You can learn most of what you need to know by taking the morning train through Silicon Valley. As the hills fade into the distance, young engineers with earbuds in and laptop screens illuminating their faces fix bugs. They leave DeepMind at Mountain View, Stanford at Palo Alto, and Nvidia at Santa Clara. Each station feeds a distinct node in a global supply chain that, three or four years ago, nearly no one outside the industry could identify. Everyone is aware now. or believes they do.
The race to create the best AI chip has grown to be the most costly project in the history of contemporary technology. By the end of the decade, Citigroup projected that spending on AI datacenters would reach $2.8 trillion, surpassing the yearly production of both Italy and Canada. In 2025, Microsoft alone invested about $80 billion in AI capital expenditures. The abstract numbers cease to feel abstract when you enter one of Digital Realty’s Santa Clara facilities, where racks of GPUs scream at 120 decibels behind black steel cages. Your ears will ring for hours after just five minutes. Strangely, that sound is the new economy.
| Detail | Information |
|---|---|
| Market leader | Nvidia, headquartered in Santa Clara, California |
| Nvidia market value | Roughly $4.3 trillion, up 30-fold since 2020 |
| Jensen Huang’s net worth | Approximately $160 billion |
| Forecast AI datacenter spend by 2030 | $2.8 trillion (Citigroup, September 2025) |
| Hyperscaler combined 2026 AI capex | About $650 billion across Alphabet, Amazon, Meta, Microsoft |
| Microsoft’s 2025 AI capex | Roughly $80 billion |
| Alibaba’s three-year AI commitment | Over $53 billion |
| Typical chip development cycle | Up to four years from theory to manufacture |
| Notable startup compressing the cycle | Chipmind (Zurich), out of ETH Zurich |
| Critical European tool | ASML’s $400M EUV lithography machine, the only one capable of the most advanced nodes |
| Key oversight body for U.S. export controls | U.S. Department of Commerce |
At the heart of it all is Nvidia, a $4.3 trillion company that, ten years ago, hardly anyone outside of semiconductor circles took seriously. With his leather jacket and effortless stage presence, Jensen Huang has emerged as the AI era’s closest thing to a well-known industrialist. Customers, including Microsoft, Google, Meta, and even Chinese cloud players where export regulations permit, purchase his company’s GPUs—which are used to train all major models—faster than factories can manufacture them. The odd thing is that nobody really believes this will last forever. Before settling on something more resilient, Tesla went through multiple cycles of belief and skepticism in the face of similar doubts.
Demand isn’t actually the bottleneck. Now is the moment. The four-year design cycle in the industry has become a silent drag on everything happening upstream, which is why Harald Kröm, a Swiss founder from ETH Zurich, founded Chipmind. His proposal is simple: automate the repetitive 40% of chip development, free up engineers to work on more challenging issues, and save a year. In October, he secured $2.5 million in pre-seed capital. The established behemoths Cadence and Synopsys are vying for the same position. As this develops, there’s a feeling that the companies that figure out how to speed up chip design may be the true winners of the next phase, rather than the chip designers themselves.
China comes next. Due to U.S. export restrictions and significantly lower funding, Chinese AI labs are several months behind the American frontier. Despite this, they have responded with ingenuity, as constrained engineers typically do. To get more performance out of fewer failures, DeepSeek, Alibaba, and Moonshot AI have all leaned toward mixture-of-experts architectures and aggressive quantization. Leapfrogging Nvidia’s customers is insufficient. However, Beijing’s full-stack push for semiconductor self-sufficiency continues to advance, fab by fab, and the gap is smaller than the export-control hawks would like.
The true fragility of the entire supply chain is obscured by the trillion-dollar headlines. One Dutch company produces the lithography equipment used by ASML. Taiwan continues to produce the majority of the most sophisticated fabrication. Talent clusters for chip design exist in perhaps five cities across the globe. It’s difficult to ignore how a single disruption, such as a pandemic, a Strait crisis, or an unexpected export decision, could affect all layers simultaneously. It appears that investors think the boom will last forever. Perhaps it is. Or perhaps, as one DeepMind engineer recently put it, everyone is constantly working and no one is quite sure where the natural stopping point is. Maybe there isn’t. The part that ought to cause people to stop is that.
