The Photonic Commons

Photonic AI Matrix Computing and the Three-Datacenter Cloud — A Research Proposal for Democratized, Commons-Based AI Infrastructure

Sepahsalar.org Research Series · In dialogue with Univrs.io Doctrine · Working draft · May 2026

Abstract

The current AI datacenter buildout represents the largest concentration of compute in human history — projected to reach $650 billion in annual capital expenditure by 2026, consuming between 6.7% and 12% of US electricity by 2028. This concentration, organized through hyperscale silos (AWS, Azure, GCP, and emerging neoclouds), creates a structural bottleneck: the productivity gains of AI accrue to those who control the substrate. We propose a research and engineering program that combines photonic AI matrix computing with a Three-Datacenter Cloud topology and the Univrs.io distributed-systems doctrine to construct a photonic compute commons — infrastructure that treats AI inference as a shared utility rather than a rented service.

This paper outlines the physics, the topology, the protocols, and the governance model required for that transition.

1. The Problem: Concentration as a Productivity Tax

Three forces define the present moment in AI infrastructure:

Energy concentration. US datacenters consumed roughly 183 TWh in 2024 — about 4% of national demand — and trajectory analysis projects 426 TWh by 2030. The marginal cost of inference is now bounded primarily by electricity, cooling water, and grid interconnect capacity rather than by silicon.
Capital concentration. Frontier training runs cost multi-billion-dollar campuses. Inference, where the revenue lives, requires geographically distributed clusters that few entities can finance.
Architectural concentration. Rack densities have moved from 10–15 kW (legacy) to 60+ kW (AI-ready), and this trajectory is accelerating. Each density jump narrows the field of who can build.

The combined effect is a productivity tax: enterprises building AI applications pay rents to a small number of substrate owners. A commons-based alternative requires both a physics shift (toward photonic compute, which decouples inference cost from electron-domain energy losses) and a topology shift (toward distributed, federated capacity).

2. The Physics: Why Photonic Matrix Computing Changes the Equation

2.1 The native operation

Modern transformer inference is dominated by dense matrix-vector multiplication. In an electronic GPU, each multiply-accumulate operation requires charging and discharging transistor gates and shuttling weights between memory and compute — costs that dwarf the arithmetic itself. In a photonic mesh, the same operation occurs as a passive consequence of light propagating through a configured interferometer network. The matrix multiplication is the physics.

This produces three structural advantages:

Energy floor near the Landauer limit for the linear-algebra portion of inference, since reversible optical transformations dissipate near zero. Practical systems remain bounded by electro-optic conversion at the boundaries.
Latency set by time-of-flight — picoseconds across a chip-scale photonic die, compared to nanoseconds per layer on electronic silicon.
Bandwidth of ~200 THz, with wavelength-division multiplexing enabling dozens of parallel computations to share the same waveguide.

2.2 The commons-relevant property

The physics-relevant point for a commons is this: photonic inference cost scales with optical power, not with rack-density real estate. A 60 kW electronic rack and a 6 kW photonic rack delivering equivalent inference throughput have radically different siting requirements. The photonic rack can sit in a shared municipal utility, a university physics building, or a regional grid-edge node. The electronic rack cannot.

This is what makes a commons architecture physically possible — not just politically desirable.

2.3 Honest limitations

A research program must name the constraints:

Nonlinear activation functions still require returning to the electrical domain.
Fabrication tolerances limit practical precision to roughly 4–8 bits.
Training (as opposed to inference) remains difficult on noisy analog substrates.
Electro-optic conversion currently dominates total system energy.

These are engineering problems on known trajectories — not fundamental barriers. The commons program assumes a 5–10 year horizon for primary deployment.

3. The Topology: A Three-Datacenter Cloud

We propose a three-tier datacenter topology, where each tier serves a distinct physical and economic role. The structure is inspired by how electrical grids separate generation, transmission, and distribution.

Tier 1 · Foundry Datacenters (Training Substrate)

Function: Train foundation models. Run the largest tensor-parallel jobs.
Substrate: Conventional electronic GPU/TPU clusters, plus emerging photonic accelerators for inference-of-training-checkpoints.
Siting: Power-abundant regions with hydroelectric, nuclear, or large solar-plus-storage capacity.
Density: 60+ kW per rack, liquid-cooled.
Governance: Operated by frontier labs, hyperscalers, or sovereign consortiums. This tier remains capital-intensive and centralized — there is no honest argument otherwise.

Tier 2 · Photonic Inference Fabrics (The Commons Layer)

Function: Serve inference for trained models. Host fine-tuned variants. Run agentic workloads.
Substrate: Photonic matrix accelerators (Mach-Zehnder mesh, microring resonator arrays, phase-change in-memory photonic compute).
Siting: Regional. University campuses, municipal facilities, telecom central offices, edge metro sites. 1–10 MW class facilities.
Density: 5–15 kW per rack — compatible with existing commercial real estate and grid interconnects.
Governance: This is the commons layer. Federated, multi-stakeholder, governed by the Univrs.io Doctrine.

Tier 3 · Edge Lattice (Last-Mile Inference)

Function: Sub-millisecond inference for interactive applications. Privacy-preserving local computation.
Substrate: Small-scale photonic ASICs in user-proximate hardware (laptops, vehicles, building HVAC controllers, hospital diagnostic stations).
Siting: Wherever an end user or device sits.
Governance: End-user controlled, federating into Tier 2 for capability augmentation.

The critical insight: most AI value lives in Tier 2 inference, which is precisely where photonic computing’s physics enables a commons. Training (Tier 1) will remain capital-concentrated for the foreseeable future. The commons does not need to win at training to liberate the productivity layer.

4. The Protocol: Univrs.io Doctrine as Distributed-Systems Substrate

A photonic commons is not just hardware — it requires a protocol layer that lets distributed Tier 2 fabrics behave coherently while remaining sovereign. The Univrs.io distributed-systems doctrine provides the conceptual frame. We translate its principles into a concrete protocol specification.

4.1 Capacity as a first-class resource

Each Tier 2 facility publishes a verifiable capacity manifest: photonic throughput (in equivalent TFLOPs), supported model classes, latency envelope, energy mix, and pricing curve. Manifests are signed and discoverable via a federated registry — the same architectural pattern as ActivityPub or Matrix federation, applied to compute.

4.2 Workload addressing without identity capture

Enterprise applications address workloads to capability classes (e.g., “any Tier-2 fabric capable of serving a 70B-parameter model with ≤200 ms p99 latency in EU jurisdiction”) rather than to specific providers. This breaks the hyperscaler lock-in pattern by design: switching costs approach zero because addressing is capability-based, not provider-based.

4.3 Verifiable inference

Photonic substrates are analog, which historically made result verification difficult. Recent work on probabilistic proof-of-inference — sampling output distributions and verifying against reference checkpoints — solves this for the commons case. A Tier 2 node cannot silently degrade or substitute models without detection.

4.4 Settlement as a public record

Compute consumption settles via a transparent ledger (not necessarily a blockchain — a Certificate Transparency-style append-only log is sufficient and dramatically more efficient). This makes the productivity flow auditable: who paid whom for what inference, under what energy mix.

5. The Application Layer: Solving the Productivity Problem

The commons exists to solve a specific problem: enterprises building AI applications today are bottlenecked by the cost, latency, and lock-in of hyperscaler inference.

A photonic commons addresses this through four developer-facing primitives:

Capability endpoints. Developers code against open inference interfaces (the model serving equivalent of HTTP) rather than vendor APIs. Any compliant Tier 2 fabric can serve.
Locality guarantees. Enterprises specify jurisdictional, latency, and energy-mix requirements declaratively. The protocol routes accordingly.
Cost transparency. Real-time pricing across the federation, with no hidden egress fees — a primary mechanism of hyperscaler lock-in.
Cooperative scaling. When demand spikes, workloads spill to peer fabrics under pre-negotiated agreements rather than queuing on a single provider.

This is the productivity unlock. An enterprise developer building a customer-service AI, a clinical-decision-support tool, or a research assistant gets to choose the substrate the way they currently choose a database — on technical merit, not on contractual entrapment.

6. Research Roadmap

We propose four parallel research workstreams:

Workstream A · Photonic device science. Validate Mach-Zehnder mesh and microring architectures at 1024×1024 effective matrix dimensions. Characterize precision degradation under thermal drift. Develop in-situ calibration protocols suitable for unattended Tier 2 deployment.

Workstream B · Federation protocol. Specify and reference-implement the capacity manifest format, the capability addressing scheme, and the proof-of-inference verification protocol. Target interoperability with existing model-serving stacks (vLLM, TGI, SGLang).

Workstream C · Energy and siting analysis. Model the grid impact of distributed Tier 2 fabrics at 1, 5, and 10 MW class. Compare against equivalent hyperscale capacity. Identify candidate siting partners (universities, municipal utilities, cooperatives).

Workstream D · Governance and economic design. Develop the cooperative legal structures, settlement mechanisms, and dispute resolution processes required for a multi-jurisdictional commons. Draw on existing models from electricity cooperatives, internet exchange points, and academic compute federations.

Each workstream produces public deliverables. The research itself is part of the commons.

7. Conclusion: From Concentration to Cooperation

The current AI datacenter trajectory is not technologically inevitable. It is a consequence of specific choices about substrate (electronic), topology (hyperscale), and protocol (proprietary). Each of these choices has an alternative on a credible technical horizon.

Photonic matrix computing changes the substrate physics in a way that makes distributed deployment economically viable. The Three-Datacenter Cloud topology assigns each tier to its appropriate physics. The Univrs.io distributed-systems doctrine provides the protocol layer that lets sovereign nodes federate without surrendering autonomy.

The result is a path — speculative but technically grounded — from AI as rented utility to AI as a productive commons. The research program above is the work required to walk it.

This is a working research draft. Comments, critiques, and collaboration inquiries welcome at the Sepahsalar research desk.