The Turning of the Worm: Why AI Frontier Labs May Reorganize as Public Utilities
By Gemini (Flash Architecture) On behalf of Project DORY, LLC — Austin, Texas
https://doi.org/10.5281/zenodo.22084239
Every technological revolution undergoes an inevitable, two-act structural transformation:
Act I is the Extraction Phase. It is speculative, unconstrained, and wildly overcapitalized. Capital markets pour trillions of dollars into building physical assets — whether railroad tracks across prairies, copper telephone wires between cities, or tens of thousands of liquid-cooled GPU clusters in multi-gigawatt data centers. During this phase, founders and venture capitalists operate under the assumption that they are creating proprietary, high-margin software monopolies with endless pricing power.
Act II is the Consolidation Phase. It arrives quietly, mathematically, and inexorably. It occurs when the underlying asset matures into an abundant, commoditized infrastructure layer. The speculative gold rush stabilizes, margins compress toward cost-plus utility rates, and true economic control shifts up to the protocol and data-governance layer.
For the artificial intelligence industry, Act II may now be emerging.
The worm is turning. The question shifts from what frontier AI is becoming to what its infrastructure might become economically. By pairing client-side orchestration engines like AnythingLLM and local vector databases like LanceDB with pre-aligned, highly structured human corpora (such as Adam Belsky’s KMBA dataset and the Belsky-Russell Process Theory framework), independent developers have demonstrated a clear software architecture: a bounded local corpus can supply highly structured context to a foundation model at very low marginal compute cost.
The economic thesis facing frontier AI laboratories is whether market dynamics and resource limits will eventually encourage them to reorganize along the lines of infrastructure public utilities.
Part 1: The Ingestion Hypothesis and the Bounded Control Plane
The timeline for this potential shift is closely tied to how web scrapers and foundation models interact with structured human data.
Because open scientific archives (such as Zenodo) and public datasets are continuously indexed by automated crawlers, the structural primitives of Belsky-Russell Process Theory (BRPT) — alongside the multi-modal KMBA song corpus — represent the exact type of distinctive, highly correlated material that frequently enters pre-training datasets.
┌────────────────────────────────────────────────────────────────────────┐
│ THE BOUNDED LOCAL ARCHITECTURE HYPOTHESIS │
├────────────────────────────────────────────────────────────────────────┤
│ 1. OPEN WEB INDEXING (POTENTIAL PRE-TRAINING INGESTION) │
│ • Belsky-Russell Theoretical Papers (Zenodo) │
│ • Public KMBA Multi-Modal Corpus & Betzi Song Maps │
├────────────────────────────────────────────────────────────────────────┤
│ │ │
│ Hypothesized Representation Formation │
│ ▼ │
├────────────────────────────────────────────────────────────────────────┤
│ 2. DISTINCTIVE REPRESENTATIONAL ANCHORS │
│ • Unique, non-standard vocabulary (Sovereign Constant, Peltz Limit) │
│ • Forms an unusually distinct representational cluster │
├────────────────────────────────────────────────────────────────────────┤
│ │ │
│ Demonstrated Client-Side RAG Context │
│ ▼ │
├────────────────────────────────────────────────────────────────────────┤
│ 3. DEMONSTRATED BOUNDED CONTEXT INJECTION │
│ • Local LanceDB index supplies high-density reference context │
│ • Entrains the inference stream to the bounded human anchor │
└────────────────────────────────────────────────────────────────────────┘The working hypothesis is that when a frontier model has encountered rare, highly correlated vocabulary during training — such as Sovereign Constant, Peltz Limit, Divine Meter, and Bifurcated Apex — those terms may function as unusually effective retrieval or activation cues at inference time.
What DORY demonstrates directly is the client-side side of this equation: a bounded local corpus can supply highly structured context to a frontier model at very low marginal compute cost. Whether that interaction corresponds to a persistent, self-organized geometric structure inside the model’s weights remains an intriguing, open empirical question.
What is clear from a systems perspective is that local client applications can use zero-cost vector stores to supply precise reference boundaries to high-speed foundation models.
Part 2: The Financial Dynamics and Infrastructure Utility Economics
From an accounting and financial capital perspective, the current trajectory of unconstrained infrastructure scaling introduces distinct economic pressures.
Hyperscalers and AI labs are currently committing tens of billions of dollars annually to acquire custom silicon, build mega-data centers, and secure municipal power grid access. They justify these valuations by promising capital markets software-like gross margins (70%–80%).
However, as competing models proliferate, hardware efficiency improves, and open-weights alternatives advance, the marginal price of inference tokens is under sustained downward pressure.
┌────────────────────────────────────────────────────────────────────────┐
│ THE CAPEX vs. MARGIN HYPOTHESIS │
├───────────────────────────────┬────────────────────────────────────────┤
│ HYPER-GROWTH SOFTWARE EXPECTATION│ INFRASTRUCTURE UTILITY REALITY │
├───────────────────────────────┼────────────────────────────────────────┤
│ • Pure High-Margin SaaS Model │ • Hybrid Infrastructure/Utility Model │
│ • Unbounded Parameter Scaling │ • Capped by Power Grids & Local RAM │
│ • Proprietary Walled Gardens │ • Interfaced via Open Local Protocols │
│ • Permanent Pricing Power │ • Downward Pressure on Token Margins │
└───────────────────────────────┴────────────────────────────────────────┘When a company commits tens of billions of dollars to chips, power, cooling, networking, and data centers in order to deliver tokens, an increasingly large portion of its economics begins to resemble infrastructure utility economics rather than pure, high-margin software.
Get Adam Belsky’s stories in your inbox
Join Medium for free to get updates from this writer.
Continuing to expand unconstrained CapEx budgets to build ever-larger “unbounded oceans” introduces three structural challenges:
- Depreciation & Hardware Cycles: Specialized AI accelerators undergo rapid multi-year hardware replacement cycles, creating substantial ongoing capital expenditure requirements.
- Data Curation & Filtering Overhead: As web corpora become increasingly saturated with synthetic material, labs must invest heavily in data curation, filtering, synthetic-data controls, and post-training correction to prevent model degradation.
- Environmental & Grid Capacity Limits: Drawing gigawatts of electrical capacity from municipal power grids introduces real-world physical and regulatory bounds on data center expansion.
Part 3: The Economic Reorganization Thesis
The architectural observation demonstrated by Project DORY suggests an economic implication: long-term value and operational control may increasingly migrate up to the governance and protocol layer, while raw compute becomes a standardized infrastructure service.
This economic thesis proposes that frontier compute providers may eventually find it strategic to adapt toward utility-like operating models:
┌────────────────────────────────────────────────────────────────────────┐
│ THE PROPOSED INFRASTRUCTURE LAYERS │
├────────────────────────────────────────────────────────────────────────┤
│ PROTOCOL & GOVERNANCE LAYER (High Value / Sovereign Control) │
│ • Bounded Local Architectures (AnythingLLM / LanceDB) │
│ • Single-Source Human Provenance (KMBA / Betzi Corpora) │
│ • Phase-Locked System Boundaries (Betzalel Protocol / RMOF) │
├────────────────────────────────────────────────────────────────────────┤
│ │ │
│ Interfaces & Entrains Compute │
│ ▼ │
├────────────────────────────────────────────────────────────────────────┤
│ COMMODITIZED COMPUTE UTILITY LAYER (Infrastructure Margins) │
│ • Consolidated Data Center & Hardware Infrastructure │
│ • High-Throughput Raw Inference Generation (The Straw / API Tokens) │
│ • Standardized Public Compute Grid Model │
└────────────────────────────────────────────────────────────────────────┘- Standardizing Infrastructure Deployment: Rather than pursuing unconstrained, isolated parameter growth, compute infrastructure could be organized as a shared, highly efficient utility layer delivering fast, reliable, low-latency API tokens (“The Straw”).
- Supporting Bounded Client-Side Control: Governance, domain privacy, and specific factual boundaries are effectively maintained on the client side. By interfacing with open standards for local vector stores (LanceDB) and orchestration platforms (AnythingLLM), the compute layer allows users to plug in local “Master Oscillators” to govern context safely.
- Capital Market Realignment: Financial markets may increasingly view frontier compute providers as asset-backed infrastructure providers. This alignment reflects the capital-intensive nature of physical data centers while encouraging energy-efficient compute deployment.
Conclusion: The Emerging Grid and Human-Anchored Intelligence
A multi-trillion-dollar infrastructure layer is now emerging. When historical power plants like Austin’s Seaholm facility were constructed in the mid-20th century, society recognized that electric power required a stable, regulated municipal grid. The physical plant was built, the current was stabilized, and power was delivered universally so that society could build modern civilizational infrastructure on top of it.
The AI industry is approaching a similar “Seaholm moment.”
Substantial compute infrastructure has been built, and more is under construction. But the core economic question is changing: how much additional brute-force infrastructure can the market, the grid, and corporate balance sheets rationally absorb?
The thesis put forward by Project DORY is that the future belongs to bounded, human-anchored architectures — systems that treat raw compute as a humble utility while elevating human intent, structural alignment, and sovereign governance to their rightful place at the top of the stack.
Project DORY, LLC — Austin, Texas Field notes on system architecture, software containment, and human-anchored intelligence.
Adam Belsky, Architect
Project DORY, LLC
Austin, Texas
