---
title: "Building Explainers - DeepSeek Inference"
description: "Another little explainer, this time on DeepSeek's DeepInfer inference engine, refined over three passes to strip out language-model tics."
type: note
pubDate: 2026-10-21T00:00:00.000Z
topics: ["learning", "llms", "infrastructure", "visualisation", "experimentation"]
canonicalUrl: "https://actuallymaybe.com/blog/building-explainers-deepseek-inference/"
---
> Another little explainer, this time on DeepSeek's DeepInfer inference engine, refined over three passes to strip out language-model tics.

Made another little explainer for myself.

DeepSeek says margins of 70-80% on their inference framework, gives you a sense of how good they are at infra.[^margins]

**[The work behind the next token](https://deepseek-inference.actuallymaybe.com)**

Been trying to refine different explainers, made one for myself on this. Comes out ok, obviously it's a more personal explainer in the style I prefer, but I think it's not bad for three passes.

Source: [deepseek-ai/DeepInfer](https://github.com/deepseek-ai/DeepInfer)

## How it was made

This time I tried to gather inspiration from around the web for different classic explainers, so influence from [Bret Victor](https://worrydream.com/ExplorableExplanations/) or [3Blue1Brown](https://www.3blue1brown.com/). There are a few other pieces involved in there.

I think the key to it was to have multiple passes to improve the language. There are a lot of tics for the language models that really annoy me. I have [a repo where I keep a collection of skills](https://github.com/mtreilly/maybe-agentic) that I use, and I've updated that with some of the pieces I've removed recently. These are [LLM-generated release notes of the updates](https://github.com/mtreilly/maybe-agentic/releases/tag/v2026.10.21), but they're correct.

## Related Thoughts

- [Building Explainers Random Thoughts - Drone 101](/blog/building-explainers-random-thoughts-drone-101/): the first explainer experiment, also drawing on Bret Victor's explorable explanations.
- [Cheaper Code Should Mean More Experiments](/blog/cheaper-code-more-experiments/): why it's worth trying a narrow idea, seeing what works, and trying another version.

[^margins]: DeepSeek's own published figure is $38 of revenue per accelerator-hour at off-peak prices, released alongside DeepInfer on 20 October 2026. The margin is an estimate built on that figure: [Latent Space](https://www.latent.space/p/ainews-deepseek-deepinfer-open-source) puts it at about 80%, assuming $5 to $8 of all-in cost per card-hour, and other estimates range from roughly 74% to 96%. These leave out cache storage, networking, staff, idle capacity, and training.

## Further reading
- [Building Explainers Random Thoughts - Drone 101](https://actuallymaybe.com/blog/building-explainers-random-thoughts-drone-101/) — An experiment in interactive explainers: drone control, Fourier, Bret Victor, philosophical prompts, accessibility, and learning with LLMs.
- [Mountain to Climb Updates - Convergence Explorer](https://actuallymaybe.com/blog/mountain-to-climb-updates-convergence-explorer/) — New updates to the Convergence Explorer: a tidier sidebar and footer, fresh World Bank data, and energy transition data from Ember.
- [Notes on Visual Interfaces, AI Tooling, Fast LLMs and Decision Models](https://actuallymaybe.com/blog/notes-visual-interfaces-ai-tooling-fast-llms-decision-models/) — Why agentic work is stuck in text, and what canvases, fast models, branching version graphs, and decision models might add.
- [Cheaper Code Should Mean More Experiments](https://actuallymaybe.com/blog/cheaper-code-more-experiments/) — When working with agentic tools, the instinct to solve hard problems inside your main project is usually wrong. Isolate the core problem first, experiment freely, then bring what works back.
- [Mountain to Climb](https://actuallymaybe.com/blog/mountain-to-climb/) — A tool for visualising how long economic catch-up growth actually takes between any two countries or regions