Why TSMC’s capacity crunch is reshaping AI chip supply chains
The immediate problem is straightforward: demand for AI accelerators has outstripped TSMC’s ability to produce them. Every major hyperscale cloud provider and chip designer wants wafers on TSMC’s most advanced nodes, and the Taiwanese foundry’s production lines are fully committed. That bottleneck has forced Google and Nvidia, two of TSMC’s largest and most loyal customers, to look elsewhere for the first time in years. Google’s response has been concrete and large. The company ordered 3 million TPUs from Intel, with delivery expected as early as 2028. That volume is significant by any measure. It signals that Google views Intel not as a stopgap but as a production-scale partner capable of handling a meaningful share of its AI chip needs, and it gives Intel a clear demand signal as it ramps its contract manufacturing business. Nvidia’s approach is more cautious but still telling. The company is evaluating Intel’s advanced packaging capabilities alongside the 18A process node, according to reporting on backup manufacturing. Nvidia has not placed a confirmed production order, but the fact that it is running tests on Intel’s technology reflects a calculation that relying on a single foundry carries growing risk as AI workloads scale. Other coverage underscores how both companies are probing alternatives. Reports that Google and Nvidia are contemplating Intel-made chips frame the move as a strategic hedge rather than a full-scale relocation away from TSMC. At the same time, analysis noting that the pair are exploring Intel as a backup highlights how fragile the current supply-demand balance has become. If TSMC’s ramps on its next-generation 2 nm and 3 nm nodes slip by even one quarter, Intel could capture a meaningful share of AI accelerator wafer starts from at least two of the industry’s biggest buyers. The hypothesis is testable: Intel’s ability to win and hold these customers depends on whether its 18A process can match TSMC’s yield and performance within a narrow window. A delay on TSMC’s side widens that window considerably and could lock in long-term dual-sourcing patterns.Google’s 3 million TPU order and Nvidia’s 18A tests
The scale of Google’s commitment stands out. Three million TPUs is not a trial run or an engineering sample. It is an order large enough to anchor a production line and give Intel the revenue visibility it needs to justify continued investment in its foundry services business. Intel has struggled for years to attract outside customers to its manufacturing operations, and Google’s order is the strongest validation yet that the effort is gaining traction with top-tier, technically demanding buyers. Strategically, the order also gives Google leverage. By demonstrating that it can move critical AI hardware designs to a second foundry, Google reduces its exposure to TSMC’s allocation decisions and potential geopolitical shocks. If Intel proves capable of delivering on time and at competitive performance, Google gains a credible alternative that can be dialed up or down depending on cost, availability, and roadmap alignment. Nvidia’s testing phase carries different implications. The company designs the most widely used AI training chips in the world, and any shift in its manufacturing strategy would ripple across the entire supply chain, from memory suppliers to data center builders. Nvidia is testing Intel’s technology to determine whether it can serve as a viable backup for future GPU designs, assessing not only transistor performance but also packaging density, thermal behavior, and integration with high-bandwidth memory. No public timeline exists for when Nvidia might move from testing to ordering, but the evaluation itself puts pressure on TSMC to prioritize Nvidia’s allocation and roadmap needs. It also signals to other foundries that the door is open, at least in principle, for advanced GPU production if they can meet Nvidia’s rigorous technical and volume requirements. Both moves reflect a shared calculation. TSMC’s capacity constraints are not a temporary blip tied to a single product cycle. AI chip demand has been climbing quarter after quarter, and the foundry’s expansion plans, while aggressive, have not kept pace with hyperscale build-outs and model training budgets. Google and Nvidia appear to have concluded that waiting for TSMC to catch up is riskier than qualifying a second source now, even if Intel’s technology is not yet proven at the same scale.Open questions about Intel’s yield, timeline, and staying power
Several gaps in the evidence make it difficult to assess how durable this shift will be. No public data confirms Intel’s yield rates on the 18A process. Yield, the percentage of usable chips produced per wafer, is the single most important metric for any foundry customer. If Intel’s yields are significantly lower than TSMC’s, the cost advantage of diversifying disappears quickly, and any theoretical capacity gain may be offset by higher effective prices per working chip. Neither Google nor Intel has released official statements confirming the specific process node or packaging technology that the 3 million TPU order will use. The absence of that detail matters because it determines whether Intel is competing at the true leading edge or filling a tier just below TSMC’s most advanced offerings. A design taped out on a slightly older node might be easier for Intel to manufacture but would reduce performance per watt and could constrain Google’s most demanding AI workloads. Nvidia’s internal test results and qualification timelines for the 18A process have not been disclosed. Without those benchmarks, the market is left to infer Nvidia’s level of confidence from the fact that testing is underway, not from hard performance data. If Nvidia ultimately decides not to proceed with Intel for production, it would raise questions about whether Intel can meet the stringent requirements of the highest-performance AI accelerators. TSMC’s own capacity expansion plans add another variable. The company has announced new fab construction in Arizona and Japan, and any acceleration of those timelines could reduce the urgency that is currently pushing customers toward Intel. Conversely, delays in bringing those fabs to volume, or challenges in staffing and equipment installation, would reinforce the case for dual sourcing and could prompt additional large orders from other AI chip designers. There is also the question of Intel’s staying power as a contract manufacturer. Building a competitive foundry business requires not just one or two marquee wins but a sustained pipeline of designs across multiple nodes and packaging generations. If Intel cannot maintain its technology cadence or if it struggles to scale up capacity profitably, early customers could find themselves once again relying primarily on TSMC, with all the concentration risk that entails. For companies building products that depend on AI chips, whether cloud services, autonomous vehicles, or enterprise software, the practical consequence is clear. The supply of advanced AI silicon is no longer controlled by a single company’s production schedule. A second credible source changes the negotiating dynamics for every buyer in the chain, introduces new technical and commercial trade-offs, and marks the beginning of a more multipolar era in leading-edge chip manufacturing-provided Intel can turn today’s tests and headline orders into reliable, high-yield production at scale. More from Morning Overview*This article was researched with the help of AI, with human editors creating the final content.
