During the May 2026 semiconductor sector-wide selloff, NVIDIA’s stock declined by just 0.8%, while Broadcom plummeted 15% and AMD fell 3.6%. This divergence was no random fluctuation—it reflected deep structural differences in business models, customer concentration, and technological trajectories. The market is voting with capital: the “winner-takes-most” logic of AI infrastructure has entered its second phase, shifting from technical feasibility to commercial sustainability.
NVIDIA’s core advantage lies in its full-stack AI ecosystem. The CUDA software moat is far harder to replicate than raw hardware performance. Even as AMD launched MI300X and Broadcom rolled out custom AI accelerators, enterprise customers remain unwilling to bear migration costs. While cloud providers like Microsoft Azure, Meta, and Oracle have experimented with alternatives, actual deployments still rely overwhelmingly on NVIDIA’s H100 and B100 GPUs. According to Q1 2026 earnings, NVIDIA’s data center revenue grew 78% year-over-year, with 90% coming from hyperscalers—a high concentration that paradoxically acts as a buffer against cyclical volatility.
Broadcom’s sharp drop, by contrast, exposed its reliance on an acquisition-driven growth model. Despite strengthening its enterprise software portfolio through the VMware deal, its AI chip strategy remains reactive—designing application-specific ASICs for individual clients like Google’s TPUs or Amazon’s Trainium chips. This approach thrives during capital expenditure booms but falters when cloud spending slows. In early 2026, AWS announced cuts to non-core AI projects, immediately undermining Broadcom’s order visibility. Crucially, Broadcom lacks a unified software abstraction layer; each chip requires its own toolchain, preventing economies of scale.
AMD faces a more complex dilemma. It seeks to challenge NVIDIA in general-purpose GPUs while maintaining cash flow from traditional CPUs. It champions open ecosystems like ROCm yet compromises with closed-system partnerships, such as co-developing AI PC chips with Microsoft. This strategic ambiguity becomes fatal when market sentiment turns risk-averse. Although MI300 series specs appear competitive with H100 on paper, real-world training efficiency lags due to immature software optimization. In April 2026, a major North American bank canceled its planned MI300X deployment and instead ordered additional H200 units, citing “unacceptable delays in model iteration caused by debugging overhead.”
Notably, the selloff was triggered not by weakening demand but by tightened U.S. export controls on advanced computing. New Treasury rules require additional licenses for any AI chip containing more than 25% U.S.-origin technology destined for China. NVIDIA, with its compliant H20/H200 product lines tailored for the Chinese market, has already established regulatory pathways. Broadcom and AMD have not, raising investor concerns about revenue disruption. In 2025, China accounted for roughly 18% of Broadcom’s total revenue and 22% of AMD’s—but neither discloses AI-chip-specific regional breakdowns, amplifying uncertainty.
From a capital allocation perspective, NVIDIA reinvests over 70% of its free cash flow into R&D and wafer capacity commitments—primarily through TSMC’s CoWoS advanced packaging lines in Taiwan, China—creating a virtuous cycle. Broadcom, meanwhile, returned $12 billion to shareholders via buybacks in 2025. AMD, constrained by scale, pursues a “focused breakthrough” strategy but cannot cover the full stack. These divergent approaches are magnified during downturns.
I judge that the next 12 months will serve as a stress test for the AI chip landscape. If the global economy achieves a soft landing and cloud capex rebounds, Broadcom could recover on the strength of its customization edge. If recession deepens, NVIDIA’s ecosystem moat will further narrow the competitive window. As for AMD, unless ROCm achieves a breakthrough in developer adoption, its share in the high-end training market may remain stuck in single digits. The real wildcard lies ahead: as AI inference costs approach the limits of Moore’s Law, will the market reassess the cost-efficiency of general-purpose GPUs? That could mark the next structural inflection point.
The question today is not who will replace NVIDIA, but whether rivals can build sustainable differentiation without relying on a “second CUDA.”