⬡ Layer 1 — Physical & Frontier Foundation

Detailed view of the deepest layer only — cables, power, cooling, GPU factories, and how capacity flows upward
1A — BACKBONE Global Connectivity How bits physically cross oceans and continents
Submarine Cables
Undersea fiber mesh
Landing Stations
Ocean meets land
▼
1B — CAMPUSES Hyperscale Data Centers The buildings, power, and cooling that house everything
Hyperscale Campuses
The physical buildings
Power Delivery
The limiting resource
Cooling Systems
Removing the heat
▼
1C — FRONTIER COMPUTE GPU Clusters & Fabrics Where training and high-end inference actually run
GPU Clusters
The accelerators
Scale-Up Fabric
Inside the domain
Scale-Out Fabric
Rack to building
▼
1D — TRAJECTORY Where Layer 1 Is Going AI factories, denser racks, optical scale-up, multi-GW campuses
Frontier Trajectory
Next 2–5 years
↑ Upward flow — Capacity from this entire layer is what gets exposed as cloud regions, managed services, and model APIs in Layer 2 and above. Layer 2 itself is not shown here.
Glossary scroll for all terms used on this map
Accelerator
A chip specialized for AI math (usually a GPU). Faster at the matrix operations LLMs need than a general CPU.
Activation
The intermediate numbers produced by each layer of a neural network during a forward pass. Moved between GPUs during training.
AI Factory
A data-center campus designed from the ground up for dense AI compute — power, cooling, and networking co-optimized for multi-hundred-MW to multi-GW GPU clusters.
Bandwidth
How much data can move per second (Gbps / Tbps). Higher bandwidth keeps GPUs fed instead of waiting.
Coherent Optics
Advanced fiber technology that encodes data on the phase and amplitude of light, enabling very high capacity per wavelength.
Co-packaged Optics (CPO)
Putting optical engines next to the switch or GPU silicon to cut power and increase density versus traditional pluggable modules.
Data Parallelism
Training technique: each GPU (or group) sees different data batches, then they average gradients so the whole cluster learns one model.
East-West Traffic
Data moving sideways between servers/racks inside a data center (as opposed to north-south traffic to users). Dominates AI training.
Gradient
The signal that tells each weight how to change during training. Must be synchronized across GPUs.
Hyperscaler
A company that runs enormous cloud/AI campuses at global scale (Amazon, Google, Microsoft, Meta, etc.).
Inference
Running a trained model to get an answer right now (as opposed to training, which builds the model).
InfiniBand
A high-performance network technology widely used for scale-out in AI and HPC clusters. Low latency, high bandwidth, RDMA support.
Landing Station
Building where an undersea cable comes ashore and connects to terrestrial fiber networks.
Latency
How long a message takes to arrive. Critical inside scale-up domains (hundreds of nanoseconds) and still important for scale-out.
Liquid Cooling
Using liquid (usually water or dielectric fluid) to remove heat directly from chips. Required once racks exceed ~40–50 kW.
Multi-core Fiber
Optical fiber with multiple independent cores in one strand, multiplying capacity without a thicker cable.
NVLink / NVSwitch
NVIDIA’s high-speed interconnect that lets GPUs inside a domain talk to each other at terabit rates — the heart of scale-up.
NVL72
A dense NVIDIA rack design that packs 72 GPUs with very high internal bandwidth; typical power ~100–130+ kW.
Pipeline Parallelism
Training technique: different layers (or stages) of the model live on different GPUs and the data flows through them in a pipeline.
PUE (Power Usage Effectiveness)
Total facility power ÷ IT equipment power. 1.1 means only 10% overhead for cooling and distribution. Lower is better.
RDMA
Remote Direct Memory Access — lets one GPU write into another GPU’s memory with almost no CPU involvement. Keeps scale-out fast.
Scale-Across
Extending the cluster network between buildings or campuses when one building can no longer hold the whole job.
Scale-Out
Networking that connects many racks/buildings so they act as one large cluster. Higher latency than scale-up, still very fast.
Scale-Up
Networking that makes a small group of GPUs (a domain) behave like one giant accelerator. Ultra-low latency, ultra-high bandwidth.
Submarine Cable
Undersea fiber-optic cable that carries >99% of intercontinental internet traffic.
Tbps / Pbps
Terabits / Petabits per second. Units of network capacity. Modern cables are measured in hundreds of Tbps; large fabrics in Pbps.
Tensor Parallelism
Training technique: individual weight matrices are split across GPUs that must talk to each other constantly — requires scale-up fabric.
Terrestrial Fiber
Fiber-optic cables that run on land (between landing stations and data centers, or between cities).
Training
The process of adjusting a model’s weights on huge datasets so it learns. Extremely compute- and communication-heavy.
Wavelength-Division Multiplexing (WDM)
Sending many different colors (wavelengths) of light down the same fiber at once, multiplying capacity.
How to read this: This is a zoomed-in view of only Layer 1 from the main pyramid. Click any card to light related pieces and see real scale numbers, how the systems connect, and how capacity eventually flows upward.