INFRASTRUCTURE / AI SYSTEMS
STRUCTURAL OBSERVATION
FILE THR-OBS-011 / OBSERVED / 2026
The Infrastructure Is Becoming Composable
Open RAN, IOWN, distributed GPUs, and AI-native networks are turning previously fixed infrastructure layers into resources that can be dynamically recombined.
Core observation
The AI infrastructure race is moving beyond the accelerator.
Competition has focused heavily on GPUs and AI accelerators, but the surrounding layers are now changing at the same time.
Radio access is becoming open and software-defined.
Optical transport is moving deeper into computing systems.
GPU resources can increasingly be distributed and accessed remotely.
Networks are beginning to participate in decisions about where AI computation should happen.
The result is not simply a faster communications stack.
The infrastructure itself is becoming composable.
“The AI race is moving outward from the chip, while light is moving inward toward it.”
“At the same time, the infrastructure between them is becoming programmable.”
Structural model
Layer 01
APPLICATION / PHYSICAL WORLD
- Robots
- Autonomous systems
- Factories
- Mobility
- AI services
Layer 02
COMPUTE PLACEMENT
- Distributed GPU
- Remote inference
- Edge compute
- In-Network Computing
Key shift: Computation no longer has to live next to the device using it.
Layer 03
ACCELERATOR
- NVIDIA
- AMD
- EdgeCortix
- other AI accelerators
Key shift: Compute becomes specialized.
Layer 04
INTERCONNECT / OPTICS
- Silicon photonics
- CPO
- Optical switching
- Photonic-electronic convergence
Key shift: Light moves progressively closer to the processor.
Layer 05
TRANSPORT
- IOWN
- All-Photonics Network / APN
- AIOWN
Key shift: Transport becomes optical and increasingly software-controlled.
Layer 06
RADIO ACCESS
- Open RAN
- vRAN
- AI-RAN
- 5G / 6G
Key shift: Radio access becomes programmable and disaggregated.
Layer 07
SUPERVISION
- Policy
- Safety boundaries
- Human review
- ALLOW / HOLD / DENY
Key question: Who decides when programmable infrastructure is allowed to reconfigure itself?
Accelerator names mark the class of specialized compute. They are not recorded as parties to the cited demonstrations.
Observed signals
Signal 01 — OPEN RAN
OBSERVED
NTT DOCOMO continues to promote an Open RAN ecosystem designed to enable flexible networks through open and virtualized radio access infrastructure.
The radio layer is shifting away from tightly integrated proprietary hardware toward modular software-controllable infrastructure.
NTT DOCOMO — 5G Open RAN EcosystemSignal 02 — IOWN / AIOWN
OBSERVED
NTT is evolving the IOWN concept toward AIOWN, describing an AI-era infrastructure that optimizes resources including GPUs, networks, power, and edge operations.
Optical networking is no longer only a communications layer. It is becoming part of how computational resources are organized.
NTT — Medium-Term Management Strategy / AIOWNSignal 03 — DISTRIBUTED GPU
OBSERVED
In March 2026, NTT DOCOMO and NTT demonstrated AI inference using remotely distributed GPU resources connected to a 5G network through the IOWN All-Photonics Network. In-Network Computing controlled both communications and AI processing.
The network begins to influence not just where data travels, but where computation happens.
NTT / NTT DOCOMO — March 2, 2026Signal 04 — PHYSICAL AI
OBSERVED
NTT has also demonstrated remote operation and automated control of heavy machinery using IOWN APN together with local 5G and wireless networks.
Composable network and compute infrastructure is moving toward physical systems, where latency, reliability, and control directly affect machines in the real world.
NTT / NTT EAST / Taisei — April 10, 2026
The Boundary Is Moving
Network optics
City ↔ City
Data Center ↔ Data Center
Rack ↔ Rack
Board ↔ Board
Chip ↔ Chip
Inside the chip
Network architecture
Dedicated hardware
Virtualized network functions
Open interfaces
Software-defined infrastructure
AI-assisted orchestration
Autonomous reconfiguration
Two transformations are occurring simultaneously.
Light is moving inward.
Control is moving upward into software.
From stack to resource pool
FROM FIXED STACK
TO RESOURCE POOL
BEFORE
Device
Base station
Network
Data center
Server
Accelerator
Each layer relatively fixed.
AFTER
- Radio
- Compute
- GPU
- Optical path
- Storage
- AI model
- Power availability
treated increasingly as resources that can be selected, allocated, and orchestrated.
Analysis
For decades, infrastructure architecture was largely determined when equipment was installed.
A base station had its radio equipment. A server had its processors. A data center had its compute. A network carried traffic between them.
That separation is weakening.
Open RAN decomposes radio infrastructure. Virtualization separates network functions from dedicated hardware. IOWN pushes optical connectivity deeper into the computing environment. Distributed GPU architectures separate AI computation from the device requesting it. In-Network Computing begins to coordinate communication and computation.
These changes point toward a different infrastructure model: resources are discovered, selected, combined, and reassigned according to workload.
This is composable infrastructure at network scale.
Why this matters for AI
AI scaling is increasingly constrained by factors outside model architecture.
- GPU availability
- interconnect bandwidth
- latency
- power availability
- cooling
- network topology
- physical location
As these constraints become dominant, competitive advantage may shift from owning the fastest processor to orchestrating the surrounding infrastructure efficiently.
“The next AI infrastructure advantage may come from controlling the relationships between resources, not simply owning the resources themselves.”
When Composable Infrastructure Meets Physical AI
For a chatbot, additional network latency may be inconvenient.
For a robot, autonomous vehicle, industrial machine, or remotely operated excavator, latency and availability can become operational constraints.
Physical AI therefore strengthens the need for coordinated control across radio, transport, compute, accelerators, edge resources, and safety systems.
This connects programmable infrastructure directly to the physical world.
- radio
- transport
- compute
- accelerators
- edge resources
- safety systems
Programmability Creates a New Boundary
As infrastructure becomes dynamically configurable, a new question appears:
Who is allowed to reconfigure it?
- Can an AI workload move to another GPU cluster automatically?
- Can network capacity be reassigned autonomously?
- Can radio resources change according to inferred user requirements?
- Can a physical machine continue operating when network conditions deteriorate?
- Can an AI agent trigger infrastructure changes without human approval?
SHIRO & Co. STRUCTURAL INFERENCE
OBSERVE
ASSESS
DECIDE
ALLOW / HOLD / DENY
ALLOWillustrativeHOLDillustrativeDENYillustrativeRECONFIGURE
Conceptual connection only. Current Open RAN, IOWN, and AIOWN systems are not recorded as using SHIRO & Co.’s ALLOW / HOLD / DENY model.
Observed / inference boundary
OBSERVED
- Open RAN development
- IOWN / APN deployments
- AIOWN strategy
- remote GPU demonstration
- In-Network Computing
- physical machinery demonstrations
INFERENCE
- infrastructure is becoming composable
- computation is becoming location-independent
- control is moving into orchestration layers
- supervision will become more important as autonomous reconfiguration increases
Structural inference. Not established fact.
Related observations
The Supervisory Layer
CAPABILITY → DECISION. Programmability creates actions that can be taken. The Supervisory Layer is this observatory’s model for deciding whether an action proceeds.
ALLOW / HOLD / DENY
RECONFIGURATION → AUTHORITY. The chain on this page is a structural inference. It is not a description of Open RAN, IOWN, or AIOWN control planes.
When AI Agents Gain a Hardware Layer
NETWORK → MACHINE. Physical AI appears here as an existing observation: an agent that can act on equipment. Composable infrastructure is one path by which compute and radio reach those machines.
Identity Is Not Supervision
AGENT → ACTION. Shadow AI Agent supervision separates what an agent may access from whether a specific action should proceed. The same gap appears if an agent can trigger infrastructure changes.
The Human Reserve
AUTONOMY → RESERVE. If infrastructure can reorganize itself, some reconfiguration decisions may still remain with a person. The link is conceptual.
Sources
- NTT DOCOMO — 5G Open RAN Ecosystem
NTT DOCOMO · 2021-02-03
- NTT — Medium-Term Management Strategy / AIOWN
NTT · 2026-05-08
- NTT / NTT DOCOMO — Remote GPU and In-Network Computing
NTT / NTT DOCOMO · 2026-03-02
- NTT / NTT EAST / Taisei — Remote Heavy Machinery over IOWN APN
NTT / NTT EAST / Taisei · 2026-04-10
OBSERVATION
AI infrastructure is no longer evolving layer by layer. Compute, radio, optical transport, and orchestration are beginning to converge into a programmable resource system.
OPEN QUESTION
When infrastructure can reorganize itself around an AI workload, where should the authority to reconfigure it end?
SHIRO & Co.
Observatory Network
Structural observation — 2026