00Opening read
The physical AI economy has spent the past year discovering its bottlenecks. Today’s signal is that the bill is arriving.
Victoria wants new data centers to secure their own renewable power and storage, pay for grid connections and cover network upgrades. Bond investors are demanding more compensation from AI borrowers as hyperscalers return to debt markets again and again. India’s Lohum is looking to buy nickel mines rather than rely only on recycled material.
China is responding differently. Alibaba is trying to own more of the stack itself, from models and AI chips to data centers.
These are not separate stories. They mark a change in how the buildout works.
Power, financing, materials and supply security are moving from external conditions into the project itself. The next winners may not be the companies with the biggest capacity announcements. They may be the ones that can carry the cost of their own constraints.
01What moved
Data centers may have to become power developers
Victoria has proposed some of Australia’s clearest rules yet for new data centers.
Developers would need to secure renewable generation and storage, pay their connection costs and fund required grid upgrades. They would also face tighter siting and water rules.
The significance goes beyond Australia. If this model spreads, the economics of a data center change. Compute developers would increasingly need the balance sheet and expertise to develop power infrastructure at the same time.
That favors scale.
Credit markets are getting less generous with AI
AI stocks are rallying again. The debt market is more cautious.
Reuters reports that investors are demanding wider spreads from AI-related borrowers as the volume of infrastructure debt grows. Goldman Sachs expects hyperscaler issuance to reach a record $420 billion next year.
The concern is not that the largest technology companies are about to default. It is that investors are being asked to absorb enormous amounts of debt without perfect visibility into the returns generated by the infrastructure behind it.
Capital is still available. It is becoming more expensive to assume that it always will be.
Alibaba builds deeper into the physical stack
Alibaba unveiled a new AI processor, the Zhenwu V900, and says it plans to put the chip into mass production in early 2027.
The company is also targeting more than 20 GW of global data-center capacity by 2032.
That combination matters more than the model-size headline. China’s technology restrictions are pushing companies toward vertical control: chips, models, cloud infrastructure and deployment inside the same ecosystem.
The AI network becomes an investment category
Ligent Technologies raised about $723 million in Hong Kong and finished its first trading day above its offer price.
Ligent makes optical transceivers and related equipment. Those components move data between the servers and switches inside AI infrastructure.
As compute density rises, moving data efficiently becomes a constraint of its own.
The AI trade is spreading from the processor into the network around it.
India’s mineral strategy moves upstream
Lohum built its critical-minerals business around recycling. Now it is looking for nickel mines in Indonesia and the Philippines.
The company wants to increase annual nickel capacity tenfold over the next 18 months and is raising capital to support its plans.
That shift is revealing. Recycling can reduce dependence on imported raw materials, but it cannot always provide enough feedstock for rapid growth.
Resource security eventually reaches the mine.
02The throughline
The constraint stack is becoming a responsibility stack.
Who supplies the power? Who pays for the grid connection? Who finances the next data-center campus? Who owns the mineral feedstock? Who qualifies the AI system before it enters a factory, vehicle or aircraft?
Those questions used to sit around the edge of the AI story. They are moving toward the center.
A useful way to track the buildout now is not only by announced capacity, but by constraint ownership.
Projects become more credible when the difficult dependencies already have an owner, a budget and a timeline.
That may be the difference between a gigawatt on a slide and a gigawatt that actually turns on.
03What to watch
- September 24: Trump and Xi are expected to meet in Washington. AI, export controls and critical minerals remain connected negotiating surfaces.
- Q1 2027: Alibaba says its Zhenwu V900 AI chip will enter mass production and commercial release.
- 2027: hyperscaler debt issuance could reach $420 billion, making credit-market appetite an increasingly important infrastructure signal.
- Next 12-18 months: Lohum plans to scale nickel capacity tenfold while pursuing mine acquisitions in Southeast Asia.
04Source trail
Reuters, Victoria data-center power rules
https://www.reuters.com/business/energy/australias-victoria-proposes-new-data-centres-must-secure-their-own-renewable-2026-09-22/
Reuters, AI debt markets
https://www.reuters.com/legal/transactional/corporate-bond-buyers-get-picky-with-flood-ai-debt-2026-09-22/
Reuters, Alibaba AI chip and data-center capacity
https://www.reuters.com/business/retail-consumer/alibaba-plans-ai-model-with-5-trillion-10-trillion-parameters-unveils-new-chip-2026-09-22/
Reuters, Taiwan export orders
https://www.reuters.com/world/asia-pacific/taiwan-august-export-orders-hit-record-high-value-strong-ai-demand-leads-2026-09-22/
Reuters, Ligent Hong Kong listing
https://www.reuters.com/world/asia-pacific/ligent-technologies-jumps-hong-kong-debut-outpacing-key-indices-2026-09-22/
Reuters, Lohum nickel strategy
https://www.reuters.com/world/india/indias-lohum-seeks-buy-nickel-mines-indonesia-philippines-2026-09-22/
Reuters, Dassault Rafale AI tests
https://www.reuters.com/business/aerospace-defense/dassault-aviation-tests-ai-algorithms-rafale-fighter-jet-2026-09-22/