The press release said two things in one breath, and one of them had to be false.
On one line, the Vanguard International Semiconductor joint venture in Singapore had reached "full production." Four lines later, the same document promised mass production in early 2027. Full production is a terminal state — a fab running at designed capacity, the yield curve flattened, the customer base committed. Mass production is a future event, contingent on tool qualification, defect density, and a demand signal that has not arrived. You cannot be finished and also not yet started.
I have spent most of my adult life reading technology claims the way other people read contracts, hunting for the clause that contradicts the other clause. In early 2017, I sat in an Austin hackathon with four young developers and audited early ERC-20 implementations against the original Ethereum whitepaper. We found a gas-optimization defect that would have cost the projects deploying it real money — not because anyone lied, but because everyone was too excited to subtract. That is the discipline I bring to a press release about a semiconductor fab in Singapore, which is not, on its face, my beat.
Except it is. It turns out the thing I have been building for a decade has a floor made of silicon, and that floor is being poured right now, eight thousand miles from where I sit.
The contradiction is not a typo. It is a tell. When a promotional document cannot choose between "we made it" and "we will make it," the honest reading is that the milestone is a construction milestone, not a commercial one. So I pulled the seam apart. What came out is a story about geopolitics, valuation narratives, and the invisible mature-node chips that make verifiable AI physically possible — three things the crypto industry is currently getting wrong in exactly the same way.
Why a consensus person is reading a foundry update
Let me explain why someone whose daily work is data availability sampling is reading a wafer fab update at all.
For most of my career, the physical world was somebody else's problem. The chain abstracted it: you do not need to trust the counterparty because you can verify the state transition. That abstraction held up beautifully as long as the assets were digital and the adversaries were other people's servers.
Then in 2024 I started a pilot connecting autonomous AI agents to decentralized identity protocols, and the abstraction cracked along a line I had not anticipated. The problem was not the cryptography. The problem was attestation. To prove that an AI agent is who it claims to be — to prevent a deepfake from inheriting a verified credential — you need a root of trust that a remote attacker cannot rewrite. That root lives in hardware: a secure element, a trusted execution environment, a physically unclonable function. And every one of those parts is fabricated on a mature process node.
Now consider the thing that triggered my curiosity.
Vanguard International Semiconductor — VIS, a Taiwanese specialty foundry — is building its first Singapore wafer fab as a joint venture with NXP, the Dutch automotive and industrial semiconductor company. The entity is called VSMC. The ownership is instructive: VIS holds 60%, NXP holds 40%, and TSMC holds a little under 20% of VIS itself. The project is roughly 7.8 billion dollars for something in the range of 55,000 wafers per month on 300mm substrates, initially at 130nm down to 40nm and extensible toward 28nm. Two fabs are planned; the second depends on demand.
Notice what that cap table is. A lead operator. A captive customer with a strategic stake. A silent minority holder with world-class optionality. I have seen that structure a hundred times — in token distributions, in grant programs, in DAO treasuries. It is the shape you choose when you want alignment without control, and control without exposure.
And the lithography is DUV. Dry and immersion, not extreme ultraviolet. There is no EUV anywhere in the building.
That single fact is the entire geopolitical thesis of the project, and almost nobody reporting on it says so out loud.
The 99% yield is not the news
Let me start with the number the release wants me to remember: 99% yield at trial production.
In mature-node manufacturing — anything at 40nm and above — a first-pass trial yield in the low nineties is already acceptable. After ramp, 97 to 99% is ordinary. TSMC's mature lines hold 98-99% as a matter of routine. So a 99% trial figure is a competent result, not an extraordinary one. It is the fab-world equivalent of a testnet hitting ten thousand transactions per second in a lab: technically true, structurally uninformative, and extremely useful as a marketing asset.
I have watched this pattern in crypto for years. A metric gets promoted from "expected" to "announced" without anyone marking the transition. In DeFi, that is TVL counting the same dollar twice across composable protocols. In semiconductor public relations, it is a yield number quoted without its comparison baseline.
The number that actually matters is not peak yield. It is yield retention — the ability to hold 98% six months into volume production when the customer mix shifts, the tool set drifts out of spec, and the process window narrows. That number does not exist yet, because volume production does not exist yet. Which brings us back to the contradiction.
Chasing the frontier where code meets belief has taught me one thing above all: the interesting claim is never the headline metric. It is the metric the headline was designed to hide.
What the fab is actually for
Here is where I part company with the framing everyone reached for.
The release attributes the expansion to "AI infrastructure demand." That is technically defensible and strategically misleading, in the same way that describing a stablecoin as "a DeFi primitive" is technically defensible and strategically misleading.
The demand base bound to this fab is automotive and industrial, because NXP is the customer and NXP's business is automotive and industrial. Electric vehicles carry three to five times the semiconductor content of an internal combustion vehicle, and most of that content is mature node: power management ICs, gate drivers, battery management, sensor interfaces, body control. That is the foundation. It is boring, it is durable, and it is not going away.
The AI part is real, but it is second-order, and it works through a channel almost no crypto coverage explains. AI training clusters get all the attention because they consume the most advanced logic. But a single AI rack is not only GPUs. It is dozens of power management ICs, voltage regulator modules, clock generators, retimers, interface chips, fan controllers. Those parts are mature node, they are consumed in volume, and the per-server count is several times that of a general-purpose server.
So the correct statement is not "AI demand drives mature-node capacity." It is: the AI buildout multiplies demand for the least glamorous chips in the stack, and the market prices that demand through the most glamorous narrative available. The industry has been doing this since the first GPU-adjacent token, and it works because nobody wants to underwrite a power management IC.
The tokenized supply chain that isn't
I want to address the thing my own industry will inevitably try to do with this story, because I have been in the room when it happens.
Within a year, someone will propose tokenizing the fab's capacity. A tranche of wafer starts, securitized and traded on-chain, with a smart contract releasing payment on verified delivery. There will be a whitepaper. There will be a governance token. Someone will call it real-world asset infrastructure and someone else will call it the future of supply-chain finance.
Most of it will fail, and I want to be precise about why.
Semiconductor capacity is not a commodity that benefits from permissionless transfer. It is a relationship: a qualified process, a design kit, a yield-learning loop that runs for years, a customer who has signed away their second source. The value is not in the token, it is in the qualification, and qualification does not transfer on a blockchain. When I audited contracts back in 2017, the failure mode was always identical — the token modeled the flow but not the trust.
Where on-chain provenance does earn its keep is narrower and less exciting. Verifiable bills of lading for equipment moving between fabs. Attested provenance for conflict-mineral sourcing. Audit trails for export-control compliance that a regulator can verify without trusting the exporter's spreadsheet. Those are real. They are also not a ten-billion-dollar market, and anyone telling you otherwise is selling the token, not the supply chain.
The depreciation cliff nobody prices
Let me do the arithmetic the release omitted entirely. No investment figure, no capacity figure, no revenue, no timeline to profitability.
Working from the disclosed project background rather than the article: roughly 7.8 billion dollars of capital expenditure against roughly 55,000 wafers per month is on the order of 140 million dollars per thousand wafers per month of capacity. That is inside the normal band for a mature-node greenfield fab. It is also, unambiguously, a heavy asset.
Now apply semiconductor depreciation. Process equipment typically depreciates over five to seven years. A greenfield fab's depreciation charge begins before the revenue does, which means the first two to three years of operation run at gross margins in the low teens or worse, recovering toward the 25 to 35% band that specialty foundries achieve at maturity. Depreciation-adjusted breakeven for a project like this lands somewhere around 2028 or 2029.
I know this pattern intimately because I have watched it in DeFi. A protocol launches with a large emissions budget, buys liquidity with it, and reports TVL as traction. The depreciation charge in a fab is the emissions schedule in a liquidity mining program: a real, unavoidable cost of acquiring position that the marketing department universally declines to mention.
A project that buries its capex is a project that knows its capex is the story.
Singapore is the product
Here is the part I find genuinely fascinating, and it has almost nothing to do with semiconductors.
Every advanced-node discussion today runs through export controls. The Netherlands restricts immersion DUV shipments to China. The United States restricts advanced logic, high-bandwidth memory, and the tools that make them. Japan restricts a specific list of equipment and materials. The entire architecture of restriction is built around a node threshold.
VSMC is building below that threshold. That is not an accident; it is the point of the site selection. A mature-node fab in Singapore can buy ASML immersion scanners, Applied Materials deposition tools, Lam etch chambers, and Tokyo Electron coaters with none of the licensing friction a comparable Chinese project now faces. When founders ask me about jurisdictional strategy, this is the structure I describe: you do not fight the restriction, you build in the space the restriction does not reach.
But the deeper move is the location itself. Singapore is not Taiwan and it is not China. For a European customer like NXP, or an American Tier 1 supplier, that matters more than any technical parameter on a datasheet. The value proposition is not "the best mature-node foundry." The value proposition is "a mature-node foundry that will still be shipping in the scenario where the alternatives are not."
That is an insurance product wearing a manufacturing costume. And insurance, as anyone in this industry should understand, is priced on the tail, not the mean.
In the silence of the chain, we hear the future. The same is true of a fab: the most important thing about it is the scenario in which it is the only one still running.
Everyone is building the same thing
This is not a single project. It is a pattern, and the pattern has a name: friendshoring.
The United States is spending CHIPS Act money to pull advanced and mature capacity onshore. Europe wants 20% of global production by 2030. Japan is subsidizing TSMC's Kumamoto build while rebuilding a mature-node base. Singapore is positioning itself as a regional specialty hub with VSMC as the anchor tenant. China, meanwhile, is running the largest mature-node capacity expansion in the history of the industry through its state fund programs.
Look at that list as a DeFi person and you will see something familiar. Every chain in the modular thesis wants to be the place where execution happens, and every one of them is subsidizing liquidity to get there. The competition is not settled on architecture. It is settled on who convinces the most projects to deploy first, and the same is true of foundries — VIS is not winning on process physics, it is winning on who signs the multi-year capacity agreement.
The result is not efficiency. The result is fragmentation with better branding. Duplicated capacity, higher structural costs, lower utilization at the margin, and a long, grinding price war in mature nodes as Chinese supply comes online. The efficiency loss is real and it is permanent. Resilience is not free, and the invoice arrives in the form of gross margin.
The convergence nobody is hedging
Which brings me to the part that genuinely worries me.
My current work sits at the intersection of autonomous agents, verifiable credentials, and privacy-preserving AI. The thesis is straightforward: if an agent is going to transact on someone's behalf, it needs a verifiable identity, and that identity needs a hardware root of trust — a secure element, a TEE, an attestation key burned into silicon that neither the agent nor its operator can forge.
Every one of those components is mature node. Not 3nm. Not 2nm. Secure microcontrollers, hardware security modules, secure elements in phones and wallets, TPMs, the controller silicon in a validator's machine — mature node, mostly 40 to 130nm, fabbed by exactly the kind of facility VSMC is building.
Now stack the two trends. The AI buildout is consuming mature-node capacity at the same moment that the verifiable-AI stack is going to need it in volume. And where is new mature-node capacity being built? Taiwan, Singapore, Japan, Arizona, Dresden, and China. Meanwhile the industry's financial attention is entirely on 2nm.
This is the arbitrage. The scarcest input to trustworthy AI is not compute. It is trusted silicon on an unglamorous node, and almost nobody building agent infrastructure is thinking about supply.
I have watched this movie. Curiosity is the only leverage in DeFi Summer, and it is the only leverage now.
The cycle, and who pays for resilience
One more thing before the part where I disagree with everyone.
Timing. The semiconductor inventory cycle runs roughly three to four years. We came out of a deep destocking in 2023 into a mild restock, and automotive and industrial inventories are still elevated going into 2025 and 2026. Normalization lands somewhere in that window. VSMC's volume production lands in 2027. Depending on whether Chinese capacity arrives before or after demand recovers, that timing is either fortunate or catastrophic, and the operator does not control it. In a capital-intensive industry, timing is not a variable you manage. It is a variable you inherit.
And there is a human ledger to this geopolitical insurance policy that rarely makes it into the analysis. Fabs consume water and power at industrial scale, and the communities hosting them absorb those costs while the benefit is abstracted into the phrase "supply chain security." The engineers who staff these lines will spend a decade tuning a process that will be obsolete before they retire. Art is the glitch that proves we are human; the wafer is the substrate that proves we are industrial. Both are true. Only one of them gets a press release.
Where I disagree with everyone
The counterargument goes like this: mature-node capacity is abundant, fabs are fungible, and the AI framing for a specialty foundry is a rounding error in a costume. Fair. I would push it further than the skeptics do.
The strongest contrarian read is not that the AI framing is exaggerated. It is that the AI framing is a distraction from a more important and less comfortable fact: the crypto industry's own hardware dependency is unhedged and unexamined. We argue endlessly about execution environments and data availability while every one of our trust anchors is manufactured in a supply chain we do not model, do not monitor, and would not recognize if it broke.
And here is the part that stings. The people who benefit most from verifiable AI are the same people least likely to think about where secure elements come from. We celebrate the abstraction and ignore the substrate. I did that for years. The pilot I ran in 2024 only worked because someone else had already fabbed the silicon.
So no, I will not argue that VSMC is an AI play. I will argue that the release's own terminology is broken, that its headline metric is routine, that its financials are absent by design, and that its real value sits in a geopolitical option the industry that needs it most is not pricing.
That is a less exciting story. It is also the true one.
Takeaway
The protocol is cold; the evangelist is warm. But a warm evangelist running on cold silicon is still running on somebody else's fab.
When the AI cycle turns — and it will — the question will not be who has the most GPUs. It will be who secured the boring wafers three years earlier, back when the press release said "full production" and quietly meant "not yet."
The next decade of verifiable intelligence is being poured in Singapore right now, at 40 nanometers, by people who will never write a whitepaper. The real question is whether anyone in this industry is paying attention — or whether we will discover our dependency the way we always do, at the moment it fails.