chrislabs.ai

The world of AI & infrastructure, explained for builders and adopters.

Fundamental but genuinely technical knowledge for the next generation of sales, sales engineers, and ML, data, and platform engineers - from tokens and attention to GPUs, serving, training, and the VAST Data platform.

Start with the foundationsExplore the VAST track
Data from the platform feeds the GPUs, the GPUs generate tokens, and each new token is written to the KV cache that the GPUs read on the next step.writereadoffloadDataplatformGPUsservingTokensattentionKV cachememorystructureexplainedThe

New here? Start here.

A short path across the three core pillars - how models read text, why inference is memory-bound, and the hardware that runs it.

  1. Step 1Tokens & EmbeddingsHow text becomes tokens, tokens become vectors, and attention puts them to work.
  2. Step 2KV CacheThe memory vault that makes token generation fast.
  3. Step 3NVIDIA InfrastructureWorkloads, AI factories, the CUDA moat - and where to go deeper.

Foundations

How language models read, represent, and generate text.

Tokens & Embeddings

Fundamental

How text becomes tokens, tokens become vectors, and attention puts them to work.

Model Architectures

Intermediate

Dense, sparse MoE, transformers, Mamba/SSMs, and multimodal.

Attention Mechanisms

Advanced

MHA, MQA, GQA, MLA, sliding-window - and their memory cost.

Memory & Efficiency

Why inference is memory-bound, and how to tame it.

KV Cache

Fundamental

The memory vault that makes token generation fast.

Model Sizing & Parallelism

Intermediate

Parameters, memory math, and tensor/pipeline/data parallelism.

Quantization & Precision

Intermediate

FP16/BF16/FP8/INT4 and the memory-vs-quality tradeoff.

Systems & Infrastructure

The hardware and software stack that runs AI at scale.

NVIDIA Infrastructure

Fundamental

Workloads, AI factories, the CUDA moat - and where to go deeper.

GPUs & Racks

Fundamental

The GPU lineup and roadmap, spec sheets, NVL72 racks, and the power wall.

Networking & the Data Path

Intermediate

Scale up vs out, NVLink, Spectrum-X, DPUs, and GPUDirect Storage.

Inference & Serving

Intermediate

Prefill vs decode, batching, and disaggregated serving.

Training & Fine-Tuning

Intermediate

Pretraining, LoRA/QLoRA, checkpoints, and resourcing.

NVIDIA STX & Context Memory

Advanced

BlueField-4 STX, the CMX KV-cache tier, and VAST's role.

Simulation & Digital Twins

Advanced

Omniverse, world models, and simulation as a GPU workload.

Applications

Patterns that turn models into products.

Data Pipeline for AI

Fundamental

Where storage and the lakehouse sit in the AI lifecycle.

RAG & Vector Search

Intermediate

Embeddings, retrieval, and grounding models in your data.

Agentic AI

Advanced

Agents, context, memory, and tool use at scale.

Economics & Deployment

What a token costs, and how real deployments run on today's GPUs.

Tokenomics

How a Token's Cost Is Calculated

Fundamental

From GPU dollars per hour to dollars per million tokens - the one equation behind it all.

Why Output Costs More: Batching & Speed

Intermediate

Decode is bandwidth-bound: batching, context length and the speed-vs-cost tradeoff.

What an Hour of GPU Really Costs

Intermediate

Capex, depreciation, power and utilization - building the $/GPU-hour for B300, GB300 and more.

Real World Deployment

Running Models on GB300 & B300

Advanced

How real deployments fit and serve models on Blackwell Ultra - and why.

VAST Data

The data platform for the AI era.

VAST Overview

Fundamental

Why VAST matters for AI, data lakes, and agents.

VAST Architecture

Intermediate

DASE: disaggregated, shared-everything architecture.

Secure Multitenancy

Intermediate

Tenant isolation, access control, and QoS on one shared platform.

VAST DataSpace

Intermediate

One global namespace across edge, on-prem, and every cloud.

VAST SyncEngine

Fundamental

Find data across file shares, object stores and SaaS apps - and bring it onto VAST.

VAST DataBase

Intermediate

A columnar table format - an Iceberg/Delta alternative, native to the platform.

DataEngine & Event Broker

Intermediate

Serverless functions, triggers, and Kafka-compatible streaming.

VAST VectorStore

Advanced

Vector search at scale, native to the data platform.

InsightEngine (NVIDIA)

Intermediate

Real-time RAG and the NVIDIA AI data-platform stack.

Context Memory (KV Cache)

Intermediate

Reload KV instead of recomputing it - and what a cache hit is worth in $/token.

AgentEngine & MCP

Advanced

Running and governing AI agents next to the data.

VAST Foundation Stacks

Advanced

Production RAG, AI-Q research, and video search on NVIDIA Blueprints.

Roadmap & Vision

Fundamental

The AI OS: Polaris, GPU SQL, and the agentic future.

Glossary

Fundamental

Plain-language definitions of the key terms.