2026 OCP APAC Summit | NTT Brings AICC to OCP: AI's Next Bottleneck Isn't Compute, It's Linking Distributed Compute with Optics

2 days ago

7 min read

Updated: 20 hours ago

For the past few years, AI infrastructure has had one priority: "bigger." Whoever builds the biggest cluster and packs in the most GPUs wins. But the AI Computing Continuum (AICC), presented at OCP by NTT's Masahisa Kawashima on behalf of the IOWN Global Forum, is about something else. As the industry shifts from "developing AI" to "using AI," compute is forced out to rural areas with power and land while users are concentrated in cities. What really holds back the next phase isn't compute density, but stitching data centers and user sites scattered across regions into a single computing space with high-bandwidth, low-latency optical connectivity. With three models — Scale-Across, Sync-Across and Reach-Across — AICC formally elevates "connectivity" to the same rank as "compute." For the optical supply chain, this isn't a new technology launch; it's a roadmap that moves DCI (data center interconnect) from a supporting role to center stage.

1. Why Now: As AI Shifts from "Development" to "Use," Connectivity Becomes the New Bottleneck

Over the past three years, compute infrastructure has had a single technical theme: scalability. That made sense — to feed exponentially growing model parameter counts, building bigger clusters was the only answer, and that path has been extremely successful.

But the inflection point has arrived. The turn Kawashima highlights is that the industry's center of gravity is moving from "developing AI" to "using AI". The people who use AI — enterprises, the public sector, research institutions, every industry — are overwhelmingly in metro areas, while AI data centers, constrained by power and land, can only be built in suburbs and the countryside. So a problem that didn't exist before has surfaced:

Compute is in the countryside and users are in the city — how do you connect them?

That is exactly the problem AICC sets out to solve. Its stance is blunt: scale is not everything. Connectivity and digital ID are becoming as important as compute. This matches the earnings signals we've been tracking — in its latest quarterly call, Google explicitly drew the next battleground for optical interconnect "across data centers," which we break down in full in Earnings Highlights: Alphabet | FY2026 Q2 — The Next Battleground for Optical Interconnect Is "Across Data Centers". AICC effectively upgrades this cross-data-center need from an internal engineering problem at a single CSP to a cross-industry standardization effort.

Behind it lies an even harder driver: sovereign AI. Many countries pouring money into AI infrastructure were never really after "a few more data centers"; the goal is to strengthen their domestic digital industries without being beholden to foreign equipment suppliers or infrastructure operators at critical moments. That force turns "distributed, autonomous, controllable" into a hard requirement — and the glue of a distributed architecture is a low-latency optical network.

2. What AICC Is: The Scale-Across, Sync-Across and Reach-Across Models

AICC is a joint initiative of OCP and the IOWN Global Forum. Its goal is to tie "data centers of many sizes" together with "regional user sites spread across geographies" through high-bandwidth, low-latency connectivity into a unified accelerated computing space.

Familiar scale-up (stacking compute within a rack) isn't repeated here. What AICC actually adds are three "horizontal" models, all keyed on the word across:

Reach-Across: Make cloud AI compute resources available to user sites "as if they were sitting in their own server room." It does this with on-demand, high-bandwidth, low-latency connections linking user sites to AI data centers. This is AICC's enterprise-facing front door.

Sync-Across: Provide low-latency interconnect between data centers so AI users can build resilient, scalable data infrastructure. Kawashima emphasized this piece in particular — data infrastructure is the real foundation of AI applications; compute is just the layer on top.

Scale-Across: Extend computing to geo-scale to support AI workloads that must run across multiple distributed data centers. This lines up perfectly with what system vendors have been pushing this year — Arista has bet its whole year on scale-across networking, a pivot we covered in Earnings Highlights: Arista | Q2 FY2026 — All-In on Scale-Across. The difference is that Arista is talking about products, while AICC wants to turn the concept into an open, multi-vendor specification.

To sum up all three in one line: Reach-Across solves "getting there," Sync-Across solves "keeping data in sync," and Scale-Across solves "computing together across regions". All three are jobs for optical networks.

3. Two Use Cases That Turn Abstract Concepts into Business

AICC doesn't stop at architecture diagrams; Kawashima offered two quite practical deployment scenarios.

Use case 1: Zero-Trust GPU Sharing. Today, to use GPU-as-a-Service, an enterprise must first "preload" data to the GPUaaS site, and before that pass a rigorous data-leakage risk assessment — the single biggest barrier to GPUaaS adoption. AICC's answer is to connect GPUaaS sites and users with on-demand, high-bandwidth, low-latency links, so AI applications pull data directly from the user site via fast remote storage access during computation, instead of moving the whole dataset over first. Once data reaches the GPU node, it is protected with confidential computing. Enterprises no longer have to bet on the integrity of the infrastructure operator, and leakage risk drops sharply — hence "zero-trust."

The more important consequence: enterprises are no longer locked into a single GPUaaS site, and can instead pick the best-value site for each AI job based on GPU availability and price. That effectively gives rise to a GPU-as-a-Service marketplace where GPU compute trades like a spot commodity. For a market with chronic compute shortages and extremely opaque pricing, that's a step with a lot of upside.

Use case 2: Physical AI in logistics warehouses. Why warehouses? Because warehouses are to logistics what data centers are to IT — the foundation of efficiency. Smart-warehouse applications need real-time local control, but staffing every warehouse with highly skilled IT personnel isn't realistic. AICC uses Reach-Across to offload AI compute to the cloud and, through a sensible function split, shrinks the on-site computer into "a simple, small, remotely operated device," dramatically cutting on-site operations burden. In Kawashima's words, it brings the cloud computing experience to environments that "simply moving to the cloud can't handle."

4. Why OCP × IOWN — and a Necessary Reality Check

The division of labor is actually clear. The IOWN Global Forum owns "use cases and architecture": it works with industry leaders such as financial institutions and broadcasters to identify the most painful problems, defines end-to-end solution architectures, and uses PoCs (proofs of concept) and techno-economic analysis to show the designs hold up technically and economically. OCP owns the "building blocks": breaking those solutions into open hardware specifications. One defines what to solve and what makes economic sense; the other turns it into hardware anyone can build — in theory, a highly complementary pairing.

But a reality check is in order. The IOWN flag has been waving for years, from the APN (All-Photonics Network) to various PoCs; the vision has always been grand and deployment always slow. AICC's maturity today is still at the stage of "inviting OCP members to review use cases and hoping everyone builds the open hardware specs together" — it's a call to action, not a stack already in volume production. From architecture definition to open specification to someone actually shipping against it, every step carries a risk of breaking down.

Another tension to watch: how much of the sovereign AI drive is technical rationale, and how much is political sentiment? Could sovereignty demands lead every region to roll its own specs, ultimately carving the "unified computing space" into isolated islands? That would be the exact opposite of AICC's "unified" intent. A standard's value lies in being adopted, not in being announced — true of every forum-driven initiative, and AICC is no exception.

5. Conclusion

First, AICC's positioning: it isn't a "component-level" breakthrough like yet another CPO design or optical engine, but a system- and network-level architectural initiative — rewriting AI from a "single hyperscale data center" mindset into one of "distributed data centers plus user sites, stitched together with optics." Its real contribution is promoting "connectivity" from an accessory of compute to a first-class citizen alongside it.

For Taiwan's supply chain, the signal cuts both ways. The upside: all three "across" models rest on high-bandwidth, low-latency optical links — DCI, coherent optical modules, pluggable optical modules, and long- and mid-reach optical transport between data centers and in metro networks. These segments, long overshadowed by flashier scale-up (in-rack) themes, will be repriced by cross-data-center demand. Taiwan's existing strengths in optical transceivers, connectors, passive components and system assembly sit right on this demand curve. To quickly map where you stand, start with our 2026 AI Infrastructure: The Full Optical Communications and CPO Supply Chain Map.

The caveat: AICC is an initiative led by Japan's NTT/IOWN with OCP's endorsement; its home narrative is in EMEA and APAC enterprise markets, and the say over hardware specs may not rest with Taiwanese suppliers. The lag from initiative to orders could well be measured in years. What's worth tracking isn't the vision in the press release but three concrete signals: whether OCP formally launches a corresponding open hardware spec next year, whether Sync-Across low-latency DCI gets beyond PoC, and whether any CSP or telecom operator jumps in to test a GPUaaS marketplace. Until those happen, AICC is a roadmap worth understanding for Taiwanese suppliers, but not one to bet on just yet.

The bottom line: the real weight of this keynote isn't any new hardware it announced, but that it changed the question for AI infrastructure from "is there enough compute?" to "can the compute be connected?" — and the latter is exactly where optics plays at home.

This article is for technology and industry trend analysis only and does not constitute investment advice.