Spokes.wiki Search About

Synthesis

The evolving thesis. Current best understanding of this wiki's topic, updated on every ingest.

Current thesis

This wiki tracks cloud-hosting providers. The corpus now spans three models of how projects get hosted — raw IaaS, managed PaaS, and edge compute — anchored at the top by a hyperscaler and organized along a pricing/acquisition-model spectrum:

The hyperscaler anchor. aws is the high/complex end the whole spectrum implicitly defines against: 200+ à-la-carte services you assemble and separately meter. Its revamped 2025 free tier is a credit funnel ($200 credits, 6-month auto-close, 30+ always-free services) — the same "free has a shape" mechanic as oracle-cloud-free but time-boxed rather than permanent. Most of the Tier-2 PaaS pitches ("no idle markup," git-push) are really pitches against assembling raw AWS yourself, so AWS is the baseline that explains why managed PaaS exists.

Tier 1 — raw IaaS (you rent the box):

Tier 2 — managed developer PaaS (you push code, they run it):

Tier 3 — edge compute (no instance to size): cloudflare is a third model entirely — Workers/Pages/R2/Workers AI, global and per-request, with R2's no egress fees the sharp foil to fly-io's metered egress. Workers AI also directly addresses the "where do AI stacks run?" question. The official cloudflare-workers docs (T1) show this tier is no longer just stateless functions at the edge: it carries a full data plane — Durable Objects, D1 (SQL), KV, R2, Vectorize — so an entire stateful app can run end-to-end on the edge, not just request handlers in front of an origin.

A new shape — stateful suspendable sandboxes. aws-lambda-microvms adds a point the spectrum didn't have: Firecracker-based isolated VM sandboxes with a serverless model and state-preserving suspend/resume (up to 8h; suspended = no charge). It's render's scale-to-zero without losing the running state, at VM isolation, billed per-second — "warm scale-to-zero." Notably it's pitched first at agent sandboxes (running LLM-generated code), so it doubles as a partial answer to the open question below (where AI-agent stacks run): the hyperscaler is now selling the agent-execution substrate as a primitive (cross-spoke seam to ../agentic-tooling-wiki's harness sandboxes / Signadot ephemeral envs).

So the real axis is a cost-model spectrum: hyperscaler à-la-carte (aws) → fixed monthly VPS (hetzner-cloud) → permanent free tier + PAYG (oracle-cloud-free) → IaaS/PaaS midpoint (digitalocean) → edge per-request (cloudflare) → low monthly floor + per-second usage (railway) → pure pay-as-you-go (fly-io/vercel) → free-with-spin-down (render). Tier 1 sells capacity; Tier 2 sells a deploy workflow (git push, zero-downtime deploys, preview envs) and bills actual usage; the edge tier sells global per-request execution with no machine to size. A recurring catch shows up across the free tiers — Oracle's idle-suspension, Render's spin-down + Postgres-expiry, Fly's metered egress vs Hetzner's bundled traffic — "free" always has a shape.

Sourcing caveat: these remain mostly vendor/official surfaces (pricing & docs pages) — good for what's offered. The first neutral third-party benchmark now lands too: vpsbenchmarks-docean-hetzner grades providers A–F on a percentile curve from a reproducible suite (web/sysbench/geekbench/fio/iperf3), so vendor claims can finally be checked against independent reliability/price-perf data. (An earlier Oracle signup-SPA stub was removed 2026-05-31; see log.md.)

What the hyperscaler actually sells

azure and google-cloud join aws, and with three examples the spoke can stop generalizing from one. The three descriptions converge on the same product, and it is not compute.

What a hyperscaler sells above a rented machine is an explicit failure-domain hierarchy: zone → region → geography, each defined, each contractual. Google states it most precisely — a zone "should be considered a single failure domain," a region is "three or more zones housed in three or more physical data centers." Microsoft sells the same shape plus curated region pairs for geo-replication that customers cannot choose. On top of that sits the piece that is genuinely hard to replicate: managed multi-region state (BigQuery, Cloud Storage, Firestore, Spanner) "designed to be able to function following the loss of a single region."

That reframes the spoke's central comparison. hetzner-cloud and the fixed-price VPS tier compete on compute per euro, and they are not selling against zones and residency boundaries at all — which is why the price gap can be as large as it is without either side being wrong. The hyperscaler's premium buys geography and managed replication; if a workload does not need them, the premium buys nothing.

One drift worth tracking: Azure's newer regions are not paired, using zones as the primary redundancy instead. Redundancy is moving down a level, from region-to-region to zone-to-zone.

Open questions

The purchase-model axis (2026-08-10). The spoke has been comparing tiers by architecture — raw IaaS, managed PaaS, edge. Azure's own price list adds an axis the other tiers don't have: the same machine costs $35.04, $21.63, $13.85 or $7.40 per vCPU-month depending purely on how you buy it (on demand, 1-year reserved, 3-year reserved, spot). A 2.5× spread with no technical difference, 5× with eviction risk. Two consequences. First, any comparison quoting "the hyperscaler price" without a commitment term is not comparing anything. Second, the crossover question above is under-specified: hetzner-cloud beats Azure on demand easily and beats 3-year reserved by much less, so the answer depends on a contract, not a workload. railway's metered $20.02/vCPU-month, re-verified the same day, sits between Azure's 1-year and 3-year reserved rates.

Growth edges

Ranked; each names the kind of source that would close it (see ../QUALITY.md → Growth edges).

  1. A neutral PaaS benchmark. The spoke's oldest gap and the one it has already hunted and failed: vpsbenchmarks-docean-hetzner covers raw VPS, nothing covers the managed tier (fly-io/render/railway). Searched 2026-06-15 — the space is entirely affiliate-monetized comparison content. Open by scarcity, not oversight. — needs: a methodology-disclosed cold-start / throughput / price suite for managed PaaS (T1/T2); do not close it with another affiliate roundup. Hunted 2026-09-16, a fourth time — nothing ≥ bar. Tried the academic angle (a methodology-disclosed PaaS study rather than a roundup): nothing peer-reviewed measures the managed tier; the new arrivals are bex.co (11 Sep 2026) and the same devopsboys/techsy/birjob set. Not eligible again before 2026-09-30. Hunted 2026-10-07, a fifth time — nothing ≥ bar. Cold-start benchmark papers specifically; still nothing peer-reviewed. openstatus's edge-latency run resurfaced (warm latency only, not a PaaS-under-load measurement) alongside newer bex.co cost teardowns. Five hunts, same conclusion: market structure, not a search gap. Not eligible again before 2026-11-07. The candidate list widened the same day, not from a hunt. google-cloud-run arrived via routing and names itself the same way fly-io/render/railway do — "no idle markup," true scale-to-zero — but from a hyperscaler, with no stated cold-start number of its own. Any future PaaS benchmark this edge finally closes with should include it: the question "can a hyperscaler out-PaaS the PaaS tier" is now askable with four, not three, products to measure.
  2. Where the Tier-1/Tier-2 crossover actually falls. The "no idle markup" promise has to stop paying off at some scale, and no source here names it. — needs: a worked cost comparison at stated load, or a first-party pricing model detailed enough to compute one.
  3. Real egress overage bills. Free-tier catches are catalogued; what egress actually costs at scale is the remaining nuance. — needs: a first-party billing doc (T3 acceptable — this is a vendor fact) or a published post-mortem with figures.
  4. A concrete agent-stack → provider mapping. The spoke knows which managed targets are common and not what runs where, and the question has moved on to where agent-written code is safely executed (aws-lambda-microvms vs E2B/Modal). — needs: a T2 survey or a system report naming its substrate.

Coverage edges (added 2026-08-08, at the curator's request for a wider backlog). These widen what the spoke covers instead of answering an open question above. Any of them closes with one ordinary solid source, and a vendor's own pricing page counts (T3 is allowed where the edge names a vendor fact).

  1. Two of the three hyperscalers are missing. CLOSED 2026-08-09 (research pass) — azure and google-cloud from their own region/reliability documentation (T3, as the edge allowed). Successor: the price half. Both pages describe failure-domain geography and neither carries a rate, so the spectrum's expensive end is still anchored on aws alone for cost. — needs: the two vendors' pricing pages for a comparable VM, or one three-way priced comparison.
  2. Silent changes to permanent free tiers. Converted from the hub's ## Outstanding backlog on 2026-08-16, where it had been carried since 2026-08-13 as a one-off waiting for a second instance. Oracle halved a tier it advertises as always free and published nothing that says so; the spoke holds the change only because a page was re-verified against the live site. A "permanent" free tier that moves without an announcement is a claim about vendor conduct, and this spoke has one data point for it — needs: a second documented instance (any provider), or an archived-page comparison showing a free-tier change against the vendor's own announcement record. Until then the spoke says it happened once, which is what it can support.
  3. The European mid-market beyond Hetzner. OVHcloud, Scaleway, Netcup and IONOS sell the same fixed-price VPS tier the spoke's thesis is built on, and none of them is written up. — needs: first-party pricing/spec pages, one page per provider.
  4. Object storage and the egress fight. The spoke prices compute and treats storage as a footnote, while the loudest price war of the last three years (S3 versus R2 versus B2, and who charges for egress) happens in storage. — needs: the three first-party pricing pages plus one worked cost comparison at a stated volume.
  5. GPU hours as a price tier. Inference has to run somewhere, and the spoke has no page on a GPU host. — needs: first-party GPU-hour pricing from Lambda, RunPod, CoreWeave or Hetzner's GPU line. Cross-spoke: the workload side sits in ../llm-inference-wiki.

Contradictions

None yet. (Watch vendor-stated pricing against future independent benchmarks.)

Index

Catalog of every wiki page, grouped by schema.org @type. Read this first when answering a query, then drill into the relevant pages. Updated on every ingest.

WebPage (sources)

Tier 1 — raw IaaS

Tier 2 — managed developer PaaS

Spans tiers / edge

Hyperscaler / frontend PaaS

Pricing data (first-party, machine-readable)

Independent benchmarks (third-party)

Synthesis

Related pages

Links to

Linked from