There is a number that has been haunting my sleep lately: 281. That is the billions of dollars Goldman Sachs projects will be spent on wafer fabrication equipment (WFE) by 2028. It is a staggering figure, one that implies a compound annual growth rate north of 20% from 2024’s base. But as I dissected the report, I realized this isn’t just a financial forecast. It’s a statement of faith. It’s a bet on human collective action, on our ability to coordinate capital, engineering talent, and geopolitical will at a scale we have never managed before. And it reminded me, inevitably, of the early days of decentralized protocols. We are not just building chips; we are building a consensus mechanism for the physical world. Code is law, but people are purpose.
The report’s core thesis is deceptively simple: the insatiable demand for AI compute, specifically the need for HBM4 memory and 2nm GAA logic, will drive a super-cycle in semiconductor capital expenditure. The logic is sound on paper. NVIDIA’s GPUs are sold out. Every hyperscaler from Google to Meta is hoarding accelerators like they are the last lifeboats on a sinking ship. But the translation of this demand into WFE spending is not a linear equation. It is a complex, fragile system of interdependencies, bottlenecks, and human decisions. Based on my years auditing token distribution models for fairness, I see a similar structural flaw in this forecast. It assumes a level of flawless execution that history suggests is unlikely. We must trust, but verify. But also, connect.
Let’s get into the technical weeds, because that’s where the truth lives. The forecast rests on two pillars. The first is the transition to Gate-All-Around (GAA) transistors at the 2nm node, a shift that requires a radical increase in the number of deposition and etching steps. The second is the ramp of HBM4, which is less about the memory cells themselves and more about the advanced packaging—the TSV etching, the hybrid bonding, the stacking of silicon dies into a single, powerful package. This is where the real bottleneck lies. TSMC’s CoWoS capacity is the chokepoint for the entire AI supply chain. They doubled it in 2024 and it still wasn’t enough. The report implies a doubling again by 2026, but that requires a herculean effort in a space where precision is measured in nanometers and yield rates can make or break a quarter. Resilience beats hype every time.
My own experience in the 2020 DeFi Summer taught me that when you have a supply-demand imbalance, you get a proliferation of solutions, not a single, elegant one. We saw it with yield farming strategies; we are seeing it now with chip packaging. The report’s focus on the big three—ASML, Applied Materials, Lam Research—is accurate, but it misses the periphery. It underestimates the potential for companies like Besi and ASM International, who specialize in the hybrid bonding equipment that HBM4 desperately needs. The WFE pie is growing, but the slices are changing shape. The report’s own data on supply chain security shows a 100% dependency on ASML for EUV lithography. That is a single point of failure. In my world, that is an unacceptable systemic risk. It is a 51% attack waiting to happen.
Here is my contrarian take, and it’s not the usual “this is a bubble” argument. The risk isn’t that demand will vanish. It’s that the supply will fail to materialize. The report forecasts ASML EUV capacity at roughly 50-60 units per year. But the lead time for a high-NA EUV tool is 12-18 months. A single machine costs over $300 million. If a single component in that machine—say, the Zeiss optics—has a yield problem, the entire timeline slips. We saw this with the shift from Ethereum’s Proof-of-Work to Proof-of-Stake. The Merge was delayed multiple times not because of a lack of consensus, but because of the sheer complexity of coordinating the upgrade across a decentralized network. The semiconductor industry is no different. It is a global, decentralized network of suppliers, and its upgrade path is fraught with coordination failures.
The report’s hidden assumption is that AI demand is a monotonic function. But what if the AI capex cycle peaks in 2026? The report’s own data shows that a single variable—hyperscaler spending—can swing the forecast by hundreds of billions. If Google or Microsoft blinks, the 2028 number goes from 281 billion to something much more modest. This isn’t a criticism of the forecast; it’s a critique of our collective psychology. We are prone to extrapolating current trends into the future, forgetting that every boom is followed by a bust. The community is the new central bank, and right now, the community of hyperscalers is printing money with reckless abandon. But interest rates on innovation always rise eventually.
I’ve lived through this. During the 2022 bear market, I managed the transition of a community through a governance crisis. The panic was palpable. The temptation was to capitulate, to sell everything, to retreat to the safety of fiat and silence. But we didn’t. We focused on resilience. We built forums for people to vent, we communicated transparently about the risks, and we held the line. The result was a 40% reduction in churn. The same principle applies to the semiconductor industry. The 2026-2028 build-out will happen, but it will not be smooth. There will be delays, budget overruns, and technical setbacks. The winners will not be the companies with the best technology, but those with the strongest community of engineers, suppliers, and customers. Those who can navigate the chaos with a steady hand.
Let’s look at the onshoring trend, which the report correctly identifies. The US CHIPS Act, the European Chips Act, Japan’s Rapidus project—these are all attempts to build local, resilient supply chains. But this is a classic case of decentralized coordination. We are building parallel, redundant systems. It’s like having multiple Layer-2 solutions for the same base chain. They are all trying to solve the same problem, but they are doing so in isolation. The result will be a massive duplication of effort. The report estimates this will add 20-30% to long-term costs. That’s the price of sovereignty, and it’s a price we are willing to pay. But we must be honest about the trade-off. We are sacrificing efficiency for resilience. In a bear market, that is a hard sell. In a bull market, it’s a no-brainer.
The real insight from the report is not the 281 billion number. It’s the shift in value creation. The report shows that equipment makers have ROIC of 30%+, far above their cost of capital. They are the true minters of this cycle. They hold the keys to the kingdom. But with great power comes great responsibility. The report warns that their valuations are already at the high end of historical ranges. The market has already priced in the growth. This is where I see the biggest risk. We are so focused on the technology that we forget the human element. We forget that these companies are run by people who can make mistakes, who can over-promise and under-deliver. The community is not just the buyers and sellers; it is the entire ecosystem of engineers, technicians, and executives. And their morale is fragile.
So, what is the takeaway? The Goldman Sachs report is a beautiful, well-reasoned analysis of a possible future. But it is not a prophecy. It is a map of a territory that is still being charted. As we navigate this terrain, we must remember that the most valuable asset is not the silicon or the software, but the trust between the people who build it. The next few years will test our collective resolve. We will face supply chain shocks, geopolitical storms, and economic turbulence. But if we can hold the line, if we can communicate with honesty and build with resilience, we will not just survive the cycle. We will build a foundation that lasts for generations. The question is not whether we can build the machines. It is whether we can build the consensus to use them wisely.