Photonics is the ceiling case of the Cognis bearer model: the same Grain cell that crosses a 5 bps radio whisper terminates a multi-terabit photonic fabric without translation. This page states, with citations, where the photonic layer stands in 2026 — and precisely what Cognis Weave does and does not add on top of it.
Every number below is sourced (references at the bottom of the page). If you find one out of date, tell us — this page is maintained as a review document, not a brochure.
IEEE P802.3dj defines 200 Gb/s per lane — 800 GbE as 4×200G and 1.6 TbE as 8×200G — with completion targeted in 2026 [1]. NVIDIA ships co-packaged optics in Quantum-X InfiniBand switches in early 2026 and brings CPO to Ethernet with Spectrum-X Photonics in H2 2026, on TSMC's COUPE process with 3.2 Tb/s photonic engines [2][3]. Broadcom's Bailly platform takes the open-ecosystem route to CPO at 51.2 T [4].
400ZR is deployed at scale. OIF published the 800ZR Implementation Agreement in October 2024 [5]; analysts project >200,000 800ZR-class ports shipping in 2026. The 1600ZR+ digital baseline was approved in Q4 2025 with the IA targeted for Q3 2026 and products expected 2027–28 — symbol rates head toward ~240 GBd, which strains every analog component in the chain [6][7].
Microsoft's hollow-core fiber recorded 0.091 dB/km at 1550 nm — below the ~0.14 dB/km floor of solid silica — published in Nature in September 2025 [8]. Light in an air core propagates ~45% faster than in glass (~3.4 µs/km vs ~4.9 µs/km), and Azure has committed to 15,000 km of deployment, with Corning and Heraeus contracted for volume manufacturing [9].
Nothing on this page claims to beat the nonlinear Shannon limit of fiber [10], the attenuation floor of the medium, or the propagation delay of light. Cognis Weave is a protocol contribution: it changes what rides the photons, not the photons. Claims scoped accordingly — see what we don't claim.
A Bearer Functor is the protocol's contract with a physical link: it declares reach, line format, FEC class, loss model and a latency budget, and the Weave adapts cell pacing, redundancy and flow control to that envelope. The Grain cell — identity, encryption, content addressing — is byte-identical across every row of this table. That is the entire point.
| Photonic bearer | Typical reach | Line format | FEC class | Envelope: what the functor negotiates |
|---|---|---|---|---|
| CPO scale-out fabric (Quantum-X / Spectrum-X class, Bailly class) [2][3][4] | <100 m | 8×200G PAM4 per port [1] | RS(544,514) with inner-code concatenation at 200G/lane | Sub-µs latency budget; cell coalescing off; congestion signals mapped from fabric telemetry; loss treated as fault, not weather. |
| Pluggable intra-DC 800G / 1.6T [1] | 0.1–2 km | 4× / 8×200G PAM4 | RS(544,514) KP4-class | Standard datacenter envelope: shallow retransmit window, ECN-style pressure, corpus placement biased to rack locality. |
| Coherent DCI — 800ZR / ZR+ [5] | 80–1,000 km | DP-16QAM coherent, ~120 GBd | Concatenated soft-decision (oFEC/CFEC class) | Latency budget includes DSP + FEC (~µs–tens of µs); envelope exposes pre-FEC BER as link weather so the Weave derates before the link dies. |
| 1600ZR-class (2027+) [6][7] | metro/regional | ~240 GBd coherent | SD-FEC, IA pending | Declared as a future envelope: the functor abstraction is how Cognis adopts it without touching any layer above. |
| Hollow-core route [8][9] | metro/long-haul | as carried (ZR-class) | as carried | Same envelope as the coherent link it carries, with the latency term reduced ~31% per km (3.4 vs 4.9 µs/km) — the Weave's path scoring picks it up automatically from measured RTT, no configuration. |
| PON access (50G-PON, ITU-T G.9804) [11] | ≤20 km ODN | 50G TDM-PON | LDPC per G.9804 | Asymmetric envelope: upstream burst-grant latency modeled explicitly; Chorus multicast maps naturally onto PON's shared downstream. |
| Free-space optics (Taara-class links) | ≤20 km air | ~20 Gb/s | vendor | Weather is a first-class envelope variable: the same session degrades toward a radio bearer and returns, without re-keying or address change — the failover demo we lead pitches with. |
A photonics vendor's product occupies one row. Integrating with Cognis Weave means publishing an envelope for it — a declarative document, not a firmware change. Everything above the bearer (identity, cells, flows, corpora, swarms) is already done and identical across rows. See the architecture and the envelope registry — or ask the Bearer Finder which row fits your link and generate the envelope live in the Envelope Lab below.
Describe the link — what connects the endpoints, how far, what you move, what you optimize for — and the finder ranks the bearers in this pack using the same physics as the lab below. Bearers that cannot close the link are shown as exactly that: outside their preset range, or over their unamplified loss budget. Every result opens in the lab at that distance, and comes with a starting Weave in the composer.
Pick a bearer, drag the distance, choose a payload. The lab computes the one-way latency budget from the physics — propagation, FEC/DSP, cell serialization — checks optical loss against the class budget, and writes the live envelope declaration: the exact document that binds this link into the Weave.
Switch from the CPO fabric to the radio whisper with the same Grain-cell payload: eleven orders of magnitude in line rate, one envelope schema, one protocol. Watch the loss margin go negative on a long FSO hop — that is the moment the Weave derates to radio instead of dropping your session.
The demo we stage for partners, as a page: a single Cognis Weave session stepping from a co-packaged-optics fabric down to a 5 bps radio whisper and back — identity, keys and cell format unchanged the whole way. Click any stage, or let it play.
Every bearer on this page ships as a plain-TOML envelope, served from this site. No SDK, no registration — curl one, hand it to the daemon, open a flow.
# grab the 800ZR DCI envelope straight from this page curl -fsSL https://cognis.network/envelopes/zr800-dci.toml -o zr800-dci.toml # declare the bearer to your node cognis bearer add ether --envelope zr800-dci.toml # open a flow that prefers it — and watch it derate gracefully when you pull the fiber cognis flow open cognis://<peer>/corpus --bearer zr800-dci --watch
curl -s https://cognis.network/envelopes/index.json lists all eight declarations — cpo-fabric, dc-800g, zr800-dci, zr1600, hcf-route, pon-50g, fso-20g and the lora-whisper control case. Each file is byte-for-byte what the Envelope Lab generates at that bearer's reference distance — margin_db included — and the build regenerates the pack from the lab's own preset table, so the two cannot drift.
An envelope is ~20 lines of TOML: class, line rate, the latency terms, a loss budget, and behavior (weather signal, degrade_to chain). If you make photonic hardware, your integration with Cognis Weave is this file — see install, then admin@cognis.digital for the pilot.
You build the fastest physical layer in history. What the stack above you does with it is mostly 1980s-shaped. The partnership thesis is simple: your optics differentiate on physics; Cognis makes the difference legible end-to-end.
The Weave's path scoring consumes per-cell timing and pre-FEC signal quality, so a lower-latency or lower-loss link wins traffic because it is better, not because an operator re-engineered routes. Hollow-core's ~31%/km latency edge, for example, is picked up automatically [8].
One session, live, stepping from an 800G photonic fabric to a coherent metro link to a 20 Gb/s FSO hop to a LoRa whisper — same identity, same encrypted cell stream, no renegotiation. This is the demonstration we stage with partners, on partner hardware.
1600ZR-class links land as a new envelope declaration, not a protocol revision [6]. When your next part doubles the baud rate, the integration cost with Cognis is a document, and your customers' sessions simply get faster.
Envelope-aware transport means the protocol can say which physical property bounded a transfer — loss, latency, or ceiling — per flow. That is field data for your product teams that today dies inside someone else's TCP stack.
The AI plane (checkpoint distribution, KV-cache handoff, portable collectives — see For AI) is designed for exactly the fabrics CPO is being built to serve [2][3]. If your silicon targets AI clusters, the workloads Cognis carries are your benchmark workloads.
Fixed scope: one partner link type, one envelope authored jointly, the degradation demo on your hardware, and a written measurement report you can put in front of your own customers. Contact admin@cognis.digital.
This section exists because the fastest way to lose a room of photonics engineers is to imply the protocol changes the physics. It doesn't.
Cognis Weave does not increase the information capacity of any optical channel. The nonlinear Shannon limit [10] binds us like everyone else; we ride standard PHYs (802.3dj lanes, ZR-class coherent DSPs) as they are [1][5].
Propagation delay is distance over group velocity. The Weave can choose the lower-latency path (and does, automatically), but only physics — e.g. an air core [8] — actually reduces µs/km.
Cognis Weave today is a protocol specification and reference software (cognisd) in private preview. There is no Cognis ASIC. Envelope declarations for the CPO-class rows above are validated against published interface specs, not yet against partner hardware — that is what the pilot is for.
Where this page says Cognis "binds" to a bearer class, it means the bearer model expresses that link's envelope and the reference stack runs over any IP/Ethernet service on it. Native (non-encapsulated) bindings are per-partner engineering.
Cognis network