The US-China Chip War: AI Infrastructure at a Crossroads

The semiconductor ecosystem is not merely expanding; it is undergoing a structural re-alignment driven by the AI supercycle, an accelerating US-China technological separation, and nationalistic campaigns to secure chip supply chains. For Broadcom, a leading designer of networking, broadband, and custom AI accelerators, these converging forces define a high-growth environment freighted with systemic complexity ^23,25. The most significant shift is the broadening of AI infrastructure demand from training to inference, a trend that plays directly to Broadcom’s strengths in high-bandwidth connectivity and custom silicon. Yet geopolitical fragmentation, domestic capacity buildouts, and supply chain bottlenecks are reshaping the competitive landscape and the trajectory of capital expenditure.

The Decoupling Tightens Beyond Chips

The US-China technology rivalry has metastasized into a full-spectrum containment effort. Export controls that began in 2022 as limits on advanced AI chips ¹⁶ have now multiplied to target compute-as-a-service models, electronic design automation tools, and even equipment servicing ²². The so-called “small yard, high fence” architecture is not static; proposed legislation would give Congress direct authority over AI chip exports, tightening the mesh further ¹⁶. Chinese officials denounce these measures as suppressive ¹⁴, but procurement data tells a more complex story: illicit flows through third-country channels in Malaysia and Singapore persist, revealing demand-side pressure that formal restrictions cannot entirely choke off ^16,17. Paradoxically, despite US Commerce approvals, no H200 chips have been shipped to Chinese entities—not because of a legal barrier, but due to domestic Chinese directives that prioritize indigenous capacity over licensed imports ¹⁰. The decoupling, then, is now two-sided: a mutual disengagement that leaves both supply chains less efficient and more exposed.

China’s Capacity Buildout Echoes Historical Overbuilds

China’s response to external pressure is a national-scale sprint to 12-inch wafer fab construction ¹. These fabs are expected to begin operations in 2026 and reach full capacity by 2027, a timeline that mirrors earlier industrial push cycles ¹. The precedent is not comforting. The ramp bears an uncomfortable resemblance to the photovoltaic market crash, where sudden overcapacity destroyed margins globally ¹. The present buildout aims to slash import reliance, but it could unleash a supply glut that compresses pricing across mature and even leading-edge nodes ¹¹. More strategically significant is Huawei’s Tau Scaling Law roadmap, which targets 1.4nm-equivalent chip density by 2031 without extreme ultraviolet (EUV) lithography ¹⁰. This sidesteps the most critical US-controlled tool, potentially bifurcating the industry into EUV and non-EUV ecosystems. Domestic Chinese AI chip suppliers are already projected to capture over 75% of the local market by 2030 ¹⁰, with Huawei replacing Nvidia where sanctions cut deepest ¹⁰. The performance parity of Chinese AI models—now 90–95% of leading US models at 10–20% of the cost—underscores that innovation is not solely dependent on access to advanced fabrication ⁹. This raises a structural competitive threat to Western AI firms that may be underestimated ⁹.

Inference Reshapes Demand, But the Bottleneck Is Physical

The AI market is transitioning from a training-centric phase to an inference-dominant era, a shift that redistributes demand toward low-latency networking, custom accelerators, and edge deployment ^25,27. Inference workloads now account for approximately 70% of total AI demand and are growing independently of training buildouts ^1,4. This is structurally significant for Broadcom, whose Ethernet, Fibre Channel, and custom ASIC products sit precisely at the intersection of inference infrastructure. Hyperscalers are already deploying custom chips like Google’s TPU and recognizing the energy-efficiency advantages of purpose-built designs ^24,26. Yet the underlying constraint is physical: compute supply scarcity remains the primary bottleneck, and the AI memory shortage is projected to persist through 2030, limiting the rate at which inference capacity can be deployed ^5,7.

The Margin of Error Is Thin Across Power, Construction, and Foundry Risk

AI infrastructure growth is colliding with the slow-moving realities of electrical grids and construction. Transformer shortages and turbine backlogs are extending lead times for data center power delivery, forcing utilities like Duke Energy and American Electric Power to revise connection timelines downward ¹². On the construction side, permitting delays and union negotiations in the US are adding friction that Asian building cycles do not face ^4,13. The US fab buildout itself runs roughly 1.5 years behind Asian benchmarks, and sparse service networks for advanced equipment cause cascading repair delays ⁴. TSMC’s Arizona expansion is a live case study: water scarcity, labor shortages, and attrition have repeatedly pushed back ramp schedules ^4,20.

Geopolitical risk around Taiwan remains the system-level fault line. The consensus among supply chain analysts is that a disruption—whether blockade or invasion—would halt output for 6–18 months, an interruption that would propagate across every fabless semiconductor company ⁸. Broadcom, with its heavy reliance on TSMC’s leading-edge foundry services, is exposed not because of a unique dependency but because no alternative foundry can currently replicate TSMC’s scale and performance ¹². The recent pause of a US arms package for Taiwan only underscores the fragility of the status quo ¹⁸. Yet some investors deem invasion fears overstated, noting the catastrophic economic consequences for China itself ⁴. The margin of error here is unquantifiable—a binary risk that cannot be hedged by diversification alone.

Diversification Efforts Are Real, But Late

Regional blocs are attempting to build semiconductor sovereignty, but timelines stretch to the end of the decade and beyond. The EU Chips Act reboot envisions €120 billion in investments by 2035, and India’s partnerships with the US on manufacturing and supply chain resilience are beginning to attract infrastructure commitments, such as the Tata-ASML memorandum of understanding for fab infrastructure ^2,3,15,19. These efforts, while strategically rational, will not alleviate near-term concentration risk. The recent US-China summit concluded without major chip deals, confirming that the divide is durable and institutions are adapting to a bifurcated world ²¹.

Broadcom at the Nexus: Opportunity Meets Systemic Risk

Broadcom’s custom ASIC engagements—Google TPU, Meta, and others—position it to capture disproportionate value from the inference expansion and hyperscaler capex. As model complexity grows, demand for high-bandwidth switching and routing will rise, directly benefiting Broadcom’s networking portfolio. The rise of open-source and Chinese AI models, while potentially compressing margins at the application layer, could broaden AI adoption and thus increase infrastructure demand ^6,9. Broadcom’s hardware layer remains indispensable irrespective of which AI model provider dominates, making it a relative safe haven amid software disruption.

But this growth is contingent. Data center buildout bottlenecks—power, transformers, construction—are the binding constraints on revenue recognition. If these prolong, even secured design wins will translate into delayed shipments. China’s decoupling reduces Broadcom’s addressable market for advanced chips in a historically significant region, though accelerated Western capex may partially offset the loss. The non-EUV trajectory China is pursuing could eventually diminish the premium for EUV-based nodes, altering the cost structure across the industry. Foundry concentration risk, particularly around TSMC and Taiwan, remains the single point of failure that no product roadmap can fully mitigate. Broadcom’s increasing engagement with Intel Foundry Services and Samsung represents the start of a diversification strategy, but these alternatives are not yet at parity. The window for migration to a more resilient foundry mix is open, but the fab build timelines suggest it will not be wide for several years. The structural re-alignment is still underway, and the enterprises that map their exposure to these physical, contractual, and geopolitical bottlenecks now will be the ones that navigate the coming inventory pressures with the least damage.

The Constraint Lies Beneath the Software

The lesson of the current cycle is unchanged from the earlier infrastructure eras: software ambitions rest on physical substrates. The AI supercycle is no exception. The wafer starts, the transformer lead times, the water permits—these are the actual determinants of what gets deployed and when. For Broadcom, the strategic imperative is not merely to win design slots but to secure the supply chain that converts those slots into installed capacity. The margin here is measured in quarters of inventory, kilowatts of power delivery, and clauses in licensing agreements. As China’s fab buildout accelerates and the inference transition deepens, the next phase of growth will belong to those who treat infrastructure not as a background condition but as the primary architecture of competitive advantage.