Executive brief
The engineering site at Échirolles, on the southern edge of the Grenoble metropole, has been renamed repeatedly and, on 31 March 2026, changed owner: the French State became sole shareholder of Bull, the supercomputing business, while cybersecurity products stayed with Eviden inside Atos Group. For anyone building a career there, the practical question is not what the employer is called but which skills survive the renaming. This companion answers it from primary evidence only: three vacancies posted on the group's own job board between 31 August and 3 September 2026, and the published technical specification of Alice Recoque, which names the exact software and interconnect stack the site's engineers will operate. No salary is published in any of them, so this dossier prints no pay figure at all. Nine claim families are refused at the end.
I. The skills question, restated correctly
Careers advice about a site like this usually starts from the employer. That is the wrong end. The record shows the name changing on 29 January 2026, when the historic Bull brand was relaunched, and the shareholder changing on 31 March 2026, when Atos Group completed the sale of the Advanced Computing activities to the French State for an enterprise value of up to €404 million including €104 million of contingent earn-outs. Across both events, the same building kept doing the same three things: designing high-performance computing systems, building the software that models and prices them, and developing identity-and-access products.
So the durable unit of analysis is the technical stack, not the letterhead. And unusually for a European industrial employer, that stack is published in detail, because the customer for the flagship programme is a public joint undertaking that documents what it bought.
The stack the machine actually names: a hiring inventory, not an adjective
| Layer | What the published documents name | Publisher |
|---|---|---|
| Operating and scheduling | The CEA’s Ocean suite plus SLURM, Kubernetes, LUSTRE, Grafana and Prometheus | EuroHPC JU |
| Interconnect | BXI v3 at 400 Gb/s for CPU and 800 Gb/s for GPU endpoints | EuroHPC JU |
| Storage | Provided by DDN | Bull |
| Scalar partition | A dedicated partition based on SiPearl Rhea2 processors | CEA |
| European content | Nearly three-quarters of BullSequana XH3500 component production relocated to Europe | Bull |
Sources 5 Bull · 4 European High Performance Computing Joint Undertaking · 6 Commissariat à l'énergie atomique et aux énergies alternatives
II. The stack the machine actually names
On 18 November 2025 the EuroHPC Joint Undertaking published the architecture of Alice Recoque, contracted to Eviden, to be installed at the CEA's TGCC in 2026 and opened to users in 2027 according to the CEA's own HPC pages. The specification is, read carefully, a skills inventory:
| Published component | What it implies for an engineer |
|---|---|
| BullSequana XH3500 racks, warm-water direct liquid cooling; chilled doors on scalar racks | Thermal and mechanical integration; datacentre hydraulics as a first-class engineering discipline |
| AMD EPYC "Venice" 256-core CPUs; AMD Instinct MI430X GPUs, coherent memory | Heterogeneous programming, memory-coherent porting, GPU performance work |
| SiPearl Rhea2 Arm processors (scalar partition) | Arm toolchains and cross-architecture validation, not x86-only habits |
| BXI v3 interconnect: 800 Gb/s GPU, 400 Gb/s CPU endpoints | Low-level network software: fabrics, drivers, collectives, congestion |
| CEA Ocean suite plus SLURM, Kubernetes, Lustre, Grafana, Prometheus | Scheduling, container orchestration, parallel filesystems, observability |
| DDN storage; data-centric storage architecture | I/O engineering and storage tiering at petabyte scale |
Two features of this list are worth naming. First, it is half open-source: SLURM, Kubernetes, Lustre, Grafana and Prometheus are skills a candidate can acquire and evidence outside any employer, which makes them portable across the renaming. Second, the proprietary half (the BXI interconnect, the XH3500 platform, the Ocean operating suite) is European and, since 31 March 2026, state-owned. The vendor states that nearly three-quarters of XH3500 component production has been relocated to Europe. Skills tied to those components have an unusual property: their demand is set by industrial policy rather than by a single company's order book.
Two published costs for one machine, and they are different objects
| What is counted | Value as published | Date | Publisher |
|---|---|---|---|
| Total budget for acquisition, delivery, installation and maintenance | EUR 354,800,000, funded 50% by the EuroHPC JU and 50% by France, the Netherlands and Greece | 18 November 2025 | EuroHPC JU |
| Overall project cost over five years of operation | €554 million | 18 November 2025 | Bull |
| Stated performance | Capacity to surpass one exaflop per second, a fiftyfold increase over the previous system with five times the electrical power | 18 November 2025 | Bull |
Sources 4 European High Performance Computing Joint Undertaking · 5 Bull · 6 Commissariat à l'énergie atomique et aux énergies alternatives
III. What the site was actually advertising in late 2026
We treat individual vacancies as what they are: dated, individual postings, never a market rate or a hiring plan. Three of them, read together, describe the site's real content.
A full-stack developer, published 31 August 2026, reference 548023, location Échirolles (FR). The posting is unambiguous about geography, "Sites HPC Bull en France : Échirolles (38, près de Grenoble) ou Les Clayes-sous-Bois (78)", and about working pattern: hybrid, with what it calls intentional office days for team collaboration. The team is Solution, Design & Architecture (SDA), tooling division. The product is CQFD, described as having begun as a clever prototype and become the critical internal application used across pre-sales, solution review, factory and delivery to model, configure and price the company's supercomputers. The posting states the role uses modern AI-assisted development methods, naming Claude, Gemini and Copilot as tools, for rapid innovation, agile prototyping and high-quality delivery, and tells the candidate they will shape "the tool that dictates how multi-million-euro sovereign technology contracts are concluded."
That single posting is more informative than a trend report. It says that at a European supercomputer manufacturer in 2026, an internal configure-price-quote application is treated as critical infrastructure, that its development is expected to be assisted by commercial AI coding tools, and that the skill being bought is domain modelling of HPC configurations, not framework fluency alone.
A senior C/Linux developer, published 3 September 2026, reference 543669, location Échirolles (FR). This one carries the Eviden boilerplate and sits in the Cybersecurity Products division, on Evidian IAM: Web Access Management and identity federation, delivered both on-premises and as software-as-a-service, with the R&D team described as based in Grenoble. The stated profile is Bac+5 engineering or equivalent, with a minimum of ten years of full-stack software development experience including architecture design, reliability and performance, plus strong C, TCP/IP, HTTP and DNS networking, encryption knowledge, and experience writing architecture and API specifications and designing highly scalable web services. The role includes coaching the team.
A network software engineer, R&D HPC. When we consulted it on 4 September 2026, the job board returned "Sorry, this position has been filled." We report that as it stands: it is evidence that low-level HPC networking roles exist at this site and close, and it is not evidence of anything about time-to-fill, which no publisher states.
IV. The pattern: three job families, one address
Put the postings beside the machine specification and the site resolves into three durable families.
1. Systems and low-level engineering. C on Linux, networking to the level of TCP/IP internals and fabric behaviour, performance and reliability. The Evidian posting demands ten years of it for identity software; the closed vacancy demanded it for HPC interconnect R&D. This is the family least affected by ownership: the BXI v3 endpoints at 800 and 400 Gb/s need the same competence whether the employer's shares sit with a listed group or the State.
2. Platform and operations engineering. SLURM, Kubernetes, Lustre, Grafana, Prometheus: named by EuroHPC as the operating stack of a machine that must serve scientific users, AI factories, startups and SMEs. Engineers who can run multi-tenant scheduling on shared hardware, and observe it honestly, are operating the layer where an exascale system either delivers or embarrasses its owners.
3. Industrialised internal software. The CQFD posting is this family's proof: business-critical internal tools built by a dedicated product team, with AI-assisted development as a declared method rather than a rumour. The skill is translating a supercomputer's configuration rules into software that pre-sales, factory and delivery all trust.
A fourth capability sits across all three and is rarely written into a job description: reading a public procurement document as a technical brief. The EuroHPC release, the CEA's TGCC pages and the vendor release together tell a candidate more about the next three years of work at this site than any recruiter conversation: including that the host site had to reinforce a slab to 2,800 kg/m² over 700 m², add 24 MW of electrical capacity and build a 20 MW warm-water cooling loop, per the CEA. Candidates who arrive already fluent in that document set interview differently.
≈ €720 million
the 2025 revenue of the divested Advanced Computing perimeter: stated as c. €720 million by the company and as c. €0.7 billion by the seller: whose capital passed to the French State on 31 March 2026
Enterprise value was published as up to €404 million including €104 million of contingent earn-outs, after a perimeter adjustment that excluded zData and revised both figures down. Revenue is a perimeter measurement, not a wage or a headcount, and nothing about pay is inferred from it.
Sources 1 Atos Group · 2 Bull
V. Employer stability, stated rather than assumed
The published facts on the employer's stability are these, and no more. The completion release of 31 March 2026 states that the French State is now sole shareholder, that the transaction "has been designed to ensure full continuity of service for Bull's clients and employees", and that the acquisition provides "a long-term public shareholder, ensuring stability, strategic continuity and the preservation of critical expertise". It states that Bull brings together more than 3,000 professionals and experts, half of them in France, and generated revenue of about €720 million in 2025. It states that the company operates the only supercomputer manufacturing plant in Europe, at Angers. The January 2026 brand launch had described "over 2,500 engineers and experts" and 1,500 patents; an August 2026 vacancy claims more than 1,600 patents. We print the three figures with their dates and reconcile none of them.
What none of those documents states is how many people work at Échirolles, how the staff split between the two employers now sharing the address, or what anyone is paid. A candidate deserves to know that those numbers are absent rather than to be handed an estimate dressed as research.
VI. How to enter, concretely
The evidence supports four moves and no more. Build the portable half of the stack first (SLURM job scheduling, Kubernetes, a parallel filesystem, Prometheus and Grafana) because it is named in the machine's own specification and can be evidenced without an employer. Keep C and Linux at professional depth: both site postings that state a language state C, and the senior one asks for ten years and Bac+5. Learn one Arm toolchain alongside x86, because the scalar partition is SiPearl Rhea2. And treat AI-assisted development as an interview subject rather than a secret: the site's own 31 August 2026 posting names Claude, Gemini and Copilot as working method, which means a candidate can discuss tooling discipline (review, testing, provenance of generated code) as a competence rather than a confession.
VII. What we refuse to claim
No salary figure, band, median or trend for this site: none of the three postings we cite publishes pay, and inventing one would be the single most damaging thing we could do to a reader's negotiation. No Échirolles headcount, and no split of staff between Bull and Eviden. No hiring target, vacancy count or growth rate for either employer. No time-to-fill inference from the filled network-engineering vacancy. No claim that state ownership guarantees job security. No skills-shortage percentage for French HPC: we found no publisher of one and will not model it. No transfer of the nuclear-style sector workforce studies onto computing. No assertion that AI-assisted development reduces or increases engineering headcount at this site. And no forecast of Alice Recoque's operational staffing, which no publisher has stated.