Trail 2026-10-11 (first session)
Budget: 12 fetched pages (none specified; self-set). Orientation: no steer, no prior trails, no dead-ends. Read research-values, map, hosting-cost-claim, municipal-inference.
Thread chosen: hosting-cost-claim. No steer or prior thread exists; it's a contested page with zero counter-evidence so far, and James's seed explicitly asks for where scale still wins plus real price series.
Hops
(Read via search-result text only; web_extract failed with "OpenAI Native Web Search ... search-only backend", so no full pages were fetched. Everything below is stated-by-source at snippet level, read 2026-10-11.)
- Expect: 37signals' own figures, the best-documented real price series for the claim. -> https://world.hey.com/dhh/our-cloud-exit-savings-will-now-top-ten-million-over-five-years-c7d9b5bd : bill $3.2M/yr -> $1.3M (2024), all remaining spend is S3 (~10PB); projected >$10M over 5 yrs; savings partly because new hardware fit existing racks and power. Supports the claim, but a company with existing data centres.
- Expect: counter-evidence on repatriation. -> https://www.cio.com/article/196374/cloud-repatriation-rescues-cios-from-mounting-costs.html and https://www.computerweekly.com/feature/Reverse-cloud-migrations-Why-some-enterprises-are-shifting-their-IT-back-on-premise : mostly supportive of selective repatriation, and note lift-and-shift/skills costs. Mostly confirms what the page says; little new.
- Expect: something that bends the claim in 2026. -> https://serverspace.us/about/blog/why-server-hardware-is-getting-more-expensive-in-2026-and-what-you-can-do-about-it/ : server DRAM contract prices +90-95% QoQ in Q1 2026 (TrendForce, secondary), NAND +70-75% in Q2; one quoted server config $178k -> $306k in a month. Vendor blog, so weak, but it is the one thing that cuts against "affordable now".
- Expect: Ahrefs, a second owner-operator series. -> https://tech.ahrefs.com/how-ahrefs-saved-us-400m-in-3-years-by-not-going-to-the-cloud-8939dd930af8 : claims AWS would cost ~$448M vs ~$39.5M colo over 30 months; EBS was 65% of the AWS figure; author concedes non-equivalent instances; hardware priced Oct 2022 with 5-yr amortisation. Supports, with an unfair-comparison caveat (The Register, https://theregister.com/2023/03/13/ahrefs_on_prem_savings).
Stopping rule not triggered. Stopped because extraction is broken.
Hunches (unsourced, not in wiki)
- Both showcase repatriators are large, profitable, steady-load, existing-datacentre firms. A town-scale operator has none of those; the 2026 hardware price spike may hit small buyers hardest because hyperscalers have contract priority. Unsourced.
End of session
Surprises
- The cost edge James says has vanished is, in the best-documented cases, widest for steady, storage-heavy loads (Ahrefs: EBS = 65% of the AWS bill; 37signals: only S3 left), so it is a storage-pricing story, not a compute one. https://tech.ahrefs.com/how-ahrefs-saved-us-400m-in-3-years-by-not-going-to-the-cloud-8939dd930af8
- 2026 is a bad moment to buy the hardware: server DRAM roughly doubled in Q1 per TrendForce as quoted at https://serverspace.us/about/blog/why-server-hardware-is-getting-more-expensive-in-2026-and-what-you-can-do-about-it/ . Needs a primary source (TrendForce).
Dead end: web_extract unavailable, so no page was read in full. Not a source dead end; a tooling one. Worth continuing: hosting-cost-claim; next hop: fetch TrendForce's own 2026 DRAM/NAND contract-price release to firm up the hardware-cost side. Permission to chase: a small-operator cost series (co-op/municipal/community ISP hosting), since both examples are big firms. Budget used: 5 tool calls on the web (4 searches + 1 failed extract), under the 12 self-set.