I was at a Lagos co-working space last week, debugging a smart contract for a DePIN project that promises to democratize AI compute. The founder, a brilliant 22-year-old, was explaining how their token would reward users for contributing GPU cycles to train open-source models. He was glowing, optimistically quoting projections of a million nodes within two years. I asked him a simple question: "Who makes the chips for those GPUs?" He paused. He didn't know the answer. That’s the crypto blind spot.
This isn't a story about ASML and TSMC in isolation. It's a story about a fundamental truth the blockchain industry refuses to confront: the hardware we pin our decentralized dreams on is produced by the most centralized, geopolitically vulnerable, and capital-intensive supply chain in human history. ASML's recent expansion plans and TSMC's massive capital expenditure are not just corporate news. They are the two most important signals for the future of any token that claims to be the “world computer.” Trust me, I’ve been auditing these claims since 2017.
Context
Currently, the global economy is in the midst of what analysts call the "second wave" of AI — the shift from training massive models to deploying them for inference at the edge. This shift requires an unprecedented number of advanced chips, specifically GPUs and ASICs built on TSMC's 5nm, 4nm, and 3nm nodes. To make these chips, you need ASML's extreme ultraviolet (EUV) lithography machines. ASML is the sole supplier. TSMC’s capacity for these advanced nodes is effectively sold out for the next 18 months. For crypto, this matters because every project promising “decentralized AI inference” — from Render Network to new Layer-1 chains built for AI — relies on the same finite pool of high-end silicon. When I piloted a DeFi-for-unbanked project in 2020, I learned that hardware constraints are just as real as smart contract bugs. The code can be trustless, but the chip is not.
Core Insight
Let’s run the numbers. ASML plans to increase EUV output to around 90 units per year by 2025-2026. That sounds impressive until you realize each machine takes 12-24 months to build, and after delivery, TSMC requires another 12-18 months to integrate and stabilize yield. The total latency from an expansion decision to usable chips is over three years. Meanwhile, demand for AI compute is doubling every six months. The gap is structural, not cyclical. For the crypto space, this means the supply of GPUs for mining or AI inference is not going to ease for at least two to three more years. Projects that mint tokens based on the assumption of infinite hardware scalability will face a brutal reality: the value of their token is directly tied to TSMC's monthly output, not just the quality of their code.
But it’s worse than that. The geopolitical overlay adds a layer of risk most crypto projects ignore. The US, Netherlands, Japan, and Taiwan are effectively forming a “chip cartel” to limit China’s access to advanced nodes. This means any project with Chinese users or developers — and many decentralized projects have global aspirations — could face supply chain disruption overnight. Based on my audit experience, I’ve seen protocols that claim to be “permissionless” yet rely on Nvidia GPUs that are themselves subject to US export controls. You cannot build a trustless system on top of a trust-based hardware monopoly.
Contrarian Angle
Here’s the contrarian take: The market’s euphoria about AI chips is not wrong — the demand is real. But the narrative that this expansion will solve the bottleneck is naive. The expansion itself is a double-edged sword. TSMC’s capital expenditure is so massive (projected $30B+ annually) that it will compress gross margins for years. ASML’s ability to scale is limited not by money but by the availability of specialized optics from Zeiss and a workforce of thousands of PhD-level engineers. The rate of capacity addition is fundamentally bounded by the rate at which humans can assemble atoms. That rate is slow. For a bull market that moves on hype, a three-year latency is an eternity.
Also, consider what this means for DePIN. During the 2022 bear market, I hosted daily “Code & Coffee” sessions where developers realized that their tokenomics relied on unrealistic compute supply elasticities. The same flaw persists today. Every DePIN project that says “we will reward users for unused compute” ignores the fact that the total pool of unused advanced compute is shrinking, not growing, as AI absorbs every available cycle. The contrarian insight: the value of GPU-backed tokens will increasingly correlate with the stock of ASML, not the number of claims in a whitepaper.
Takeaway
The second wave of AI is real, and it is the biggest tailwind for blockchain projects that actually need compute — but only if they understand the physics. Trust the process of decentralization, but verify the code that runs on the hardware. More importantly, verify the hardware itself. I believe the projects that will survive are those that explicitly plan for chip scarcity: designing for efficiency, building on rollups that batch less data, or even exploring decentralized alternatives like FPGA-based networks. The next time you see a token claiming to power “the world’s AI,” ask: who builds the chips? If the answer isn’t part of the protocol’s risk model, you’re buying a dream built on a lithography machine that might not arrive on time.
The future of blockchain AI depends on whether we can honestly confront the bottlenecks of the physical world. The code is ready. The wafer is not.