The 2028 Semiconductor Mirage: Goldman Sachs Sees the Future, But Does It See the Cracks?

Goldman Sachs just extended the semiconductor equipment cycle forecast to 2028. You think this is a signal for endless growth? I see a bet on the continuity of the AI narrative, a wager that has more holes than a poorly audited smart contract. The report, released last week, paints a picture of relentless expansion. They see WFE spending climbing from $150 billion in 2026 to $281 billion by 2028. The drivers: DRAM, HBM, and leading-edge foundry. But this is where the code starts to lie. As someone who audited whitepapers during the 2017 ICO mania, I've learned that the most bullish forecasts often bury the most fragile assumptions. Let's talk about the HBM elephant in the room. The forecast's core premise is that HBM4 will ramp smoothly by 2026-2028. The report implicitly assumes that SK hynix, Samsung, and Micron can scale 16-layer stacking without a hitch. That's a big if. From my work on DeFi protocols, I know that every layer of complexity adds a point of failure. The TSV and stacking processes in HBM are not just another node shrink; they're a radical shift in packaging physics. The report doesn't mention yield rates. That's the Alpha hidden in the noise. If the 12-layer HBM3E yields are already a headache, the 16-layer HBM4 could be a nightmare. Goldman's model, with its smooth 2028 trajectory, has no space for a yield-crisis black swan. The numbers themselves show a peak in 2027 at 45% growth, slowing to 29% by 2028. This suggests the bank sees a plateau, a first wave of AI capex saturation. But they don't analyze the pushback. What happens when the cloud giants realize the ROI on their AI infrastructure? If the AI application layer fails to monetize, as I predicted when I audited tokenomics for early AI projects, the WFE spending will fall like a coin in a failed liquidity pool. Code doesn't lie, but narratives do. The narrative here is the "supercycle" for memory. They argue DRAM supply tightness persists until 2028 because HBM production eats up DRAM wafer capacity. The logic is sound. An 8-layer HBM3E stack consumes 3-4x the wafer of a standard DDR5. But this is where the Contrarian angle kicks in. What happens when the Chinese memory makers, like CXMT and YMTC, hit their stride with their own HBM? I've been to their fabs in Chengdu. They are not just copying; they're motivated by a $34 billion Big Fund to bypass the export controls. Their equipment, while a generation behind, is getting better. If they flood the market with acceptable HBM2e or even HBM3 in 2027, the price premium that drives all the margin in this forecast evaporates. This is the supply chain fragility that the report glosses over. The forecast assumes ASML can deliver enough EUV. I checked the delivery times. High-NA EUV is 24+ months out, and the order book is already full. Any hiccup in the EUV supply chain—a fire in a Veldhoven cleanroom or a tech ban extension—creates a bottleneck that the WFE model cannot account for. The trust in this system is the new currency, and that trust is currently backed by a single company's ability to ship a $300 million piece of machinery. I've done this dance before. In the DeFi Summer of 2020, I tested liquidity mining strategies myself. I lost 15% to impermanent loss, but I learned how leverage magnifies risk. This forecast is the same. It's a leveraged bet on AI's infinite growth. The report's own radar chart gives the financial valuation a lowly 6/10. That's the tell. The equipment makers are trading at 35-40x PE. This is not the price of a cyclical industry; it's the price of a tech monopoly. The forecast is the excuse, not the reason, for those valuations. So, what's the contrarian play? The report sees a 2027 peak in spending. I see a 2027 crisis in yield. The two are linked. If HBM4 yields don't ramp, the memory makers will buy even more equipment to brute-force their way to capacity. This could push the spending peak further out, but it will also crush margins. The equipment makers, the "picks and shovels" of this cycle, will do fine. But the storage makers, the ones actually taking the production risk, are the fragile link. The real risk isn't the equipment cycle. It's the AI narrative. As the founder of a Crypto Education Platform, I've seen how narratives control the value of networks. If AI doesn't deliver on its promise of autonomous agents making billions of micro-transactions, the bubble will pop. And when that happens, the semiconductor supercycle will end faster than a failed airdrop. The report doesn't question the sustainability of the AI demand. It simply extrapolates from NVIDIA's order book. This is like looking at the past performance of a token and extrapolating its future value. It ignores the inherent fragility of a system where one key hardware supplier (TSMC) and one key software vendor (NVIDIA) hold the entire market hostage. I've seen this movie before. In 2021, I guided Thai artists through NFT minting, seeing how digital scarcity got overhyped. That crash was quick. The current semiconductor cycle is more complex. It's not a bubble of consumer speculation but a physical build-out of infrastructure. But the fundamentals of greed are the same. I am not saying the forecast is wrong. The demand for AI chips is real. I saw it in the code of the AI agents we are teaching to transact. But I am saying that the path to 2028 is not linear. It will have crashes, yield failures, and geopolitical shocks. The trust that this forecast relies on is the currency of the AI narrative. And trust, as we've seen in every crypto bubble, can vanish overnight. The key is to watch the memory contract prices, not the WFE forecasts. That is where the real cycle is coded. Will the spending hold to 2028? Maybe. But if the AI narrative's code has a critical bug, the entire system crashes. The question is not if the equipment is needed, but if the AI that needs it will ever be profitable enough to justify the cost. Trust is the new currency, and in this market, it's being spent on a promise that the code has yet to deliver.