TSMC says packaging capacity is limiting customer growth. The $874 million Commerce proposes funds the bottleneck after this one.
AI accelerator supply depends on more than the number of leading-edge wafers a foundry can print. Logic and high-bandwidth memory still have to be assembled into a working package. TSMC CEO C.C. Wei put the constraint plainly on the company's July 16 earnings call: "Our packaging capacity is so tight that now it's limiting my customers' growth." For leading accelerators that depend on CoWoS-class integration, advanced packaging is one of the clearest binding constraints in the AI supercomputing supply chain today.
That does not make high-bandwidth memory secondary. CoWoS is where the accelerator die and HBM stacks converge, so a shortage or qualification delay on either side stops the finished product. SCN's earlier analysis found that the HBM allocation question had already moved into long-term agreements and system design. Packaging is where the queue forms, with HBM coupled to the same output.
Three developments in late July mark the queue and its possible exits. TSMC said the constraint is limiting customer growth. Intel promoted EMIB-T as a US-based packaging path for larger AI devices. On the same day Intel spoke, the US Department of Commerce announced up to $874 million in letters of intent across seven companies, with a minority, non-controlling federal equity stake attached to each proposed award. The three moves do not run on the same clock. Intel has an engineering option but no disclosed volume, the federal portfolio proposes to fund inputs rather than lines that clear the current queue, and reported industry estimates say the queue is already scheduled to shorten as capacity comes on.
Wei's next answer on the earnings call matters almost as much as the first. He welcomed more packaging suppliers because their capacity would support growth in TSMC's front-end wafer business. That is unusually direct evidence of where output is getting stuck. A finished leading-edge wafer cannot become a saleable accelerator until it is packaged and tested.
Package size makes the constraint hard to read from wafer-per-month figures. TSMC's CoWoS-L line is in mass production at a package size the company describes as 5.5 times reticle area, according to TrendForce's account of its roadmap disclosures. A reticle is the largest area a lithography tool exposes in one shot, roughly 26 by 33 millimeters, and neither CoWoS-L nor Intel's bridge packaging prints a die larger than that. Both compose several reticle-limited dies inside one much larger package. Package area, edge exclusion, test structures and product mix all shape how many finished units a wafer yields, and for the largest CoWoS-class parts that count falls into the single digits per 300-millimeter wafer. As each package grows, a rising wafer-capacity number does not translate proportionally into more accelerators. The binding constraint is often the part missing from the roadmap slide, as the skilled-trades queue in AI data center construction has already shown at a different layer of the stack.
TSMC says CoWoS-L uses a redistribution-layer interposer with embedded local silicon interconnect bridges where compute dies and HBM need dense links. The large interposer fans signals and power across the package while the local bridges carry the high-density die-to-die connections. TSMC's disclosed roadmap, as relayed by TrendForce, moves from 5.5 times reticle area now to 9.5 times in 2027 and 14 times in 2028. Those multiples are TSMC's own roadmap figures, not independent measurements.
Intel's EMIB starts from a different structure. Its technical brief describes small silicon bridges placed in the organic package substrate only where adjacent dies need high-density lateral connections, avoiding a full interposer across the package. EMIB-T adds through-silicon vias to the bridge itself, so power travels from the substrate up through the bridge rather than routing around it, which Intel says reduces the voltage droop of earlier designs and supports higher HBM and UCIe signaling rates.
What Intel claims for that architecture is a scaling roadmap, not a shipped capacity. Intel describes EMIB-T as part of a US packaging platform able to scale packages to more than eight times a current industry-standard reticle area during 2026 and more than twelve times by 2028. The finer envelope figures Intel presented at this year's Electronic Components and Technology Conference — package dimensions, bridge counts and signaling rates — differ from one trade account to the next, so they read as a vendor roadmap relayed with some spread rather than settled specifications. Intel and TSMC both frame their scaling in reticle-area terms, but the figures do not describe directly comparable manufacturing capacity, so the useful comparison is architectural rather than numerical.
On commercial readiness, Intel discloses more than the volume skeptics assume and less than a direct comparison would need. On its Q2 2026 earnings call the company said it has a growing EMIB-T backlog and is preparing for customer production ramps in 2027. It has not disclosed monthly EMIB-T capacity or named an external customer, so its ability to absorb the current overflow remains unquantified even as the commercial pipeline is no longer hypothetical.
The sharpest recent sign of how tight packaging has become comes from reporting on the incumbent itself. The Information reported on July 30, citing two people with direct knowledge, that TSMC is developing an embedded-bridge path of its own, described internally as "EMIB-like" or "quasi-EMIB," with substrate maker Kinsus; the account was relayed across trade coverage over the following day. The reported design embeds silicon bridges in the organic substrate and leans less on the redistribution-layer interposer that CoWoS-L depends on; the report frames it as a defensive move to keep customers said to be weighing Intel's approach inside TSMC's supply chain. TSMC and Kinsus have not acknowledged the work, and no official name, structure or production timeline has been disclosed, so it stands as reported rather than confirmed. It is also separate from the glass-substrate research Wei referenced on the July 16 call. Treated as a market signal rather than a proof point, it is one more indication that packaging capacity is the pressure point.
Packaging may be the tightest link today without being a durable wall. TrendForce reported industry estimates put the CoWoS supply-demand gap near 20 percent in June and around 10 percent by the end of 2026, with further improvement expected in 2027, on TSMC capacity of 120,000 to 140,000 wafers per month and another 50,000 to 60,000 at outsourced providers. These are reported estimates and definitions vary, but they describe a remedy already under construction. The 2027 numbers scatter more widely, from roughly 112,000 wafers per month in one bank's forecast to figures above 200,000 in market chatter, so the size of the exit is less settled than its direction.
Amkor's Arizona facility supplies the later part of that remedy. The company says its $7 billion Peoria campus will complete construction in mid-2027 and begin production in early 2028, with TSMC collaboration covering InFO and CoWoS. TSMC is separately developing square-panel CoPoS packaging, which could use more usable area than a round wafer, although reported timelines place volume production in 2028 or later.
Packaging is also starting to look less like a back-end step than an architectural layer. At this year's ECTC, CEA-Leti's Pascal Vivet laid out a 3D-stacking roadmap that pairs die-to-wafer and wafer-to-wafer hybrid bonding toward sub-micron pitch with memory stacked on top of compute and backside power delivery, as an answer to limits that adding more HBM does not solve. "HBM means high bandwidth, but it's not wide enough," Vivet told EE Times, describing a move toward memory that is "slower, wider, closer" to the logic it feeds. The limiter he names is power density rather than power budget: "water cooling defines the maximum power density of the chip." That extends the same HBM allocation dynamic SCN has tracked into system design.
The dated capacity plan gives the scheduled-queue reading real weight, and so does the uncertainty about demand before the new lines arrive. If accelerator orders soften in 2027, capacity built for today's shortage could land in a balanced market or an overhang. If demand holds, the constraint may migrate to HBM qualification, ABF dielectric film, photonic materials, or the thermal-density limit CEA-Leti describes. Packaging is where the queue has formed. Another link may bind once the new lines arrive.
The Commerce package addresses that moving target one layer upstream. NIST's seven proposed awards span integrated photonics, novel memory, advanced packaging, substrates and component provenance, and are letters of intent subject to further diligence and final agreement rather than disbursed funds. GlobalFoundries would receive up to $300 million for co-packaged optics research, Kepler Computing up to $245 million for 3D and ferroelectric memory, Multibeam up to $140 million for multi-column electron-beam lithography aimed at fine-pitch packaging flows, and Thintronics up to $50 million for low-loss inter-layer dielectrics. Multibeam comes closest to the active packaging process, and even there NIST describes research and development work, not a production line with disclosed output. None of the scopes describe new CoWoS-equivalent capacity that would ship packages in 2026 or 2027.
The GlobalFoundries award targets a different constraint. Co-packaged optics moves optical engines closer to switching or compute silicon to shorten the electrical distance a signal must travel; it affects how accelerators connect, not how many packages leave a CoWoS line, though future CPO can also create new packaging requirements of its own. SCN's analysis of switch-side and compute-side co-packaged optics explains why that distinction matters. Commerce would be funding a technology that can improve the usable scale of AI systems even though it does not clear the current packaging queue.
Thintronics addresses a less visible dependency. Ajinomoto Fine-Techno controls more than 90 percent of the build-up film market used in advanced chip substrates, according to MIT Technology Review, and a domestic low-loss dielectric would diversify one of the inputs sitting under many high-end organic-substrate packages. The award also gives the US a concrete answer to the substrate gap SCN identified in Europe's advanced-materials strategy.
The dollars do not compare directly, but their scale is instructive. TSMC said it plans $60 billion to $64 billion in capital spending during 2026, with 10 to 20 percent going to advanced packaging, testing, mask making and related work. Against that range, the entire $874 million portfolio equals roughly 7 to 15 percent of one year of TSMC's packaging-and-adjacent spending — not packaging alone — and it is proposed R&D money spread across seven paths. Commerce would be funding inputs that may widen future supply choices rather than trying to match TSMC's current equipment build.
The policy change sits in the funding terms. Commerce says the department will receive a minority, non-controlling equity stake in every recipient as a condition of the proposed awards. In May it used the same language for nine quantum and supercomputing letters of intent worth up to $2.013 billion, and in August 2025 the federal government invested $8.9 billion in Intel common stock for a 9.9 percent passive stake; counting $2.2 billion in previously paid CHIPS grants, total federal investment and support in Intel reached $11.1 billion. Three packages in twelve months make minority ownership look like a repeatable instrument rather than a one-off condition. What the latest NIST announcement does not publish is the valuation basis, conversion mechanics or milestone terms that will decide what the stakes are worth and how they affect a recipient's later fundraising.
The same sovereignty logic runs through the supply paths. Intel is presenting New Mexico packaging as a US alternative to TSMC's CoWoS capacity. Samsung and Broadcom are building a separate integrated stack, extending the vertical-integration strategy SCN has tracked in Samsung's silicon photonics program. Samsung says the memorandum could cover more than $200 billion of memory and foundry business through 2030 and expects the relationship to extend into advanced packaging; its announcement does not disclose how much of that projected value, if any, is attributable to packaging. The deal is evidence of a non-TSMC stack forming, with packaging a stated part of the collaboration rather than a quantified commitment.
For accelerators that depend on CoWoS-class integration, advanced packaging is one of the clearest binding constraints today, because the supplier operating the largest line says it is limiting customer growth. Reported capacity estimates and the disclosed 2027–2028 ramps at TSMC, Amkor and Intel suggest the queue should loosen rather than harden into a permanent wall, and HBM stays coupled to the same constraint. The reported quasi-EMIB effort, if confirmed, would be one more sign of how tight packaging has become — supplementary to that direct evidence, not a substitute for it. The $874 million federal portfolio works farther upstream and on a longer clock. Its most durable near-term consequence may be the financing model: Commerce is making minority ownership part of how the United States buys options on the next semiconductor supply chain.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
