In a few weeks, I will make my fourth trip to Davos for the annual World Economic Forum. Undoubtedly, AI will dominate the agenda, but not because of better models or larger valuations. The real question hanging over Davos this year is whether a technology that depends on radical global coordination can keep scaling in a world that’s a politically and economically is becomeing more divided. AI depends on an intensely physical, global supply chain—and Davos is where that reality is about to collide with geopolitics.
It may be a good sign that the U.S. House has established a prominent presence just off the promenade, positioning itself as a hub for American business leaders, policymakers, and allies. Like so many Trump administration initiatives, it is funded by corporate sponsors, including Microsoft, McKinsey, and Ripple. But what message will we be sending? We are open for business? We come in peace? We’re here to help? Perhaps all of the above—or maybe something more transactional, even adversarial.
The message will decide the future of the technology that everyone–the U.S., China, really the whole planet–is resting its hopes on. Because the AI boom everyone is hoping for may require a networked, globally connected world to make it happen.
AI Depends on a Global Hardware Supply Chain
Geopolitical strategist Peter Zeihan has been making this point in his characteristically blunt way for some time. AI runs on semiconductor chips–a lot of them. The semiconductor supply chain is not a tidy ladder; it’s a sprawling ecosystem with too many single points of failure to hand-wave away. In one of his sharpest formulations, Zeihan argues that the “high-end processing chips that Taiwan is famous for” require “100,000 steps… 9000 companies,” and that they are “scattered around the world.”
That framing reframes Taiwan risk as something larger than geopolitics–Taiwan is the CPU of the AI economy. Even when the U.S. celebrates a domestic milestone—like Nvidia and TSMC unveiling the first U.S.-made Blackwell wafer in Arizona—fabrication is only one link in the chain. And even that flagship project is not going well, TSMC is currently running a $440M loss on the project, and that will get worse before it gets better.
The remainder of the system remains important, especially the parts that are hardest to replicate quickly. Nvidia CEO Jensen Huang has been unusually direct about where today’s constraint sits: “Packaging has remained a bottleneck,” he said, noting that while capacity has grown “probably four times” in under two years, it remains a limiting factor.
Then there’s high-bandwidth memory, the quiet co-star of the AI boom. According to Counterpoint estimates cited by Reuters, HBM market share is heavily concentrated: SK Hynix at 53 percent, Samsung at 35 percent, and Micron at 11% as of Q3 2025. Those fabs are located worldwide, only a small fraction are made in the U.S.
Above it all sits the lithography toolchain. Extreme ultraviolet lithography—the gateway to advanced nodes—runs through ASML, headquartered in Veldhoven, Netherlands. No other company in the world can match ASML’s capabilities.
Zeihan’s deeper concern isn’t just fragility—it’s time. Writing about AI hardware, he argues that after research breakthroughs, it can take “another decade+ for production and supply chains to get sorted out” especially for specialized chips still in development. That timeline collides directly with a political moment where putting America first involved tariffs, protectionism and an economic withdrawal from he world stage. Ironically, the U.S. seems incredibly open to international military adventures, as long as they are short-lived.
But What if AI is Just Software?
If Peter Zeihan sees AI as an industrial system that can’t survive without globalization, venture capitalist Marc Andreessen offers a fundamentally different reading. Andreessen’s core claim is that AI should be understood not as a fragile hardware stack, but as the next great phase of software development. In Why AI Will Save the World, he characterizes AI as a “universal problem solver,” arguing that its economic and social impact will ultimately resemble past software revolutions more than those of heavy industry. In that framing, today’s GPU shortages and supply-chain chokepoints are real, but temporary.
Andreessen’s confidence rests on a familiar dynamic: software compounds, and abstraction wins. Compute gets cheaper. Models get more efficient. Training techniques improve. Inference costs fall. Over time, the advantage shifts away from whoever controls the most hardware toward whoever builds the smartest systems on top of it. This is not a denial of physical constraints so much as a bet on their eventual irrelevance. As Andreessen has argued elsewhere, the history of technology is the story of doing more with fewer resources, repeatedly.
Where Zeihan worries that AI’s dependence on chips, memory, and geopolitics could constrain its growth, Andreessen assumes those limits will be circumvented. AI, in this view, doesn’t need perfect global coordination forever—just long enough to bootstrap itself into something more abstract, more distributed, and less capital-intensive. The Internet is built, 5B people have access to it from the phone in their pocket, AI will find a way to get there.
So who’s right? Maybe both? Andreeson is right that the software stack is adapting quickly, squeezing more performance out of less compute. Zeihan is right that scaling AI as infrastructure—cheap, ubiquitous, and reliable—depends on a level of global coordination that now looks politically fragile.
That’s the tension Davos can’t escape this year. AI may eventually behave like pure software, abstracted from borders and bottlenecks, but today it is still deeply physical, capital-intensive, and dependent on a level of global coordination that’s increasingly under strain. Davos has always been where the future gets workshopped before the world events play out, often messily, in real time.
This year, the question is simple and unsettled: can the most global technology ever built still thrive in a world that seems determined to pull itself apart?
