AI, Decoded.
Everything you need to discover, understand, compare and build with AI — from silicon to applications.
- Know what to use.
- Know how it connects.
- Know what it runs on.
- Know what it costs.
From silicon to software. From models to applications. From discovery to action.
- 1,116
- products
- 2,416
- sourced answers
- 1,033
- GPU price readings
- 901
- dependencies
Tell us what you are building and we name the layers you actually need — plus the ones you can skip. No account, and it says when we hold too little to answer.
One trip through the stack, with Infra
Seven stops, from a vague idea to a number you can check.
01A vague task meets the whole stack
Building with AI means choices on 10 layers, from silicon to applications. We list 1,116 products across them and say where a layer is still thin.
See the layers →02What a GPU hour costs
For the H100-SXM-80GB, TensorDock posts the lowest rate of 13 sellers today: $2.25 per GPU-hour.
See the spread →03What a million tokens cost
The cheapest input rate per maker, side by side: granite-4.0-h-micro at $0.02 and gpt-oss-20b at $0.02 per million tokens.
All LLM prices →04Picking the tools
8,106 MCP servers are listed, each with what it can reach and the credential it wants. Pick what your stack needs.
Browse the tools →05Learning to use it
The Academy has 3 hands-on labs: catch the invented claim, fix the prompt, clean it before you paste. Free with an account.
Open the Academy →06Why you can trust the numbers
Every value is read from a source page and stamped with the date it was read. After 21 days without a fresh read, a price is hidden rather than shown stale. Newest reading: 2026-10-27 10:16 UTC.
How we read our data →07When a number is wrong
A company can claim its own listing with an email address at its own domain. When a figure here is wrong, the correction is published and the earlier value stays on the record rather than being quietly overwritten.
See the corrections record →
End of the trip
Where next?
Every layer, price and tool from the trip is one click away, or answer a few questions and get a stack back.
The ten layers · silicon up
Silicon first · count of listed productsA cross-section of the whole stack. The deepest coverage today sits at orchestration / agents and data / rag; silicon / hardware is thin. A layer opens at 12 listed products. Below that we say it is thin rather than pretend it is surveyed.
- 018Silicon / Hardware
- 02108Compute / Cloud
- 03144Model / Foundation
- 04186Inference / Serving
- 0554MLOps / Lifecycle
- 06189Data / RAG
- 07252Orchestration / Agents
- 0868Observability / Evals
- 0978Security / Governance
- 1029Applications
Resources: read, learn, plug in
All resources →Compare before you choose
Two GPUs, two models, or any two products in the same category, side by side, with the page every value was read from. 61 categories have a pair to compare.
From the blog
What is moving in AI, written up with the pages it was read from.
- LLM costs · 2026-10-27Batch pricing: when it fits and when it does notA batch tier prices tokens lower in exchange for a delayed answer, so it fits work nobody is waiting on and fails anything a person or a pipeline is blocked by.
- Renting GPUs · 2026-10-27Compare GPU providers fairly: same card, same terms, same unitA fair GPU provider comparison holds three things fixed: the exact card, the commitment and the unit. Then it adds what each rate leaves out.
- LLM costs · 2026-10-27How to estimate a monthly LLM billCount requests, measure average input and output tokens per request, multiply each by its rate per million, and add the hidden calls before trusting the total.
Questions we can answer right now
Each answer cites the page it was read fromTensorDock, at $2.25 per GPU-hour — the lowest of 13 sellers posting a rate for the H100 SXM 80GB. The spread across the rest is the part worth looking at.
See the spread →We hold no throughput benchmark, so we will not pick for you — but 47 inference engines are listed with their licence, their open-source health and what each one runs on.
Compare the engines →Only for some retrieval patterns — 27 are listed here, alongside the cases where a Postgres extension is enough and a separate service is not.
See what is listed →Cheapest model from each maker, by input rate. One per maker — ranking raw rates fills the card with one family.
$ PER MILLION TOKENS · INPUT THEN OUTPUT · READ FROM 16 SELLERS
H100 SXM 80GB · on-demand · what changed this week
EVERY PRICE LINKS TO THE PAGE IT WAS READ FROM · PRICES OVER 21 DAYS OLD ARE HIDDEN
Prices that moved this week
Full changelog →- OpenRouter2026-10-27deepseek/deepseek-v4-flash-0731 ▼ 73.0% — $0.440 to $0.119
- OpenRouter2026-10-27deepseek/deepseek-v4-flash-0731 ▼ 82.0% — $1.32 to $0.238
- OpenRouter2026-10-27deepseek/deepseek-v4-flash-0731 ▲ 70.0% — $0.014 to $0.024
- OpenRouter2026-10-27moonshotai/kimi-k3 ▲ 24.3% — $0.152 to $0.189
- OpenRouter2026-10-27moonshotai/kimi-k3 ▼ 42.5% — $20.00 to $11.50
Weekly. What moved in pricing, one tool worth knowing, one community benchmark. 500 words.
NOTHING SENDS YET — WE ARE ONLY SAVING ADDRESSES