Nvidia Is Reportedly Funding CoWoS Capacity at SPIL. The Package Is the Bottleneck Now.

In short
Key takeaways
- Reuters reported on 6 November 2026 that Nvidia is financing CoWoS expansion at SPIL, a unit of ASE Technology, from one anonymous industry executive.
- Nvidia did not respond to a request for comment and ASE said it does not comment on market rumours involving its customers, which is neither a denial nor confirmation.
- Yole Group expects advanced packaging revenue to overtake traditional packaging revenue for the first time in 2026, with capital spending approaching $28 billion.
- Pat Gelsinger estimates advanced packaging could reach half the value of a chip product, with some processors integrating 10 or more chips and sometimes 20.
- Powertech's $400 million panel-level joint venture with Broadcom in Singapore could begin production in 2028, two years after the constraint is most acute.
- Microsoft's Azure hardware chief said panel-level packaging would cut material waste but that Microsoft has not decided whether or when to adopt it.
On 6 November 2026, Reuters reported that Nvidia is providing billions of dollars in financing to expand CoWoS advanced packaging capacity at Siliconware Precision Industries, a unit of ASE Technology. The claim was attributed to "an industry executive familiar with the matter, who sought anonymity for lack of authorisation to talk to media". Nvidia did not respond to a request for comment. ASE said it does not comment on market rumours involving its customers.
That belongs at the top of any account of the story, because it is the entire evidentiary basis for the most quoted sentence in it. There is no disclosed amount beyond "billions", no terms, no confirmation from either company, and no named source. The report says the scale had not previously been reported. It does not say how the number was established.
The same Reuters feature, which the report describes as drawing on interviews with more than a dozen industry executives and analysts, carries a second layer that is better sourced and more useful. Advanced packaging has become the binding constraint on accelerator supply. The bottleneck moved out of the lithography bay and into the room where dies and memory get joined.
What CoWoS actually is
CoWoS stands for chip-on-wafer-on-substrate, and it is TSMC's process for assembling an accelerator out of parts that were manufactured separately.
A large AI chip is not one piece of silicon. The logic die is one, and the high-bandwidth memory is a stack of DRAM dies that comes from a different supplier entirely. CoWoS places the logic die and its memory stacks side by side on a silicon interposer, a piece of silicon whose only job is to carry very dense wiring between them. The interposer is then mounted on an organic substrate, which distributes power and connects the finished module to the board.
The reason that matters is bandwidth. Training and inference both spend most of their time moving data between the processor and the memory beside it. The physics of that link is set at packaging: the width of the bus, the distance a signal travels, and the energy cost per bit. Route those signals off the package and across a circuit board and the bandwidth falls while the power bill climbs. Keeping memory close is what makes it fast, and close is a packaging decision, not a circuit design one.
So the package sets what an accelerator can be. How many memory stacks fit around the logic die, whether the interposer can be built at that size, and whether the assembled module yields well enough to ship all bound the product before a wafer is started. None of those limits yields to another lithography tool, because none of them is lithography.
The constraint is a building, not a design
A fab constraint and a packaging constraint behave differently. When demand outruns wafer starts, you can in principle add a fab, though it takes years and a great deal of capital. When demand outruns CoWoS, you need cleanroom space, bonders, interposer supply, substrate supply, and a yield curve that improves with volume. The equipment is specialised, the process is difficult, and the ramp is measured in quarters rather than weeks.
TSMC dominates the most sophisticated advanced packaging, including CoWoS, and its output is used by Nvidia, AMD and Broadcom, and by the cloud companies designing their own accelerator silicon. Reuters reports that TSMC's capacity limits have spurred some chip designers to seek additional supply. That sentence is the story in miniature: the customers of the most important packaging process are looking for a second source, and the second source is not a new process. It is more of the same one.
This is also why packaging capacity gates shipped units in a way that die area does not. A wafer can be probed and binned. A package either assembles and passes test or it is scrapped, and the scrapped module has already consumed a leading-edge logic die and several memory stacks that were themselves in short supply. Yield at the assembly step therefore multiplies against everything upstream. A few points of package yield is not a rounding error on a product where the components inside it are the scarcest things in the supply chain.
The numbers that are named
The figures Reuters carries for the shift are forecasts, and they should be read as such. Yole Group expects advanced packaging revenue to overtake traditional packaging revenue for the first time in 2026, and expects capital spending on the technology to approach $28 billion. Those are two different quantities. The first is revenue crossing over. The second is spending on equipment and capacity. Neither measures CoWoS capacity in wafers per month, which is the number that would actually say how tight the constraint is.
Pat Gelsinger, Intel's former chief executive and now a partner at Playground Global, gave the shift a shape in Taipei. "Packaging was always that thing that you just didn't worry about," he told reporters. He estimates advanced packaging could eventually account for as much as half the value of a chip product, as some graphics and central processing units integrate 10 or more chips, sometimes as many as 20.
That is a directional estimate from a former competitor's chief executive, not an analyst forecast, and it should be carried with that weight. It is still a useful description of the direction: more of the value of the product sits in the assembly, and less in the transistor.
Why a customer funds a supplier's competitor
If the Reuters financing claim holds, the interesting part is not the money. It is the direction of the cheque.
Nvidia is the largest buyer of the process that TSMC controls. SPIL is an outsourced assembly and test house, and a unit of ASE Technology. Financing capacity there instead of waiting for allocation at TSMC is what a customer does when the queue, not the price, is the problem. It buys a claim on output, and it buys it at a supplier the incumbent does not own.
The report does not say whether the arrangement includes exclusivity, preferential allocation, or a minimum volume. That gap matters, because preferential capacity is the entire point of paying for capacity in advance. Without it, the financing is a subsidy. With it, it is a moat.
The pattern is not unique to Nvidia. AMD announced more than $10 billion of investment across Taiwan's semiconductor ecosystem in May 2026, including partnerships with ASE, SPIL and Powertech on advanced packaging. Powertech is putting $400 million into an advanced packaging joint venture with Broadcom in Singapore, with production potentially beginning in 2028. When several designers pre-fund capacity across several suppliers at once, they are not shopping. They are hedging against the possibility that none of the suppliers can give them enough.
Lisa Su, AMD's chief executive, said it plainly in Taipei in October 2026. CoWoS and advanced packaging are "super important for our high-performance computing going forward", she said, and TSMC "has been an incredible partner in advanced packaging, and we're going to continue to do more together. And yes, we do need more capacity overall."
That quote is worth holding next to the framing of the story. It is not a customer describing a supplier in decline. It is a customer describing a supplier that cannot build fast enough.
The process race underneath
Three paths are in play, and the timing is what separates them.
CoWoS is the incumbent, and it remains TSMC's main AI packaging technology. TSMC's own answer to the panel question is chip-on-panel-on-substrate, CoPoS. The company said in April 2026 that it had built a pilot line, with production expected within a couple of years.
Panel-level packaging is the challenger, and its advantage is arithmetic. Instead of round wafers, it uses large rectangular panels, which fit more chips into a production run and waste less material at the edges. The advantage grows as AI packages get larger, which is the direction the products are moving. Powertech is working with AMD and Broadcom on it.
The strongest evidence about the challenger came from an unexpected place: a buyer's non-decision. Rani Borkar, Microsoft Azure's hardware chief, told Reuters that panel-level packaging would reduce material waste when packaging larger AI chips, and that Microsoft has not decided whether or when to adopt it. Microsoft is among the largest buyers of the silicon in question. If the process does not get adopted by the buyer with the most volume, the case for it is a case about cost, not a settled outcome.
Seiji Nakama, president of Tokyo Electron Taiwan, gave the equipment view: the advanced packaging equipment market is growing much faster than the chip-making gear market, but it remains smaller. Sree Kesapragada, a vice president at Applied Materials, framed panel-level as near-term rather than a decade out. "It's not like, 10 years, 15 years off," he said.
Near-term and 2028 are not the same thing. Powertech's Singapore joint venture, the only panel-level capacity with a public date attached, could begin production in 2028. The constraint is binding now. Panel-level packaging is an answer to the 2028 and 2029 problem. Something has to carry 2026 and 2027.
That is what the SPIL financing, if it happened, is for. It is not a bet on the challenger process. It is money into the process that works today, at a supplier that is not TSMC.
What would settle the claim
ASE Technology and SPIL are listed companies. A customer financing arrangement of the reported scale would normally be the kind of thing that surfaces in a material information disclosure on the Taiwan Stock Exchange, or at minimum in quarterly investor materials. Whether such a disclosure exists was not established here, and it is the single check that would do the most work.
If ASE has disclosed an arrangement, the story upgrades from one anonymous executive to a primary record. If it has not, that absence is evidence, though not proof, about the reliability of the source. A denial from Nvidia, ASE or SPIL would settle it in the other direction. A second report with named sources would do the same. None of those exists as of publication.
One adjacent claim is sometimes merged into this one and should not be. An SDxCentral report dated 21 October 2026 said Nvidia had confirmed plans with SPIL to expand optical-engine capacity 15 times by mid-2027. That is a different product and a different process, and it is being carried here at one remove rather than read from the original. It is not evidence about CoWoS.
Where this could be wrong
The financing claim is the weakest element and should not be repeated as fact. One anonymous industry executive, quoted at second hand, no company confirmation, no amount beyond a rounded "billions", no terms, no exclusivity, no contract structure. ASE's response was that it does not comment on market rumours involving its customers. That is a policy statement. It is not a denial, and it is not confirmation.
The motive is an inference, not a reported fact. The report documents capacity limits, and it documents separately that customers are funding packaging capacity at multiple suppliers. Reading the Nvidia item as an attempt to buy priority at a second source is consistent with that pattern. It is not something any of the parties said.
The quantification is thin. The report is narrative about capacity without figures. It does not give CoWoS capacity in wafers per month, it does not say how much of the gap SPIL's expansion would close, and it does not say when the added capacity arrives. Without those numbers, the claim that the constraint eases in any particular quarter cannot be tested.
The Yole figures are forecasts, and they measure different things: revenue crossover and capital spending. Gelsinger's "half the value" is an estimate from someone with a stake in the answer being large. Borkar's material-waste point is directional. Cost per packaged die, yield at panel scale against wafer scale, and the equipment cost difference are the variables that decide the process race, and none of them is in the reporting quoted here.
Panel-level packaging may arrive sooner than 2028, or a hyperscaler may commit to it and pull the timeline forward. Either would weaken the timing argument that near-term relief has to come from more CoWoS. Equally, TSMC's own expansion could ease the constraint without any customer financing, in which case the SPIL story is a non-event dressed as a structural shift.
Finally, this is a snapshot of one report on one day. A supply chain story sourced largely to interviews ages quickly, and the version of it that survives will be the one that shows up in a filing.
On this page
Sources
- AI boom turns chip packaging into next battleground for semiconductorsWen-Yee Lee, Reuters, 2026
- AMD says it will invest over $10 billion across Taiwan's AI ecosystemReuters, 2026
- AMD plans to substantially increase supply in 2027, CEO saysReuters, 2026
Frequently asked questions
What is CoWoS?
Chip-on-wafer-on-substrate is TSMC's process for assembling an AI accelerator from separately manufactured parts. The logic die and its high-bandwidth memory stacks are placed on a silicon interposer that carries the dense wiring between them, and the interposer is mounted on an organic substrate that handles power and connects the module to the board.
Did Nvidia confirm the SPIL financing?
No. Reuters attributed the claim to one industry executive who sought anonymity for lack of authorisation to talk to media. Nvidia did not respond to a request for comment. ASE said it does not comment on market rumours involving its customers, which is a policy statement rather than a denial.
Why is packaging a bottleneck rather than lithography?
AI chips are limited by how fast they can move data between the processor and the memory beside it, and that link is built at packaging. Adding lithography capacity does not add interposer, bonder or substrate capacity, and a packaging line ramps over quarters. TSMC's capacity limits have pushed chip designers to look for additional supply.
Does panel-level packaging solve the shortage?
Not in the near term. The only panel-level capacity with a public date is Powertech's $400 million joint venture with Broadcom in Singapore, with production potentially starting in 2028. The constraint is binding now, so the 2026 and 2027 relief has to come from more CoWoS.