Understand new things quickly and in-depth, so you can move faster.

Lectogram plans a whole lecture around what you already know, then writes it while you get on with your day.

Get started Read a lecture first ↓

A whole lecture on any subject, written for you.

Share a little about you, once.

Pick a topic, and what you want to learn.

Planned around you, written in the background.

15–45 minutes.

Not a chat reply. A whole lecture.

Chapters with diagrams and worked examples.

New ideas, explained through the ones you know.

Ask anything, right in the lecture.

Every lecture stays in your library.

One place, on your phone or laptop.

app.lectogram.com

Generate

Your Lectures

From PDFs

FDA Guidance for Our Wearable

From Topics

How Transformers Work

What CRISPR Actually Edits

Control Theory for Robot Arms

Why Rome Fell

Reading a Balance Sheet

Learning ProfileCreditsInvitesSettings

Learning Profile

Your context

Engineer turned founder, angel investor, PC gamer since the 90s. GPUs are a black box. I cook big weekend dinners for friends.Engineer turned founder, angel investor, PC gamer since the 90s. GPUs are a black box. I cook big weekend dinners for friends.

Save

Generate New Lecture

From TopicFrom TopicFrom PDFsFrom PDFs

Choose PDFs…PDF only • max 50 MB per file

solid_state_battery_review.pdf34 pages • 3.8 MB×

fusion_seed_deck.pdf22 pages • 9.4 MB×

Topic / Title

Why AI Inference Costs What It DoesWhy AI Inference Costs What It Does

What do you want from the lecture?

Details & Scope

Teach me the chemistry behind this battery paperTeach me the chemistry behind this battery paperTeach me the physics this deck is betting onTeach me the physics this deck is betting onHow chips are made, from sand to finished processorHow chips are made, from sand to finished processorVenture term sheets, before I sign my first oneVenture term sheets, before I sign my first oneOur API bill is our biggest cost. I want to know what actually drives it.Our API bill is our biggest cost. I want to know what actually drives it.

Model

Balanced (Opus 5.5)▾

*I understand that lectures are generated by AI and may contain errors, omissions, or other inaccuracies. I am responsible for the generation cost, which is deducted from my credit balance regardless of whether I’m satisfied with the result.

Generate Lecture →

Your Lectures

From PDFs

FDA Guidance for Our Wearable

From Topics

Why AI Inference Costs What It Does

Generating lecture

How Transformers Work

What CRISPR Actually Edits

Control Theory for Robot Arms

Why Rome Fell

Reading a Balance Sheet

Why AI Inference Costs What It Does

Personalized for you

Chapter 3: The Real Bottleneck Is Memory, Not Math

Why AI Inference Costs What It Does

Take our 70B reference model, stored at 8-bit (70 GB), on one H100. To produce a token, every one of the 70 billion weights must travel from memory into the compute units.

Moving the weights takes about 150 times longer than doing the math on them.

Time budget for one output token (70B model, 8-bit, one H100)

Reading the weights20.9 ms

Doing the math0.14 ms

Almost all the time goes to moving weights

Worked Example: Check It on Your Own Rig

An RTX 4090 has ~1,008 GB/s of bandwidth. An 8B model at 8-bit is ~8 GB.

Predicted ceiling: 8 GB ÷ 1,008 GB/s ≈ 7.9 ms per token → ~126 tokens/s

The same model at 4-bit (~4.5 GB): ≈ 4.5 ms → ~220 tokens/s ceiling

Local-inference benchmarks typically land at 70–85% of these ceilings.

From Your World: The Pantry Door

Your weekend dinners: an absurdly fast chef, one narrow pantry door. Dinner comes out at the pace of the door.Your weekend dinners: an absurdly fast chef, one narrow pantry door. Dinner comes out at the pace of the door.

The Tensor Cores are busy 0.14 ms out of every 21 ms, under 1% of the time.

×

What would you like to know?So would a GPU with more FLOPs help?

→

The GPU’s idle compute is where that difference comes from, and Chapter 4 shows how providers put it to work.

3. Arithmetic Intensity and the Roofline

There’s a single number that tells you which limit a workload hits.

Your Lectures

From PDFs

FDA Guidance for Our Wearable

From Topics

Why AI Inference Costs What It Does

How Transformers Work

What CRISPR Actually Edits

Control Theory for Robot Arms

Why Rome Fell

Reading a Balance Sheet

×

What would you like to know?So would a GPU with more FLOPs help?

→

Questions×

PThe Tensor Cores are busy 0.14 ms out of every 21 ms, under 1% of the time…

So would a GPU with more FLOPs help?

Barely, at small batch sizes: decode is bandwidth-bound. The H200 kept H100's FLOPs, added bandwidth, and decodes faster.

Ask a follow-up…→

Lectogramnow Your lecture is ready

Any subject. Written for you.

Pay per lecture. No subscription.

lectogram.com

The app, start to finish. Read full lectures free below.

0:00 / 1:19

The app, start to finish. Read full lectures free below.

Every lecture starts with a plan for its reader.

Lectogram does the extended planning and drafting your ChatGPT or Claude subscription won’t, using the same underlying models. It maps the whole subject around your profile, then writes each chapter to that plan.

Written for an engineer turned founder whose API bill is their biggest cost

Why AI Inference Costs What It Does

How it opens

Your API bill is a per-token price multiplied by a token count, but both of those numbers come from physical limits you can measure: how fast a GPU can pull bytes out of memory, how many customers’ requests it can fit side by side, and how much memory each conversation takes up while it’s running. You’ve built gaming rigs and backend systems, so you already have most of the intuition.

Read the whole lecture →

Ask once. It’s written while you get on with your day.

Tell it about you, once. Your background, what you already know, how you like to learn.
Ask for a lecture. A topic and what you want from it, or upload PDFs.
Get on with your day. Writing takes 15–45 minutes. We email you when it’s ready to read in your library, on your phone or laptop.

The writing takes its time so the reading doesn’t.

Ask about any paragraph, right in the lecture.

The answer opens right alongside it, so you go deeper without losing your place.

A lecture with the paragraph 'The Tensor Cores are busy 0.14 ms out of every 21 ms, under 1% of the time' highlighted, and the question 'So would a GPU with more FLOPs help?' answered in a side panel.

Read a real lecture.

Unedited, exactly as written for each reader. Any lecture you make can be shared the same way: a link anyone can read free, no account needed.

Pay per lecture.

Usually $5–15 a topic lecture.
PDF lectures usually $9–27, and the most capable model costs more.
Questions: a few cents each (6–22¢).
Add $10 of credit to start; each lecture is paid from it. No subscription.

You pay for what a lecture uses. If it fails before writing anything, you aren’t charged. If credit runs low partway, it pauses until you top up.

Prices follow model costs, so lectures get cheaper as models do.

Get started

Good to know.

How accurate is it?

Lectures are written by Anthropic’s Claude models, and you choose the model for each one. Any AI can be wrong, so read a sample in a field you know, and ask about any paragraph that looks off.

Is the lecture emailed to me?

No. The email just tells you it’s ready. You read it in your library at app.lectogram.com, on your phone or laptop.

Can I use my own PDFs?

Yes. Upload papers, chapters or documents and say what you want from them. PDF lectures cost more.

Who sees my lectures?

Only you, unless you share one as a public link. Your PDFs are deleted once the lecture is written, and nothing you send is used to train AI. Privacy policy

What if a lecture isn’t what I hoped for?

You pay for what a lecture uses, even if it misses, so read a sample first. If it fails before writing anything, there’s no charge. If your balance runs low partway, the lecture pauses until you top up, then carries on.

Questions? support@lectogram.com

What do you want to understand next?

Get started