The Great AI Bottleneck Debate: 2027 Peak Constraint vs. 2028 Capacity Release
TREE NEWS reports: As AI compute arms race accelerates, a critical inflection point is approaching. With Broadcom, Nvidia and other tech giants signing massive long-term supply lock-in agreements between late 2025 and early 2026, the semiconductor industry is undergoing its largest capital expansion in history. The real impact of this investment will ignite a fundamental debate over supply-demand dynamics by 2028.
On September 9, tech analysts Ben Bajarin and Jay Goldberg faced off on the AI chip cycle during the deep industry podcast The Circuit. Bajarin argues that 2027 will mark the most extreme supply constraint “bottleneck point” across the industry, with 2028 bringing massive new capacity that will still be absorbed by exponentially growing AI demand, keeping supply-demand slightly tight without a cliff-edge oversupply. Goldberg, citing 50-year semiconductor cycle patterns, warns that the massive capacity being simultaneously advanced by TSMC, Micron, Samsung and SK Hynix will trigger a pricing power collapse and brutal margin mean reversion when it all comes online in 2028.
Despite their fundamental disagreement on the 2028 endpoint, both analysts agree on one clear point: the era of “buying AI supply chain blindly for explosive gains” is over. As bottlenecks evolve from single segments to multi-polar fragmentation, investors must apply stricter standards to identify core assets with cross-cycle pricing resilience.
Systemic Constraints: Shortboard Effect Dominates Chip Shipment Ceilings
The premise for understanding this debate lies in a basic logic of the semiconductor supply chain: actual shipments of the entire system are determined by the weakest single bottleneck. Even if GPUs or memory supply is ample, any shortage in substrates, MLCCs, or wafer testing can prevent total output from increasing. This “shortboard effect” directly explains the current supply chain tension. Substrate, power and analog semiconductor, foundry, and memory suppliers have historically adopted conservative investment strategies, refraining from greenfield capacity until customers provide clear long-term demand commitments. Only after late 2025 to early 2026, when Broadcom, Nvidia and major cloud service providers submitted substantial, stringent long-term lock-in agreements, did major suppliers fully launch a new round of capital expenditure. Since building a wafer fab, advanced packaging line (like TSMC’s CoWoS) or high-end ABF substrate plant naturally requires 24 to 36 months for construction and cleanroom certification, the actual batch output of this investment will only form effective supply by 2028 at the earliest. This timeline is the common starting point for both sides of the debate.
Ben Bajarin: 2027 Supply Peak, 2028 Moderate Return
Ben Bajarin’s core argument is built on the incompressibility of physical construction cycles. He believes that between 2026 and 2027, old capacity will be fully locked while new lines are not yet operational, placing global advanced processes and packaging in the most extreme supply bottleneck in history. Major cloud providers will have no choice but to accept any available compute quota—2027 may become the year with the most severe supply constraints in semiconductor history. For 2028, Bajarin disagrees with the pessimistic “capacity flood” narrative. He points out that as AI models shift from pure text to high-definition video, real-time audio, 3D generation and long-range reasoning, token consumption grows exponentially, and demand-side expansion will be fast enough to quickly absorb newly released capacity. In his forecast, the market at best exits “absolute shortage” in 2028, returning to a benign state of slightly tight supply-demand (demand/supply around 105% to 110%), without cliff-edge oversupply. He also suggests that current supply bottlenecks act as a “safety valve,” delaying overinvestment. He compares this AI infrastructure cycle to historical railway, canal, 3G and mobile internet construction booms, noting that these previous areas all experienced severe overbuilding. The key difference in this AI compute cycle is that even with clear demand, the construction speed of production facilities and data centers is equally constrained by physical limitations, and supply constraints themselves are curbing overinvestment. He adds that in enterprise AI deployment, high token costs and limited compute access are also objectively slowing the migration from experimental to production environments.
Jay Goldberg: Historical Cycle Iron Law, 2028 Pricing Power May Backfire
Jay Goldberg’s reverse warning comes from a historical summary of 50 years of semiconductor capital cycles. His core judgment: in every upcycle, all participants shout “this time is different,” but every time it ends with severe overcapacity. The scale of current projects under construction is alarming:
- TSMC is advancing nearly 20 new wafer fabs and packaging bases globally;
- Micron has 7 large memory plants under construction worldwide;
- Samsung and SK Hynix are equally aggressive in HBM investment;
- Additionally, government-subsidized capacity from Intel and others.
Goldberg also highlights the historical pattern of “double/triple booking.” During chip shortages, customers like Apple, Qualcomm and cloud providers often submit overlapping or even doubled capacity intentions to multiple suppliers to ensure delivery. Once supply begins to release and lead times shorten, these inflated “ghost orders” are instantly cancelled. On pricing, Goldberg believes Nvidia’s current ultra-high gross margins above 70% and Broadcom’s high ASIC premiums are essentially scarcity premiums created by supply shortages. By 2028, when massive fab machines are all running and fixed asset depreciation creates enormous cost pressure, foundries and chip designers will inevitably engage in fierce price wars to maintain utilization rates, and hardware gross margins will face brutal mean reversion.
Investment Consensus: Era of Selective Picking Replaces Indiscriminate Buying
Despite significant divergence on the 2028 endgame, both analysts agree on a key market judgment: the window for “blind buying of AI supply chain for excess returns” has closed. Over the past two years, any company with business tied to advanced packaging, HBM, GPUs or high-speed optical modules saw share prices benefit greatly from sector-wide premiums driven by a single bottleneck. But from now to 2027-2028, supply chain bottlenecks are evolving from a single segment (like early CoWoS packaging) to multiple dimensions. As custom compute players like MediaTek and Marvell accelerate their layouts, and massive capacity is released sequentially in 2028, hardware suppliers lacking irreplaceable software ecosystems or core IP moats will face dual pressures of margin compression and inventory destocking from late 2027 to 2028. The common conclusion of both analysts: investors can no longer view the semiconductor supply chain as a homogeneous whole. They must focus on core segments with true cross-cycle pricing resilience to secure an advantageous position in the divergent landscape after the largest capacity expansion in history lands.
