OCP Global Summit 6/6 | Radix Sets the Scale, OCS Stays in the Slow Layers
OCP Global Summit paid series 6/6: network architecture from scale-up to scale-across — how big a single-hop domain really is, what sets radix and plane count, how to connect across buildings and campuses, and which layers OCS is actually used in
Free section
Part 4/6 of this series showed how multi-plane design keeps scale-out running through failures, and 5/6 argued that at 448G the 1 to 3 m range goes to active copper first. This final piece asks about topology: what decides how big one network tier can get, and which layer does optical circuit switching (OCS) actually belong in?
The market's standard answer: scale-up grows from 72 accelerators in one rack toward thousands, scale-out adds more tiers, DCI handles links between data centers, and OCS is Google's secret weapon. During ECOC we cited Cignal AI's view that about 95% of OCS shipments go to Google; see ECOC 2026 | 95% of OCS Shipments Are One Customer: Google.
At the 2026 OCP Global Summit, the people drawing the topologies were the buyers and the switch-silicon vendors.
First, Eridu and Delos Data put the claimed scale-up domain next to the single-hop domain: NVL576 is 72 accelerators in a single hop, OpenAI Jalapeño has 128 in a single hop out of 2,048, and Microsoft Maia 200 has 48 in a single hop out of 6,144.
Second, Google split its network into three layers: the Virgo backend is a two-tier packet-switched fabric, Apollo OCS sits in the Jupiter front end, and the TPU ICI is built on OCS. For three or more buildings, Google said OCS has lower latency and lower cost than an electrical-switch spine, but this is still a design concept under study.
Third, scale-across now has an order of magnitude: Meta's cross-campus architecture connects tens of thousands of 800G ZR modules at each end, and Google estimates building-to-building bandwidth at the 10 to 100 Pbps scale.
STT's view: how big one network tier can get is set by switch radix and plane count; OCS only goes into the layers where traffic changes slowly and comes in large chunks.
▲ Where OCS sits across Google's four network tiers and GPU scale-up: deployed in TPU ICI and the Jupiter frontend, not seen on the Virgo backend slides, under evaluation for cross-building campus links, and with no buyer commitment for GPU scale-up (Image source: Simple Tech Trend original; data from Google, Marvell, Lightelligence and NVIDIA 2026 OCP Global Summit presentations and panels; Google official blog)
Why Ethernet is moving into scale-up: 2026 OCP APAC Summit | Ethernet Moves Into Scale-Up.
In the paid section
Single-hop breakdown: how far apart the claimed domain and the single-hop domain are across five systems
Radix breakdown: how many endpoints one 102.4T switch connects in a single hop versus two tiers
Plane breakdown: what sets the plane count, and what flattening pushes onto fiber
Protocol breakdown: how far ESUN, SUE-T, UALink over Ethernet and Tahan have come
Cross-campus breakdown: the optics switch at 10 km, and the numbers behind the deep-buffer and shallow-buffer camps
OCS breakdown: six possible positions, two deployed and one under evaluation
Where we stand on our June Virgo article: which sentence we revise and which we keep
Differences from our ECOC calls: one partial contradiction, one partial revision, and five continuations
Taiwan supply chain: Taiwanese names on the slides and the two real bottlenecks
Counterarguments and eight signals to watch
──────────────────────────────
That is the summary of this piece.
STT's full analysis — the math behind why radix sets the single-hop domain, the actual status of OCS in six positions, a sentence-by-sentence revision of our June Virgo article, and the differences from our ECOC series — is in the paid section, with 7 original charts.
👉 Subscribe to the STT paid section to read the full version:
──────────────────────────────
This article is for technology and industry trend analysis only and does not constitute investment advice.
Comments