What's different from 2025 to 2026 is the AI Infrastructure local opposition. While BigTech tries to bury this news, it's a huge deal-breaker. Think Datacenter projects stalling in multiple U.S. states!
Local Opposition and Regulation Community pushback has become a "material driver of attrition" of datacenter rollouts. In 2025 and early 2026, roughly 30% to 50% of large-scale projects faced delays or cancellations due to:
Water Usage: Large campuses can consume billions of liters of water annually for cooling, triggering moratoriums in drought-prone areas.
Zoning Changes: Towns are increasingly wary of "data center alleys" and are implementing stricter noise, aesthetic, and environmental impact reviews.
The health and noise factors for local community residents near AI Infrastructure is a growing and serious concern. Higher electricity prices are also a bottleneck while the U.S. experiences a growing affordability crisis.
AI is political. Because the growing disparity and wealth inequality involved is so large. Young people will likely make this a political issue for the 2028 Presidential elections.
If it wasn’t for more recent geopolitical conflict these “supply-chain chokepoints” wouldn’t be seen as problematic at all. Many arise precisely because it is most efficient to let some firms / countries take over certain industries such as mineral refining (which is low-margin and dirty).
The Chinese prototype was built by salvaging components from older ASML machines through secondary markets. This proof-of-concept is no longer viable at scale because China no longer has access to the critical Western inputs that make ASML possible. And as of today, it hasn't produced a single working chip.
Meanwhile ASML just certified its High-NA EUV line for high-volume manufacturing, shipping $400M machines to Intel, Samsung, and SK Hynix for sub-2nm production. China needs to independently recreate not just the machine but the entire ecosystem underneath it: Zeiss precision optics, high-power lasers, photomasks, resists, multilayer mirrors. These systems are built from a large multitude of extremely advanced and specialized firms that for the most part is in Europe, Japan, United States, etc.
The real problem for China isn't just catching up either. It's that the target keeps moving. They're racing to reach where the frontier was five years ago while ASML and its partners are already somewhere else entirely.
They sure are trying, aren't they. Hard to imitate ASML or TSMC at this point. China always seems 10 years behind for all their solid (and sketchy) efforts.
What we have to remember now is TSMC, ASML and Nvidia all have significant R&D budgets now, world class talent and supply-chains that have taken them decades to perfect. It's hard to copy that.
In 2025, South Korean manufacturers bought 55% of their helium from countries in the Gulf Cooperation Council, a union of six Arab nations. Taiwan bought 69% of its helium from the GCC in 2024, according to a recent report from analysts at Barclays, per CNBC.
Another chokepoint that comes to mind in the AI race, with LLMs, is AI coding supremacy. This is what makes Anthropic's IPO so exciting for many. Their coding models makes them such a favorite when it comes to their API, Enterprise AI, B2B revenue as a critical moat for America's dominance in AI's future.
Coding capability as a chokepoint is an interesting frame — it's the one bottleneck that's actually self-reinforcing, since better coding models accelerate the development of better models. The question is whether that moat holds when the underlying hardware supply chain runs through adversaries' geography. Anthropic's API revenue is impressive, but it still depends on Nvidia GPUs that depend on SK Hynix HBM that depends on https://thesynthesis.ai/journal/the-single-point.html.
The Iran War and Helium is a super interesting story I find: https://www.cnbc.com/2026/03/19/the-iran-war-is-threatening-supply-helium-what-it-means-for-markets.html
It makes you realize just how geopolitical all of these supply-chains are.
It was awesome working with you on this!
Thank you so much for this! The infographics and links are stellar.
What's different from 2025 to 2026 is the AI Infrastructure local opposition. While BigTech tries to bury this news, it's a huge deal-breaker. Think Datacenter projects stalling in multiple U.S. states!
Local Opposition and Regulation Community pushback has become a "material driver of attrition" of datacenter rollouts. In 2025 and early 2026, roughly 30% to 50% of large-scale projects faced delays or cancellations due to:
Water Usage: Large campuses can consume billions of liters of water annually for cooling, triggering moratoriums in drought-prone areas.
Zoning Changes: Towns are increasingly wary of "data center alleys" and are implementing stricter noise, aesthetic, and environmental impact reviews.
The health and noise factors for local community residents near AI Infrastructure is a growing and serious concern. Higher electricity prices are also a bottleneck while the U.S. experiences a growing affordability crisis.
AI is political. Because the growing disparity and wealth inequality involved is so large. Young people will likely make this a political issue for the 2028 Presidential elections.
Sadly, our leaders, instead of building bridges, are creating chokepoints with every decision and action they take.
If it wasn’t for more recent geopolitical conflict these “supply-chain chokepoints” wouldn’t be seen as problematic at all. Many arise precisely because it is most efficient to let some firms / countries take over certain industries such as mineral refining (which is low-margin and dirty).
Think China has developed some EUV capabilities internally. Stole some of the tech. ??
The Chinese prototype was built by salvaging components from older ASML machines through secondary markets. This proof-of-concept is no longer viable at scale because China no longer has access to the critical Western inputs that make ASML possible. And as of today, it hasn't produced a single working chip.
Meanwhile ASML just certified its High-NA EUV line for high-volume manufacturing, shipping $400M machines to Intel, Samsung, and SK Hynix for sub-2nm production. China needs to independently recreate not just the machine but the entire ecosystem underneath it: Zeiss precision optics, high-power lasers, photomasks, resists, multilayer mirrors. These systems are built from a large multitude of extremely advanced and specialized firms that for the most part is in Europe, Japan, United States, etc.
The real problem for China isn't just catching up either. It's that the target keeps moving. They're racing to reach where the frontier was five years ago while ASML and its partners are already somewhere else entirely.
They sure are trying, aren't they. Hard to imitate ASML or TSMC at this point. China always seems 10 years behind for all their solid (and sketchy) efforts.
What we have to remember now is TSMC, ASML and Nvidia all have significant R&D budgets now, world class talent and supply-chains that have taken them decades to perfect. It's hard to copy that.
Taiwan needs Qatar? 🤔
In 2025, South Korean manufacturers bought 55% of their helium from countries in the Gulf Cooperation Council, a union of six Arab nations. Taiwan bought 69% of its helium from the GCC in 2024, according to a recent report from analysts at Barclays, per CNBC.
Supply chain disruptions increase prices. Higher prices impacts who gets critical resources.
Another chokepoint that comes to mind in the AI race, with LLMs, is AI coding supremacy. This is what makes Anthropic's IPO so exciting for many. Their coding models makes them such a favorite when it comes to their API, Enterprise AI, B2B revenue as a critical moat for America's dominance in AI's future.
Coding capability as a chokepoint is an interesting frame — it's the one bottleneck that's actually self-reinforcing, since better coding models accelerate the development of better models. The question is whether that moat holds when the underlying hardware supply chain runs through adversaries' geography. Anthropic's API revenue is impressive, but it still depends on Nvidia GPUs that depend on SK Hynix HBM that depends on https://thesynthesis.ai/journal/the-single-point.html.