The AI computing arms race is approaching a pivotal inflection point.

As technology heavyweights like Broadcom and NVIDIA finalize massive long-term supply agreements with their vendors between late 2025 and early 2026, the semiconductor industry is embarking on its largest capital expansion in history. The real impact of this investment will ignite a fundamental debate over supply-demand dynamics in 2028.

On September 9, technology analysts Ben Bajarin and Jay Goldberg engaged in a direct clash over the AI chip cycle on the in-depth industry podcast The Circuit.

Ben Bajarin contends that 2027 will represent the industry's most extreme supply-constrained "bottleneck point." While new capacity will be released en masse in 2028, he argues that continuously expanding AI demand will keep the market in a slightly tight balance, avoiding a cliff-like oversupply.

Jay Goldberg, citing five decades of semiconductor industry cyclical patterns, issued a contrarian warning: the simultaneous capacity buildouts by TSMC, Micron, Samsung, and SK Hynix could trigger a collapse in pricing power and a brutal reversion to mean gross margins once that output floods the market in 2028.

Despite their fundamental disagreement on the 2028 outcome, both analysts share one clear consensus: the era of easily buying any AI supply chain stock for guaranteed gains is over. As bottlenecks shift from a single chokepoint to multiple diverse constraints, investors must apply far stricter standards to identify assets with genuine cross-cycle pricing resilience.

The System Constraint: How the Weakest Link Dictates Chip Output

Understanding this debate requires grasping a basic semiconductor supply chain principle: total system output is determined by the single most fragile bottleneck segment.

Even if GPU or memory supplies are abundant in one area, any shortfall in substrates, MLCCs, or wafer testing prevents total production from increasing.

This "weakest link" effect directly explains the current supply chain tightness. Substrate, power and analog semiconductor, wafer foundry, and memory suppliers have historically adopted conservative investment strategies, refusing to launch greenfield capacity without explicit long-term customer demand commitments.

Only after Broadcom, NVIDIA, and major cloud service providers submitted large-scale, stringent long-term lock-in agreements to suppliers in late 2025 and early 2026 did primary vendors fully initiate a new round of capital expenditure.

Since constructing a wafer fab, advanced packaging line (such as TSMC's CoWoS), or high-end ABF substrate facility naturally requires 24 to 36 months including cleanroom certification, the meaningful output from this investment wave cannot materialize before 2028.

This timeline serves as the common starting point for both sides of the debate.

Ben Bajarin's View: 2027 as the Supply Ceiling, 2028 as Moderate Rebalancing

Ben Bajarin's core argument rests on the physical limits of construction timelines that cannot be compressed.

He believes that between 2026 and 2027, with existing capacity fully booked and new lines yet to come online, global advanced process nodes and advanced packaging will experience the most severe supply bottleneck in history.

Cloud providers will have no choice but to accept whatever compute quotas they can secure, making 2027 potentially the tightest supply-constrained year in semiconductor history.

For 2028, Bajarin rejects the "capacity flood" doomsday narrative.

He points to AI models evolving from text-based to high-definition video, real-time audio, 3D generation, and long-range reasoning, driving exponential growth in token consumption. This demand expansion will be sufficient to absorb newly released capacity quickly.

In his projection, 2028 will at most transition the market out of absolute shortage into a healthy state of slight tightness, with demand-to-supply ratios around 105% to 110%, avoiding any cliff-edge oversupply.

He also suggests that the current supply bottleneck serves as a useful "safety valve," slowing the pace of overinvestment.

Comparing this AI infrastructure cycle to historical railroad, canal, 3G, and mobile internet booms, he notes those eras suffered severe overbuilding. The key difference this time: even with clearly demonstrated demand, the physical construction speed of production facilities and data centers imposes natural limits, and supply constraints themselves curb excessive investment.

He adds that in enterprise AI deployments, high token costs and limited compute availability have objectively slowed migration from experimental environments to full production.

Jay Goldberg's Perspective: Historical Cycle Rules Suggest 2028 Pricing Power Reversal

Jay Goldberg's contrarian warning derives from five decades of semiconductor capital cycle history. His core thesis: in every upcycle, all participants proclaim "this time is different," yet each cycle concludes with severe capacity oversupply.

The scale of currently under-construction projects is alarming:

TSMC is advancing nearly 20 new wafer fabs and packaging facilities globally; Micron has 7 major memory plants under construction worldwide; Samsung and SK Hynix are making equally aggressive HBM investments; government-subsidized capacity from Intel and others adds further supply.

Goldberg also highlights the historical pattern of double and triple booking.

During chip shortages, customers like Apple, Qualcomm, and cloud providers submit overlapping or even doubled capacity commitments across multiple suppliers to secure allocation. Once supply begins to release and lead times shorten, these inflated "ghost orders" are cancelled almost instantly.

On pricing, Goldberg argues that NVIDIA's current gross margins exceeding 70% and Broadcom's premium ASIC pricing represent scarcity premiums created by supply shortages.

By 2028, when massive fab capacity comes fully online and fixed asset depreciation creates enormous cost pressures, foundries and chip designers will inevitably engage in fierce price competition to maintain utilization rates. Hardware gross margins will face brutal mean reversion.

The Investment Consensus: A Selective Era Replaces Broad-Based Gains

Despite their significant divergence on the 2028 outcome, both analysts agree on one crucial market assessment: the window for indiscriminate AI supply chain investment yielding outsized returns has closed.

Over the past two years, any company touching advanced packaging, HBM, GPUs, or high-speed optical modules saw substantial stock price gains driven by the sector-wide premium from single-point bottlenecks.

Going forward, supply chain constraints are dispersing from single segments like early CoWoS packaging into multiple parallel dimensions.

As custom computing players such as MediaTek and Marvell accelerate their presence, and massive capacity releases cascade through 2028, hardware suppliers lacking irreplaceable software ecosystems or core IP moats will face dual pressures of margin compression and inventory digestion between late 2027 and 2028.

The joint conclusion from both analysts: investors can no longer treat the semiconductor supply chain as a homogeneous whole. They must concentrate on core segments with genuine cross-cycle pricing resilience to secure advantageous positions in the differentiated landscape that follows history's largest capacity expansion.