Teravolt looks to cannibalize older industries to meet AI power demand
- ID
- 24207
- Status
- summarized
- Published
- 14 Sep 2026, 9:45 PM
- Fetched
- 14 Sep 2026, 10:19 PM
- Provider
- The Register
- Category
- technology
- Original URL
- https://www.theregister.com/ai-and-ml/2026/09/14/teravolt-looks-to-cannibalize-older-industries-to-meet-ai-power-demand/5296014
- Source URL
- https://www.theregister.com/headlines.atom
Summary
- Score
- 6.5
- Created
- 14 Sep 2026, 10:22 PM
- Tags
- Audience
- developersai_ml_learnerssaas_founders
What happened
London-based AI infrastructure company Teravolt argues that repurposing existing industrial sites—bitcoin farms, aluminum smelters, old thermal power plants—is faster and more lucrative than building new grid infrastructure for AI datacenters. Gartner projects global datacenter power demand rising from 104 GW (2025) to 290 GW by 2030, while Teravolt forecasts a 240 GW shortfall by 2036 against 410 GW of AI demand. The economics are stark: an aluminum smelter generates $170-190/MWh gross revenue, while the same site as an AI datacenter yields $450-900/MWh with ~$300 EBITDA.
Why it matters
If AI compute capacity is constrained by grid buildout timelines (5-15 years) rather than datacenter construction (1-3 years), cloud costs for inference and training could rise or shift to regions with surplus power. Malaysian builders should watch whether Malaysia's industrial-site repurposing and grid capacity in Johor and elsewhere positions it as a net winner or loser in this reallocation, since the article's logic implies AI workloads will migrate to wherever spare energy exists.
Discussion angle
The aluminum-smelter-to-datacenter revenue math ($170-190/MWh vs $450-900/MWh) is a concrete signal of how aggressively old industrial sites will be cannibalized—does Malaysia's industrial base and grid situation make it a likely destination for this repurposing wave, or do planning and permitting timelines here match the 5-15 year bottleneck described?