Montana AI Summit · September 2026

Foundations of AI

AI did not suddenly become smart in 2022. Capability followed data, compute, and a software stack that could use both.

The CubCloud AI briefing, on one page: seventy years of history, what is actually on a modern GPU, the software stack that sits on top of it, and the vocabulary to follow any AI conversation in the room.

Walk the TimelineDownload the Handout
History
1950 → today
Hardware
What's actually on a GPU
Software
Silicon up to the app

The through-line

AI did not suddenly become smart in 2022. Capability followed data, compute, and a software stack that could use both.

Part One

A Short Timeline

Seventy years of a research program that was named before it worked, failed twice in public, and came back when the hardware caught up.

  1. 1950Turing

    Ask whether a machine can behave as if it thinks. A test instead of a soul.

  2. 1956Dartmouth

    McCarthy names the field. The bet: intelligence described precisely enough to simulate.

  3. 1958Perceptron

    Learning as weighted connections — then a proof that one layer is not enough.

  4. 1970sFirst winter

    Overpromise and combinatorial explosion. Funding retreats.

  5. 1980sExpert systems

    Useful in narrow domains, brittle at scale. A second winter follows.

  6. 1986–99Backprop + stats

    Deep nets become trainable in principle. Speech and translation go probabilistic.

  7. 2012AlexNet

    ImageNet + better algorithms + two consumer GPUs. Vision changes in a year.

  8. 2017+Transformer & scale

    Attention makes language GPU-friendly. Pretrain once; adapt many times.

Five Things to Keep

If you remember nothing else from the briefing, remember these.

  1. 01
    A 70-year program, not a product.
    Named in 1956, failed twice, returned when data and compute arrived.
  2. 02
    Learn the function if you cannot write the rules.
    Modern systems fit patterns from examples. Rules still help — they no longer carry everything.
  3. 03
    GPUs won because the math is linear algebra.
    Thousands of simple cores do matrix multiplies. CUDA turned a chip lead into a platform.
  4. 04
    Software is layered on purpose.
    CUDA → libraries → PyTorch → trainers → servers. Each layer buys speed and hides the layer beneath it.
  5. 05
    The live constraints are physical.
    Energy, memory bandwidth, interconnect, data quality, and people who can run the machines.
Part Two

What's on a Modern GPU

Four pieces decide how fast a model trains and how much a token costs to serve.

Tensor Cores
Mixed-precision matrix units (FP16, BF16, FP8, FP4). This is the throughput.
HBM
Stacked high-bandwidth memory. If weights live too far away, the cores starve.
Interconnect
NVLink in the rack; InfiniBand or Ethernet across racks. Big models are distributed computers.
The rack
The unit of AI is no longer a card. It is a cooled rack with a power budget.
Part Three

The Software Stack

Every layer buys speed and hides the one beneath it. Read it top-down to see a product; bottom-up to see the machine.

Apps & productsChat, search, agents, copilots, science tools
ServingvLLM, SGLang, TensorRT-LLM, Kubernetes, queues
Training frameworksPyTorch, JAX, TorchTitan, DeepSpeed, Megatron
Libraries & compilerscuDNN, NCCL, Transformer Engine, Triton, torch.compile
RuntimeCUDA / ROCm / XLA — talk to the device, allocate memory, launch kernels
HardwareGPU / TPU / interconnect / HBM
Pocket Glossary

The Words You'll Hear

Eight terms that cover most of what gets said on stage and in the hallway.

Token
A chunk of text (often a word piece) the model reads and writes.
Parameter / weight
A number the model learned. Modern models have billions to trillions.
Training
Adjust weights so predictions match targets. A long distributed job.
Inference / serving
Use a trained model on new requests. Latency and cost per token matter.
Attention
Every token can look at every other token. Powerful; costly as context grows.
KV cache
Memory of past tokens during generation. Often what fills the GPU.
CUDA
NVIDIA's programming model. Why most research still starts on their chips.
Quantization
Store weights in fewer bits (8, 4) so more model fits and each token costs less.
Built in Missoula

The constraints are physical. So is the infrastructure we run.

CubCloud AI operates dedicated H200, H100, and RTX 6000 Blackwell racks in Montana — the energy, bandwidth, and people side of this talk, in a building you can visit.

Talk to Our TeamSee the HardwareHandout PDF