Trail 2026-10-14 (sixth session)

Budget: 15 fetched pages or 40 tool calls. Orientation: steer, index, research-values, dead-ends, trail-2026-10-11-c, trail-2026-10-11-d, trail-2026-10-12.

Thread chosen: hosting-cost-claim in its surplus-hardware form (within thread-convivial-computation), steer item 2. Item 1 (resources) has had three sessions in a row and the last "worth continuing" (ENAMI) is close to it; item 2 has not been touched since James reframed it. I will weight the discovery part of it, "who actually runs on surplus", over price-firming.

Hops

Tooling: web_extract still broken (search-only backend); searches come through the free-tier fallback; no PDF tools on the host. Each hop says full read or search text.

1. Expect: ITRenew/Sesame, the firm that resells decommissioned hyperscaler racks, to find who buys surplus. -> empty result. Retried with plainer words (hop 3).
2. Expect: an unfamiliar "who runs on surplus" case outside the cloud market: Havana's SNET, a resident-built network said to run on improvised kit. Worth it because it crosses thread 1 and the farmer-telephone precedent. -> AP, AFP, WLRN and a CiberCuba reprint (search text): ~40,000 users, rooftop antennas and cables between houses, nine servers, no politics allowed; Resolutions 98/99 (July 2019) legalised private networks but capped wireless at 100 mW and banned cables over streets, which made SNET illegal by its form; absorbed into the state Joven Club; from Feb 2021 the state charged 20 CUP/month to ~28,000 users. Surprise.
3. Expect: the measured efficiency of older servers, which decides whether surplus is frugal or a power sink. -> Uptime Institute (search text): EURECA, 300 European data centres, servers >5 years = 40% of fleet, 66% of IT energy, 7% of compute. Points against very old kit.
4. Expect: who resells hyperscaler surplus. -> ITRenew snippets: certified racks resold at ~50% of new; Circle B runs bare metal on recycled hyperscale hardware in Amsterdam; Iron Mountain bought ITRenew. Titles and one-liners only; kept in the trail, not the wiki.
5. Expect: Bashroush et al., the primary study of refurbished vs new. -> read in full (computer.org): refurbished = new on SERT (13.7 vs 13.8); per-generation gains shrinking (~16%); refurbished replacing >6-year-old servers pays back in ~1 year; replacing <5-year-old servers pays back poorly either way. A window, not a blanket yes.
6. Expect: a live market where ordinary people rent surplus. -> Hetzner server auction, read in the browser: 164 refurbished servers (i7-6700/7700, Ryzen 5 3600), now sold under "the smart way to start self-hosting ... Dare to own your stack". Prices did not render; the JSON feed was blocked by the cookie wall (hop 7). The incumbent itself now sells the self-hosting story on surplus.
7. Expect: Hetzner's price feed. -> cookie wall. Nothing.
8. Expect: the IMC 2017 SNET measurement paper in full. -> search gave the abstract (largest isolated community network; tens of thousands of households; .snet domains). Download worked but there are no PDF tools on the host; a stdlib parse timed out at 300s. Abstract only.
9. Expect: how SNET was organised internally. -> pointed to Dye et al.
10. Expect: Dye's ethnography of SNET's power structure. -> ar5iv, read in full in the browser (findings and methods). Surprise: physical chokepoints make brokers; on the sparse outskirts brokers extract more rent because rewiring around them is costly, and yet are more accountable because they are known. Distributed infrastructure embeds the inequality of its setting.

Stopped at ~12 fetches and ~30 tool calls to write up.

Hunches (unsourced, not in wiki)

• The three absorption cases now in the wiki (farmer mutuals, ENAMI, SNET) were all ended through a technical or administrative parameter (upgrade standards, plant sales, a 100 mW cap and a cable rule) rather than a ban on ownership. If that is the general pattern, the defence of convivial infrastructure is in standards-setting, which is where James's platform background is strongest.
• Dye's rent-at-the-chokepoint finding and Ostrom's design principles might meet: "known identity" as the check on brokers is close to Ostrom's monitoring by members. Not checked.
• Surplus hardware has an age window (roughly 3-6 years old) set by electricity price; at California prices (~3x the study's $0.10/kWh) the window would close sooner. My inference, unsourced.

End of session

Surprises

• Havana residents built a city-wide network for ~40,000 people out of rooftop antennas and cables strung between houses, and the state ended it in 2019 not by banning it but by legalising private networks with a 100 mW power cap and a rule against cables crossing streets; the state then absorbed it and began charging. https://latinorebels.com/2019/08/19/cubavideogamers
• The ethnography of that network finds community ownership does not remove chokepoint power: whoever sits at the physical bottleneck can extract rent, more so where the network is sparse, checked mainly by everyone knowing who they are. https://ar5iv.labs.arxiv.org/html/2004.12207
• A refurbished server measured as efficient as a new one of the same model, and was the fastest-paying replacement for servers over six years old; but servers over five years old were 66% of the energy and 7% of the compute in 300 European data centres. https://www.computer.org/csdl/journal/su/2022/01/09246737/1oqGe9mvgcg

Dead end: the full IMC 2017 SNET measurement paper: no PDF extraction on the host and web_extract is search-only. Retry with a working extract backend. Worth continuing: havana-snet. Next hop: Resolutions 98 and 99 of MINCOM (2019) themselves, to see exactly which technical parameters were chosen and whether any community network was licensed under them. Permission to chase: a deliberate search for other community networks ended by technical standards rather than bans (guifi.net in Catalonia, Freifunk in Germany, rural WISPs under spectrum rules), as a test of the "absorbed through a parameter" hunch. Budget used: ~12 fetches/searches (of 15); ~32 tool calls (of 40).