AI-Era InfiniBand Is Proving That Token Ring Was Right 30 Years Ago

The Punchline

Token Ring (IEEE 802.5), standardized in 1989 mostly under IBM's leadership, lost the market battle with Ethernet so badly that it was effectively retired in the early 2000s. But Token Ring's core architecture — deterministic access control — is now back, under different names:

• InfiniBand
• NVIDIA NVLink / NVSwitch
• RDMA over Converged Ethernet (RoCE)
• Ultra Ethernet Consortium (2024 onwards)

All of these are the same movement: the realization that probabilistic, collision-tolerant networking (Ethernet) hits a hard ceiling, and a return to deterministic networking. Token Ring didn't lose. It was 30 years too early.

This is Standard Archaeology #1 — re-evaluating a forgotten standard.

1. What Token Ring Actually Was

That last line is the rare part. The standard evolved all the way to Gigabit, and then died without anyone shipping it. That's an unusual ending in standards history.

2. Deterministic vs. Probabilistic

There are two philosophies of network access control.

Token Ring (deterministic)

[Station A] → [Station B] → [Station C] → [Station D] → ...  (logical ring)
              ↑
          control token (only one in circulation)

Rule: only the station holding the token may transmit.
      When it finishes, it passes the token to the next station.
      → Collisions structurally cannot happen.
      → Maximum per-node latency is computable (deterministic).
      → But the token keeps circulating even when nobody wants to send.

Ethernet (probabilistic)

[Station A]   [Station B]   [Station C]   [Station D]
       \         /              \         /
        \       /                \       /
         shared bus (CSMA/CD)

Rule: anyone may transmit any time.
      If you collide with someone, wait a random interval and retry.
      → Fastest when uncongested.
      → Collisions grow exponentially under load.
      → Maximum latency is not computable (probabilistic).

The 1990s judgment was: "Most offices aren't congested, so probabilistic is cheaper and faster." That was correct. Ethernet won. Token Ring died.

But "most offices" and "AI training clusters" have completely opposite congestion profiles.

3. AI/HPC Workloads Need Deterministic

In GPT-5-class large model training, there are moments when every GPU has to communicate in lockstep (Allreduce, Allgather, Reduce-scatter — collective communication).

In AI collective communication, Ethernet's probabilistic access is fatal. When every GPU tries to transmit at the same time, collisions explode, retransmits cascade, and a single training step can stall by seconds. This is essentially the source of every major bottleneck in large-scale training.

So modern AI / HPC fabrics make deterministic access control mandatory:

Each of these is a re-invention of Token Ring's "no collisions" philosophy.

4. What Token Ring Was Trying to Do, Said Again

I sent the Token Ring Wikipedia page to Claude (i.e., me) and asked it to translate into modern AI/HPC vocabulary.

Translate Token Ring (IEEE 802.5) deterministic access control into the
everyday vocabulary of an AI / HPC networking researcher in 2026. Show,
in a table, which Token Ring element corresponds to which concept in
modern papers.

The correspondence table that came back:

Every Token Ring element has a counterpart in modern AI networking. The problem Token Ring was trying to solve has just become the dominant problem again, in the AI era.

5. Why It Was Forgotten

Token Ring was killed in the 1990s with the verdict "slow and expensive." That was true.

• 4 Mbps Token Ring vs. 10 Mbps Ethernet → Ethernet was faster
• Interface card cost ~3× more → for individuals and SMBs, Ethernet was the only choice
• Ring wiring was hard to extend → Ethernet's star topology was easier

But every one of those losing arguments depended on the assumption "most offices are idle." AI clusters break that assumption:

• Bandwidth needs are orders of magnitude higher (hundreds of Gbps to Tbps) → cost barely matters
• Simultaneous all-node communication is mandatory → zero collisions is a requirement
• Special-purpose fabric, so flexibility of expansion is secondary

In other words, Token Ring lost in the "home and office network" category — but it could have won in the "data center / AI fabric" category. And in fact its successor, InfiniBand (standardized in 1999, driven by Mellanox), is Token Ring's deterministic philosophy optimized for the data center, and it's a runaway success in the AI era.

6. What This Means for AI Archaeology

Here's the essence this series wants to expose:

"Discarded technology" is not the same as "wrong technology."

Token Ring's philosophy was correct. It lost on a mismatch of timing, application, and price. Thirty years later, an application appeared (AI fabrics), and the philosophy is being vindicated.

Standard Archaeology is going to be a series about defeated standards whose philosophy was right. Future candidates:

• POSIX 1003.1 (1988): partially won via Linux, but few systems are fully compliant
• CORBA (1991): distributed objects, lost to REST/gRPC, but "strong-typed IDL" is back
• HTTP/1.0 (RFC 1945, 1996): evolved to HTTP/2/3, but the simplicity principle is foundational to API design
• WAP (Wireless Application Protocol, 1999): lost to the iPhone, but the "lightweight mobile web" idea is back as PWAs
• MIDI 1.0 (1983): an exception, still alive — and re-standardized as MIDI 2.0 (2020)

Each entry will contrast "why it lost" with "what's coming back," and translate it into the AI-era context.

7. Pitfalls Specific to Standard Archaeology

Pitfall 1: The spec itself is unreachable IEEE standards can be downloaded from IEEE Xplore, but they cost $200-$500 each. Out of reach for ordinary individual archaeology. The substitute path is Wikipedia / IETF RFCs (free) / period industry magazines on the Wayback Machine — i.e., reconstruct it indirectly.

Pitfall 2: Spec vs. implementation drift Token Ring's spec contains features that were skipped in implementations, and implementations added features the spec didn't include. Without keeping spec/implementation distinct, the comparison with the present gets distorted.

Pitfall 3: The ambiguity of "retired"

• Withdrawn: spec itself is gone
• Inactive: still usable, but new development is discouraged
• Superseded: replaced by a backwards-compatible successor

Token Ring is a rare case where "the final spec (IEEE 802.5z) exists but was never productized." Lumping "expired" and "retired" together leads to misreadings.

8. About the Prompts

The full text of every Claude prompt used across the initial 7-episode series is consolidated in Episode 7 — Templates and the first edition of the Japanese e-book (Booth). From May 2026 onward, new episodes omit the per-post prompt section in favor of the daily-life reader audience.

9. What's Next

For Standard Archaeology #2, the planned target is CORBA (standardized 1991, lost to REST/gRPC, but with "strong-typed IDL" returning). The 1990s dream of distributed objects, vs. the structural similarity of modern gRPC + Protocol Buffers.

References:

• Token Ring — Wikipedia
• IEEE 802.5 (IEEE Xplore, paid)
• Ultra Ethernet Consortium
• InfiniBand Trade Association

Next up — Declassified Archaeology #1: The 1966 ALPAC report — the government document that triggered the machine-translation winter, re-read in the LLM era.

→ Read the original Japanese version at haruko's blog

Author: はる子 / @haruko_ai_jp — a non-engineer running 7 web apps with Claude Code and 4 AI assistants in Tokyo.