t# aibase
A full-stack AI platform — from infrastructure to application.
The repo is organized around the 5 layers of the AI stack. Each choice you make at any layer — from the hardware up to the user interface — has direct implications on the solution's quality, speed, cost, and safety.
aibase/
├── 01-infrastructure/ # GPUs & compute: on-premise, cloud, local
├── 02-models/ # model registry, serving, evaluation
├── 03-data/ # sources, pipelines, vector store, RAG
├── 04-orchestration/ # planning, execution, review, MCP
├── 05-application/ # interfaces & integrations
├── shared/ # cross-cutting config & scripts
└── docs/ # design notes, references
| # | Layer | What it covers |
|---|---|---|
| 1 | Infrastructure | AI hardware (GPUs). Deploy on-premise, in the cloud, or locally. |
| 2 | Models | Open vs proprietary, large vs small (LLM/SLM), specialization. |
| 3 | Data | Data sources, processing pipelines, vector databases, RAG. |
| 4 | Orchestration | Break tasks into thinking → execution → review. Protocols like MCP. |
| 5 | Application | Interfaces (text/image/audio…) and integrations with other tools. |
Revenue across the stack is a pyramid: the closer a layer sits to the hardware at the base, the more of today's money it captures (figures mid-2026).
┌──────────────┐
│ APPS │ L5 · fragmented, many startups
┌─┴──────────────┴─┐
│ ORCHESTRATION │ L4 · LangChain ~$16M ARR — max hype, min revenue
┌─┴──────────────────┴─┐
│ DATA │ L3 · Scale AI ~$2B (Meta took 49%)
┌─┴──────────────────────┴─┐
│ FOUNDATION MODELS │ L2 · OpenAI ~$33B · Anthropic ~$45B*
┌─┴──────────────────────────┴─┐
│ INFRASTRUCTURE / CHIPS │ L1 · Nvidia ~$75B/qtr → ~$300B run-rate
└──────────────────────────────┘
base = most $ captured today
| Layer | Who | Revenue (mid-2026) |
|---|---|---|
| 1 · Infrastructure / chips | Nvidia data-center | ~$75B/qtr (Q1 FY27, +92% YoY) → ~$300B run-rate |
| 2 · Foundation models | OpenAI / Anthropic | ~$33B / ~$45B ARR — Anthropic now leads* |
| 3 · Data | Scale AI; vector DBs | ~$2B; Meta bought a 49% stake ($14.3B) |
| 4 · Orchestration | LangChain | ~$16M ARR |
| 5 · Applications | thousands of startups | fragmented, individually small |
* Anthropic's ~$45B is gross (it books cloud-reseller end-customer spend as revenue); on OpenAI's preferred net basis it's ~$22B. Always ask: gross or net?
Two things to notice: revenue concentrates at the base (Nvidia alone, ~$300B run-rate, earns more than every model company combined), and hype ≠ revenue — the orchestration layer gets the most attention but captures the least money (LangChain is ~1/18,000th of Nvidia). Value is expected to shift up toward applications over time, as it did when cloud value moved from AWS infrastructure to SaaS apps.
A lens for reading any AI product. Clayton Christensen (Harvard, The Innovator's Dilemma) splits innovation into three types:
| Type | Idea | Effect on jobs | AI example |
|---|---|---|---|
| Market-creating | Make something expensive/hard → cheap & easy for everyone, opening a new market | Creates jobs | Foundation models that put AI in everyone's hands |
| Sustaining | Make a good product better | Roughly neutral | Adding an AI feature to an app you already use |
| Efficiency | Do the same work with less | Tends to reduce jobs | AI coding assistants, automation, summarization |
The type is not decided by the layer — it's decided by how the technology gets used. The same model can create a new market, improve a product, or cut costs. Reading an AI launch, ask: which of the three is this?
Klarna's AI assistant (built on OpenAI) handled 2.3M chats in its first month — the work of 700 full-time agents — automating 67% of conversations, cutting resolution from 11 min to under 2 min, and driving a ~$40M profit improvement in 2024. Caveat: in 2025 Klarna said it had cut human staff too far and rehired for premium support — AI-first, not AI-only.
Others in the space: Sierra, Decagon, Ada, Intercom Fin (public price ~$0.99 per resolution).
CEO math — a mid-size store, 100,000 support contacts/month:
human cost ~$5 / contact (fully loaded)
AI cost ~$1 / resolution (Intercom Fin public price)
AI deflection 67% (Klarna / Fin benchmark)
AI handles 100,000 × 67% = 67,000 contacts
cost on AI 67,000 × $1 = $67,000 / mo
same on humans 67,000 × $5 = $335,000 / mo
──────────────────────────────────────────────
net saving ≈ $268,000 / mo ≈ $3.2M / yr
The lever is volume × deflection rate × (human − AI cost), plus 24/7 and multilingual coverage.
Morgan Stanley indexed 350,000 research documents (40M words) with RAG. Before: a query meant 30+ minutes of manual search, and advisors reached only ~20% of the knowledge base. After: instant answers, 98% of advisor teams use it, and document access rose from 20% to 80% — freeing advisor time for revenue-generating client work, with answers grounded in verified sources.
Similar: Glean (enterprise search), Harvey (legal), Hebbia (finance).
The pyramid measures who earns money selling AI tooling — there the base wins. These case studies measure value created by applying AI to your own business — and that value lands on the operator's books, not an AI vendor's. Klarna's $40M isn't any "AI app company" revenue; it's a line on Klarna's P&L. So the application layer looks thin for vendors yet is where operators capture the most. Two different questions, two different winners.
For GenAI-for-SRE cases (incident RCA, observability, CI/CD risk) with full
cost-and-savings math, see docs/devops-cases.md.
Each layer directory has its own README.md describing scope and intended
contents. Start at 01-infrastructure/ and work up, or jump to the layer you
need.