Final post in the series. In the previous one, we built the 6-phase adoption framework. This one is the cheat sheet.

You already speak infrastructure fluently. AI is not a foreign language. It is infrastructure with worse naming and more hype. This glossary maps each AI term to something you already understand.

How to use this

Every entry has: the AI term, the infra analogy in parentheses, a concise definition, and when you’ll encounter it in your work. It is split into 6 categories so you can find things fast instead of pretending you remember all of it.

Core AI concepts

AI TermInfra AnalogyDefinitionWhen it shows up
ModelCompiled binaryThe trained artifact, deployable to serve predictions. Contains the learned parameters.Managing deployments, versioning artifacts, sizing storage (models range from MBs to hundreds of GBs)
TrainingBatch jobThe process of teaching the model by feeding data and adjusting parameters. Long-running, GPU-intensive.Provisioning GPU clusters, estimating job duration, planning compute bursts
InferenceAPI endpointRunning a trained model on new data to generate responses. Real-time, latency-sensitive.Every time a user or system calls an AI service. This is the production workload you monitor and scale
LLMSpecialized API service (text-in, text-out)Foundation model trained on a massive text corpus to understand and generate human language. GPT-4, Claude, LLaMA.Deploying Azure OpenAI endpoints, sizing token quotas, capacity planning
Fine-tuningConfiguration customizationAdapting a pre-trained model to your domain by training on your data.When teams need the model to understand internal terminology or specific processes
Foundation ModelBase image / golden imageLarge pre-trained model (GPT-4, LLaMA, Mistral) meant to be adapted for many downstream tasks.Selecting which base model to deploy or fine-tune. It’s the starting artifact in most AI projects
Parameters / WeightsConfiguration valuesThe internal numerical values that define how the model processes input and generates output. Large models can have enormous parameter counts, though vendors often keep the exact number private.Sizing infra: more parameters = more memory, more compute, more storage
EpochFull backup cycleOne complete pass through the entire training dataset. Fine-tuning jobs often talk in epochs; large-scale pretraining usually talks in tokens instead.Estimating duration and cost of training jobs
Batch SizeChunk sizeNumber of samples processed before updating weights. Larger = more GPU memory, more efficient.Tuning training jobs and troubleshooting OOM errors (reducing batch size is usually the first fix)
Transfer LearningTemplate reuseUsing a model pre-trained on one task as a starting point for another, preserving learned knowledge.When teams want faster, cheaper results by starting from a foundation model

Data and storage

AI TermInfra AnalogyDefinitionWhen it shows up
DatasetData source / storage volumeStructured or unstructured data used to train, validate, or test a model. GBs to PBs.Provisioning storage, planning data pipelines, managing access controls
EmbeddingHash / index keyNumerical vector representation of text/images that captures semantic meaning. Enables similarity search.Deploying RAG architectures, sizing vector databases
TokenizationSerializationBreaking text into smaller units (tokens) that the model processes. Similar to object serialization.Calculating costs (you pay per token like you pay per byte transferred), estimating context window usage, optimizing prompts
Vector DatabaseSearch indexSpecialized database that stores embeddings and searches by similarity (nearest-neighbor).Deploying RAG, provisioning Azure AI Search with vector capabilities
Feature StoreCaching layer for ML inputsCentralized repository of pre-computed features for training and inference.Architecting ML platforms that need low-latency access to features
Data DriftSchema change / input distribution shiftedWhen statistical properties of production data diverge from training data. Degrades accuracy.Model performance degrades without code changes. The silent killer of ML accuracy

Compute and hardware

AI TermInfra AnalogyDefinitionWhen it shows up
GPUCoprocessorProcessor designed for massive parallel computation, offloading matrix math from the CPU.Everywhere in AI infra: provisioning VM SKUs, monitoring utilization, managing costs
CUDAGPU instruction set / SDKNVIDIA’s parallel computing platform that enables code to execute on GPUs.Installing drivers, configuring GPU containers, troubleshooting “CUDA out of memory”
HBMGPU RAMHigh-bandwidth memory stacked on the GPU die. A100 has 80 GB HBM2e.Selecting GPU SKUs: HBM capacity determines the maximum model size a GPU can support
InfiniBandHigh-speed node-to-node networkingUltra-low-latency, high-bandwidth interconnect for distributed training. Much faster than Ethernet.Provisioning multi-node GPU clusters (ND-series VMs) for large training jobs
NVLinkGPU-to-GPU interconnectHigh-speed link connecting GPUs within a single node. It provides much more bandwidth than plain PCIe, though the exact ratio depends on the generation.Sizing multi-GPU VMs: GPUs with NVLink exchange data fast enough to matter for model parallel workloads
Tensor CoreSpecialized matrix math unitDedicated hardware in NVIDIA GPUs optimized for matrix multiply-and-accumulate operations that dominate AI.Evaluating GPU generations: Tensor Cores are why A100 is dramatically faster for AI than gaming GPUs

Model operations

AI TermInfra AnalogyDefinitionWhen it shows up
CheckpointSnapshot / backupSaved copy of model state during training: weights, optimizer state, progress.Managing storage (checkpoints can be tens of GBs each), designing fault-tolerant training
GradientError signalMathematical value indicating direction and magnitude of weight adjustments to reduce error.Troubleshooting training instability: “exploding gradients” and “vanishing gradients”
HyperparameterTunable config valueValue set before training that controls the process: learning rate, batch size, layers. Like thread count or pool size.When data scientists ask for multiple training runs with different configs: each combo is a separate job
MLOpsDevOps for modelsApplying DevOps practices (CI/CD, versioning, monitoring, automation) to the ML lifecycle.Building ML platforms, designing model deployment pipelines
Model RegistryContainer registry for modelsVersioned repository for storing and managing trained model artifacts.Implementing MLOps pipelines that need to version, promote, and rollback model deployments

Deployment and serving

AI TermInfra AnalogyDefinitionWhen it shows up
PromptAPI request bodyText input sent to the model to guide output: instructions, context, examples, question.Every LLM interaction. Prompt design impacts quality, token consumption, and cost
CompletionAPI response bodyOutput generated by the model in response to a prompt.Parsing responses, calculating output token costs, monitoring quality
Context WindowMaximum request payload sizeMaximum tokens the model processes in one request (prompt + completion combined).Designing prompts and RAG systems: exceeding context window truncates input or causes errors
Inference EndpointAPI endpoint serving predictionsDeployed model exposed as an HTTP API that accepts input and returns predictions.Provisioning, scaling, and monitoring the production-facing AI service
PTUReserved capacity (like reserved instances)Provisioned Throughput Unit. Pre-allocated capacity for Azure OpenAI models that gives you more predictable latency and throughput inside the capacity you bought.When workloads need predictable performance and you can justify reserved capacity
RAGDynamic prompt enrichment with external dataPattern that retrieves relevant documents from a knowledge base and injects them into the prompt before generation.Building enterprise AI solutions that need to answer using specific, up-to-date data
TPMBandwidth / throughput quotaMaximum tokens processed per minute for a deployment. Primary throughput metric for LLM endpoints.Sizing deployments, estimating costs, diagnosing throttling
RPMRequest rate limitMaximum API calls per minute for a deployment. Independent of token quotas.Capacity planning and troubleshooting HTTP 429

Advanced concepts

AI TermInfra AnalogyDefinitionWhen it shows up
Data ParallelismSharding data across GPUsStrategy where the dataset is split across GPUs, each processing a different batch with a full model copy.Scaling training to multiple GPUs. The simplest approach to distributed training
Model ParallelismSharding model across GPUsSplitting a model across multiple GPUs when it doesn’t fit in one GPU’s memory. Each GPU holds part of the layers.Deploying very large models (70B+ parameters) that exceed a single GPU’s HBM
LoRALightweight fine-tuningTechnique that trains a small set of adapter weights instead of the entire model, often a tiny fraction of the total parameters.When teams want to customize a foundation model without the cost of full fine-tuning
Mixed PrecisionVariable data type optimizationTraining with a mix of FP32 and BF16/FP16, using lower precision where possible to reduce memory and increase throughput.Optimizing training jobs: mixed precision can nearly double throughput on modern GPUs
QuantizationCompressionReducing model precision (FP32 → INT8 or INT4) to shrink size and speed up inference. Trades small accuracy for large efficiency.Deploying models with cost or latency constraints: quantization can cut memory usage by 4x+
Prompt InjectionSQL injection for AIAttack where untrusted input is crafted to override model instructions, causing unintended behavior.Securing AI endpoints exposed to user input. One of the first security problems you have to deal with in LLM applications
ZeROMemory optimization for distributed trainingZero Redundancy Optimizer. A family of techniques that partitions optimizer states, gradients, and parameters across GPUs to eliminate redundancy.Training large models that do not fit in GPU memory even with data parallelism. A standard DeepSpeed approach

Quick reference: top 20 terms

Pin this card. You will need it.

#AI TermInfra Translation
1ModelCompiled binary, deployable training output
2TrainingLong-running batch job that produces a model
3InferenceReal-time API call against a deployed model
4GPUCoprocessor that offloads matrix math
5LLMText-in/text-out API service
6PromptAPI request body
7CompletionAPI response body
8TokenMinimum processing unit (you pay per token like you pay per byte transferred)
9Context WindowMaximum request payload size
10Fine-tuningCustomize a base image with your data
11RAGDynamic prompt enrichment with fetched data
12EmbeddingNumerical hash/index key for similarity search
13CheckpointSnapshot/backup of training state
14TPMBandwidth quota (tokens per minute)
15PTUReserved capacity with predictable throughput
16CUDANVIDIA’s GPU SDK/instruction set
17LoRAAdapter-based fine-tuning with a tiny parameter slice
18MLOpsDevOps applied to model lifecycle
19Data DriftInput distribution shifted, model degrades
20QuantizationModel compression (4x less memory)

Series wrap-up

15 posts. From the first concept (why infrastructure engineers matter for AI) to this glossary. If you made it through the whole series, you now have enough context to talk to data scientists, architects, finance, and security without treating AI like a magic trick.

The full book, in English, is available for free at ai4infra.com. If this series saved you a few hours, send it to another infra engineer who is getting dragged into AI projects.

AI does not replace what you know. It leans on it. Networking, storage, compute, security, automation, cost control, and ugly incident response at bad hours are still the job. Now you have the vocabulary to map those instincts onto AI systems.