A field guide to the machinery behind AI
Behind the answer,
there is a system.
Meet the companies. Follow the machines under them. Open the chip. Follow a prompt. Read the scoreboard. Four interactive experiences connect the names you hear to the systems you use.
Enter the atlasBy Micah J. Miner · For curious adults and educators
01 / Company constellations
The names are familiar.
Their roles are different.
Select an organization to reveal its place in the system. Compare two histories on the same year scale. Then follow the machines: who builds the model, whose data centers run it, and which chip does the arithmetic.
Organizations can work across several layers. These placements emphasize a useful teaching distinction, not exclusive categories.
Compare the paths
Same years. Different beginnings.
Selected milestones, not complete company histories or a latest-model ranking. Drag the year control backward to reveal what had happened by then.
Follow the machines
Who stands on whose chips?
Follows the organization selected above and the year control. Boxes appear in the year this atlas first has a source for them, not a founding date. Relationships come from the companies’ own announcements; this is not a complete supplier list.
On a narrow screen, swipe the drawing sideways. Select a lit box to follow that organization.
Try a distinction / 01
Which statement best describes Google DeepMind?
Choose an answer to test the relationship.
02 / Inside the chip
One task.
Several ways to compute.
A CPU coordinates work. Accelerators specialize in different kinds of calculation. Click a compute engine to see a labeled schematic, where its memory sits, and its practical tradeoffs.
Go deeper: brands, architectures, and memory
A brand is not an architecture
NVIDIA is a company. GPU is a processor role. CUDA is NVIDIA’s programming platform. Tensor Cores are specialized units inside supported NVIDIA GPUs. These names describe different layers.
ARM64 and x86-64
These instruction-set families describe how software addresses a CPU. They do not tell you whether a device has a powerful GPU, enough memory, or a supported inference backend.
Shared and separate pools
An integrated GPU may share system memory. A discrete GPU commonly has dedicated VRAM. System RAM and VRAM are not automatically one fully interchangeable pool; offloading can add transfers and constraints.
Training and inference
Training updates model weights and often needs extra memory for gradients and optimizer state. Inference uses learned weights to generate outputs. Our memory lab models inference only.
Primary reading: NVIDIA CUDA guide · llama.cpp backends · Google TPU documentation · AWS Neuron SDK
Try a distinction / 02
A laptop advertises an NPU. What should you check next?
Choose an answer to test the hardware claim.
03 / Follow a prompt
Where does your
question actually go?
Switch between local and cloud inference. Step through the path, then try fitting an illustrative model into a memory budget.
This is an explainer. It does not run a model or send a prompt. Diagram spacing and animation do not represent measured speed.
Local LLM memory lab
Can it fit?
Change one variable at a time. Watch the budget move.
Illustrative memory budget
Fit is the first gate.
Speed also depends on memory bandwidth, compute, model architecture, backend support, context length, and thermal limits. A model that fits may still feel slow.
What this estimate assumes
Raw weight storage = billions of parameters × bits per weight ÷ 8, in decimal GB. We add 10% to weight storage as a teaching allowance for quantization metadata and format overhead. Actual files vary.
You set the runtime and context allowance yourself. This is not a prediction of KV-cache size. The cache depends on layer count, attention design, cache precision, context length, and batch size. OS reserve is another explicit assumption.
We treat the selected memory as one usable pool. Dedicated VRAM and unified memory behave differently. Disk storage and swap do not provide equivalent performance. These are generic examples, not specifications for Micah’s computers.
For real model files and backend support: llama.cpp documentation.
Try a distinction / 03
Will lowering weight precision guarantee the same answer quality?
Choose an answer to test the tradeoff.
04 / Read the scoreboard
Every launch has a scoreboard.
Check who kept score.
Labs publish benchmark tables with each new model. Choose a model to set the lab’s own figure beside an independent run of the same test. Then see what happened when five AI agents were asked to gather these numbers.
On a narrow screen, swipe the drawing sideways. The table below lists every value.
Five agents, one request
Same numbers. Different reading habits.
Micah gave five AI agents the same request to gather these benchmarks, each on its own page. Their figures mostly agree. How they sourced them does not.
Try a distinction / 04
Anthropic reports 66.4% for Claude Opus 5.5 on Terminal-Bench 4.0. Artificial Analysis reports 60%. What should you check first?
Choose an answer to test the claim.
Bring it back to practice
Ask about the system,
not just the answer.
Who controls the model?
Distinguish the model developer from the app and infrastructure provider.
Where is the information processed?
A local app can still call cloud services. Check the actual configuration.
What evidence matters?
Compare performance on your task, with your constraints, rather than brand recognition.
0 of 4 distinctions explored. Progress lasts for this page session.
Evidence and scope
Follow the source.
Checked October 12, 2026. Timelines select historical turning points; they do not inventory every release. Company announcements establish dates and stated positions, not independent proof of performance or safety. The machine map draws announced relationships from company sources and dates each link; it is not a complete inventory of suppliers, and an announced path is not proof of deployment. Explanations and teaching analogies are editorial synthesis. The scoreboard was checked October 14, 2026: a lab’s benchmark table is a claim about its own model, and an independent run is one more measurement with its own settings, not a final verdict.
Scope, accessibility, and privacy
Four buildless experiences by Micah J. Miner. All schematics are simplified teaching diagrams, not physical chip layouts. All controls work with a keyboard; animations can be paused and follow reduced-motion preferences. The machine map has a text equivalent that updates with the selection. No sign-in, analytics, prompt submission, cookies, or saved personal data. Interactions run in your browser. External source links require internet access.
X is the social platform; xAI developed Grok. Their roles remain distinct even when ownership changes. The 2026 SpaceX acquisition is included as a dated ownership milestone. The atlas is unaffiliated with the companies described.