That NDRC justification is the tell: electricity billed as 70% of operating costs, when construction and chips dominate the real bill. It's the same reason US data centers cluster in the exurbs next to big metros instead of out where power is cheapest. The technicians and the fiber matter more than the electrons.
EDWC is a king case study of companies doing whatever they were going to do anyway and then post-hoc gushing to the media about how this brilliant policy definitely guided their actions.
“injustice” may be too harsh a word. This sounds as if those poorer regions are stupid, but the reality is nice-sounding projects like these are what those regions very willingly embrace. If nobody “wastes” money on them, what are they left with? And if money and resources are taken away from rich regions and redirected towards poor regions, which party is actually on the receiving end of “injustice”?
Good to see you here, the blurb from Noah Pinon on the join membership page made me wonder if it was not another NED sponsored organ. Now I'll give it a few tries.
I rarely find myself pushing back on ChinaTalk, but I think its reading of China’s “Eastern Data, Western Computing” project misses something important.
In Chinese economic policy, “east” and “west” are not just geographic labels. The real goal is to move more data-center capacity away from crowded coastal cities and major economic hubs, where people and AI infrastructure are already competing for electricity and water.
In this context, “the west” often means less-developed regions, not simply China’s geographic west.
Overbuilding is a real risk. But the policy is not just about moving every data center westward. It is about shifting computing tasks that are less sensitive to latency into regions with cheaper land, energy, and resources, while using the national power grid to distribute pressure more evenly.
The key idea is not relocation for its own sake. It is a national system for matching different kinds of computing demand with the places best able to support them.
The governance parallel is the most underrated framing here. EDWC wasn't a failed energy policy — it was a successful local official incentive: build data center parks, show capital formation in GDP accounts, secure central approval, justify land acquisition. The mechanism is structurally identical to the property build-out: capital expenditure is visible and countable in the statistics that matter for cadre evaluation; utilization is abstract and deferrable.
The semiconductor chokepoint section is critical and undersold. A sub-50% rack rate on already-installed servers is one problem. The harder constraint is that the AI training chips needed to justify full-scale western compute hubs aren't obtainable at scale under current export controls — the policy was effectively designed before it was clear that domestic alternatives couldn't substitute for H100-class clusters at frontier workloads. Western provinces broke ground on infrastructure predicated on a compute stack that was already being foreclosed by export controls they had no visibility into.
The talent geography point is the most structurally durable. Data center operations require an ecosystem — HVAC technicians, cooling specialists, network engineers, experienced facilities managers — that formed organically around eastern tech clusters over two decades. Central mandates can fund server racks in Gansu; they cannot mandate that operational culture into existence.
At Blue Lotus Research we track where China's AI industrial strategy collides with hard physical and geopolitical constraints. This piece is one of the cleaner empirical illustrations we've seen of that collision. — bluelotus1618.substack.com
The governance parallel is the most underrated framing here. EDWC wasn't a failed energy policy — it was a successful local official incentive: build data center parks, show capital formation in GDP accounts, secure central approval, justify land acquisition. The mechanism is structurally identical to the property build-out: capital expenditure is visible and countable in the statistics that matter for cadre evaluation; utilization is abstract and deferrable.
The semiconductor chokepoint section is critical and undersold. A sub-50% rack rate on already-installed servers is one problem. The harder constraint is that the AI training chips needed to justify full-scale western compute hubs aren't obtainable at scale under current export controls — the policy was effectively designed before it was clear that domestic alternatives couldn't substitute for H100-class clusters at frontier workloads. Western provinces broke ground on infrastructure predicated on a compute stack that was already being foreclosed by export controls they had no visibility into.
The talent geography point is the most structurally durable. Data center operations require an ecosystem — HVAC technicians, cooling specialists, network engineers, experienced facilities managers — that formed organically around eastern tech clusters over two decades. Central mandates can fund server racks in Gansu; they cannot mandate that operational culture into existence.
At Blue Lotus Research we track where China's AI industrial strategy collides with hard physical and geopolitical constraints. This piece is one of the cleaner empirical illustrations we've seen of that collision. — bluelotus1618.substack.com
The 32% average utilization is the number that travels. It reframes this from a China-planning story into a global-overbuild one: the US is reaching similar empty-rack risk from the opposite direction, not a planner misreading power costs but vendor-financed capex racing ahead of proven demand, with Nvidia now helping fund the buyers of its own chips.
Excellent piece. One thing that really stood out to me is that AI infrastructure seems to behave much like previous industrial revolutions. Governments can encourage where it goes, but ecosystems still matter. Electricity and land are necessary, yet they're only part of the equation. Skilled labor, suppliers, networking, cooling, maintenance, and existing industrial capabilities appear to pull AI infrastructure toward places where those ecosystems already exist. It also made me wonder whether some of the western buildout is beginning to resemble the old "build it and they will come" approach to real estate development. Fascinating analysis.
Nice work! Agreed. I think it’s more about the migration of energy East, rather than of data West. Should be ‘Eastern Data, Western Energy,’ resembling the 2000s ‘West-to-East Power Transmission’ initiative
I like the interactive microsite, great maps!
That NDRC justification is the tell: electricity billed as 70% of operating costs, when construction and chips dominate the real bill. It's the same reason US data centers cluster in the exurbs next to big metros instead of out where power is cheapest. The technicians and the fiber matter more than the electrons.
EDWC is a king case study of companies doing whatever they were going to do anyway and then post-hoc gushing to the media about how this brilliant policy definitely guided their actions.
“injustice” may be too harsh a word. This sounds as if those poorer regions are stupid, but the reality is nice-sounding projects like these are what those regions very willingly embrace. If nobody “wastes” money on them, what are they left with? And if money and resources are taken away from rich regions and redirected towards poor regions, which party is actually on the receiving end of “injustice”?
Good to see you here, the blurb from Noah Pinon on the join membership page made me wonder if it was not another NED sponsored organ. Now I'll give it a few tries.
I rarely find myself pushing back on ChinaTalk, but I think its reading of China’s “Eastern Data, Western Computing” project misses something important.
In Chinese economic policy, “east” and “west” are not just geographic labels. The real goal is to move more data-center capacity away from crowded coastal cities and major economic hubs, where people and AI infrastructure are already competing for electricity and water.
In this context, “the west” often means less-developed regions, not simply China’s geographic west.
Overbuilding is a real risk. But the policy is not just about moving every data center westward. It is about shifting computing tasks that are less sensitive to latency into regions with cheaper land, energy, and resources, while using the national power grid to distribute pressure more evenly.
The key idea is not relocation for its own sake. It is a national system for matching different kinds of computing demand with the places best able to support them.
The governance parallel is the most underrated framing here. EDWC wasn't a failed energy policy — it was a successful local official incentive: build data center parks, show capital formation in GDP accounts, secure central approval, justify land acquisition. The mechanism is structurally identical to the property build-out: capital expenditure is visible and countable in the statistics that matter for cadre evaluation; utilization is abstract and deferrable.
The semiconductor chokepoint section is critical and undersold. A sub-50% rack rate on already-installed servers is one problem. The harder constraint is that the AI training chips needed to justify full-scale western compute hubs aren't obtainable at scale under current export controls — the policy was effectively designed before it was clear that domestic alternatives couldn't substitute for H100-class clusters at frontier workloads. Western provinces broke ground on infrastructure predicated on a compute stack that was already being foreclosed by export controls they had no visibility into.
The talent geography point is the most structurally durable. Data center operations require an ecosystem — HVAC technicians, cooling specialists, network engineers, experienced facilities managers — that formed organically around eastern tech clusters over two decades. Central mandates can fund server racks in Gansu; they cannot mandate that operational culture into existence.
At Blue Lotus Research we track where China's AI industrial strategy collides with hard physical and geopolitical constraints. This piece is one of the cleaner empirical illustrations we've seen of that collision. — bluelotus1618.substack.com
The governance parallel is the most underrated framing here. EDWC wasn't a failed energy policy — it was a successful local official incentive: build data center parks, show capital formation in GDP accounts, secure central approval, justify land acquisition. The mechanism is structurally identical to the property build-out: capital expenditure is visible and countable in the statistics that matter for cadre evaluation; utilization is abstract and deferrable.
The semiconductor chokepoint section is critical and undersold. A sub-50% rack rate on already-installed servers is one problem. The harder constraint is that the AI training chips needed to justify full-scale western compute hubs aren't obtainable at scale under current export controls — the policy was effectively designed before it was clear that domestic alternatives couldn't substitute for H100-class clusters at frontier workloads. Western provinces broke ground on infrastructure predicated on a compute stack that was already being foreclosed by export controls they had no visibility into.
The talent geography point is the most structurally durable. Data center operations require an ecosystem — HVAC technicians, cooling specialists, network engineers, experienced facilities managers — that formed organically around eastern tech clusters over two decades. Central mandates can fund server racks in Gansu; they cannot mandate that operational culture into existence.
At Blue Lotus Research we track where China's AI industrial strategy collides with hard physical and geopolitical constraints. This piece is one of the cleaner empirical illustrations we've seen of that collision. — bluelotus1618.substack.com
Thanks for writing this. Underwater and space data centers are the other siren songs.
The 32% average utilization is the number that travels. It reframes this from a China-planning story into a global-overbuild one: the US is reaching similar empty-rack risk from the opposite direction, not a planner misreading power costs but vendor-financed capex racing ahead of proven demand, with Nvidia now helping fund the buyers of its own chips.
Excellent piece. One thing that really stood out to me is that AI infrastructure seems to behave much like previous industrial revolutions. Governments can encourage where it goes, but ecosystems still matter. Electricity and land are necessary, yet they're only part of the equation. Skilled labor, suppliers, networking, cooling, maintenance, and existing industrial capabilities appear to pull AI infrastructure toward places where those ecosystems already exist. It also made me wonder whether some of the western buildout is beginning to resemble the old "build it and they will come" approach to real estate development. Fascinating analysis.
Nice work! Agreed. I think it’s more about the migration of energy East, rather than of data West. Should be ‘Eastern Data, Western Energy,’ resembling the 2000s ‘West-to-East Power Transmission’ initiative