01 Trigger Event

36Kr citing CLS News: Musk-led Terafab manufacturing project claims capacity for 1 terawatt of compute power. Semiconductor industry experts pushed back—getting fab equipment to high-volume production yield requires at least 30 years, and "achieving stable yield at Terafab may be harder than putting Starship on Mars."

02 What This Really Means

What this really means isn't something I just figured out today; it's something Musk's rhetoric pushed back onto the main stage: the supply-side bottleneck in AI compute has evolved from "whether there's GPU capacity" to "who controls the capacity to manufacture GPU capacity."

These two layers are not the same.

The first bottleneck—the entire industry has been watching this for the past three years—involves H100/B200 allocation, Colossus-scale clusters, and Microsoft/Google/Meta capex growth rates. This is a chip design and packaging layer bottleneck. NVDA is near-monopolistic here, but theoretically it can be circumvented at the architecture layer by ASIC/TPU/MI300 series.

The second bottleneck—the one Terafab is trying to punch through directly—is the fab itself. Lithography, etching, thin-film deposition, CMP: each step's yield is a decade-long accumulation process, not something you can buy with money. An ASML EUV lithography machine costs $200 million, but the real value isn't the machine—it's TSMC's recipe for running these machines.

What Musk is doing is essentially pulling the "vertical integration" bet from the chip design layer (Tenstorrent, Dojo, Trainium) up to the fab layer. This is an order-of-magnitude leap, and it's a wall that even NVDA itself hasn't crossed.

03 Historical Analogy / Structural Comparison

The closest historical parallel, I think, is Samsung's foundry ambition versus TSMC.

Samsung started grabbing foundry orders in the mid-2010s, "leading" process node announcements multiple times (14nm, 10nm—Samsung talked about them first), but customers—including Qualcomm and NVIDIA's early nodes—kept flowing back to TSMC. Yield is the invisible moat.

A more recent example is Intel's IDM 2.0 strategy. Pat Gelsinger's promise back then: Intel 18A would reclaim process leadership while opening up foundry to external customers. Three years in, Intel 18A still hasn't stabilized yield, and foundry customers have largely defected. Intel itself is living proof of how hard it is for top-tier fab players to replicate TSMC's model.

The 30-year number might be exaggerated (Intel's foundry transition didn't start from scratch, and it's already taken 5+ years without stabilizing), but the order of magnitude is right—the barrier to entry at the fab layer is measured in decades, not quarters. I might quibble with the expert's wording, but I agree with the conclusion.

04 What This Means for AI Builders

Short term (6-12 months), no adjustments needed. Even if Terafab runs at full speed, wafers won't come out until 2028+. Near-term capacity remains locked at TSMC + Samsung + SK Hynix; the HBM/CoWoS/SoIC bottlenecks won't ease because of this.

Medium term (12-36 months), a few non-obvious calls:

First, stop modeling "Nvidia being circumvented" as a high-probability event. Even if Grok/xAI gets a self-developed chip, it'll be an internal cost-reduction tool at the same level as Amazon Trainium or Meta MTIA—not a variable that changes the supply landscape. What can change the supply landscape is a 10% capacity addition at the fab layer, and that's something Elon can't solve either.

Second, for companies like opcx doing model routing/token gateways, upstream concentration is a short-term tailwind, not a headwind—the scarcer model APIs are, the larger the gateway's arbitrage space becomes. The path where Musk actually threatens opcx's business isn't self-developed fab; it's if Grok API's long-term pricing stays below 70% of industry rates—that's the metric worth monitoring.

Third, if you're stress-testing long-term AI capex assumptions (for model deployment/privatization/on-prem), don't bet on "fab diversification will reduce NVIDIA premium." It hasn't happened in the past 5 years, and it probably won't happen in the next 3 either.

05 Counterarguments / Risks

I could be wrong in three places.

First, I might be underestimating Musk's ability to weaponize "capital + speed" as an asymmetric advantage. He pulled off the SpaceX script—"traditional aerospace says 30 years, I did it in 5"—and the Tesla factory script—"traditional automakers say 10 years, I did it in 3." Can fab be broken by the same playbook? I can't judge, because I don't have internal data on whether fab's accumulated knowledge is truly "incompressible." The 30-year expert judgment might be industry conservatism, or it might be physical reality.

Second, "30 years" itself might be a strawman. Semiconductor experts saying zero-to-stable yield takes 30 years doesn't mean Terafab from-announcement-to-wafer takes 30 years. Musk's goal might not be first-time-right, but a rapidly iterative learning curve—something repeatedly validated in Starship iteration. If Terafab accepts "first 5 years burning cash on 60% yield," it might not need 30 years.

Third, and where I'm most likely to misjudge: I've been assuming "chip = strategic bottleneck" is the core narrative of 2025-2028. But if any model lab releases a new architecture by end of 2026 strong enough to significantly compress compute demand—say O(N) attention, or an MoE variant matching GPT-5 quality at 10x lower training cost—then the strategic value of any Terafab-class project goes to zero within 12 months. I haven't modeled this scenario internally, but it sits in my own tail-risk list.

One addendum: this 36Kr flash news itself has limited information density—30 years is a single-source retelling, and I can't cross-verify the number. If any readers come across specific semiconductor analyst quotes or more detailed yield curve data, I'd welcome the pointer.