Physical Infrastructure Control Plane

The control plane for
physical AI infrastructure.
Starting at the fiber and topology layer.

RackTwin turns racks, switches, GPUs, NICs, optics, panels, fibers, ports, and paths into the live physical model underneath the AI datacenter. Starting at the fiber and topology foundation, RackTwin gives software, operators, and eventually agents a trusted control surface for coordinating deployment, change, capacity, failure reasoning, and operations.

Physical awareness becomes the coordination layer.

RackTwin starts from the physical world — racks, optics, ports, panels, fibers, and paths — then turns that ground truth into a control plane for the logical layers above it: topology policy, deployment planning, network intent, failure reasoning, capacity decisions, and operational automation.

Layer 1 · Physical Awareness

Capture the real physical substrate

Represent what actually exists: racks, GPUs, NICs, switches, optics, patch panels, fibers, ports, positions, and end-to-end paths.

ports optics fibers placement
Layer 2 · Fabric Understanding

Build a live model of topology, constraints, and failure domains

Use the physical model to reason about topology structure, pathing, oversubscription, drift, blast radius, and the consequences of change.

topology constraints blast radius drift
Layer 3 · Logical Coordination

Coordinate the logical layers from physical ground truth

Feed physical awareness upward into network intent, deployment workflows, capacity planning, and the software systems that decide how the cluster should be wired, changed, and operated.

network intent capacity planning change windows cluster workflows
Layer 4 · Control Plane

Create a control plane for operations and eventual automation

Once the logical stack is grounded in the physical model, RackTwin becomes the control surface for review, validation, deployment, operational change, and agent-driven reasoning over the datacenter.

review validation operations agents
Thesis: the physical model is not the last layer. It is the foundation layer that lets the logical and operational layers coordinate safely and eventually automate the datacenter.

The bottleneck has moved into the physical layer.

AI clusters are no longer ordinary datacenter deployments. The physical shape of the infrastructure now affects deployment speed, failure domains, utilization, placement, and capital efficiency. Yet the physical layer is still managed with spreadsheets, PDFs, stale diagrams, and tribal knowledge.

01

Fiber and optics gate capacity

GPU capacity only matters when the cluster can actually be connected, verified, and brought online without avoidable physical-layer delays.

02

Design intent drifts from reality

As-built state changes during deployment and operations. Without a living model, every change window reintroduces uncertainty.

03

Automation needs ground truth

Agents, schedulers, and operators can only reason about infrastructure if the physical world is represented precisely enough to trust.

Before AI datacenters can manage themselves, they need to understand themselves.

RackTwin creates a live physical twin of the datacenter fabric. It is not a drawing tool. It is a model that connects physical objects, design intent, operational state, and the artifacts teams use to build and maintain the cluster.

Layer 1

Physical system of record

Represent racks, GPUs, NICs, switches, optics, patch panels, trunks, fibers, ports, and end-to-end paths as explicit infrastructure data.

what exists
Layer 2

Validation and review

Compare design intent to observed state, surface drift, check constraints, and review risk before physical changes happen.

what changed
Layer 3

Control plane

Expose the model to operators, software systems, and eventually agents that need physical truth to plan, debug, and automate.

what to do next

Starting at the fiber and topology foundation.

RackTwin starts at the fiber and topology layer because it is the foundational layer of the physical fabric. It helps teams move from fragmented planning artifacts to a reviewable model that survives design, build, handoff, and operations — then carries that physical awareness upward into the rest of the stack.

F

Fiber and topology design

Plan physical connectivity for AI cluster fabrics and keep the topology connected to real components, ports, and paths.

design intent · topology · physical paths
Δ

As-designed vs. as-built

Compare intended state to what is observed or imported, then make drift visible before it becomes operational risk.

desired state · observed state · review diff
A

Build artifacts from one model

Generate the artifacts teams need to install, review, and maintain physical infrastructure without forking the source of truth.

BOMs · labels · work orders · handoff
R

Risk and blast radius

Understand which racks, switches, links, and workloads are exposed before a planned change or physical failure hits production.

failure domains · path impact · change risk
3D

Reviewable visual proof

Give design, operations, and executive stakeholders a shared view of what is being built and why the topology matters.

rack view · floor view · path drill-down
>

Software-facing physical truth

Make the physical layer available to tools that need to reason about placement, capacity, maintenance, and change planning.

queryable model · APIs · agents later

From design review to autonomous operations.

The long-term opportunity is not just better diagrams or cleaner cabling spreadsheets. It is a physical foundation that software, operators, and AI systems can reason over.

SR

System of record

One place to understand the physical state of an AI datacenter: what exists, how it is connected, what changed, and what depends on it.

racks, ports, paths, optics, panels, devices, and dependencies
OP

Operational reasoning

Support change planning, failure analysis, capacity planning, install validation, and maintenance workflows from the same physical model.

plan, validate, diff, simulate, reconcile, operate
AI

Agent-ready infrastructure

As datacenter operations become more automated, agents need a trusted representation of the physical world before they can act safely.

reason over the physical substrate before taking action
CP

Physical control plane

The end state is a programmable physical layer that connects design, deployment, operations, workload placement, and lifecycle management.

from physical foundation to datacenter control plane

A model that produces operational answers.

RackTwin is meant to turn physical infrastructure into something teams can ask questions of, review, and safely change. The public story is simple; the hard part is turning messy physical reality into reliable operational truth.

Physical model
1cluster: "ai-fabric" 2model: physical_system_of_record 3 4objects: 5 - racks 6 - switches 7 - gpus 8 - nics 9 - optics 10 - panels 11 - fibers 12 - ports 13 - paths 14 15questions: 16 - what changed? 17 - what is exposed? 18 - what can deploy? 19 - what should happen next?
Review output
Primary category Physical infrastructure control plane
Foundational layer Fiber, topology, and deployment validation
Core model Racks, ports, optics, panels, paths, dependencies
Operator value Faster deployment and safer change windows
Platform value Physical truth for software and agents
Long-term direction AI datacenters that reason over themselves

Where the physical control plane starts earning trust.

The first users should feel immediate value in today’s workflow while seeing the path to a much larger operational system.

AI cluster planning
Model the foundational fiber and topology layer before procurement, install, and handoff.
design · review · validate
Brownfield import
Convert messy cut sheets and observed state into a reviewable model.
import · normalize · reconcile
Change planning
Understand physical impact before the change window starts.
diff · risk · blast radius
Deployment handoff
Generate operator-facing artifacts from the same model reviewers approved.
BOM · labels · work orders

The physical layer becomes programmable.

RackTwin starts at the fiber and topology layer because it is foundational. The broader platform is the layer that lets teams and software reason about the real-world machine underneath AI.

01
Fiber and topology foundation
02
Physical system of record
03
Operational reasoning layer
04
AI datacenter control plane

RackTwin is building the physical control plane for AI datacenters.

Physical awareness for the AI infrastructure stack

RackTwin · Physical infrastructure control plane

RackTwin exists because AI datacenters are becoming too large, expensive, and automated to manage the physical layer with disconnected spreadsheets, diagrams, and tribal knowledge. The physical fabric needs a live model that software and operators can trust.

The company starts at the fiber and topology foundation: where GPUs, NICs, optics, switches, ports, panels, and paths become explicit infrastructure data. From there, RackTwin coordinates review, deployment, change planning, operational reasoning, and eventually agent-driven datacenter automation.

RackTwin is led by Brian Guarraci, who previously worked on the OpenAI supercomputing team after building infrastructure and platform systems at Workday, Twitter, LinkedIn, and Microsoft.

Physical model Fiber & topology Validation Operations Automation

Building or operating AI infrastructure?

Request access to RackTwin.

We are working directly with design partners who are planning, building, importing, or operating physical AI datacenter infrastructure.

The best fit is a team with real physical-layer pain: fiber and optics planning, topology validation, deployment handoff, change windows, drift, or brownfield cleanup.

  • AI datacenter operators and builders
  • Network, fiber, and physical infrastructure teams
  • GPU cluster architects and platform teams
  • Strategic partners interested in physical infrastructure automation
Name is required
Valid email is required
Company is required
Please select a role

Invite-only while we onboard design partners directly.

Request received.

Thanks. We received your request and will follow up directly at the email address you provided.

If you need to reach us separately, email [email protected].