Working paper · AI infrastructure
AI Compute: Bottlenecks and Alternatives
Which chips and software can train and run AI models when cost, supply or power is a constraint? We compare alternative hardware and software options by the workloads they can serve, their total cost and the practical limits of deploying them.
Which hardware and software choices relieve the bottlenecks?
AI computing capacity depends on more than the chip. We examine chip manufacturing, advanced packaging (how chips and memory are assembled), high-bandwidth memory (HBM), networking, software support and available electricity. An option is useful only if it can run the intended workload and be deployed at the scale needed.
The investigation asks where supply becomes constrained, which alternatives are credible, and whether a substitute removes a bottleneck or creates another dependency. Comparing purchase prices alone misses software migration, utilisation and operating costs.
Shared knowledge and the AI-agent team are the research foundation for this work. The research method covers comparisons with our team's existing drafts and new questions raised by changes in the market.
Fast-changing claims need dates and source boundaries.
- primary records separated from attributed estimates
- vendor interest and commercial relationships disclosed
- time-stamped claims and revision history
- uncertainty preserved when capacity or pricing cannot be verified
Research is separate from commercial offers.
Ancapex also has a separate commercial business, the OTC Compute Desk, which arranges access to computing capacity. Its current offers do not determine Research conclusions. A technology discussed here is not necessarily available through the desk; availability and prices require separate verification.
Have a source, contradiction or correction?