Content delivery engineering
Glesso re-engineers content delivery around throughput per node. Fewer machines, each carrying far more traffic, orchestrated on Kubernetes and built to survive failure without a standby fleet.
THE JUMP
Four times the traffic through a quarter of the fleet.
4×
fewer nodes for the
same delivered volume
8 nodes
at 400 GbE — where competing fleets top out today
2 nodes
at 1.6 Tb/s — our design standard
Both columns deliver the same volume. Today 800 GbE per node is in production, which already halves the fleet; 1.6 Tb/s is the standard the architecture is built to, not a delivered figure.
3.2 Tb/s
the per-system ceiling the architecture scopes
SCOPED, NOT YET BUILT
02 — THE PROBLEM
Most cache fleets are sized the way they were a decade ago: one workload per box, one interface per box, and a second box standing by in case the first one fails. Traffic doubles, so the fleet doubles — and with it the rack space, the power draw, the switch ports, the spares inventory, and the number of machines somebody has to patch at two in the morning.
The hardware is not the limit. Most delivery nodes are retired with the majority of their network, memory-bandwidth and storage capability never used. The limit is the architecture wrapped around them.
WHAT DOUBLES WITH THE FLEET
03 — HOW IT WORKS
Cache nodes stop being pets. Delivery runs as an orchestrated fleet — scheduled, load-balanced, health-checked, and rolled forward without taking capacity offline. Adding a node becomes a scheduling decision rather than a deployment project.
We tune the whole path the bytes actually travel: NIC queues and receive-side scaling, IRQ affinity, NUMA placement, PCIe lane budget, the NVMe read path, kernel TLS offload, and page-cache behaviour. A single system carries 800 gigabits today and is architected toward multi-terabit.
Failure is absorbed by the fleet rather than by a duplicate of every box. Health-based steering and automatic rescheduling mean N+1 headroom instead of 2N — which is the single largest block of hardware most estates buy and never use.
04 — THE LADDER
The node count is the machines it takes to deliver 3.2 Tb/s at each tier. That progression is the whole argument in one column.
100 GbE
LEGACY32
nodes400 GbE
INDUSTRY BASELINE8
nodes800 GbE
IN PRODUCTION4
nodes1.6 Tb/s
OUR DESIGN STANDARD2
nodes3.2 Tb/s
SCOPED1
nodeFaded tracks are architectural targets, not yet built.
05 — THE ARITHMETIC
The comparison that matters is not against decade-old 100-gigabit hardware. It is 1.6 Tb/s measured against the 800 gigabits we run in production today, and 800 measured against the 400 the rest of the market sits on. Each step roughly doubles what one machine delivers while adding well under double its cost.
That is the whole mechanism. Throughput scales faster than the bill of materials, so cost per delivered gigabit falls at every rung.
The second-order savings are larger:
Every one of those is a line on a delivery budget. Every one of them is driven by how many machines are in the fleet.
THE MODEL
Indexed to the 800 GbE node in production today.
| PER NODE | 400G | 800G | 1.6T |
|---|---|---|---|
| Throughput | 0.5× | 1× | 2× |
| Hardware cost | 0.75× | 1× | 1.4–1.6× |
| Cost per Gb/s | 1.5× | 1× | ~0.75× |
Modelled from hardware cost ratios, not a measured customer result. The 400G and 800G columns describe built hardware; 1.6 Tb/s is the architectural target and its figures are projections.
06 — BILLING
Our cost per delivered gigabyte falls as a fleet fills up. Legacy tiered contracts keep that saving. We pass it back: every additional gigabyte you deliver lowers your rate on all of it, continuously, with no tier to reach and no cliff to fall off.
Your rate as volume grows
RATE / GB
Curve shapes are illustrative. Actual rates and the decay constant are set per contract against your delivered volume.
A single blended rate that recalculates as volume grows, applied to everything you delivered that period — not just the gigabytes above a line.
Tiered pricing punishes the traffic that sits just under a threshold and makes forecasting a guess about which bracket you land in. There are no brackets here.
Denser nodes make each additional gigabyte cheaper for us to serve. Sharing that back means we both want the same thing: more traffic through fewer machines.
07 — SERVICES
01
Reference designs for high-throughput delivery: node architecture, cache tiering, request routing, TLS termination, and the orchestration layer that holds it together.
02
We run the fleet. Deployment, upgrades, capacity planning, incident response and performance-regression tracking against agreed delivery targets.
03
Hardware selection and validation for delivery workloads: NIC and PCIe budgeting, NUMA topology, NVMe tiering, memory bandwidth, thermal and power envelope.
04
The control plane beneath the caches: scheduling, eBPF service load balancing, network policy, rollout strategy, observability, and explicit failure domains.
05
Getting the request to the right node: authoritative DNS, geographic and latency-based steering, health-aware withdrawal, and anycast design.
06
Per-node throughput, cache hit ratio, origin offload and cost per delivered gigabyte, measured continuously, so capacity decisions rest on data.
START HERE
A fixed-scope audit of an existing delivery estate. You get a measured throughput ceiling, the constraints holding it there, and a modelled cost per gigabyte before and after.
08 — PLATFORMS
09 — ENGAGEMENT
Fixed scope, fixed price. We measure what your fleet can actually do, and model what it should cost.
We design and build the delivery architecture, then hand it over with the documentation and runbooks needed to operate it.
Your hardware, your data centres, our team operating the delivery layer.
Our infrastructure, our operations, delivered as a service. For teams who want the outcome without owning the fleet.
10 — CONTACT
Tell us what you are delivering, how much of it, and what it costs you today. We will tell you what we think it should cost.
hello@glesso.comNo form. No backend, no spam surface.