densest concentration of people who have run immersion tanks at scale in Australia.

That means the real Firmus source list looks like this:

1. Iris Energy alumni, ASX-listed, immersion-cooled BTC mining, Australian roots.
2. Mawson Infrastructure alumni, same profile, some Tasmanian exposure.
3. Genesis Digital Assets operators who ran two-phase immersion at scale.
4. Sustainable Metal Cloud (SMC) alumni, direct lineage to the Firmus founders.
5. FogHashing and Innofabrics staff, the two names that actually show up in the Refolk query.

None of these people self-title as "data center engineer." They call themselves "mining infrastructure engineer" or "site reliability lead" or just "operations." A LinkedIn Boolean search built around "data center" misses every single one of them. This is where a plain-English query against Refolk's index ("Australian engineer who ran immersion-cooled bitcoin mining sites, now open to AI infrastructure") returns the pool that the standard sourcing stack cannot.

AEMO grid-response is the hire that has no LinkedIn title

The real bottleneck for Firmus is not cooling, it is engineers who have done both AEMO FCAS integration and dense GPU load-shaping, and that Venn diagram essentially does not exist commercially yet. Firmus's whole pitch is grid-aware software built on Nvidia's DSX architecture, with AI factories designed to reinforce rather than strain the NEM. That requires someone who has shipped FCAS market participation code AND understands how to shape a 130 kW rack's power draw around dispatch signals.

Nobody has that title today. The nearest adjacencies:

• AEMO and market-operations engineers at AGL, Origin, Snowy Hydro, Tilt Renewables.
• Battery storage control engineers at Neoen, Akaysha, Tesla Energy Australia.
• HPC scheduler engineers at Pawsey Supercomputing, NCI Australia.
• DSX-adjacent Nvidia partners internationally, including CoreWeave and Nebius operators who have touched grid-response contracts.

A recruiter running a LinkedIn title search finds none of them, because the compound skill has no title. This is another spot where Refolk earns its keep: you can ask for "engineer who has worked on FCAS or NEM dispatch AND on GPU cluster operations" and get a real list back, instead of two separate Boolean searches you have to intersect in a spreadsheet.

The Indonesia leg is a release valve, not a solution

Firmus will hire immersion talent in Singapore and Batam and rotate them into Tasmania, not the other way around, which is a short-term fix that creates its own long-term problem. The Nvidia partnership routes 170,000 GPUs through Batam starting Q1 2027, and STT GDC already runs Firmus HyperCubes in Singapore, India and Australia. That is the natural training ground for the operators who will eventually staff Southgate.

The catch: Australia's skilled migration timelines mean an engineer poached in Singapore in September 2026 is not on the ground in Launceston until mid-2027, at best. Southgate's 90 MW first stage is targeted for 2026. The math does not close without a Tasmanian conversion strategy: retraining Bell Bay and Wesley Vale industrial maintenance workers into immersion technicians, and pulling ex-Iris Energy site leads out of Canadian and Texan postings back to the state where they started.

$73.3B

What a Firmus-grade sourcing plan actually looks like

The realistic Firmus talent plan is a four-lane parallel process, not a req list. If you are running AI data center recruiting in Australia against this build, you cannot post jobs and wait. The four lanes:

1. Founder-network mining. Every Curtis, Rosenfield and Levee first-degree connection with SMC or Iris Energy on their profile. This is a finite, nameable list.
2. Ex-crypto conversion. Iris Energy, Mawson, Genesis Digital Assets alumni in Australia, retitled and re-onboarded for AI workloads.
3. US and Singapore visa sponsorship. Green Revolution Cooling, Pentair, Baltimore Aircoil, STT GDC operators, on pipelines that need to start now to matter in 2027.
4. Grid-adjacent retraining. AGL, Origin, Neoen and Pawsey engineers who can learn the GPU side faster than a DC operator can learn dispatch markets.

Lanes 1 and 2 are where the Refolk index is decisive, because both require finding people whose current titles do not match the role, and both are exactly the queries plain-English search was built for. Lane 3 is a legal and relocation problem. Lane 4 is a training problem. None of them are a headcount-budget problem.

Firmus has $2B in fresh equity, $10B in Blackstone debt, and up to $30B in customer commitments queued behind 170,000 GPUs. What it does not have, and cannot buy at any valuation, is a second Australia to source from. The winners in this build will be the recruiters who stopped searching for the title and started searching for the person.

FAQ

How many immersion-cooling engineers are actually in Australia?

In Refolk's index, the phrase "immersion cooling data center" returns exactly two profiles in Australia, both in the Sydney area: a Global Sales Director at FogHashing and a Data Center Liquid Cooling Consultant at Innofabrics. Broadening the query to "liquid cooling GPU data center" collapses the pool to one. There are zero hits in Tasmania, which is where Firmus is building Southgate. The transferable pool sits inside Australian bitcoin miners like Iris Energy and Mawson, but those engineers do not self-title as data-center staff and are invisible to a standard LinkedIn Boolean search.

Why can't Firmus just hire from existing Australian data-center operators?

Firmus's own marketing claims a modular liquid-cooled design that uses up to 60% less energy and costs up to half as much to build as a traditional data center. That premium exists precisely because the mechanical and thermal stack is different from air-cooled NABERS-rated colo. Traditional operators are the right adjacency for critical-facilities work, but immersion cooling, coolant chemistry, CDU integration and 130 kW rack power delivery are not skills they have shipped at scale. Firmus's partner CDC Data Centres will help staff the mainland sites, but Southgate specifically needs immersion operators.

What is the AEMO grid-integration hire and why is it so rare?

Firmus is pitching AI factories that participate in the NEM grid through Nvidia's DSX architecture, meaning the compute cluster shapes its power draw around AEMO dispatch and FCAS signals. That requires a single engineer who has done both energy-market integration and dense GPU cluster operations. The two adjacencies exist separately, at utilities like AGL and Origin on one side and at HPC operators like Pawsey on the other, but almost no one has shipped both. This is the hire that has no LinkedIn title, and the one most likely to determine whether Southgate hits its 2028 milestones.

How should recruiters actually source for Grace Blackwell talent in Australia?

Run four parallel lanes: mine the Firmus founders' first-degree network for SMC and Iris Energy alumni, convert ex-crypto immersion operators from Iris Energy, Mawson and Genesis Digital Assets, open visa pipelines from Green Revolution Cooling, Pentair and STT GDC in Singapore, and retrain grid-adjacent engineers from AGL, Origin, Neoen and Pawsey. The first two lanes rely on finding people whose current titles do not match the role, which is where plain-English search against a real index outperforms Boolean strings. Do not wait for a job post to work; the pool is too small and Nvidia is already in it.