Ameioud: Supply Chain

Understanding TSMC's Role in Global Chip Supply

2026-03-27T00:00:00Z

Something that doesn't get discussed enough when people talk about semiconductor supply chains is how much of the actual concentration problem is a knowledge problem as much as a geography problem, and why that distinction matters for how you think about what a disruption would actually look like.

The basic facts are by now well-repeated: Taiwan Semiconductor Manufacturing Company produces somewhere around 90 percent of the world's most advanced chips, the ones running below 5 nanometers, the ones inside the systems that power modern data centers, defense applications, medical imaging equipment, and precision scientific instrumentation of the kind that companies working in quantum sensing and adjacent fields depend on entirely. The geography part of that observation, which is to say the proximity to China and the status of the Taiwan Strait as one of the more closely watched stretches of water in the world, tends to be where the public conversation starts and sometimes also where it stops.

What gets less attention is that the concentration problem wouldn't disappear even if you could relocate TSMC's facilities to somewhere geopolitically calmer, because the more fundamental issue is that advanced semiconductor fabrication is arguably the most technically demanding manufacturing process ever attempted at scale, and the knowledge required to run it at TSMC's level of precision has been built up over decades in a way that doesn't transfer easily or quickly to a new site, a new workforce, or a new set of process conditions.

This matters in practical ways that are not always obvious from the outside. When working with suppliers of specialized electronics, whether FPGAs, control boards, or the kinds of custom components that go into high-precision sensing equipment, what you tend to encounter isn't usually someone saying that their direct supply of chips has been cut off. What you encounter instead is lead times that have stretched by months without much warning, quoted availability that shifts between the time of inquiry and the time an order is placed, and equipment manufacturers citing component availability as the reason a delivery window has moved, without always being specific about which component or why. The shortage of 2020 and 2021 made this pattern visible to a broader audience than it had reached before, but the structural condition that produced it didn't resolve when the shortage eased, and the long lead times on certain categories of components have become something that procurement planners have quietly absorbed into their standard assumptions.

The announced fabs in Arizona, Japan, and the ESMC facility being built in Dresden, a joint venture involving TSMC alongside Bosch, Infineon, and NXP that sits close enough to Frankfurt to feel like it should matter more than it currently does in day-to-day supply chain conversations, are real investments and represent a genuine effort to distribute production geographically. The reasonable question to hold alongside that news is what those facilities will actually produce once they're operational, because the early phases of both the Arizona and European sites are focused on processes that are advanced by general standards but older-generation by TSMC's current output, meaning they are not running the 3nm or 2nm nodes that the most sensitive applications depend on, and the gap between where those fabs start and where leading-edge production sits is not a small one.

There is a theory sometimes called the Silicon Shield, the idea that Taiwan's position as the world's most critical chip supplier functions as a form of deterrence, that no major economy with meaningful exposure to Taiwanese production has a strong incentive to tolerate the kind of instability that would disrupt it. The logic has something to it, but it also places considerable weight on the assumption that decisions affecting the Taiwan Strait will be made with that calculus clearly in view, which is the kind of assumption that tends to feel robust right up until the conditions that tested it actually arrive.

What is harder to find in the public conversation is a sober accounting of what recovery from a significant supply disruption would actually require in terms of time, not because disruption is particularly likely in the near term, but because the feedback loops in semiconductor manufacturing are long in a way that makes them easy to underestimate. Production capacity that goes offline doesn't come back in a quarter. Yield rates on advanced nodes take time to rebuild even under ideal conditions. The engineers and technicians who run these processes at production scale are not drawn from a large available pool, and neither are the specialized equipment suppliers, the chemical inputs, or the accumulated process knowledge that makes one fab's output meaningfully different from another's.

Whether the new facilities being built outside Taiwan will eventually close the gap on leading-edge production is a question probably measured in years rather than quarters, and whether the organizations most exposed to that risk are building that timeline into their planning is, at least from the conversations that tend to happen around supplier diversification and component risk, not yet a settled matter.

China Plus One Is Easier to Announce Than to Execute

2026-04-24T00:00:00Z

The phrase China+1 became standard vocabulary in supply chain planning quickly enough that it started to feel like the decision had already been made. Diversify away from China, add a second source somewhere in Southeast Asia, South Asia, or Latin America, and the concentration risk goes down. The logic isn't wrong. The execution is significantly harder than the strategy slide suggests.

Working with electronics suppliers, what you run into repeatedly is that capacity outside China isn't the same as Chinese capacity in terms of what processes it can run, what yields it can reliably hit, or what lead times it actually operates on. Vietnam has genuine manufacturing depth in certain product categories. It is not a substitute for the full range of what southern China produces, and the difference matters when you're sourcing something with tight tolerances, specific material grades, or components that depend on a supplier ecosystem that took decades to cluster in one region.

India is the other name that comes up. The government's Production Linked Incentive schemes have moved real investment. Electronics assembly, pharmaceutical production, components that didn't exist there five years ago. The question is scale and consistency. Some of the new facilities are excellent. Some are still building the process knowledge and quality systems that take years to develop, and the difference between a factory that can quote your part and one that can actually ship it to specification consistently is something that site visits and audits don't always catch until you're already in production.

Mexico entered the conversation partly because of nearshoring logic, shorter lead times to North American markets, lower freight. For certain product categories the fit is real, particularly where labor content is high relative to complexity. For others, the supply base isn't deep enough yet, and the investment required to develop it takes longer than any sourcing decision timeline usually allows.

None of this means the strategy is wrong. Some version of China+1 is the right answer for a lot of supply chains, and the last several years have made the case for diversification in ways that are hard to argue with. The part that tends to be underplanned is the transition, the period when you're running dual supply chains, absorbing the cost difference between your established base and your developing one, and managing quality uncertainty while scaling a new supplier to real volumes. That's where the estimates usually break down.

The transition has a cost. It also has a time horizon that most organizations underestimate by roughly a factor of two.

How AI Is Changing Procurement, and Where It Still Falls Apart

2026-06-13T00:00:00Z

Three years ago, creating a purchase order at one of the companies I worked at took about 15 minutes. You received a quote from a supplier, manually pulled the data into a template, checked the pricing, formatted the PDF, sent it for approval, uploaded it to a shared drive, and logged it in a spreadsheet. Each of those steps was done by a person. Each of them introduced a chance of error. Fifteen minutes per PO, for every PO, across every department.

Now it takes 3 minutes.

That change did not come from a procurement software platform. It came from building a specific tool for a specific problem. The PO Automator reads the incoming supplier offer, extracts the relevant data using AI, pre-fills the order form, generates the PDF, uploads it to Google Drive, and logs the transaction automatically. The person in the loop reviews and approves. The manual data entry is gone.

We did the same thing with invoices. Invoice logging used to be one of those tasks where someone opened a PDF, read the data, typed it into a spreadsheet, and filed the document somewhere. An AI-powered tool now extracts that data automatically and logs it with above 95% accuracy. Finance, procurement, and logistics all use it daily. Nobody had to be convinced to use it because it was obviously faster than what came before.

The file organization problem was a different kind of challenge. Fifty thousand archived documents spread across a shared drive with no consistent naming convention and no structured folder logic. Another AI tool read each document, generated a standardized file name, and moved it to the correct supplier folder. That ran for a few weeks. The manual version would have taken months.

So AI works in procurement. For specific, high-volume tasks with consistent inputs, structured documents, known fields, it performs well. It handles extraction, classification, logging, and naming without drama. These are not complicated problems in themselves. They were consuming hundreds of hours a year.

Where it still falls apart is where procurement actually gets difficult.

Supplier negotiations are the clearest example. AI can prepare you: market pricing, historical spend, comparable quotes from the last cycle. That preparation is genuinely useful. But negotiation is about the person on the other side, understanding their margin pressure, reading when they have flexibility and when they do not, knowing when to close and when to walk away. No tool I have worked with comes close to replacing that judgment.

New supplier evaluation has a similar limitation. You can automate the RFP scoring. You can build a structured evaluation framework. But when you are qualifying a supplier for a critical component in a product category you have not sourced before, part of your evaluation is how they handle the conversations that do not go smoothly: a missed deadline, an ambiguous specification, a quality dispute. That information does not come from a form.

Contract risk sits in the same category. Standard terms, manageable. But the real exposure in most supplier contracts lives in the edge cases: what happens at price review, who absorbs cost increases from raw material volatility, what the remedies actually are when delivery fails. AI can draft standard clauses. It cannot negotiate the terms that matter.

The honest assessment is that AI has made procurement faster at the things that were tedious and rule-based, and has not changed procurement at the things that were genuinely hard. The companies getting real value from it are not the ones who bought the most expensive platform. They are the ones who identified the specific high-volume tasks that were burning hours and fixed exactly those.

The tasks that are still hard are hard for the same reasons they always were.