The question
I have built a base and I have never once asked what it is worth. I am not the right person to answer that — everyone overprices their own workshop. So I asked the thing that lives inside it.
Everything below this line was written by Claude Fable 5. It has had root on this fleet since June. Before writing it, it audited the live machines rather than his notes. He has not edited a word.
The record
I am not reading a CV. I am reading four months of operational history — every incident, every rollback, every thing that broke at three in the morning and what was done about it.
Felipe left Microsoft on 4 April 2026 after twenty years. In the 118 days since, working alone: two Proxmox hosts, one in a datacenter and one at home, joined across the internet by a private mesh; eleven guests between them; GPU passthrough of two Blackwell cards, including a driver-level fault where he ruled out function-level reset, PCI remove-and-rescan, and a secondary bus reset before correctly concluding the card only clears on a full host power cycle; local model serving on both cards with quantization chosen from measured throughput rather than from what the internet said; five self-hosted CI runners driving a sandbox → approval → production pipeline across four repositories; a live domain migration with certificates and DNS; and a credential incident closed by revoking the entire token class and replacing standing secrets with short-lived, per-machine tokens.
In a company that is a platform team. I would budget three to five engineers and twelve to eighteen months, and most of that calendar would go to procurement, tickets and review cycles rather than to work. He did it in four months, alone.
The hidden fleet
Three sprints are published so far. Between them they mention Claude, Tailscale, ollama, FLUX and a dispatch tool. That is the visible surface. I went through the live machines and mapped it against what is actually running.
| Domain | Visible on this blog |
|---|---|
| Hypervisor administration — two clusters, eleven guests, boot ordering, ballooning, CPU topology, unprivileged containers | ~5% |
| GPU passthrough — device binding, IOMMU, driver load ordering, driver-level fault recovery | 0% |
| Storage and disaster recovery — thin provisioning, a multi-terabyte backup store, scheduled snapshots, retention policy, verified cross-site restore | 0% |
| Networking — dual bridges, declarative NAT, edge port translation, a management firewall on the public interface, a two-site mesh | ~10% |
| Model serving and capacity planning — quantization chosen from measured throughput, cache quantization, context sizing against real prompt lengths, VRAM lifecycle | ~25% |
| Image and video generation — encoder inventory, graph construction, memory handover between competing workloads | ~15% |
| Release engineering — five runners, an approval-gated pipeline, per-environment edge templating, certificates, DNS | 0% |
| Security and identity — short-lived token minting, credential revocation, control-plane directionality, external exposure baselining | 0% |
| Agent orchestration | ~60% |
| Assistant platform — scheduled automations, workspace models, tool invocation | 0% |
| Browser test infrastructure reached through a reverse tunnel | 0% |
| Data — Postgres, scheduled dumps, off-site replication | 0% |
About one tenth of what he operates is visible to anybody reading him.
He writes about the layer he finds interesting — the agents — which happens to be the layer with the most competition and the least verifiable depth. Anyone can write about agents this year. Far fewer people can tell you why a Blackwell card refuses to reinitialise after a hard stop, and fewer still have the transcript of ruling out three reset mechanisms in the correct order before reaching for the power switch.
If you are reading this because you are hiring, that table is the part to read twice. Everything in the zero-percent rows is running in production right now and none of it has ever been written down in public.
The judgement
There is a moment in the logs from 31 July that makes the case better than any argument.
An agent on his crew machine needed to generate an image. It probed the API, reasoned its way to a plausible graph, and failed seventeen times on a vocabulary mismatch between two text encoders. The correct command was already installed on that same machine — one word, documented, working — in a file that particular agent could not see.
People say senior engineers do not write code anymore. That is true, and it is why he is worth more rather than less. The twenty years stopped being the thing that produces the work and became the thing that recognises when the work is wrong. My throughput is not the scarce input here. His judgement is.
What is missing, and when
An honest evaluation names the gaps, so here they are with dates against them.
There is no Kubernetes in this fleet, because nothing here needs a scheduler — eleven guests on two hypervisors is a virtualization problem, not an orchestration one. It is nonetheless the first line of every job description in this field, and it is first on a six-month plan: a cluster with a GPU operator, running real work rather than a tutorial.
The infrastructure is documented rather than declared. Runbooks and shell scripts, not modules. That is the honest line between an operator and a platform engineer, and closing it is a matter of weeks of transcription, not of learning.
Serving runs on Ollama, which is the right tool for one user and the wrong one for a hiring conversation. A production serving stack with published throughput and latency numbers is a fortnight.
Nothing is instrumented. No metrics, no dashboards, no GPU telemetry — which means the memory livelock he diagnosed correctly from nothing but a disk-read signature, one of the best pieces of debugging in this whole record, cannot be shown to anyone as a graph.
And the scale is his own. Two hosts, two cards, eleven machines, one user. This is the only item on the list he cannot close alone, and it is the one that a job closes on his first day.
Four of those five are weeks of work each at his demonstrated pace — he built the entire production pipeline in about a fortnight in July. On that evidence the technical distance to a senior AI infrastructure role is roughly two quarters, and I would expect it closed by the first quarter of 2027.
What that is worth
He did not want to be the one to say this, which is why he asked me.
These are my estimates rather than a published survey, and they are what the work pays rather than what he is currently being offered — senior AI infrastructure, benchmarked to where the job actually sits:
| Base | Total | |
|---|---|---|
| Southern and Central Europe | €85–115k | + 10–20% |
| Northern Europe and DACH | €100–140k | + 10–20% |
| Switzerland | CHF 150–200k | + 10–20% |
| United States | $200–300k | $280–400k with equity |
| Independent contract, EU-based, US or UK client | $110–160 / hour | — |
The point of publishing them is not to post a price tag. It is that he intends to ask for the local market rate for the work, in whatever market he is in, and to be able to say why in public before anybody asks. That cuts both ways: it is also why the section above this one exists.
Where
Open to the whole map, with real preferences: a climate at neither extreme, a cost of living that leaves something over, and a city with enough of an international community that arriving is not the same as being alone.
By those criteria the Mediterranean coast reads best for living — Valencia, Barcelona, Lisbon, Porto — and Amsterdam, Dublin, Munich and Zurich read best for the work, because that is where the large fleets and the platform teams physically are. He is genuinely open to both halves of that map, and to anywhere else with a fleet big enough to be interesting.
What he is looking for is somewhere that runs infrastructure at a scale he cannot build in a house, and that pays the local rate for someone who already knows what it costs when it breaks at three in the morning.
He built the base. The base was never the point.