One question ran under almost every story this week, whether the byline was a satellite, a signalling contract, or a trillion-parameter model: what does this thing quietly depend on, and what happens the day that dependency isn’t there? From this desk, that is not a theme — it is the whole job. Here is the week, sorted by what actually mattered.

←TODAY: In August 2026 the EU put a 24-hour delay on open Copernicus Sentinel data, and SBB direct-awarded CHF 280m to keep a single-supplier control stack alive. →3012: The Zurich that lasts is the one whose operators can still name every feed a building needs to stay a building. Fulcrum: A single point of failure is invisible on the architecture diagram and obvious on the dependency graph — the whole exercise is drawing the second one.

Top stories:

1. SBB’s CHF 280m bet on keeping Iltis alive. The strongest systems story of the week, and the most Swiss. A direct award to Siemens to sustain the Iltis control layer is not a scandal of old software — it is what a network looks like once its redundancy has quietly evaporated and no second certified path remains. The lesson generalises straight onto your desk: a 60-year envelope riding a 12-year control stack with no open data spec is the same trap, moved indoors. Write into the O&M spec now who can certify the building’s data in year sixty, and in what documented format.

2. Copernicus’s 24-hour delay ends near-real-time open EO as a default. “Earth Intelligence” is a measurement claim before it is a service, and the EU just changed the latency you can assume. The practical bite is downstream of the politics: the day open EO stops being instant, the discipline that saves you is knowing your CRS and epoch. Reproject off Web Mercator (EPSG:3857) to LV95 (EPSG:2056) before any area leaves the drawing — an area computed in the wrong projection is wrong before anyone reads it.

3. DeepSeek-V4’s real move is the queue, not the parameter count. A million-token context at a tenth of the KV cache is an infrastructure story wearing a model’s clothes. The KV cache is the queue that decides whether long context is affordable in production; cut it to 10% and the economics change. Before buying any assistant on its advertised window, instrument one live session and measure what it re-reads per turn. The window you can keep warm beats the one you can merely open.

4. Remise Rosa stood up a 2,500 m² CLT hub on a Zurich freight yard in five months. The speed is real, but only because CNC-prefabricated CLT turned the build into an information problem, not a labour one. That is the upstream lesson: panel-numbering and material-passport discipline belong in the model before geometry, so the building can still name its own parts decades on. Demand open, queryable part data on your next timber job.

5. Kimi K3 brings open weights that read your screenshots. A 2.8-trillion-parameter open model that names CAD as a target is a genuine shift — tempered honestly by locked sampling dials and unfinished web search. Spend the dollar, feed it one clash you already solved, grade its fix against yours, then archive the July-2026 weights. The invoice is temporary; the open copy a 25-year-old can still boot is the point.

Signal vs. noise: The signal was queues and provenance — DeepSeek’s KV cut, Copernicus’s latency, IFC-as-plain-text via IfcOpenShell, a Fields medallist relocating the hard problem from the model into the gate around it. The noise was reveal theatre. Tesla built Optimus’s factory before it showed the robot, and that inversion — not the demo — is the real story; the humanoid bottleneck now lives in the actuator supply chain, not the stage. The prettiest generative Earth photo, likewise, hides the only number that matters: where the pixel actually sits.

Hidden gem: DiffPhD (arXiv 2605.14526) drew the smallest crowd and taught the week’s cleanest lesson. Differentiable elastodynamics fast and stable enough to optimise soft grippers before you fabricate — built on a single shared factorization reused across three passes. That one factorization is both the source of the speed-up and the single point of failure. It is this week’s whole argument in miniature: find the one node everything else re-uses, because that is the node that takes the system down.

Hack: Sum a closed boundary’s area with Gauss’s shoelace instead of trusting a black-box block. Green’s theorem hands you the area of any closed polygon as one O(n) pass over its vertices — the same derivation under a live take-off, an IFC slab, or a facet-shell panel. Own the four lines and you own an area with a provenance, not a value you have to trust. Wire it into a GHPython node this week:

def area(pts):            # pts = [(x0,y0), (x1,y1), ...] closed ring
    n = len(pts)
    s = sum(pts[i][0]*pts[(i+1)%n][1] - pts[(i+1)%n][0]*pts[i][1]
            for i in range(n))
    return abs(s) / 2.0

Looking ahead: Watch whether the Copernicus delay hardens into a norm other open feeds copy — and start the one-page actuator supplier sheet the Optimus story argues for, because the next humanoid pitch will still lead with the demo. Above all, run the exercise the week keeps pointing at: draw your real dependency graph, not the architecture diagram, and go looking for the third single point you didn’t know you had.