Learn AI Infra with Phoebe

The platform behind your whole data team

A data platform is eight layers, five personas and a lot of money. This course builds the reference architecture that lets a data engineer ingest and clean, an analyst build dashboards and reports, a data scientist train and deploy, an AI engineer ship LLM applications, and a business client ask a question and trust the answer - all on one stack. Vendor-neutral, with real options and honest cost bands at every layer, every price fetched from the vendor's own billing feed with the date attached. A live simulator computes the bill as you build: watch it climb from $0 to $11,295 a month as your five people come unblocked, then flip one lever and watch analytics egress add $21,402 a year for exactly the same work. And the sequencing rule the whole course exists to teach: self-serve is layer eight, and buying it first is the most expensive mistake in enterprise data.

8platform layers
5data personas
16sessions
45min per session

The leader track 🤝 · for CDAIOs, heads of data & platform sponsors · 6 x 45 min · no code

The investment case: what each layer buys, where the money actually goes, why the build-vs-rent literature cannot be trusted, what really happened in the platform builds that published numbers - and the sequencing that avoids shelfware.

🗺️ Leader 1 · easy

What AI infra actually is now

The eight layers in plain language, the five people they serve, and what genuinely changed in 2025-2026 - including why every major launch is now about grounding context for agents.

▶ Leaders start here 🧱 Leader 2 · easy

The layer map and what each buys you

Which persona each layer unblocks, why two layers unblock nobody at all, and the chokepoint idea that makes any vendor pitch evaluable in ten minutes.

▶ The build-out sim ⚖️ Leader 3 · medium

Build vs buy vs rent

Why every published GPU TCO analysis concludes "rent" and who wrote them, why depreciation is the assumption that decides the answer, and the finding that internal builds succeed at a third the rate of specialist vendors.

The captured literature 💰 Leader 4 · medium

The cost model

Storage versus egress asymmetry, GPU list prices versus what the market pays, per-installed-GPU licensing, and token spend - the four places platform budgets go wrong.

▶ Storage economics 📊 Leader 5 · deciding

Real launches and honest failures

Airbnb's six-year arc, GetYourGuide's 20% saving with the effort published beside it, Birmingham's £216.5M, Lidl's €500M data-model conflict - and why data platform failures never make the news.

The evidence session 🧭 Leader 6 · deciding

Your platform roadmap

Sequencing that avoids shelfware, what to fund first and why, the questions that separate a real vendor from a demo, and how to stage a platform nobody has time to stop using.

Strategy session

The practitioner track 🛠️ · for platform engineers, DE & architects · 10 x 45 min · vendor-neutral

Build the architecture layer by layer, bottom to top. Each session states the job, names the real options with honest trade-offs, finds the chokepoint, and prices it. b10 assembles your organization's reference architecture, costed and sequenced, and defends it.

🏗️ Practitioner 1 · easy

The eight layers and five personas

The whole stack in one session: who each layer unblocks, where the chokepoints are, and the live build-out from $0 to $11,295 a month.

▶ Practitioners start here 🗄️ Practitioner 2 · easy

Storage + open table formats

Egress economics that decide architecture, retention fine print, and the format war - where Iceberg won, and why XTable translates metadata rather than data.

▶ The egress trap 📇 Practitioner 3 · medium

Catalog, lineage, governance

The REST catalog spec, credential vending and scan planning - and why governance portability, not file format, is the real lock-in surface in 2026.

The chokepoint layer 🔄 Practitioner 4 · medium

Ingestion + orchestration

Airflow 3's security break, Dagster's asset model and what "success" means in each, Prefect's runtime graphs - choosing by failure semantics rather than by stars.

DE unblocked ⚙️ Practitioner 5 · medium

The lakehouse engine

Medallion honestly, the three unification axes vendors actually compete on, and the SQL and Python surface your analysts and scientists both work against.

The compute contract 📐 Practitioner 6 · hands-on

Semantic layer, dashboards, reporting

MetricFlow's join graph as the real product, why the semantic layer is a compiler not a store, and the dashboard-and-PDF surface your analysts ship from.

DA unblocked 🖥️ Practitioner 7 · hands-on

Compute + MLOps

GPU economics without the marketing, per-installed-GPU licensing, and the model registry as the narrow waist - aliases over stages, skew versus drift.

DS unblocked 🚪 Practitioner 8 · hands-on

The LLM gateway + app runtime

SDK versus proxy - one app with many providers, or many apps with one governed door. Routing, budgets, caching, and the gateway as your eval-set collector.

AI engineer unblocked 💬 Practitioner 9 · hardest

Self-serve, and why it comes last

The evidence session: 86.6% to 10.1% between demo and enterprise conditions, 9.50% to 69.23% from column descriptions alone, and why the benchmarks themselves are two-thirds mis-annotated.

▶ The confidently-wrong trap 🏆 Practitioner 10 · hardest

Capstone: your reference architecture

All eight layers chosen, costed and sequenced for your five personas - then defended line by line against the questions a sponsor will actually ask.

Ship the architecture
easy - everyone medium hands-on / deciding hardest

Choose your path 🗺️

Two doors into one architecture. Leaders decide what to fund and in what order; practitioners choose the components and defend the design.

🤝 Leader (fund and sequence) a1→ a2→ a3→ a4→ a5→ a6
🛠️ Practitioner (build the stack) b1→ b2→ b3→ b4→ b5→ b6→ b7→ b8→ b9→ b10
💸 Cost-first fast lane b1→ b2→ a4→ b7→ a3
🚀 Full journey (fund AND build) a1-a6→ b1-b10

The knowledge map 🧠

Both tracks at a glance - hover a session to spotlight its concepts, click any node to jump in.