Nearly every leading-edge AI chip comes out of one company: TSMC, as Epoch AI’s overview notes. For any business buying AI compute at scale, the silicon roadmap runs through a single merchant foundry in Taiwan, leaving one question: own fabrication, or rent access to it?
Terafab, backed by Tesla, SpaceX and xAI and anchored by a reported $16.8 billion first phase, is the live test of the “own” answer. We covered the project basics, and the Terafab overview hub frames the full picture. Here we compare its single-roof model against TSMC’s specialised fabs, and hand you a four-factor framework for weighing control and speed against yield, capex and demand certainty.
How does a vertically integrated semiconductor fab actually work, and why would anyone combine logic, memory, and advanced packaging under a single roof? #
A vertically integrated fab owns the whole stack: design, logic fabrication, memory fabrication, advanced packaging, and testing, under one roof. That is rare because logic and memory run on different physics, and putting them side by side reintroduces purity risks a specialised fab avoids.
What does “under one roof” actually change?
In a distributed chain, one company designs the chip, contracts fabrication to a foundry, buys memory elsewhere, and packages and tests it. Manufacturing Today describes the single roof as pulling together steps usually spread across specialist companies.
The single roof changes the coordination: teams test designs against manufacturing constraints sooner, with fewer hand-offs. Terafab still leans on Intel’s 14A process (Intel’s latest leading-edge node) and ASML’s High-NA EUV tools, as the project reporting shows. Integration shifts control toward the chip consumer, with Intel and ASML still in the loop.
Why HBM4 and 2.5D/3D packaging pull toward consolidation
Two forces pull logic and memory together. The first is advanced packaging: 2.5D and 3D arrangements place logic and high-bandwidth memory together to shorten data paths, because moving data is a bigger bottleneck than raw compute for AI accelerators. The packages are mechanically and electrically fiddly. Amkor’s Mike Kelly puts it simply: “Most of these packages are quite complicated mechanically, and certainly electrically.”
The second is HBM4 and HBM4E. Co-designing logic and memory shortens an integration path a distributed chain stretches across vendors. Add captive demand from Tesla, SpaceX and xAI, and you get a guaranteed baseline volume most merchant foundries cannot match, plus shorter design-to-production loops.
Where cross-contamination and process physics push back
The objections are physical. DRAM stores a bit in a one-transistor, one-capacitor cell that discharges and must be refreshed, a constraint logic transistors do not share. Even the naming misleads: a DRAM “10nm” label refers to memory-cell pitch, not a logic node.
Run those processes together and you reintroduce cross-contamination risk, where dopants or residue from one line bleed into another. A specialised fab isolates each process window by design; a single roof has to engineer those separations back in.
Then there is the money. A single EUV tool costs $150 million, and running multiple process lines multiplies the bill. Intel is the closest precedent, and its pivot toward foundry services is the cautionary tale: its foundry business generated $4.5 billion in a quarter while unprofitable. The deeper difficulty is covered in why leading-edge process is so hard.
How does Terafab’s single-roof model compare to TSMC’s specialised-fab approach? #
TSMC wins on yield, purity and scale; Terafab bets that captive demand and packaging adjacency outweigh those advantages. The better route depends on demand certainty and capital tolerance.
How TSMC’s specialised fabs spread capex and protect yield
TSMC runs a merchant foundry: purpose-built fabs, each tuned to a narrow process window to maximise yield and purity. TSMC’s Arizona investment pipeline has reached $265 billion across 12 facilities. Fab 21 is the concrete US example, where 4nm is running and 2nm is the next revenue driver.
TSMC took 38% of a record $320 billion foundry market in 2025 and grew four times faster than its rivals. The yield gap is clear: Samsung’s 2nm yields sit around 50 to 60 percent, against 80 percent or more at TSMC.
Terafab’s captive-demand counter-argument
Terafab’s answer is captive demand: Tesla, SpaceX and xAI would consume output regardless of the merchant market, giving the fab a floor of guaranteed internal demand. Add packaging adjacency, and the single roof looks like a wager on known volumes, with no need to win outside customers. Intel is a partner here: it is working with SpaceX, xAI and Tesla to support Terafab.
The verdict: which route suits which buyer
The verdict is conditional. If your demand is uncertain or distributed, foundry partnership is the lower-risk route. If captive demand is large and stable enough to keep one roof busy across a full depreciation cycle, integration becomes a considered bet on control and speed — a choice that sits inside the future of chip manufacturing.
How do you assess the build-versus-buy tradeoff for AI compute infrastructure? #
Demand certainty and capital tolerance are two of the four factors behind any build-versus-buy call. Build-versus-buy collapses into capital intensity, demand certainty, time-to-market, and opportunity cost. Owning fabrication only pays past a threshold of captive demand.
How do you model the capital intensity of a new hardware initiative before committing?
The four numbers to weigh before committing are depreciation, utilisation, yield ramp, and the cost of being wrong. Depreciation outruns utilisation as a new fab ramps, compressing gross margin when output is low. A merchant foundry spreads that ramp risk across many customers; a captive fab bears that risk on its own.
Computation per dollar is the metric that matters. Epoch AI’s overview notes that leading-edge AI chips are fabricated by essentially one firm. Nvidia’s flagship chip rose from $5,700 to $34,000 from 2016 to 2022, yet the H100 still delivers seventeen times more computation per dollar.
Rent, buy, design, or build: the four paths
You have four options: rent cloud compute, buy merchant GPUs, design a custom accelerator and outsource fabrication to TSMC, or build a captive fab. Most “builders” stop at option three, and your team will usually want to as well. Designing a chip can cost half a billion dollars, and tape-out on TSMC’s N3 can exceed $500 million. That is cheaper than a fab, which is why Anthropic is exploring its own chip but has yet to commit.
When captive demand justifies owning the stack
The build case flips when captive demand appears. SpaceX’s space-grade chip need is the clean example: chips for applications on Earth and in space cannot be fully met by merchant allocation, so internal capacity earns its keep. Terafab is the fullest build case because Tesla, SpaceX and xAI provide the volume to justify the $16.8 billion first phase — a wager that belongs to the broader chip-manufacturing story.
So which model wins? It depends on your business’s demand certainty and capital tolerance. For most buyers, renting TSMC’s specialised capacity remains the lower-risk route; integration only pays when captive demand is large and stable enough to amortise the capex. Terafab’s single-roof bet is a deliberate wager on Tesla, SpaceX and xAI demand. It is not a universal blueprint for AI chip supply. Foundry partnership and vertical integration are different answers to the same four-factor question, and you now hold the framework for choosing. Where the single-roof strategy fits the bigger picture is covered in the Terafab overview.
Frequently Asked Questions #
Why is TSMC effectively the only merchant foundry for leading-edge AI chips?
Because leading-edge fabrication has become so capital- and knowledge-intensive that almost no one else has kept investing at the scale required to stay ahead. Samsung and Intel Foundry remain credible rivals, but TSMC’s mature yields, deep process library, and large customer base give it a near-default position for buyers who need leading-edge AI silicon now.
Is it true you can just license a leading-edge chip design instead of building a fab?
No. You can license processor architectures such as Arm, but a leading-edge manufacturing process is not a licensable asset. The recipes, tooling, and yield engineering live inside the fab itself, so even a fabless design house still depends on a foundry like TSMC to physically produce its chip. The design is portable, but the process is not.
What does cross-contamination actually mean inside a fab?
It means unwanted material or process residue from one manufacturing step bleeding into another, for example a dopant used in memory fabrication reaching a logic line. A specialised fab isolates each process window by design. A single-roof fab that runs logic and memory side by side must engineer those separations back in, which adds cost and purity risk.
How much does a leading-edge fab actually cost to build?
A single leading-edge fab now commonly runs past US$10 billion, and the most advanced sites stretch well beyond that. Terafab’s reported first phase is around US$16.8 billion. That capital has to be amortised through years of high utilisation before yield and volume turn the plant cash-generative, which is why demand certainty dominates the build decision.
What happens if Terafab’s yield ramp takes longer than expected?
The capital bill keeps running while usable output stays low, so the economics deteriorate quickly. A merchant foundry spreads that ramp risk across many customers, but Terafab’s captive demand from Tesla, SpaceX, and xAI is the buffer: those internal buyers can absorb early wafers and keep the line learning even before yield fully matures.
Is building your own fab ever cheaper than buying wafers from TSMC in the long run?
Sometimes, but only under specific conditions. Ownership can lower the per-chip cost once a fab runs near capacity at mature yield, because you remove the foundry’s margin. That advantage only appears if captive demand stays large and stable enough to keep utilisation high across a full depreciation cycle, which is exactly the wager Terafab is making.
Do you have to own a fab to get real supply-chain control?
No. Control is a spectrum, not a binary. A fabless buyer can secure allocation, co-design packaging, and lock in multi-year capacity agreements with a foundry without owning a single tool. Owning the fab moves you further along that spectrum, but it is the most expensive way to buy influence and only makes sense at extreme scale.
Was Intel’s integrated model the same thing Terafab is attempting?
Not quite. Intel was the great integrated device manufacturer, but for most of its history it did not consolidate logic, memory, and advanced packaging under one roof at the scale Terafab proposes. It remains the closest historical precedent, which is why Intel’s later pivot toward foundry services is read as a cautionary tale about sustaining full integration.
What is advanced packaging, and why does it matter for AI chips?
Advanced packaging places chips such as logic and high-bandwidth memory physically close together in 2.5D or 3D arrangements, shortening the data paths between them. For AI accelerators this matters because moving data is often a bigger bottleneck than raw compute, so tighter packaging adjacency speeds the logic-to-memory integration that Terafab’s single roof is built around.
What does captive demand actually mean, and why does it change the maths?
It means guaranteed internal customers. In Terafab’s case, Tesla, SpaceX, and xAI would consume output regardless of what the merchant market does. That certainty gives the fab a demand floor, which keeps utilisation high and turns a risky greenfield plant into a wager on known volumes rather than a bet on winning outside customers.
Should a startup build its own AI chip, or rent capacity?
Almost always rent, and almost never the fab. Startups are best served renting cloud compute or buying merchant GPUs, and even those that design custom accelerators still outsource fabrication to a foundry. Owning a fab only starts to make sense once captive demand is large and stable enough to amortise billions in capital.
How long does it take to bring a leading-edge fab online?
Typically three to five years or more from groundbreaking to first volume production, with yield ramp adding further quarters after that. The lag is central to the time-to-market factor in build-versus-buy: a buyer who needs chips now is usually better renting foundry capacity than waiting for a greenfield plant to mature.