GateKPT - AI from the physical layer up.
A public research terminal for the AI stack: power, chips, data, models, software, testing, and business context.
L01 - Power
AI runs on electricity. Getting it takes years.
5+yr - Median wait to connect a new project to the US grid.
Berkeley Lab, Queued Up 2026
A data center takes about 18 months to build. The power connection takes longer.
L02 - Chips
Chips are fast. Getting data to them is slow.
3TB/s - Memory bandwidth on one NVIDIA H100 chip.
NVIDIA, H100 overview
The chip finishes its math early and waits on memory.
L03 - Data
A model can only learn what its data contains.
60% - AI projects Gartner expects to be dropped by 2026 without good data.
Gartner, AI-ready data risk
Collecting data is cheap. Deciding what is correct is not.
L04 - Models
Building the model is one cost. Running it is the recurring one.
40GB - Memory used by one long request to Llama 3 70B.
NVIDIA, KV cache offload
Everyone quotes the build cost. The bill comes from running it.
L05 - Software
The model is one part. The software around it decides if it works.
1st - Search is the first thing to check when an answer comes back wrong.
RAG systems review, 2025
If the right page was never found, no prompt can fix it.
L06 - Testing
If you cannot test it, you cannot trust it.
4steps - NIST's four steps for AI risk: govern, map, measure, manage.
NIST AI Risk Framework
It worked last week, the model updated, and nobody can prove either part.
L07 - Business
A tool nobody uses has no value.
16% - AI projects IBM says have scaled across a whole company.
IBM, AI data quality
The tech worked. Then it changed hands and the goal changed.