Where the bytes come from
Fourteen nodes, twelve countries, one reason.
A firmware image is large and a fleet is far away. Moving a 1.4 GB build across an ocean once and then serving it locally is the entire argument for having nodes at all. Everything below is capacity planning in service of that.
Node inventory
Figures are for the 30 days ending 21 September 2026. Median first byte is measured from the devices actually served by that node, not from a synthetic probe in the same building.
| Node | City | Country | Egress capacity | Edge cache | Served 30 d | Median first byte | Status |
|---|---|---|---|---|---|---|---|
| fra-a | Frankfurt | Germany | 40 Gbit/s | 96 TB | 104 TB | 18 ms | active |
| fra-b | Frankfurt | Germany | 40 Gbit/s | 96 TB | 88 TB | 21 ms | active |
| ams-a | Amsterdam | Netherlands | 40 Gbit/s | 64 TB | 71 TB | 24 ms | active |
| lon-a | London | United Kingdom | 40 Gbit/s | 64 TB | 63 TB | 27 ms | active |
| waw-a | Warsaw | Poland | 20 Gbit/s | 48 TB | 44 TB | 31 ms | active |
| sto-a | Stockholm | Sweden | 20 Gbit/s | 32 TB | 29 TB | 34 ms | active |
| mad-a | Madrid | Spain | 20 Gbit/s | 32 TB | 33 TB | 29 ms | active |
| mil-a | Milan | Italy | 20 Gbit/s | 32 TB | 31 TB | 33 ms | active |
| iad-a | Ashburn | United States | 40 Gbit/s | 96 TB | 77 TB | 22 ms | active |
| dfw-a | Dallas | United States | 40 Gbit/s | 64 TB | 52 TB | 26 ms | active |
| yyz-a | Toronto | Canada | 20 Gbit/s | 32 TB | 26 TB | 35 ms | active |
| gru-a | São Paulo | Brazil | 20 Gbit/s | 48 TB | 30 TB | 48 ms | active |
| sin-a | Singapore | Singapore | 40 Gbit/s | 64 TB | 41 TB | 38 ms | active |
| nrt-a | Tokyo | Japan | 20 Gbit/s | 32 TB | 23 TB | 41 ms | active |
| 14 nodes | 14 cities | 12 countries | 420 Gbit/s | 800 TB | 712 TB | — | — |
Two further nodes are in provisioning and are not taking production traffic; they appear in this table on the day they do. We publish city, capacity and volume. We do not publish per-node addressing details here — customers who need them for an outbound allowlist request them from support@{{DOMAIN}} and receive them under the terms in Legal.
Where the 712 TB went
Volume served per node, last 30 days. Frankfurt carries the European metering and router fleets; Ashburn carries the terminal fleet, which is why it is high on bytes and low on device count.
How a device picks a node
The candidate list
On check-in, a device receives an ordered list of three to five node hostnames under {{DOMAIN}} together with a short validity period. The order is computed from measured transfer performance for devices on comparable networks in the same region — not from a distance calculation, because the shortest path on a map is regularly not the fastest one for a device on a consumer link.
The agent tries them in order and remembers which one worked. It does not re-evaluate on every transfer; churn between nodes is worse than a slightly suboptimal choice.
Why not just one hostname
Because a device on a marginal link needs to fail over inside a single transfer, not on the next check-in. A payment terminal three hours into a 1.4 GB download cannot afford to start again because a node was drained. The candidate list means the remainder of that transfer continues from a different node, at the byte it reached, with the same content-addressed blocks.
This is also why a device holds its session open: on an unstable mobile link, the handshake is a meaningful fraction of the cost of a small delta.
Replication and cache behaviour
| Behaviour | Rule | Effect |
|---|---|---|
| Pre-warm on publish | A published release is pushed to every node that serves at least one matching device before the canary cohort starts. | ≤ 14 × image |
| Cold miss | A node that is asked for a block it does not hold fetches it from origin and keeps it. Only the first device pays the extra latency. | 1 origin fetch |
| Eviction | Least-recently-requested, with a floor: the three most recent builds of any product line with active devices at that node are never evicted. | — |
| Deduplication | Blocks are content-addressed, so two product lines sharing a base OS share cache occupancy. | −31 % cache |
| Draining | A node being taken out for maintenance stops appearing in new candidate lists and finishes in-flight transfers. | 0 restarts |
| Regional pinning | A product line can be pinned to a region so that images are replicated only to nodes in that set — used where a data-residency commitment applies. | by contract |
One replication, many deliveries
A 1.4 GB terminal image crosses the Atlantic once, to São Paulo, and is then served locally to 4,000 terminals: 1.4 GB of long-haul traffic instead of 5.5 TB. At the 30-day scale, pre-warming all 14 nodes for every release of every product line accounts for under 2 % of total egress.
Capacity and headroom
Sustained versus burst
712 TB over 30 days averages 2.2 Gbit/s across the network. That number is useless on its own: a full-fleet terminal release puts 34 Gbit/s through two nodes for six hours. Capacity is sized for the burst, and the rate cap exists so the burst is a decision rather than an accident.
Node loss
Any single node can be lost without a queue forming. Losing both Frankfurt nodes at once — 192 TB of monthly volume — spills onto Amsterdam, Warsaw and London and extends European wave durations by roughly 40 %. That is the worst modelled case and it is a slowdown, not an outage.
Private nodes
A customer with a concentrated fleet can have a node placed in a facility of their choosing and served only their product lines. It joins the same control plane and the same candidate lists, scoped to that customer. Quoted per site — see Pricing.