LLM Inference Bottleneck Registry

A versioned, community-editable list of frontier LLM-inference bottlenecks — with known solutions, adoption blockers, and dependency edges. Built so engineers, researchers, and funders can coordinate against the same map.

Live site: https://doctormilesbennettdyson.github.io/bottleneck-registry/

License: MIT for tooling, CC-BY-4.0 for content.

---

What's here

bottleneck-registry/
├── bottleneck_registry.md            # 18 bottlenecks, canonical source
├── docs/                             # static landing site (GitHub Pages serves from here)
│   ├── index.html                    # filterable registry UI
│   ├── build_registry.py             # parses ../bottleneck_registry.md → JS
│   └── assets/
└── papers/
    ├── paper_02_asic_projection/     # Bottleneck-Driven Projection of Frontier-Class LLM Inference on Dedicated ASICs
    ├── paper_03_musk_blueprint/      # The Compute-Robot-Energy Triad: Musk OS Integration Blueprint
    └── paper_04_reasoning_tax/       # The Reasoning-Tax Crisis (forthcoming)

The thesis

Public discourse on AI compute oscillates between marketing-grade "200× speedup" claims that collapse under audit, and nihilistic "we need $10 trillion in fabs" framings that paralyze coordination. The truth is in between, and it is more actionable than either.

Of the eighteen bottlenecks catalogued here:

• ~80% are coordination or engineering problems with known solutions — blocked by software lock-in, vendor incentives, benchmark theater, or missing standards.
• ~13% are algorithmic — solvable by research labs at laptop scale (latent reasoning, FP4 QAT maturation, MoE routing).
• ~7% require heavy capital (fab capacity, thermal infrastructure).

The binding constraint on AI abundance is shared map, not money or physics. This registry is one attempt at that map.

The papers

Paper 02 — Bottleneck-Driven Projection of Frontier-Class LLM Inference on Dedicated ASICs

papers/paper_02_asic_projection/

Calibrated baselines from public Groq, Cerebras, NVIDIA MLPerf data. Projection of Claude-class MoE inference on a 2027-feasible ASIC. Six-axis radar comparison. Eighteen bottlenecks taxonomy. Adversarial-critique section anticipating ten objections.

Headline: decode 10-70× faster, energy/token 50-200×, cost/M tokens 20-100×, agent density 3,000-12,000 streams per $1 M CapEx. Each range is conditional; single-scalar comparisons are misleading.

Paper 03 — The Compute-Robot-Energy Triad: A Vertical-Integration Blueprint for the Musk Operating System

papers/paper_03_musk_blueprint/

Strategic memo on cross-organizational integration of Tesla, SpaceX, xAI, X, Neuralink, the Boring Company, and the new fab venture. Three meta-flywheels (Compute↔Robot↔Manufacturing, Energy↔AI↔Energy, Cash↔Moonshot↔Cash). Five highest-leverage integrations, the most undervalued being Optimus as capex multiplier for fab and datacenter construction — the only recursive cost-reduction loop in the AI infrastructure stack.

Includes an honest treatment of the orbital-inference question under the integrated Musk stack (Starship + Starlink lasers + Megapack + Solar + own fab + Grok ASIC + Optimus orbital servicing): defensible 10-100 MW nicho, not the dominant inference layer. Five-year capital-recapitalization sequence sized to existing free cash flow.

Paper 04 — The Reasoning-Tax Crisis: How o1-Class Models Re-Based the Economics of LLM Inference

papers/paper_04_reasoning_tax/

Focused deep dive on bottleneck B16 (~6,000 words). Per-query cost on AIME-class problems has risen from ~$0.002 to $0.50-2.00 in 18 months (250-1000× increase) — the binding economic constraint on population-scale agent deployments. Dedicated silicon alone closes only 5-10× of the gap; the remaining order of magnitude requires algorithmic progress across three orthogonal solution families (speculative decoding, latent-space reasoning, adaptive depth) plus two missing coordination primitives (reasoning budget controllers, Reasoning Efficiency benchmark).

How to contribute

1. Pull request against bottleneck_registry.md. Schema is documented at the top of that file.
2. Status transitions (open → partial → resolved) require a citation.
3. Blockers must be specific. "More research needed" is not a blocker; "FP4 QAT degrades long-tail tasks by ~3% on MMLU" is.
4. Failed-attempt reports are welcome and tracked with the same rigor as wins.

After editing the registry, regenerate the site:

cd docs/
python build_registry.py
python -m http.server 8080  # preview

How to cite

@misc{rainieri2026bottleneck,
  author       = {Rainieri, Pablo Luciano},
  title        = {{LLM} Inference Bottleneck Registry},
  year         = {2026},
  howpublished = {Open community registry},
  url          = {https://github.com/DoctorMilesBennettDyson/bottleneck-registry}
}

For the individual papers, see the BibTeX in each papers/* directory.

Contact

Pablo Luciano Rainieri — independent researcher. Open issues for technical questions; PRs for content contributions.