What this is

aiinframap is an engineering reference for the physical infrastructure behind artificial intelligence — the electrons, optics, memory, silicon, systems, and models that every AI workload actually runs on. Most coverage stops at the apps. We map the stack underneath them, bottleneck by bottleneck, and keep every claim traceable to a source.

How we layer the AI industry — the 8-layer stack

Every company, topic, product, and signal is placed on one canonical model, L1 → L8, running from the power plant to the prompt:

Topics — 33 focused landscapes

Inside those layers we run 33 topic hubs, each a deep landscape on one part of the stack. A selection:

• L1 — Datacenter Power · Cooling · Buildout · Datacenter REITs · Nuclear & SMR Power · On-Site Power · Gas-Turbine Supply Crunch
• L2 — Optical Modules · AI Networking · InfiniBand vs Ethernet · CPO & Silicon Photonics · AEC & Retimers
• L3 — HBM Memory · Advanced Packaging (CoWoS / SoIC) · CXL Memory Fabric
• L4 — AI ASICs · AI Accelerator Startups · EDA & Chip-Design IP · Semiconductor Equipment · Power Semiconductors (SiC / GaN)
• L5 — GPU Cloud & NeoCloud · AI Data Stack · AI Servers & Rack Systems
• L6 — Frontier Model Labs · Robotics Foundation Models & Humanoids
• L7 — Vector Databases · AI Agent Frameworks & MCP · AI Eval & Observability · AI Gateways
• L8 — AI Coding Tools · Enterprise AI Adoption · AI Voice Agents

Companies — 100+ tracked

Each company is classified to its primary layer (with a watchlist note when it spans several). A sample of what we track: NVIDIA, AMD, Broadcom, Marvell, Groq, Cerebras, ASML, Cadence, Synopsys, Arm (L4) — SK hynix, Samsung, Micron, TSMC, ASE, Amkor (L3) — Coherent, Lumentum, Arista, Credo, Astera Labs, Ayar Labs (L2) — GE Vernova, Bloom Energy, Vertiv, Digital Realty, Oklo, Constellation Energy (L1) — CoreWeave, Lambda, Nebius, Supermicro, Dell, Scale AI (L5) — OpenAI, Anthropic, Google DeepMind, xAI, Mistral, DeepSeek, Physical Intelligence (L6) — Pinecone, Weaviate, LangChain, Arize, OpenRouter (L7) — Cursor, Cognition, ElevenLabs, Snowflake (L8). …and many more, each on its own profile page.

Products — the things they actually ship

At each layer we track the products that gate everything above them — and tie each one back to the companies and the sourced facts behind it: NVIDIA Blackwell B200 & GB200 NVL72 racks, Google TPU and AWS Trainium custom silicon, SK hynix HBM3e / HBM4, TSMC CoWoS-L 2.5D packaging, 800G → 1.6T optical transceivers, PCIe / CXL / UALink retimers, small modular reactors (Aurora, NuScale), heavy-duty gas turbines, and EUV lithography.

What the site gives you

• 8-Layer Map — the entire stack on one page.
• Topic Hubs — per-topic company rosters, sourced verified facts, and market briefings.
• Companies Database — every tracked company, layer-classified, with a profile.
• Leaderboards — ranked suppliers per layer, every row with sourced provenance.
• Direction Indices — commitment vs attention signals per layer and topic, plus forward-looking forecasts.
• Insights — cross-cutting industry synthesis and grounded long-form articles.
• Tools — engineering calculators: datacenter power, PUE, HBM lead-time, GPU TCO, inference cost, and more.
• Weekly — one engineer-focused briefing every Friday.
• Machine-readable — a JSON API and an llms.txt so agents and LLMs can read the map too.

The next 10 years — where the stack is heading

The defining pattern of this decade is that AI's binding constraint keeps migrating down the stack — and it is increasingly made of atoms, not bits. Our working view of 2025–2035:

1. Power becomes the constraint, not chips. The bottleneck has already moved from GPUs to megawatts. Grid-interconnect queues run years long, so behind-the-meter gas, fuel cells, and nuclear / SMR move from novelty to default. Single training sites cross the gigawatt line, and energy procurement becomes a core competency for anyone training at the frontier.
2. Custom silicon overtakes merchant GPUs — by volume. Hyperscaler ASICs (TPU, Trainium, MTIA, Maia) grow faster than merchant GPUs, and a healthy ecosystem of accelerator startups carves out inference. NVIDIA stays dominant in capability; the unit mix diversifies.
3. Memory and packaging stay the real chokepoint. HBM (HBM4 and beyond) and advanced packaging (CoWoS / SoIC, then glass substrates) gate how many accelerators actually ship. Whoever controls packaging capacity controls the ramp.
4. Optics goes in-package. Pluggable transceivers hit their reach and power limits; co-packaged optics and silicon photonics become the default interconnect as fabrics scale past 100k-GPU clusters and 1.6T → 3.2T.
5. Inference dwarfs training. As models deploy, inference becomes the majority of compute spend — pushing specialized inference silicon, gateways / routing, and cost-per-token to the center of the economics. Open-weight models close most of the gap with closed frontiers for everyday work.
6. Embodied AI leaves the demo stage. Vision-language-action models plus humanoid platforms move from lab videos to real deployments — the first AI layer that moves atoms, not just tokens.
7. The stack verticalizes and goes sovereign. Hyperscalers own more layers end to end, and nations stand up sovereign compute. The map of who controls which layer becomes as important as the technology in it.

The through-line: value and risk keep moving down the stack, toward the physical layer. That is exactly the layer aiinframap is built to map.

The idea

AI's binding constraint keeps moving — from GPUs to memory, to advanced packaging, to power and grid interconnects. aiinframap exists to track that supply chain end to end, with engineering rigor instead of hype: every number is sourced, every company is placed on the same stack, and every claim traces back to a citation. It is the reference layer for the people who build, supply, finance, and analyze AI infrastructure.

Engineering reference, not investment advice.