Nvidia doesn't own a single fab. So the most consequential manufacturing decisions it makes aren't about how to build a chip - they're about whose line to stand in, and how much to pay to keep a place in it.
Pairs with the Vertical-Integration Assessment — a ready-to-use strategy tool. Included in the Vertical-Integration Bet Casebook →
Nvidia does not own a fab. It never has. The company that designs the most valuable silicon on earth cannot etch a single one of its own transistors - it rents that ability, wafer by wafer, from someone else's factory. So when people say Nvidia "switched" chip makers in 2020 and "switched back" in 2022, they are describing the only vertical move a fabless company can really make: not buying a factory, but choosing whose line to stand in, and paying to hold its place.
The official story is a clean pivot and a clean return. Nvidia left TSMC for Samsung, got burned, and crawled back. Almost every part of that is either wrong or unproven. Nvidia never left. The reason it moved isn't the reason everyone repeats. And the eye-popping numbers attached to the return rest on a single anonymous report.
Two chip generations, two different factories: the flagship gaming line went to Samsung's process, then came home to TSMC's
Start with what is actually documented. When Nvidia announced the RTX 3090, 3080 and 3070 in September 2020, it confirmed they run on Samsung's 8nm process - the 8N node - not TSMC.1 Notably, its datacenter A100 from the same architecture generation stayed on TSMC's 7nm. That split is the tell: Nvidia put its high-volume consumer gaming line on Samsung while keeping its scarce, high-margin datacenter part at TSMC. Two years later, the RTX 40 series - the Ada Lovelace generation - launched on a custom TSMC 4N process, moving gaming-GPU manufacturing back to TSMC.2 On paper: out, then back.
| RTX 30 series (Ampere) | RTX 40 series (Ada Lovelace) | |
|---|---|---|
| Foundry | Samsung | TSMC |
| Process node | Samsung 8N (8nm) | Custom TSMC 4N |
| Announced | September 2020 | September 2022 |
| Positioning | Cheaper, available capacity | Denser, more efficient node |
But "out, then back" is the wrong frame, and Nvidia's own filings say so. As of its fiscal 2025 10-K, Nvidia still lists both TSMC and Samsung as foundries producing its semiconductor wafers, and still buys memory from Samsung alongside SK Hynix and Micron.3 Samsung was never dropped. What looks like a switch-and-return is really a shifting mix - Nvidia dialing the ratio between two suppliers, not abandoning one and crawling back.
Why Samsung got the gaming chips in the first place: when the constraint is volume, the cheaper open line beats the crowded prestige one
Here is the mechanism most retellings skip. A fabless company's binding constraint is rarely design - it's capacity. TSMC's leading edge is the most fought-over real estate in technology, and every major chip designer wants the same wafers at the same time. Nvidia's gaming GPUs sell in enormous unit volume at consumer prices, which means they need a lot of wafers and can't absorb premium wafer costs the way a datacenter part can. Putting the RTX 30 line on Samsung's 8N solved both problems at once: it was cheaper, and it was available. The datacenter A100 - fewer units, far higher margin, performance-critical - could justify TSMC's scarce 7nm. So the 2020 move wasn't a strategic bet on Samsung. It was a routing decision: send the price-sensitive, volume-heavy product to the open line, and keep the scarce line for the part that pays for it.
Nvidia can't build a factory into its stack, so its "vertical" lever is which foundry and which node each product gets, and how firmly it reserves the slot. That turns supplier choice into a portfolio problem: route high-volume, cost-sensitive products to whatever capacity is open and affordable, and reserve the scarce leading edge for the products whose margins can carry its price. The mistake is reading each individual product's foundry as a company-wide loyalty. It's an allocation, and allocations move every generation.
Why the gaming chips came home: the documented reason is density and efficiency, not the yield rumor everyone repeats
The story that hardened into fact is that Samsung's 8nm yields were bad, and Nvidia fled. That version began as anonymous market talk and analyst speculation about tight RTX 30 supply - never an admission from Nvidia, and never confirmed by Samsung.6 The better-documented reason is more mundane and more interesting: by 2022, TSMC's 4N process offered major improvements in transistor density and efficiency over Samsung's 8N, and Samsung had no competitive alternative node Nvidia could use without a major architecture redesign.7 In other words, the two factories were no longer offering the same kind of trade. In 2020, Samsung's open, cheaper line was worth the density gap. By 2022, with the leading edge pulling further ahead, the density and efficiency of TSMC's 4N was worth paying up for - so the routing decision flipped.
That is the whole argument in one line: nothing about Samsung necessarily got worse. The value of what TSMC offered got better, and the calculus that once favored the cheap open line now favored the dense scarce one.
The pre-booking numbers everyone quotes came from one anonymous report: the multi-billion figures are unconfirmed, and the filings tell a far smaller story
The popular version of the return has Nvidia writing enormous checks to lock up TSMC's leading-edge wafers - $1.64 billion in one quarter, $1.79 billion in another, and a $6.9 billion long-term deal to secure Ada Lovelace capacity. Those numbers are attributed to a single report from a Chinese-language outlet, relayed downstream, and confirmed by neither Nvidia nor TSMC.5 They may be directionally true. They are not documented. And the one place Nvidia does disclose its own advance payments points to a much smaller story: its total prepaid expenses rose from $142 million to $195 million across a single quarter of 2021 - company-wide, not broken out by foundry, so we don't even know whether it went to TSMC or Samsung.4
The gap between those columns is the point. A confirmed, disclosed rise of about $53 million in company-wide prepayments sits next to an unconfirmed $6.9 billion foundry deal, and the internet treats the second number as the real one. Even the Samsung side is soft: the December 2020 8nm contract was pegged at roughly $91 million by unnamed industry sources - a rough estimate, not a disclosure.6 Almost every dollar figure in this story is an estimate wearing the costume of a fact.
Wasn't this a shrewd long game after all?: reading a reactive trade-off as farsighted strategy is the flattering version, and the unprovable one
The fair objection is that this looks awfully strategic in hindsight. Nvidia used cheap Samsung capacity to flood the market during a GPU boom, then bought its way into TSMC's best node right as AI demand made guaranteed leading-edge supply priceless. Doesn't that pattern deserve to be called deliberate vertical mastery? Maybe - but the evidence pushes back. The clean narrative requires a switch that never happened, since Samsung is still a listed foundry in the FY2025 10-K.3 It requires a yield failure that was only ever rumor. And it requires pre-booking figures that Nvidia has never confirmed and its filings don't support. Strip those away and what remains is calmer and more useful: a fabless company adjusting a supplier mix as the price, availability, and quality of two foundries shifted underneath it. That's not a lesser story. It's the real one - and it explains the next move better than any tale of grand design, because the levers, not the loyalties, are what actually recur.
Nvidia's most important factory decisions are made in a building it doesn't own, on a line it has to reserve, in a queue it shares with everyone else who wants the same wafers. The company didn't pivot to Samsung and pivot back. It kept both doors open and walked through whichever one offered the better trade that year - cheap and available in 2020, dense and worth paying for in 2022. The genius, if there is one, isn't a clever check or a burned bridge. It's refusing to marry a foundry at all, and staying free to re-price the relationship every single generation.
When the real strategy lives one layer down the supply chain
Vertical-Integration Assessment
A make-vs-buy assessment for a single stage of the value chain: rate the forces that argue for owning it and the forces that argue for renting it, then read the verdict off the gap. Blank to run on a stage you're deciding now; filled as the worked example showing why the story's company pulled a stage in-house — or pushed it out.
Included, filled and blank, in the Vertical-Integration Bet Casebook. See the set → · Preview the blank →
Sources
Where this comes from — the filings, records, and reporting behind it.
- 1Nvidia officially confirmed that its Ampere-based GeForce RTX 3090, RTX 3080, and RTX 3070 (announced September 1, 2020) are built on Samsung's 8nm (8N) process node, rather than TSMC's 7nm process used for the datacenter A100.
- 2Nvidia's GeForce RTX 40 series (Ada Lovelace architecture) is built on a custom TSMC 4N process, moving gaming-GPU manufacturing back to TSMC after the Samsung-built RTX 30 series.NVIDIA, Introducing GeForce RTX 40 Series GPUs ↗ · 2022-09-20
- 3As of Nvidia's fiscal 2025 10-K, Nvidia still utilizes both TSMC and Samsung Electronics as foundries to produce its semiconductor wafers, and separately purchases memory from Samsung, SK Hynix, and Micron -- i.e., Samsung was never fully dropped as a supplier even after the RTX 40 series returned to TSMC.
- 4Per Nvidia's own quarterly SEC filings, Nvidia's prepaid expenses (a proxy for advance payments to secure manufacturing capacity, though not broken out by foundry) rose from $142 million as of January 31, 2021 to $195 million as of August 1, 2021; Nvidia does not disclose whether these specific prepayments went to TSMC or Samsung.
- 5A report attributed to Chinese outlet MyDrivers claimed Nvidia paid TSMC $1.64 billion in Q3 2021 and $1.79 billion in Q1 2022 for 5nm wafers, plus a separate long-term deal reportedly worth $6.9 billion to secure Ada Lovelace capacity -- figures that are unconfirmed by Nvidia or TSMC.
- 6Samsung's December 2020 contract to manufacture the remainder of Nvidia's Ampere GeForce RTX 30-series GPUs on its 8nm process was estimated by unnamed industry sources at roughly $91 million, amid market speculation -- never confirmed by either company -- about poor 8nm yield rates driving tight RTX 30 supply.
- 7TSMC's 4N process used for Ada Lovelace/Hopper offers major improvements in transistor density and efficiency over the Samsung 8N ('8nm Nvidia') process used for Ampere, and Samsung did not have a competitive alternative node without a major architecture redesign -- a documented technical rationale for the 2022 return to TSMC distinct from unconfirmed 'yield problem' claims.
More like this — beyond Nvidia
New Strategically analyses as they publish: the defining moves in business, checked against the record. No noise, and one click to leave.