Kushagra GoyalStack

How I think about infrastructure

Hardware choices rarely stay isolated. A processor can change software cost, a drive can change the chassis, and a GPU can change power and networking. I use this map to keep those dependencies visible.

Four parts of the infrastructure crossed by five questionsCompute, storage, networking and the rack are drawn as four rows. Configuration, lifecycle, supply, economics and AI workloads run down across all four. Where a question meets a part of the system there is a decision to make, and those decisions shape what a customer can order. Each label opens its own page.ComputeCPU MEMORY GPUStorageBLOCK FILE OBJECTNetworkingFABRIC BANDWIDTHRackPOWER COOLINGConfigurationLifecycleSupplyEconomicsAI workloadsWhat a customer can ordergoverned buildable quotable
The model

Four parts, five questions

The parts of the system run across and the questions I ask of each run down. Hover or tap a label to see where they cross, then open it to read more.

Tap a part of the system or a question to see where they cross.

What the system is made of

I think about the infrastructure in four buckets: compute, storage, networking, and the rack itself. GPUs sit inside the compute decision, because choosing one usually changes the rest of the server around it.

Four parts
Part of the system

Compute

CPU, memory, drives and accelerators, and how one choice moves the others.

CPU · Memory · Local storage · Accelerators

Part of the system

Storage

How an application accesses its data decides which platforms fit.

Block · File · Object

Part of the system

Networking

Size the compute and storage well, and the network can still be the limit.

Fabric · Topology · Bandwidth

Part of the system

Data center and rack

Power and cooling set a hard ceiling, whatever the spec sheet allows.

Power · Cooling · Density

See the hardware inside these parts

The questions I apply across the system

Whichever part I am looking at, the same five questions come up: how it gets configured, what happens as parts change, whether we can actually get them, what it costs to run, and what an AI workload does to all of that.

Five questions
Question

Configuration

Which combinations a customer is offered, and whether all of them can be built.

Question

Lifecycle

Parts change. What a customer can order should not break when they do.

Question

Supply

What happens when a CPU, memory module or SSD becomes hard to get.

Memory · CPU · SSD and NVMe

Question

Economics

What the system costs to own, not what it costs to buy.

Question

AI workloads

Training, inference and retrieval stress the same system differently.

Training · Inference · Retrieval