2026-05-05-minseok-kim-bliss-dustin

Interview: Minseok Kim — 2026-05-05

Participants: Bliss Perry, Dustin Ross, Minseok Kim

Key Themes

Capacity Reservation is Opaque and Power-Driven The process for reserving fab capacity (e.g., at TSMC) is fundamentally non-transparent. It relies on insider connections and raw market power — large consumers like NVIDIA can throw themselves around, while smaller players get pushed out. The main stakeholders who might want to reform this are the hyperscalers (Amazon, Google, Meta), who are unhappy with NVIDIA’s dominance and are building their own chips. But TSMC itself has no incentive to change: as customers compete and fight, they all still come to TSMC. “The more they fight, the better.” The same dynamic holds in memory — all roads lead to SK Hynix and Samsung for HBMs. Samsung’s market cap quadrupled in the past year on this dynamic.

SemiAnalysis as the Closest Existing Digital Twin If given $20M, Minseok would tackle supply chain visibility — essentially building a better SemiAnalysis. He views SemiAnalysis as the closest existing version of a semiconductor supply chain digital twin. Their articles reveal deep insider knowledge that almost certainly violates NDAs, but they get away with it because they’re a small company with well-placed sources. Building something better is the interesting but extremely hard problem.

Demand Opacity is the Core Problem — and It’s Structural Even the suppliers don’t know actual demand. NVIDIA has limited visibility into when HBM3 will ship from SK Hynix. Samsung didn’t know exact customer quantities until APOs arrived — Minseok personally experienced this while sending samples. The opacity is both structural (mechanically complex, hard to predict) and strategic (sharing information undermines negotiating leverage). Minseok compared it to pharma vaccine production: you don’t know the yield, you don’t know when the next batch will be production-ready.

Memory Allocation is Dynamic and Commodity-Like Memory chips are closer to a commodity than logic chips — they can be switched between customers. Allocation of wafers changes in real-time based on willingness to pay. An NVIDIA that’s willing to pay more can potentially get delivery moved from week 35 to week 33. But upstream suppliers don’t want customers to have that visibility, because it reduces their pricing power.

Capacity is Fixed in the Short Run Switching production lines within existing facilities tanks yield — different humidity, vibration, equipment calibration. SemiAnalysis estimates capacity changes via satellite imagery of fabs. Samsung’s Taylor, TX fab has been under construction since 2019 and is still not online. Strategic finance teams at these companies are extremely conservative about new capacity investment after decades of boom-bust cycles. In practice, you can treat capacity as fixed in the short run and only model long-run fluctuations.

US Fab Building Faces Structural Labor Challenges Semiconductor manufacturing is brutally labor-intensive: 24-hour shifts, thousands of workers, “graveyard” shifts (10 PM to 6 AM). Workers need significant vocational training; managers need engineering degrees. Minseok does not believe you can find 10,000 American workers willing to do this, even with imported Taiwanese/Korean expertise, because the environmental adjustment costs are real — a transplanted fab doesn’t yield like the original.

Hanmi Semiconductor vs. Samsung — Supplier Power Dynamics Hanmi Semiconductor has a monopoly on a specific manufacturing process. When Samsung tried to pressure them for a 30% discount, Hanmi refused (“Sure, f*** you”). Hanmi shifted supply toward TSMC instead. This is cited as a contributing factor to Samsung’s current competitive difficulties. It illustrates that supplier relationships in semiconductors are more about power politics than rational optimization.

Supplier Lock-In Makes Remediation Impractical at Scale Changing hardware suppliers is extremely difficult. Engineers and systems are calibrated to existing suppliers. At scale (Meta buying 20,000 GPUs), switching is almost unthinkable — and suppliers have strong incentives to keep large customers happy. The marketplace/alternate-sourcing play faces massive friction. In smaller quantities, switching is feasible.

Compliance Wedge May Not Yield Digital-Twin-Grade Data Minseok is skeptical that compliance screening (UFLPA, entity list checks) generates data granular enough for a real digital twin. Knowing whether someone is buying from a Uyghur forced-labor-linked entity is binary — it doesn’t give you production volumes, capacity utilization, or pricing. “That’s not much value — you can’t upsell that.”

Aggregate Supply/Demand Model Could Be Useful but Limited TAM A country-level aggregate supply/demand model (how much capacity exists in Taiwan, how much NVIDIA and the hyperscalers need) could serve capex planners at major firms. But the customer base is tiny — Samsung, SK Hynix, TSMC, some memory vendors, some investors. This essentially replicates what SemiAnalysis already does with their data center and accelerator models. Not a billion-dollar TAM on its own, but potentially a wedge.

Agent Workload → Supply Chain Simulator: “The Dream” The most energized part of the conversation. Minseok’s summer project at NVIDIA involves understanding hardware implications of agentic AI workloads. NVIDIA doesn’t have access to agent traces from Anthropic or OpenAI and doesn’t know what agent workloads look like. The billion-dollar question: how does enterprise agent adoption (warehouse management, inventory, etc.) translate into supply chain constraints and demand for hardware? Minseok said a “workload to supply chain vertical simulator” is something he would personally want to help build — “that’s literally the dream.” This converges with Nicole’s “5-layer cake” framework and hedge fund investor Nihar’s interest in mapping AI architecture evolution to hardware implications.

Government Unlikely to Force Data Sharing Chip supply volumes, pricing, and customer allocation aren’t considered national security information — they’re trade secrets. Governments lack both the authority and the mechanism to compel companies to hand over this data. The compliance angle doesn’t bridge this gap.

Notable Quotes

• “It’s not even one [predictable customer], it’s like the more they fight, the better.” (on why TSMC benefits from hyperscaler competition)
• “SemiAnalysis is the closest version of the supply chain digital twin that you have.”
• “It’s clear that they violated a bunch of NDAs, but they get away with it because it’s a small company.”
• “Even the suppliers don’t know. So if as a memory vendor I have visibility into what NVIDIA is going to need, how many chips by when, for what — that makes things so much more transparent and efficient. But it’s in nobody’s interest to share that.”
• “It’s both unwillingness to share plus you really don’t know. It’s a data problem… almost like pharma.”
• “There is definitely a problem. And if you can solve it, I think you’re going to be a billionaire.”
• “If you can build a workload to supply chain vertical simulator — that would be an idea that even I would want to be part of, because that’s literally the dream.”
• “We literally call it GY, which is graveyard… it’s very brutal, it’s very taxing. I don’t think you can find 10,000 American workers who are going to do that.”

Surprises

• Samsung’s market cap quadrupled in one year — a company already massive in absolute terms — driven almost entirely by the HBM demand dynamic. This is a concrete marker of how much value is accruing to memory right now.
• The Hanmi Semiconductor story is a vivid illustration that even in a consolidated industry, monopoly positions on specific process steps give small firms real power. Samsung’s competitive difficulties are partially attributed to a supplier relationship gone wrong — not a technology or market failure.
• Minseok’s skepticism about the compliance-to-digital-twin pathway is the most direct challenge to the thesis from someone with operational experience. The data you get from compliance screening is binary, not granular — it doesn’t compound into the kind of model you’d need.
• The convergence across three separate interviews (Minseok, Nicole, Nihar) on “trace AI workload evolution to hardware supply chain implications” as the most valuable analytical problem. This is a strong signal.

Open Questions

• What does Minseok’s summer project at NVIDIA actually look like? Could there be a collaboration or data-sharing angle once he’s there?
• How does the Hanmi/Samsung situation map to other supplier relationships in the semiconductor stack? Are there other single-supplier chokepoints creating similar political dynamics?
• Is the “workload to supply chain simulator” concept viable as a startup product, or does it require the kind of proprietary data (agent traces, workload profiles) that only the AI labs and hyperscalers have?
• Minseok suggested looking into cloud procurement (GCP, AWS, Azure, NeoCloud): how do they procure GPUs and what are their pain points? Is supply constraint as bad as reported?
• Connect Bliss with Nicole at NVIDIA (Minseok’s suggestion) — Bliss mentioned Nicole is already a contact, and the Nicole interview is in the vault. Follow up on the 5-layer-cake framework convergence.