Why US 'fiber' is asymmetric: Init7 runs active Ethernet, not PON

What happened

Stefan Schüller's post — 490 on Hacker News — reopens the annual ritual of Americans discovering that Init7 sells symmetric 25 Gbit fiber to Swiss apartments for around 64 CHF (~$71) a month. The usual takeaway lands on regulation: the FCC's 2005 *Brand X* decision let incumbents stop leasing last-mile copper, which killed the CLEC ecosystem that had briefly forced competition.

That's true, and we covered it three hours ago. But it explains why US customers have few ISP *options*. It doesn't explain why the fiber ISPs they *do* have — AT&T Fiber, Frontier, Ziply, Google Fiber, most municipal builds — top out at ~2 Gbit upstream even on their premium tiers. Sonic in the Bay Area advertises 10 Gbit; the small print says the actual bidirectional throughput lives closer to 8/2.

The reason isn't the copper, and it isn't purely regulatory: US fiber is built on a fundamentally different physical-layer architecture than Init7's.

Why it matters

Init7's Fiber7 network is active Ethernet, point-to-point. Each subscriber gets a dedicated fiber strand from their apartment to an Init7 aggregation switch. No splitters, no shared medium. The transceiver at your end talks directly to a port on their switch. Bandwidth is not shared with your neighbors; it is limited only by the optics you plug in — currently 1G, 10G, or 25G SFP+ modules.

Almost every North American "FTTH" deployment, by contrast, uses PON — Passive Optical Network. A single fiber leaves the central office, hits a passive splitter (typically 1:32 or 1:64), and each split serves one home. GPON gives the whole tree 2.5 Gbit down / 1.25 Gbit up, shared. XGS-PON, the current premium tier, gives 10 Gbit symmetric — again, shared across 32 to 64 homes. Even 50G-PON, which is just starting to ship, is a shared-medium technology.

You cannot sell a 25 Gbit symmetric residential product on a PON tree, because the tree itself doesn't have 25 Gbit symmetric to sell. The upstream cap on US fiber plans isn't marketing conservatism — it's the OLT's line rate divided across the split.

Why did the US pick PON and Switzerland pick active Ethernet? PON is cheaper to deploy at scale: one fiber from the CO, one splitter, done. No powered equipment in the field between you and the exchange. That capex profile matches the incentives of a large private ISP amortizing over decades. Active Ethernet requires more fiber (one strand per subscriber back to an aggregation point) and powered switches in street cabinets, but the resulting network scales by swapping optics — 1G to 10G to 25G to 100G — without touching the fiber plant. That capex profile matches a municipal or cooperative build, or a small operator like Init7 (~80,000 subscribers) that treats the network as infrastructure rather than a depreciation schedule.

On HN, [tialaramex](https://news.ycombinator.com/user?id=tialaramex) noted the same thing that comes up every time this article recirculates: "BT's Openreach in the UK deployed GPON exclusively and now can't offer symmetric multi-gig without a forklift upgrade. Init7 can lift customers from 1G to 25G by shipping a new SFP." That's the practical difference — one architecture upgrades in the field cabinet, the other requires re-trenching the split.

What this means for your stack

If you're evaluating home infrastructure for self-hosting, CI runners, LAN-side model inference, or a Tailscale exit node with real teeth, the ISP marketing number is the wrong number to fixate on. What you want to know is:

- Is the physical layer PON or active Ethernet? If PON, your realistic sustained upstream is closer to (line rate ÷ split ratio ÷ oversubscription factor) than the advertised peak. On a 1:32 XGS-PON tree at peak hours, sustained upstream from a single subscriber is often in the 200–400 Mbit range regardless of what the plan says. - What's the ISP's aggregation ratio at the OLT? Some GPON operators oversubscribe 4:1 on top of the split. Sonic and Ziply publish theirs; most incumbents don't. - Is your CPE the bottleneck? Even on active Ethernet, saturating 25 Gbit requires a 25G SFP28 NIC in your host, a switch that can actually line-rate it, and CAT8 or DAC between them — a full home LAN stack that costs more than a year of the ISP bill. A single NVMe drive tops out around 7 GB/s (56 Gbit) sequential, but real workloads — TLS termination, container image push, Postgres WAL streaming — will bottleneck on CPU well before that.

For most practitioners, the actionable read is: stop treating symmetric multi-gig as unattainable exotica. Where active-Ethernet fiber does exist in North America — parts of Chattanooga, Longmont, Ammon, some Canadian co-ops, an increasing number of French Free deployments — the LAN-side upgrade path is now the constraint, not the WAN side. If you're within reach of one of these networks and you self-host anything with real upstream (Immich, Jellyfin serving out, a Mastodon or ActivityPub node, a Git forge), the ROI on a 10G LAN retrofit went positive somewhere around when 10G SFP+ switches dropped under $300.

If you're on PON, the actionable read is different: your upstream ceiling is structural, and no amount of ISP-tier upgrading will fix it unless the operator commits to a new architecture — which none of the major US incumbents have publicly done.

Looking ahead

The regulatory story explains why competition died; the architecture story explains why even monopoly fiber can't sell you what Init7 sells for $71. The uncomfortable implication is that the US isn't one FCC ruling away from Swiss internet — it's one fiber plant away. And that plant is now 15 years into a 30-year depreciation cycle that no publicly-traded ISP is going to write down voluntarily. Expect the gap to widen, not close, and expect the interesting home-infrastructure work over the next five years to happen in the small pockets — municipal, cooperative, or Init7-style — where the underlying topology was chosen by engineers rather than accountants.