Module 18.3 ended at Edge AI—on-device models. Module 18.4 is a hardware-sizing playbook. Tier 1 is the honest default for most students and many products: no GPU, any laptop, internet + API key. Volume 18.1 SDKs (OpenAI, Anthropic, Gemini) are the runtime. Volume 06 VRAM math and Volume 12 KV/FlashAttention still matter intellectually—you just pay someone else’s cluster to host them.
Stay here until token spend, privacy, latency RTT, or offline requirements force Tier 2+. Buying a 4090 to “learn GenAI” is usually a worse classroom than this tier.
Learning Objectives
By the end of this lesson, students should be able to:
- State Tier 1 minimums: laptop/CPU, 8+ GB system RAM, browser or Python, internet, billed API key.
- List what you can ship: chat apps, RAG against a hosted embed+LLM, agents, eval harnesses.
- List what you cannot: local 7B+ weights, LoRA train, TensorRT, CUDA kernels.
- Compare API TCO vs idle silicon using a simple monthly token budget.
- Apply privacy/compliance gates that force leaving Tier 1.
- Hand off cleanly to Tier 2 when a quantized 7B on a 3060 is actually justified.
Tier 1 (API-only) is a development and production posture where all model weights live at a provider. Your machine runs the application (Python, FastAPI, a notebook, a browser), stores keys securely, and exchanges tokens over HTTPS. There is no local CUDA, no VRAM budget, and no KV cache on your laptop. The hardware requirement is: a general-purpose computer, network, and the operational discipline of Module 18.2 (auth, logging, rate limits)—not an RTX card.
Bill of Materials (what “any laptop” means)
| Item | Minimum | Nice | Not required |
|---|---|---|---|
| CPU / OS | Any x86 or Apple Silicon; Win/macOS/Linux | 16 GB RAM for browser + IDE + Docker | NVIDIA GPU, CUDA, TensorRT |
| Disk | ~20 GB for tools + git | SSD; datasets stay in cloud object store | Multi-TB checkpoint shelves |
| Network | Stable broadband; TLS | Low RTT to provider region | InfiniBand, 100 GbE |
| Secrets | API key in env / secret manager | Per-env keys, spend caps | GPU driver matrix |
| Software | Python 3.11+, SDK, HTTP client | Docker (18.2), pytest, eval scripts | torch.cuda, bitsandbytes |
What You Can Run vs What You Pretend
In scope
- Chat, tools, structured output via 18.1 SDKs
- RAG: hosted embeddings + vector DB SaaS or tiny local index
- Evals, prompt iteration, product UX
- CPU-only ONNX micro-models (optional)
Out of scope
- Local 7B+ generate() (that is Tier 2)
- QLoRA / full FT (Tiers 4–5)
- TensorRT engine builds
- Offline air-gap inference
Leave Tier 1 when
- Data cannot leave the building
- Token spend > GPU rent + ops
- Hard latency < WAN RTT
- You must customize weights
Minimal Client (no CUDA on purpose)
This is the entire “hardware stack”: HTTP + a key. Compare with Module 18.3’s torch.cuda inventory—if this script is enough for the product, do not buy a GPU.
TCO and Honesty
Stay API-only
- Classrooms, MVPs, bursty traffic, frontier quality
- Zero driver/CUDA/TensorRT on-call
- You still learn Vol. 06/12 conceptually
- Spend caps + observability (18.2) are the real ops
Do not stay out of stubbornness
- Air-gapped hospital/factory weights
- Steady high QPS where a 3060 Q4 7B is enough quality
- Need LoRA on private corpora (Tier 4)
- Sub-50 ms on-LAN latency SLOs
Related Lectures
| Lecture | Why it sits beside Tier 1 |
|---|---|
| 18.3 Edge AI | Previous: on-device ≠ API client |
| 18.1 OpenAI SDK (+ Anthropic / Gemini) | The actual runtime |
| 18.2 Authentication / Observability | Keys, traces, spend |
| Vol. 06 VRAM | What you are not buying yet |
| Vol. 12 KV Cache | Provider still pays this; you pay tokens |
| Tier 2 | First honest local GPU rung |
“I cannot learn real AI without a GPU.” Tier 1 is how most production apps start and many stay. Second: “API-only means no architecture knowledge.” You still design RAG, tools, evals, and you must understand why providers bill prompt vs completion (KV/prefill vs decode—Vol. 12). Third: “A MacBook with MPS is still Tier 1.” If you load local weights, you left Tier 1 even without NVIDIA—call it local-CPU/MPS and size it separately; this lecture’s contract is no local LLM weights.
Knowledge Check
- Short Answer: What three things does Tier 1 actually require? Answer: A general computer, internet, and a provider API key (plus basic app/runtime).
- True/False: Tier 1 requires CUDA and 12 GB VRAM. Answer: False—no local GPU is required.
- Multiple Choice: In scope for Tier 1: (a) QLoRA on Mistral-7B, (b) chat + RAG via hosted APIs, (c) TensorRT-LLM engine build. Answer: (b).
- Short Answer: Name two reasons to leave Tier 1. Answer: Privacy/air-gap, token TCO vs GPU, latency RTT, or need to train/adapt weights (any two).
- True/False: Volume 12 KV cache still exists when you call an API. Answer: True—it lives on the provider GPU; you pay in tokens/latency.
- Multiple Choice: Best classroom default: (a) force every student to buy a 4090, (b) Tier 1 SDKs + spend caps, (c) skip coding until H100s arrive. Answer: (b).
- Short Answer: Where should API keys live? Answer: Environment / secret manager—never in git.
- True/False: An API-only laptop running FastAPI is still Tier 1. Answer: True (app local, weights remote).
- Multiple Choice: Air-gapped hospital LLM: (a) stay Tier 1, (b) leave for local/VPC GPU tiers, (c) email weights to Gmail. Answer: (b).
- Short Answer: What is the next tier when you want a local quantized 7B on a 3060/4060? Answer: Tier 2 Local Basic.
Key Takeaways
- Tier 1 = any laptop + internet + API key; weights stay at the provider.
- This is the correct default until privacy, TCO, latency, or customization force silicon.
- Ops are keys, spend, and 18.2 observability—not nvidia-smi.
- Vol. 06/12 still explain the bill; you just do not own the GPUs.
- Continue with Tier 2 Local Basic.
Lab: Each student runs the SDK snippet with a spend-capped key, logs usage, and writes a one-page “stay vs leave Tier 1” memo for a fictional product (consumer homework helper vs hospital note summarizer vs factory offline kiosk).
Discussion: Is a student MacBook running a 3B GGUF on CPU still Tier 1? Force the definition: local weights vs API-only.
Recap: Tier 1 is API-only on any laptop. Continue with Tier 2 Local Basic.