The number hit my screen like a flash crash reversal: $281 billion. That's where Goldman Sachs now sees semiconductor wafer fab equipment (WFE) spending landing by 2028. Not 2026. Not a taper. An extension. I've spent 16 years watching these cycles get chopped and channeled, and this forecast, published August 25, 2025, breaks from every pattern I've tracked since I was monitoring Ethereum gas spikes during CryptoKitties.
Goldman isn't just pushing out the timeline—they're rewriting the fundamental thesis. The driver is the AI memory supercycle, and if their math holds, the semiconductor industry just entered a phase where the old cyclicality maps are obsolete.
But here's where my on-chain instinct kicks in. I don't trust headlines. I verify. And when I pulled up the underlying data, I found something that changes the entire narrative. Goldman's forecast of $281 billion by 2028 implies a capex density that most analysts haven't even begun to process. This isn't about more of the same—it's about a fundamental shift in how much capital it takes to produce a single wafer at the leading edge.
I've spent years deconstructing supply chains, from CryptoKitties congesting Ethereum to flash loan cascades on Anchor. This semiconductor cycle has the same signature: a cascading structural bottleneck that rewards the people who verify before they post.
The 2026 to 2028 Trajectory: A Breakdown
Goldman's roadmap breaks down to roughly $150 billion in 2026, climbing to $194 billion in 2027, and peaking at $281 billion by 2028. That's not just growth—that's a compound annual growth rate that outpaces anything I've seen in my 16 years of tracking this industry. The headline is the 2028 number. The underreported story is the shape of the curve.
Growth rate: 36% in 2026, 45% in 2027, then slowing to 29% in 2028. That's a peak in 2027, not 2028. The absolute dollar figure keeps climbing, but the velocity of expansion is already decelerating. This is the first tell of what I think is an AI infrastructure saturation signal hiding inside an optimistic forecast.
The three main demand pillars are: DRAM micro-shrink to the 1γ/1δ node class, HBM3E to HBM4 transitions (moving from 8- and 12-layer stacks to 16-layer), and leading-edge foundry logic moving from 3nm to 2nm GAA with High-NA EUV. Each of these is capital-intensive—and I want to focus on the memory side, because that's where the true structural change is being missed.
The HBM Memory Squeeze: A 4x Disruption
Here's the on-chain verification part. When I trace the actual wafer consumption, the math is staggering. HBM3E 8-stack consumes 3-4x the wafer area of a standard DDR5. That's not a modest uptick in demand—that's a full-scale resource squatter. This is why DRAM supply is expected to stay tight through 2028: not because the industry can't make DRAM, but because it's allocating all its capacity to the high-margin HBM.
The three DRAM/HBM players—SK Hynix, Samsung, Micron—are investing billions into HBM capacity, with SK Hynix leading the pack with a 50%+ market share. They're not just building more memory; they're building a different kind of memory that demands 4x the silicon just to produce. The economic implication is that this cycle's capacity utilization dynamics are structurally different from the 2017-2018 cycle. That was a short supply-demand imbalance. This is a structural re-rating of what memory is and what it's worth.
The Hidden Geopolitical Assumption
Here's where my 2022 Terra collapse analysis kicked in—when I learned to pivot narratives quickly when the underlying mechanics break. Goldman's forecast hinges on one massive, mostly unstated assumption: geopolitical stability. The 2026-2028 trajectory is built on a global supply chain that doesn't break. That's the biggest variable in the whole prediction.
The U.S. export controls on China have already cut off advanced process equipment (14nm logic, 18nm DRAM, 128-layer NAND). The equipment makers—ASML, AMAT, Lam, TEL—are locked out of a major portion of the Chinese market for advanced nodes. Yet the forecast still calls for record global spending. This implies Goldman is betting on a decoupling that doesn't go full-blown—a 'controlled competition' scenario.
In my view, that's a risky bet. The probability of a full tech decoupling—where the US/Japan/Korea/Taiwan operates one chain and China operates another—is higher than the street pricing. If that happens, the $281 billion forecast becomes a fantasy. You get duplicate supply chains, which is actually bullish for equipment (more total units), but the demand will come from different pockets, and the margins will be lower.
Contrarian Angle: The 2027 Peak Is a Saturation Signal
The market will read this as a linear 'go' signal for chip equipment stocks. I see something else: a 2027 peak growth rate that signals the beginning of the end of the first wave of AI infrastructure. By 2028, the 2nm transition will be complete. High NA EUV will be installed. HBM4 will be in mass production. The next step—whether that's 1.4nm, or a more radical device architecture—will require a massive new investment cycle to justify continued growth.
The question is: what happens after 2028? Goldman's forecast is remarkably quiet on this. The 29% growth in 2028 is already below the industry's historical average for the same growth period. This suggests the bank sees the current AI-driven upcycle hitting a natural plateau, and the following phase will require a new catalyst—embodied intelligence, AI agents at scale, or some other workload I haven't yet imagined.
My own analysis of the 2021 NFT metadata fragmentation taught me that the biggest data-driven trend often hides a simpler, overlooked constraint. The constraint here is: where does the next wave of demand come from? For the past 24 months, the answer has been 'AI training.' By 2028, that answer has to change.
The Contrarian Angle: The 2027 Peak Is a Signal
The market sees a straight line up. I see a curve that's already bending. The 2027 peak (45% growth) and the 2028 deceleration (29%) is the first in a series of signals that the AI infrastructure buildout is a finite event—not a permanent state.
This is what the bulls are missing. The current supply-demand squeeze is real. DRAM is below normal inventory. HBM pricing is 3-5x the standard DRAM. But the industry is already front-running the demand by building capacity. I've seen this pattern before in DeFi Summer 2020—everyone wants to 'farm' the trend, and they get a overbuild before the end users actually show up. The same thing is happening here.
The risk is a 2026-2027 AI bubble. If cloud service providers (AWS, Azure, GCP) start to pull back their capital expenditure guidance, if NVIDIA's guidance misses, the floor drops out. That's the scenario where the Goldman forecast becomes a footnote. The probability of this is 20-30% over the next 12-24 months. That's not a low-probability event—that's a tail risk with a real chance of hitting.
The Verdict: Positioning for the Peak
For the people who are long equipment makers (ASML, AMAT, Lam Research), the next 12-18 months are a good run. The order backlog is a 12-18 months visibility, and the earnings power is undeniable. But the risk reward is changing. The estimated PE for ASML is 35-40x, AMAT and Lam are 25-30x. That's priced for perfection, and perfection is not a structural condition.
Here's the contrarian play: the storage memory companies (SK Hynix, Samsung, Micron) offer a better risk/reward. They're trading at 15-20x forward PE. They have more earnings elasticity than the equipment makers because they're not just getting a higher volume of orders—they're getting a better price for their output. The DRAM price is rising. HBM is a premium product. If the AI trade holds, the memory guys will have a bigger earnings growth.
My on-chain instinct says: watch the inventory levels. Watch the DRAM contract prices. If DRAM contract prices start to flatten, that's the first signal that the super-cycle is peaking. If NVIDIA's next earnings report shows a miss, that's the second signal. This is a fast-moving market, and the time to position is now, not when the trend is confirmed.
The Bottom Line
Goldman's forecast is a macro blueprint that confirms the AI-driven super-cycle is real and that memory is the new bottleneck. But the smart money is going to be looking at the bend in the curve, not the peak.
The peak is in 2027, and the smart play is to position for that peak—not to chase the 2028 dollar figure. The play is to be in the names that have the biggest earnings elasticity: the memory manufacturers.
This cycle is real, but it's also finite. The 2028 projection is a beautiful number, but it's a target, not a guarantee. The second wave of AI infrastructure investment—the one that takes the industry to the next leg up—will be built on a different foundation than the first. And I'll be watching the on-chain data, the contract prices, and the capex guidance to see when that shift happens.
This is a time to be greedy, but smart about it. Watch the 2027 peak. Watch the storage pricing. And remember: in a sideways market, the smartest position is the one that's ready for the pivot.