Microsoft Copilot
AI across Office, Windows & GitHub
🇺🇸US
Compute & InfrastructureDeep Dive
The AI Stack: A Core Sample of the Value Chain
Every chatbot reply, and every humanoid's next step, sits on eleven layers of dependencies, from the apps people touch down to the lithography machines that print the silicon. Here is who owns each one, and where they are based.
Olympia Tech8 min read
Ask where artificial intelligence "happens" and most answers stop at the surface: a chat box, an agent, a humanoid taking its first steps. But every one of those moments rests on a deep column of dependencies. Pull a core sample straight down through the AI economy and you pass through eleven distinct strata, each owned by a different set of companies, each concentrated in a different part of the world.
The higher layers are crowded and competitive: applications, agents, model labs, the data plumbing that feeds them. The deeper you drill, the narrower the rock gets. By the time you reach the bedrock, the machines that print transistors and the materials they print on, the whole industry rests on a handful of firms, several of them with no substitute anywhere on Earth.
That shape, wide at the top and pinched at the bottom, is the single most important fact about the AI supply chain. It is why export controls, subsidies and national-security fights all converge on the same few names. Scroll down through the core sample below: the depth gauge tracks where you are, from the applications people touch to the lithography that makes them possible.
11
Layers of the stack
80+
Dominant players
12
Countries involved
01/11Applications & Agents0 – 1 km
Applications & Agents
The surface layer: assistants, copilots and enterprise software wrapped around models people use every day.
What people actually touch
Google Workspace
Gemini in Docs, Gmail & search
🇺🇸US
Salesforce
Agentforce · CRM workflows
🇺🇸US
Adobe
Firefly & creative AI tooling
🇺🇸US
ServiceNow
Enterprise workflow agents
🇺🇸US
Cursor
AI-native code editor
🇺🇸US
Perplexity
AI answer & search engine
🇺🇸US
Notion AI
AI inside docs & workspaces
🇺🇸US
Canva
Mass-market design AI
🇦🇺AU
Embodied AI & Robotics
The emerging value layer: systems that perceive, plan and act: humanoids, autonomy and physical automation.
AI that acts in the world
Tesla
Autonomy & Optimus humanoid
🇺🇸US
Figure AI
General-purpose humanoids
🇺🇸US
Boston Dynamics
Advanced mobile robotics
🇺🇸US
NVIDIA Isaac
Robotics simulation & stack
🇺🇸US
Unitree
Low-cost quadrupeds & humanoids
🇨🇳CN
UiPath
Enterprise agentic automation
🇺🇸US
Foundation Model Labs
The labs that pre-train and align the large models everything else is built on.
Who trains the frontier
OpenAI
GPT family · ChatGPT
🇺🇸US
Anthropic
Claude family of models
🇺🇸US
Google DeepMind
Gemini · UK-rooted, US-scaled
🇬🇧GB
Meta AI
Llama open-weight models
🇺🇸US
xAI
Grok models
🇺🇸US
Mistral AI
Europe's open-weight champion
🇫🇷FR
DeepSeek
Efficient frontier models
🇨🇳CN
Alibaba Qwen
Qwen open model family
🇨🇳CN
Cohere
Enterprise-focused LLMs
🇨🇦CA
Data, Vector DBs & MLOps
The operational tissue: data platforms, labeled data, model hubs and the vector search behind retrieval.
Train · serve · retrieve
Hugging Face
The hub for open models
🇺🇸US
Databricks
Data + AI platform (Mosaic)
🇺🇸US
Snowflake
Data cloud for AI workloads
🇺🇸US
Scale AI
Data labeling & evaluation
🇺🇸US
Pinecone
Managed vector database
🇺🇸US
Weaviate
Open-source vector search
🇳🇱NL
LangChain
Framework for agents & apps
🇺🇸US
Cloud & Compute
Hyperscalers and neoclouds renting out the GPU capacity that training and inference demand.
Where the models run
Amazon AWS
Largest cloud · custom silicon
🇺🇸US
Microsoft Azure
Cloud home of OpenAI
🇺🇸US
Google Cloud
TPU & GPU compute
🇺🇸US
Oracle Cloud
Fast-growing AI capacity
🇺🇸US
CoreWeave
GPU-first 'neocloud'
🇺🇸US
Alibaba Cloud
China's leading cloud
🇨🇳CN
Tencent Cloud
Major Chinese hyperscaler
🇨🇳CN
Power, Cooling & Data Centers
Increasingly the binding constraint: power delivery, dense-rack cooling and the real estate that houses gigawatt clusters.
The physical plant
Schneider Electric
Critical power & energy mgmt
🇫🇷FR
Vertiv
Cooling & power for dense racks
🇺🇸US
Eaton
Electrical distribution & backup
🇮🇪IE
Digital Realty
Global data-center real estate
🇺🇸US
Equinix
Colocation & interconnect hubs
🇺🇸US
Networking & Interconnect
At cluster scale the network is the computer; switches and links that bind thousands of GPUs into one machine.
Moving data between chipsNVIDIA
NVLink · InfiniBand (Mellanox)
🇺🇸US
Broadcom
Merchant switch silicon
🇺🇸US
Arista Networks
AI data-center ethernet
🇺🇸US
Cisco
Networking for AI fabrics
🇺🇸US
Marvell
Optical & custom interconnect
🇺🇸US
AI Accelerators & Processors
The fabless designers and ASIC owners behind the GPUs, TPUs and accelerators that do the math. NVIDIA still dominates, but AMD's GPUs and Google's TPUs are now genuine challengers, and big labs deliberately run several in parallel for leverage.
The brains · chip designNVIDIA
~80%+ of AI accelerators · CUDA moat
🇺🇸US
AMD
Credible #2 · MI450 first to 2nm · OpenAI 6GW deal
🇺🇸US
Google TPU
Merchant ASIC · trains Gemini & Claude · Anthropic, Meta
🇺🇸US
AWS Trainium
Captive cloud silicon · Anthropic co-designs
🇺🇸US
Intel
Gaudi accelerators & server CPUs
🇺🇸US
Broadcom
Co-designs custom ASICs for hyperscalers
🇺🇸US
Cerebras
Wafer-scale inference engines
🇺🇸US
Groq
Low-latency inference LPUs
🇺🇸US
Huawei Ascend
China's domestic accelerator stack
🇨🇳CN
Arm
CPU IP in nearly every chip
🇬🇧GB
Qualcomm
On-device / edge AI
🇺🇸US
Memory & Storage
AI is bottlenecked by memory bandwidth. A handful of firms make essentially all the HBM, DRAM and flash.
Feeding the accelerators
SK hynix
HBM leader · NVIDIA supplier
🇰🇷KR
Samsung
HBM, DRAM & NAND at scale
🇰🇷KR
Micron
Only US HBM maker
🇺🇸US
Kioxia
Major NAND flash producer
🇯🇵JP
Western Digital
Enterprise storage for AI data
🇺🇸US
Foundries & Advanced Packaging
Designs are nothing without fabrication, and packaging (CoWoS) is now its own first-order AI bottleneck.
Who builds the chips
TSMC
Makes ~all leading-edge AI chips
🇹🇼TW
Samsung Foundry
#2 advanced-node foundry
🇰🇷KR
Intel Foundry
US leading-edge ambitions
🇺🇸US
GlobalFoundries
Mature-node specialist
🇺🇸US
SMIC
China's top foundry
🇨🇳CN
ASE Technology
Assembly, test & packaging
🇹🇼TW
Amkor
Outsourced advanced packaging
🇺🇸US
Equipment & Materials
The deepest layer: the lithography, etch and metrology machines (and the ultra-pure materials) that make fabs possible.
The bedrock
ASML
Monopoly on EUV lithography
🇳🇱NL
Applied Materials
Deposition & etch leader
🇺🇸US
Lam Research
Etch & deposition systems
🇺🇸US
Tokyo Electron
Coaters, etch & cleaning
🇯🇵JP
KLA
Inspection & metrology
🇺🇸US
Carl Zeiss
Optics inside EUV machines
🇩🇪DE
Shin-Etsu
Silicon wafers & photoresist
🇯🇵JP
Why this map is geopolitics
The whole stack hangs on a few chokepoints
What makes this chain fragile isn't its length, it's its concentration. Several layers narrow to a single company or country with no real substitute, which is exactly where export controls, subsidies and national-security fights now land. Lose any one of these nodes and everything above it stalls.
EUV lithography · 1 firm · 🇳🇱 ASMLLeading-edge logic · ~1 fab · 🇹🇼 TSMCHBM memory · 3 firms · 🇰🇷 🇺🇸AI training GPUs · ~1 firm · 🇺🇸 NVIDIAAdvanced packaging · scarce · 🇹🇼 🇺🇸
Logos fetched live from each company's domain; flags denote primary headquarters / origin. A company may appear in several layers, that's vertical integration: NVIDIA spans accelerators, networking & robotics; Google spans models, cloud & chips; Samsung spans memory, foundry & packaging; Amazon spans cloud & silicon.
🇺🇸 United States🇹🇼 Taiwan🇰🇷 South Korea🇳🇱 Netherlands🇯🇵 Japan🇨🇳 China🇬🇧 United Kingdom🇫🇷 France🇨🇦 Canada🇩🇪 Germany🇮🇪 Ireland🇦🇺 Australia
Read the column from bottom to top and the strategic picture inverts. The apps are interchangeable; the bedrock is not. A new model lab can be funded in a quarter, but a second source for EUV lithography does not exist at any price, and a leading-edge fab takes the better part of a decade and tens of billions of dollars to stand up.
That is what to watch over the next few years. Not which assistant wins the consumer, but whether the chokepoints widen or narrow: how fast advanced packaging capacity comes online, whether HBM stays a three-horse race, and how much of the bedrock any one country can afford to control. The surface layer gets the headlines. The value, and the risk, is in the rock underneath it.
The model layer is fighting its own version of this battle: open-weight labs, several of them deliberately training on non-US silicon, have closed to within months of the frontier. We measure that race in Closing the Gap.
More from Compute & Infrastructure
Deep DiveCompute & Infrastructure
Seven Bets on a Budget Nobody Has Passed
Europe has put up to seven AI gigafactories out to tender, at €30 billion-plus, with 100,000 chips apiece at the largest sites. Only €1 billion of the EU's own money is legally committed. The rest waits on a budget that does not exist yet.
Olympia Tech · 9 min