Advanced Packaging (CoWoS / 2.5D-3D)
Chips  Demand vs supply & the price of exposure · unit of demand: CoWoS/packaging capacity (wafers/month)
TSMAMKRASXBESI
V2 · factsJun 2026
Sector scan: Semiconductors Group-level demand/supply Updated Jun 2, 2026 · data verified Facts only · no recommendation
Snapshot Product Demand Supply The gap The players The price Deep-dive next Sources

Snapshot — the group at a glance

"Advanced packaging" is the final assembly step that takes separately-made chips — a big logic die (the GPU or accelerator), several stacks of high-bandwidth memory (HBM = fast memory chips stacked vertically right next to the processor), and a connecting baseplate called an interposer (a thin silicon sheet with ultra-fine wiring that links the dies) — and bonds them into one tightly-wired package. The dominant flavor for AI is TSMC's CoWoS (Chip-on-Wafer-on-Substrate); related methods are 2.5D interposers, 3D hybrid bonding (stacking dies face-to-face with no solder bumps, for shorter connections), and chiplet integration (combining several smaller dies instead of one big one). This matters because modern AI accelerators are too big and too memory-hungry to build as a single chip — packaging is the only way to bind the pieces together with enough bandwidth. Per the scan, "every B200/B300 and competitor chip needs advanced packaging," and CoWoS is "literally the tightest bottleneck for AI GPU production today." The product is sold as packaging capacity, usually measured in wafers per month of interposer/assembly throughput.

~$40-60B est.
Rough advanced-packaging / OSAT market size, annual (not-live-verified)
~20-30%/yr est.
Approx. growth of the AI-driven advanced-packaging slice (not-live-verified)
6 scan
US-listed names the scan flags for this product (TSM, AMKR, KLIC, ONTO, BESI, CAMT); ASE/ASX is added from general knowledge
CoWoS capacity
The #1 binding bottleneck in AI chips, per the scan
Demand is running ahead of supply: the scan states "demand far exceeds supply" and CoWoS is "the #1 supply constraint in the AI chip industry." What limits supply is not raw materials but specialized capacity, equipment lead-times, and TSMC's dominance of the high-end recipe — the scan says capacity is being "tripling[ed]" yet "still cannot keep up." In money terms, the group splits into two cash shapes: capacity owners (TSMC, the OSAT assemblers AMKR and ASX) that spend heavily up front to build lines, and capital-light tool sellers (BESI, CAMT, ONTO) that sell machines. The price you pay per dollar of revenue differs sharply between the two (see Price section); the reader can weigh that against where the bottleneck sits.

The product & how money is made

The product is the physical integration of an AI accelerator: bonding the logic die to multiple HBM memory stacks on a silicon interposer, then mounting that on a substrate to form a finished package that drops into a server board. There are two business models that earn cash from it:

Source: 500-stocks scan, /Users/ravf/projects/work/.claude/worktrees/sector-hub/research/investments/500-stocks/02-semiconductors.html (Sector 14, Advanced Packaging).

Demand — how much the world will want this

Known / current (grounded): The scan states demand "far exceeds supply" today, that "every B200/B300 and competitor chip needs advanced packaging," and that CoWoS is "the tightest bottleneck within foundries." The scan also notes AI-accelerator lead times run "6-12 months" and that NVIDIA's Blackwell and next-gen Rubin architectures are "sold out years ahead." Each high-end AI GPU consumes a meaningful slice of interposer area, and every additional HBM stack per package raises the packaging burden — so packaging demand grows faster than unit chip count alone.

Who the buyers are: Chip designers who need their accelerators assembled — NVIDIA, AMD, and the hyperscalers' custom silicon (Google TPU, Amazon Trainium, etc.). They buy capacity from TSMC (CoWoS) and, increasingly, second-source it to OSATs (AMKR, ASX) to reduce reliance on a single bottleneck. The tool-makers' buyers are those same foundries and OSATs building out lines.

Forward demand (forecast): Under the premise that AGI is arriving, accelerator demand keeps compounding because (1) training the next generation needs more compute, (2) inference at scale needs vastly more deployed silicon, and (3) physical AI (robotics, autonomy) adds a new, large class of edge and datacenter accelerators — all of which require advanced packaging. More HBM per package (the move from HBM3E to HBM4) and a shift toward 3D hybrid bonding raise the packaging content per chip, so packaging demand can grow even if chip units plateau. forecast Rough market estimates put advanced-packaging-related revenue for the AI slice growing on the order of 20-30%/year over the next several years. est.

✓ VERIFIED — the following figures were confirmed from primary sources after initial publication:

Remaining caveat: some market-size and growth-rate figures not listed above are directional estimates from general knowledge (model cutoff ~early 2026), not live-verified. Company-specific financials in the Players table are from the most recent public filings or earnings. For SEC-verified deep dives on individual companies, see Stock Reports.

Supply — how much can be made, and what limits it

Current capacity & expansion (grounded): The scan is blunt — "TSMC's CoWoS capacity is the #1 supply constraint in the AI chip industry. They are tripling capacity but still cannot keep up." So the supply response is real and aggressive, but it lags demand. Amkor (AMKR) and ASE/ASX are adding advanced-packaging lines to act as second sources, and Besi (BESI) is shipping the hybrid-bonding tools needed for the next 3D step.

What actually limits supply: Not raw silicon or money alone, but a stack of constraints — (1) specialized capacity: CoWoS lines are purpose-built and take time to qualify (prove the line produces good chips reliably); (2) equipment lead-times: the bonders, platers, and metrology tools have their own order queues; (3) know-how: the high-end recipe (large interposers, tight overlay, hybrid-bonding yield) is hard to replicate, which the scan reflects in TSMC's dominance; and (4) capital + time: building and qualifying a line is a multi-quarter-to-multi-year effort. This is why the scan's capacity "tripling" still leaves the market short.

Market-share structure (estimate): TSMC controls the large majority of high-end CoWoS for leading-edge AI GPUs. est. Among OSATs, ASE/ASX is generally the largest independent assembler and Amkor the next largest; they hold the bulk of broader outsourced packaging-and-test and are growing their advanced share as second sources. est. In the hybrid-bonding tool niche, Besi is widely regarded as the leader, with Applied Materials and others competing. est. Supply is therefore concentrated: one dominant capacity owner at the high end, a small number of credible second-source assemblers, and a thin set of tool suppliers.

Source: 500-stocks scan (Sector 14 and Sector 13, Foundries); share/structure figures are general-knowledge estimates, not live-verified.

The gap — demand vs supply

Today the product is clearly short (demand above supply). The scan's own evidence: "demand far exceeds supply," CoWoS is "the tightest bottleneck," capacity is being tripled "but still cannot keep up," AI-accelerator lead times of "6-12 months," and downstream GPUs "sold out years ahead." A short market typically shows up as full utilization, long lead-times, and firm or rising pricing for the constrained step.

SignalReading todayWhat would flip it to oversupply
Capacity vs demandShort — capacity tripling, still behind (scan)Build-out catches up while accelerator demand cools
Lead times (AI accelerators)~6-12 months scanLead times collapse to weeks
Downstream GPUs"Sold out years ahead" (scan)Order cancellations / push-outs appear
Pricing on the constrained stepFirm est.Packaging prices fall as utilization drops

When could it flip? forecast A short market in a capacity industry usually flips when the multi-year build-out (which is underway now) eventually overshoots demand. Given the AGI-driven demand premise — compute and physical-AI needs still rising — a nearer-term possibility is a temporary digestion pause (a quarter or two of slower orders) rather than a structural glut. A true oversupply would require both the capacity tripling-and-more to land AND accelerator demand to plateau at the same time. Neither is visible in the grounded scan today; this remains a forecast, not a fact.

The players — who captures the money

Company (ticker)What it makesExposure to THIS productRough size est.Position
TSMC (TSM)Leading-edge foundry + CoWoS packagingDiversified; packaging is a small but critical slice of a large foundryMega-cap (~$1T-class) est.Dominant CoWoS capacity owner; the bottleneck itself (scan)
Amkor (AMKR)OSAT — assembly & test, advanced packagingPure-play packaging; advanced packaging a growing share of revenueSmall/mid-cap (~$5-10B) est.Among the largest independent assemblers; key US-listed second source
ASE / ASX (ASX)OSAT — large independent assembly & testPure-play packaging; advanced share growing within a large test/assembly baseMid/large-cap (~$20-30B) est.Generally the largest independent OSAT by revenue; broad second-source capacity (not in scan list)
Besi (BESI)Hybrid-bonding & die-attach equipmentTool maker; hybrid bonding is its core growth driverMid-cap (~$10B) est.Widely regarded as the hybrid-bonding tool leader — a chokepoint for the next 3D step
Kulicke & Soffa (KLIC)Bonding equipment (wire/advanced)Tool maker; partial exposure, broader bonding baseSmall-cap (~$3-4B) est.Established bonder vendor; advanced-packaging optionality
Onto Innovation (ONTO)Inspection & metrology, packaging processTool maker; meaningful advanced-packaging exposureSmall/mid-cap (~$8B) est.Metrology/inspection for packaging build-out
Camtek (CAMT)Inspection & metrology for packagingHigh exposure to advanced-packaging inspectionSmall-cap (~$4-5B) est.Inspection specialist tied to the packaging build-out

Source: company list from 500-stocks scan (Sector 14) — TSM, AMKR, KLIC, ONTO, BESI, CAMT; ASE/ASX added from general knowledge. Sizes and revenue-share characterizations are general-knowledge estimates, not live-verified.

The price of exposure

Plain money translation — roughly what the market pays today for $1 of this year's revenue (price-to-sales = total market value of the company divided by one year of its sales) and the cash shape of each model. est. All multiples below are approximate and not-live-verified.

Name~Market value per $1 of revenue est.Money-in / money-out shape
TSMC (TSM)~9-12x est.Capex-heavy but high-margin; large owner cash after heavy reinvestment
Amkor (AMKR)~1-2x est.Capex-heavy, thinner OSAT margins; modest owner cash, cyclical
ASE / ASX (ASX)~1-2x est.Capex-heavy OSAT at scale; broad but lower-margin cash generation
Besi (BESI)~12-18x est.Capital-light tool maker; high margins, strong owner cash per dollar of sales
Camtek (CAMT)~8-12x est.Capital-light inspection; high margin, smaller absolute cash

The arithmetic, in cash terms: the capital-light tool makers (BESI, CAMT, ONTO) convert a higher share of each revenue dollar into owner cash, and the estimated multiples above for them sit far higher (roughly 8-18x sales) than for the assemblers — so a buyer pays many dollars per dollar of sales, but those sales carry high margins. The capex-heavy assemblers (AMKR, ASX) carry estimated multiples near 1-2x revenue, because much of their cash is reinvested into plants and equipment, leaving less free cash for owners per dollar of sales. TSMC is priced as the whole foundry (~9-12x est.), not the packaging line alone, so its multiple reflects far more than CoWoS. These are order-of-magnitude figures; the reader weighs price paid against margin, growth, and where each name sits relative to the bottleneck.

Source: general-knowledge price multiples (not live-verified); business-model facts from the 500-stocks scan.

What to deep-dive next

Factual pointers for where a company-level deep-dive is most informative:

Sources & confidence

Confidence summary: HARD = product definition, players in the scan list, bottleneck identity, lead-time/sold-out qualifiers (from scan). APPROXIMATE/NOT-LIVE-VERIFIED = all dollar sizes, growth rates, market shares, and valuation multiples.