The compute supply chain
Compute is only as available as the slowest thing behind it. Chips need packaging and memory. Servers need networking. Sites need cooling, transformers and a grid connection. This map follows the chain, and also the new ideas trying to change it.
118 entries so far. Each says what the thing is, where it sits, what limits it where we found a source, and links to the source. Constraints are labelled as reported where they rest on trade press. We do not rank, we do not show prices, and nothing here is investment advice.
Related developments
- Intel says Microsoft will use its EMIB packaging for Azure chips and confirms multi-year Xeon supply
- Intel opinion piece outlines its D-series GPUs for AI and scientific computing
- Powell Industries reports a data centre order of about US$85 million
- FuelCell Energy announces Connecticut testbed for AI data-centre power systems
Earlier related developments (7)
- Vertiv signs three-year Infineon supply agreement for power semiconductors
- Nokia reports rising sales and orders from AI and cloud customers
- NVIDIA's Berlin announcements: compute source guide
- Vertiv to add about 22,000 square metres of manufacturing space in Slovakia
- Crusoe cancels a $1.25 billion gas turbine order from Boom Supersonic
- Meta announces Petal, a 1 petabit per second cable between France and the United States
- NVIDIA launches DSX Ready to qualify power and cooling products for AI factories
These news accounts have their own sources, dates and limits. They do not refresh the older evidence on this page.
Repair, reuse and retirement evidence
Three reviewed starting points distinguish guidance, a named grid-asset obligation and an operator-reported system retirement. A retirement event does not establish disposal, recovery or whole-site closure/restoration.
- Server component reuse and its practical constraints · Stage 10; Historical technical guidance.
- A grid licence requires a named line span to be removed · Stage 11; Issued grid-asset licence condition.
- Operator-reported system retirement: ASPIRE 1 · Stage 11; Operator-reported system retirement.
Certified refurbishment, secondary-GPU brokerage, municipal data-centre closure, bonds and stranded-asset liability remain unassessed here.
Chips and chipmaking
Designers, foundries, memory, packaging and the equipment that makes them.
Servers, networking and optics
The machines, racks, switches and links that turn chips into a usable cluster.
Servers and racks
Networking
Standards
Buildings, cooling and electrical equipment
Cooling, switchgear, transformers and the operators who build and run sites.
Heat reuse
Power
On-site generation, storage, nuclear projects and the grid connections that sites wait for.
Electrical
On-site power
Connectivity
Subsea cables and satellite links that decide where compute can sit.
Subsea cables
Satellite
Innovation and alternatives
New kinds of compute and new ways to organise it, with an honest view of how mature each is.
Photonic
Neuromorphic
In-memory and wafer-scale
Quantum
Edge
Space-based
Open hardware
Sovereign and public programmes
How we label evidence
- Official source the body that made the rule or the data says so
- Two sources agree independent sources match
- One source, not confirmed a single report or claim
- Watching a signal, not shown as a fact