This is my “long game” project. It’s less a polished product and more a running log of me trying to make sense of one very narrow, very important chokepoint in the semiconductor supply chain.
What this actually is (plain language)
This is my “long game” project.
I’m trying to answer a very specific question:
If the world keeps depending on Taiwan for advanced packaging, how do we make the semiconductor supply chain more resilient without just dumping the cost onto the climate?
More concretely: I’m building a tri-objective simulation–optimization model for one very narrow, very important chokepoint in the global semi chain:
- CoWoS / HBM advanced packaging (the step that turns high-bandwidth memory + logic dies into the actual AI accelerators everyone is buying),
- under disruptions (outages, demand spikes, logistics shocks),
- with three objectives in conflict:
- total cost,
- resilience (service performance + recovery time),
- life-cycle carbon emissions.
The goal is not “another generic resilience framework,” but a tool that shows, in numbers, what trade-offs are even possible when you don’t assume that all firms share data and cooperate.
Why I cared enough to go this deep
Short version: I kept reading “just collaborate more” in the literature and didn’t buy it.
Most semiconductor resilience papers end with some version of:
“firms should increase transparency, share data, and collaborate to coordinate responses.”
Sounds nice. Also sounds like something that will almost never fully happen in a world where:
- advanced chips are treated as strategic technologies,
- export controls and IP fights are normal,
- and one company (TSMC) sits in the crosshairs of both the U.S. and China.
So I wanted to flip the question:
Assume no magical cross-firm data sharing.
What can a single enterprise still do, at the CoWoS/HBM stage, to buy resilience without blowing up cost and carbon?
This thesis is my attempt to answer that question in a way that is honest about the politics, not just the math.
Where the thesis is right now
I’m still early, but the spine is there:
-
Motivation & context drafted
I’ve written the introduction on semis as “invisible infrastructure,” why CoWoS/HBM is a chokepoint, and why resilience vs decarb isn’t a fake trade-off here. -
Literature map sketched
I’ve mapped three pillars:- quantitative resilience / disruption modeling (Ivanov, Hosseini, Klos),
- decarbonization + multi-objective optimization (carbon pricing, emission factors),
- semiconductor-specific resilience (Ramírez & Le, CSET, Wu).
-
Research questions locked
Roughly:- What does the cost–resilience–carbon Pareto frontier look like at this chokepoint?
- Which levers (regional capacity split, safety capacity, transport modes) move you along that frontier the most?
- Where do resilience gains directly clash with decarb goals?
Right now I’m in the “turn this into a clean model” phase: formalizing sets, decisions, and objectives in math instead of vibes.
How I’m planning to attack it (explainer version)
1. Model the network around the chokepoint
I focus on a minimal but realistic network:
- HBM suppliers → CoWoS packaging sites (for example: Taiwan, Japan, maybe a “new” region) → downstream DC/OEM.
- Flows move through these nodes with:
- capacity limits,
- lead times,
- disruption risks,
- emission factors for each leg and mode.
Key assumption: no cross-firm data sharing.
All decisions are made with what one firm could plausibly know and control.
2. Optimization layer – design candidates
On top of that network I define a multi-objective optimization problem:
-
Decision variables, like:
- how to split volume across packaging regions,
- how much “safety” capacity to hold,
- what mix of transport modes to use by route (air vs sea, etc.).
-
Objectives:
- expected total cost,
- resilience metrics (service level, recovery time under shocks),
- expected emissions (CO2-equivalent) across production and logistics.
This layer does not spit out “the” best design.
It generates a set of good but different designs that lie along the Pareto frontier.
3. Simulation layer – stress test those designs
Then I throw disruptions at each design using a discrete-event / Monte Carlo style simulator:
- multi-week outages of a CoWoS site,
- demand spikes on AI SKUs,
- combined shocks.
For each candidate design, the simulator measures:
- total cost under disruption,
- how badly service degrades,
- how long it takes to recover,
- and resulting emissions.
That’s how I map each design into the cost–resilience–carbon space.
4. Carbon accounting and pricing
For logistics and operations, I use fairly standard ideas:
- freight emissions ≈ tonnes × distance × mode-specific emission factor,
- process emissions tied to:
- energy per unit at each site,
- regional grid carbon intensity,
- optional carbon price scenarios that convert CO2 into equivalent cost.
The goal is not to predict exact grams, but to make relative trade-offs visible and consistent.
What I expect to learn (and might be wrong about)
Honest version of the hypotheses I’m walking in with:
- You can get surprisingly far on resilience with just private levers (regional split, safety capacity, modes), even without industry-wide coordination.
- There will be zones where resilience and decarb align (for example, moving volume to cleaner but also more diversified sites).
- There will also be ugly regions where resilience basically means “pay more and emit more”, especially under tight carbon pricing and long-distance air freight.
If the model ends up saying something completely different, that’s even more interesting.
Technical appendix (for the nerds)
If you care about the guts, here’s the current direction in bullet form:
-
Problem type:
tri-objective simulation–optimization at a single chokepoint in a multi-echelon supply chain. -
Decision layer:
- Sets: regions r, time periods t, product / flow types p, transport modes m.
- Decisions (examples):
- x_pr: share of product p allocated to packaging in region r,
- c_r: additional “safety” capacity contracted in region r,
- m_prt: share of flow (p, r) shipped by mode m in period t.
-
Objectives:
- f1: expected total cost (capacity, operations, logistics, penalties),
- f2: resilience metric (for example, service loss or time to recovery aggregated over scenarios),
- f3: life-cycle CO2 of packaging plus logistics.
-
Solution idea:
- Use a multi-objective heuristic (NSGA-II/III or similar) to generate candidate designs.
- For each design, run a set of discrete-event / Monte Carlo simulations across disruption scenarios.
-
Emissions accounting (no LaTeX, just plain text):
- freight emissions: CO2 = distance × weight × emission_factor_mode
- process emissions: CO2_packaging,r = energy_per_unit_r × grid_intensity_r
- then run scenarios with different carbon prices to see how the frontier shifts.
This page will evolve as the thesis does. Right now, it’s the place where I keep myself honest about what is actually done vs still just living in my notes.