Executive summary
Training and running large AI models requires enormous amounts of specialised computing hardware. This has moved part of the technology industry from asset-light software economics toward capital-intensive infrastructure economics, changing where profits and risks concentrate.
Key findings
- 1AI workloads favour specialised processors designed for massively parallel calculations.
- 2Designing chips, manufacturing them and operating them in data centers are distinct businesses with very different capital needs.
- 3As computing becomes a major cost, the economics of AI products depend heavily on hardware efficiency and utilisation.
From software margins to infrastructure costs
For two decades, much of technology's appeal came from software economics: build a product once and sell it many times at very little extra cost. Artificial intelligence changes part of that equation. Every model has to be trained, and every answer it produces has to be computed — and both require large amounts of specialised hardware.
Why specialised chips
AI models rely on huge numbers of simple calculations performed in parallel. Processors designed for that kind of work can be far more efficient at it than general-purpose chips. That has created intense demand for accelerators, high-bandwidth memory and the networking that links thousands of chips together.
A chain of very different businesses
- Chip designers create the architecture and software ecosystem.
- Foundries manufacture chips in fabrication plants that cost enormous sums and take years to build.
- Equipment makers supply the machines that foundries need.
- Cloud and data-center operators buy the hardware and rent computing capacity to customers.
Each link has its own capital needs, competitive dynamics and risks.
Utilisation is everything
When hardware is expensive and depreciates as newer generations arrive, the key question is how fully it is used. Idle capacity destroys returns; fully utilised capacity can be highly profitable. This is an infrastructure question as much as a technology one.
The investor's lens
Instead of asking which company announced the fastest chip, a more durable question is: who has pricing power along this chain, and who is carrying the capital risk?
Risks and limitations
This is an educational overview of structural trends, not a view on any company or security. Technology cycles can shift quickly and current leaders are not guaranteed to remain so.
Facts, assumptions, estimates and opinions
TIC labels the nature of each material claim so you can weigh it appropriately.
- Verified factSemiconductor design and semiconductor manufacturing are frequently carried out by different companies.
- OpinionThe shift toward capital-intensive AI infrastructure makes balance sheets and capacity planning more important to technology investors than in the software era.
Sources and references
This piece relies on arithmetic and general principles; no external data sources are cited.
Downloads
A downloadable PDF edition of this research has not been published yet.