AI Computing Power Overcapacity Risk by 2028: The Iron Law of Historical Cycles vs. the Bottleneck Relief Debate
nashnova research
Two semiconductor analysts clashed head-on over what happens when AI chip capacity floods the market in 2028 — one says demand will absorb it; the other warns the iron law of historical cycles always wins, and pricing power will collapse.
What exactly are they arguing about?
Analysts Ben Bajarin and Jay Goldberg debated on the industry podcast *The Circuit* on September 9. The core disagreement is simple: when a wave of new fabs comes online in 2028, does the AI chip market ease gently or does pricing power collapse?
Their shared starting point is the same timeline: Broadcom, Nvidia, and major cloud providers submitted large-scale, long-term capacity-lock agreements to the supply chain in late 2025 to early 2026.
This means → new wafer fabs, advanced packaging lines, and high-end substrate plants need two to three years to build and qualify — the earliest they deliver real supply is 2028. That moment is exactly what the debate is about.
Why does Bajarin say "it won't crash"?
His core argument rests on the incompressibility of physical build cycles: between 2026 and 2027, existing capacity is fully locked and new lines are not yet online. Global advanced process and advanced packaging will be in the most extreme supply bottleneck in history.
For 2028, he rejects the "capacity flood" narrative. In plain terms = AI models are expanding from text into HD video, real-time audio, 3D generation, and long-horizon reasoning — token consumption is growing exponentially, fast enough to absorb new capacity.
He projects that by 2028 the market at most exits "absolute shortage" and returns to a mildly tight state where demand is roughly 105% to 110% of supply — not a cliff-edge glut, just breathing room.
He also draws an analogy to historical rail, canal, and 3G build-outs, noting a key difference: the production facilities themselves are physically constrained in how fast they can be built, so the supply bottleneck naturally limits overinvestment.
Why does Goldberg say "history will repeat"?
His warning comes from an induction across fifty years of semiconductor capital cycles: in every upcycle, everyone insists "this time is different," yet every single one ends in severe overcapacity. In plain terms = humans always overbuild. No exceptions.
He lays out the scale of current construction: TSMC is advancing nearly 20 new wafer fabs and packaging sites globally; Micron has 7 large memory fabs under construction worldwide; Samsung and SK Hynix are equally aggressive in HBM — high-bandwidth memory, a type of stacked memory designed specifically for AI chips; and government-subsidized Intel capacity is being fast-tracked on top of all that.
He specifically flags the historical pattern of "double-booking" — during chip shortages, Apple, Qualcomm, and cloud providers often submit overlapping or doubled capacity commitments to multiple suppliers to secure allocation. This means → the moment supply loosens and lead times shrink, these "ghost orders" vanish instantly, and real demand turns out to be far below the numbers on the books.
How long can Nvidia's 70% gross margin hold?
Goldberg argues that Nvidia's current gross margin of over 70% and Broadcom's steep custom-chip premiums are fundamentally scarcity premiums created by supply shortage — not normal returns on a defensible moat.
This means → by 2028, when massive fab capacity is fully running, fixed-asset depreciation creates enormous cost pressure. Foundries and chip designers will launch price wars to keep utilization rates — the share of capacity actually running — from collapsing.
In plain terms = an idle machine costs more than a discounted wafer, so everyone races to cut prices and ship — hardware gross margins face mean reversion.
What is the one thing they actually agree on?
Despite sharp disagreement on the 2028 endgame, both analysts share one key market call: the window where "buy anything AI-adjacent and outperform" worked is now closed.
Over the past two years, any company touching advanced packaging, HBM, GPUs, or high-speed optical modules rode a sector-wide premium driven by a single bottleneck. But from now through 2027–2028, that bottleneck is fragmenting across multiple dimensions.
This reflects a deeper shift: as MediaTek, Marvell, and other custom-compute players ramp up, hardware suppliers without an irreplaceable software ecosystem or core IP moat face margin compression and inventory drawdowns simultaneously.
Both analysts' shared conclusion: investors must focus on the handful of links with genuine cross-cycle pricing resilience — and the actual pace at which 2028 capacity lands will be the key test of which side of this debate is closer to reality.
市场有风险，内容仅供研究参考，不构成投资建议。