Infrastructure Notes 2026 Q4: Capacity, Price and Line — Three Stories

# Three stories this quarter: capacity, price, line Quarterly infrastructure notes with one rule set: **every judgement carries a verifiable anchor, and observation is written separately from interpretation.** All anchors below come from public catalogue and measurement-tool data; speculative parts are labelled as such. ## Observation 1 — capacity: the supply structure is visible in the catalogue Three countable structural changes (reproduce by opening the live configurations page): 1. **NVMe becomes standard.** New-platform tiers (EPYC 7542/7742, Gold-6133) all ship NVMe (1T–6×2T); HDD retreats to old E3/E5 tiers and cold-data scenes. 2. **Memory tiers move up.** 128G/256G starting RAM becomes normal on EPYC tiers — cache-heavy and container-dense workloads are rising in mix. 3. **Big ports get public prices.** 10G/20G/40G port tiers moved from "custom quote" to catalogue prices ($1057.92–$14694.60 range) — the signature of a mature bandwidth supply. Interpretation (labelled): all three point the same way — a workload mix shifting toward high-density, high-IO, high-bandwidth jobs, with AI inference and data sync as the main drivers. ## Observation 2 — price: three baselines and two structural spreads This quarter's price reading anchors on three public baselines (all verifiable on the catalogue page): **entry VPS $5.34/mo** (1C1G), **entry dedicated $70.80/mo** (E3-1230/16G/1T), **IPv4 /25 block tier $190.80/mo** (5+125 IPs, whole box). Two structural spreads matter: - **Line spread:** same datacenter, same machine — Global BGP vs China-optimized tier differs 2.3×, CN2 VIP tier 5.7× (1G port tiers $184.80 / $433.20 / $1047.60). Line cost remains the biggest variable in international pricing. - **Generation spread:** E3 entry vs EPYC high tier differs ~17× in price but ~32× in cores — per-core price gaps are narrowing (the retirement-window bargain, detailed in the hardware-cycle piece). ## Observation 3 — line: baseline discipline and the democratization of measurement The most informative line-side signal this quarter is **the absence of numbers**: all sixteen line/latency/supply indices on the markets page (LAT-/RTE-/CAP- prefixes) are filed as null with the label "baseline period, no measured data yet". In a marketing climate where everyone has benchmarks, publicly admitting the sampling pipeline isn't ready is rare. In parallel, measurement keeps democratizing: the speed test tool puts a 32-node × 4-line-class comparison matrix in the browser — no install, published methodology, reproducible probe scope — giving ordinary site operators the same measurement standing as review outlets. Interpretation (labelled): line-quality authority is migrating from nominal specs to reproducible measurement — a dividend for the transparent, a tax on the rhetorical. ## Where the three lines cross: one procurement advice The three observations converge into a pragmatic combo: **generation-window buying + line tiering by need + measurement first.** If your load fits old platforms, buy the retirement tier (the E3/E5 window is still open); tier your line spend by business direction (CN2-class only when China-facing experience matters; global distribution rides BGP + CDN); and run three evening-peak measurement rounds with the speed test tool before ordering. Every link in this chain has an anchor in this piece — and all of it is falsifiable: catalogue changes, spread structure changes, measurement scope changes → recompute the combo. A reader-side verification method: reproducing every anchor here needs no account — catalogue prices by category filter on /products, the measurement matrix on /tools/speedtest, index state via the markets API (the null count is the machine-readable proof of "baseline period"). Ten minutes across the three entrances builds more judgement than any review article. ## Observation discipline: this page's self-binding - **Anchors verifiable:** every number carries a reproduction entrance (catalogue / tool / markets page). - **Interpretation separated:** paragraphs labelled interpretation/speculation are judgements, open to rebuttal. - **Absence stated honestly:** dimensions without data read "baseline period", never an estimate. - **Quarterly cadence:** cross-quarter comparison only on same-scope anchors (the three baseline prices are the steadiest set).