Table of Contents
Foundry bottlenecks at TSMC and rising demand are forcing a shift to long-term contracts
In sum – what we know:
- A production ceiling – Broadcom executives note that TSMC’s capacity, previously viewed as effectively unlimited, has hit a limit that is now a primary bottleneck for the AI supply chain.
- Systemic strain – The shortage extends beyond advanced logic chips to include essential “plumbing” like laser optical parts and printed circuit boards, which are hitting their own capacity walls.
- Long-term hedging – In response to the volatility, major players are locking in supply contracts through 2028, a dramatic departure from the short-term purchasing arrangements that previously dominated the sector.
TSMC maybe be feeling the squeeze a little. Broadcom’s Natarajan Ramachandran, director of product marketing in the company’s Physical Layer Products division, recently flagged that TSMC is bumping up against the ceiling of its production capacity. On the surface, that might read like standard supply chain noise. But Ramachandran put a finer point on it — just a few years back, he would have described TSMC’s capacity as essentially “unlimited.” That era, apparently, is done.
The ripple effects are already showing up across the industry, with little reason to expect near-term fixes. Ramachandran painted TSMC’s position as a bottleneck that’s actively choking the supply chain. The foundry giant has expansion plans running through 2027, but the ramp isn’t keeping up with the demand curve right now.
Big implications
What’s driving this squeeze isn’t exactly a mystery. It’s the relentless demand for AI chips and the sprawling infrastructure architecture around them. Every major AI player is fighting over the same finite pool of leading-edge fab capacity. For AI specifically, this creates a real problem — the pace of model training, inference deployment, and next-gen hardware rollouts could hit physical constraints that no amount of clever software can route around.
Nvidia, one of TSMC’s biggest customers, has been vocal about pushing for more wafer allocation as supply tightens, and they’re hardly the only ones feeling the pressure. Google’s TPU pipeline is directly affected too, since Broadcom helps design the ASICs for Google’s TPUs on TSMC’s advanced 3nm process.
The industry response has been to lock in supply as far out as possible. Broadcom CEO Hock Tan disclosed that the company has secured capacity for critical components, like leading-edge wafers, high-bandwidth memory, and substrates, through 2028. That level of forward commitment would have been unusual not long ago. Now, three-to-five-year contracts are becoming the norm across the AI sector, a dramatic departure from the short-term purchasing arrangements that used to dominate.
A supply chain pattern
The pressure extends well beyond TSMC’s fabs. Constraints are cropping up in the laser component sector, where limited availability of certain optical parts is squeezing high-speed data center interconnects. These components don’t generate the same headlines as cutting-edge process nodes, but they’re essential plumbing for AI workloads running at scale.
Even more unexpectedly, printed circuit boards have emerged as a serious chokepoint. PCB manufacturers in both Taiwan and China are hitting capacity walls, stretching lead times and amplifying the broader supply chain strain. PCBs are about as foundational as it gets in electronics hardware, when their supply tightens, the effects cascade everywhere.
What this pattern really underscores is that semiconductor availability alone isn’t the binding constraint. The entire ecosystem needed for AI chip production, is under simultaneous stress. Scaling AI infrastructure at the pace the industry wants requires every link in the chain to expand in lockstep. Right now, multiple links are buckling at once.
So what now?
None of this should come as a total surprise to anyone watching closely. TSMC itself acknowledged back in January 2026 that capacity was tight, pointing to massive demand driven by the ongoing AI infrastructure buildout. The company said it was working to close the gap between supply and demand, but its near-term outlook remained constrained.
The gap between AI-driven demand and manufacturing capacity keeps widening, even with active expansion across the industry. TSMC is building new fabs, other foundries are investing heavily, and yet the demand curve is simply outrunning supply. Samsung Electronics has reinforced this from the memory side, noting that major customers are shifting toward three-to-five-year contracts, mirroring exactly what’s happening on the logic and foundry front.
The practical question for the AI industry is how much this actually slows things down. Forward hedging strategies, like Broadcom’s agreements running through 2028, can insulate individual companies to a degree, but they don’t conjure more aggregate capacity out of thin air. It’s worth remembering that supply crunches tend to be cyclical. The semiconductor industry has weathered them before, and the enormous capital flowing into new fabs will eventually come online. But “eventually” might mean 2028 or later. In the meantime, the companies with the deepest pockets and the longest-dated contracts will hold a meaningful edge over everyone else trying to build at the frontier of AI.
