OpenAI’s AMD & Broadcom chip deals ratchet up pressure on TSMC

OpenAI has signed large, multi‑year agreements with AMD and Broadcom to secure vast quantities of AI silicon: AMD will produce roughly 6 gigawatts of GPUs (with the first ~1GW due in late 2026), while Broadcom will deliver about 10 gigawatts of AI accelerators and networking systems starting in the back half of 2026 through 2029.

These deals aim to give OpenAI custom hardware tuned for inference and training workloads, diversify its supply away from a single vendor, and (per analysts) reduce unit cost compared with fully in‑house designs. But there’s a bigger industry story: no matter who designs the chips, almost all advanced nodes are manufactured at a single company — TSMC.

Why TSMC matters

• TSMC leads in advanced process tech (3nm and beyond) and high yield rates, making it the go‑to foundry for bleeding‑edge GPUs and accelerators.
• Most partners (AMD, Broadcom, NVIDIA, Apple and others) rely on TSMC’s fabs, so demand growth concentrates pressure on a single supplier and its capacity schedule.
• TSMC’s recent financials and comments show capacity is “very tight,” and even large customers have faced long lead times for production slots.

Technical and supply implications

Advanced chips require EUV lithography and highly refined process knowledge — capabilities that took TSMC decades to build. Competitors such as Intel and Samsung are investing to catch up, but industry analysts say it will take years, if not decades, to materially reduce global dependence on TSMC for cutting‑edge nodes.

• Foundry bottlenecks: A surge in GW‑scale procurement increases demand for wafer capacity, testing and packaging — all of which are constrained resources.
• Geopolitical risk: Concentration of production in Taiwan raises systemic vulnerability concerns for global tech supply chains.
• Cooling and power: High‑density AI racks require huge datacenter power and cooling investments; industry measures GW figures partly to reflect total power footprints, not just chip counts.

What’s next

TSMC is expanding — building fabs in Taiwan and the US (including facilities in Arizona) and planning future nodes — but capacity growth is incremental and costly. OpenAI’s deals with AMD and Broadcom will help secure wave after wave of hardware, yet they also intensify the race for TSMC slots and underscore how the AI boom stresses the entire semiconductor ecosystem.

For stakeholders this means watching three things closely: delivery schedules from AMD and Broadcom, TSMC’s capacity expansion timelines and geopolitical developments that might affect supply from Taiwan.

Discussion: Do OpenAI’s procurement deals make sense as a way to scale AI quickly — or do they simply expose the industry to higher systemic risk by concentrating production at TSMC? What steps should governments and companies take to reduce foundry dependence?