Advanced Semiconductor Packaging Costs: The Definitive 2026 Guide

Correction, October 3, 2026: Earlier versions of this guide stated "fully booked" CoWoS lines and 52–78-week booking windows as fact, attributed "fully booked through 2025 into 2026" to TSMC's Q3 2025 earnings call, and dated a "three times short" remark to that call. TSMC publishes neither lead times nor CoWoS capacity. On the Q3 2025 call its CEO called AI-related capacity "very tight"; the "three times short" remark was made at a November 2025 industry event and concerned wafer capacity. This version labels the estimates and adds what TSMC has said and filed. A source check the same week found six source links that led to pages that do not exist; they now point to the reports they described, and where a report says something different, the text follows the report: JPMorgan's end-2027 CoWoS estimate is 112,000 wafers per month (the 170,000 figure came from other investors), a quoted JPMorgan line on CoWoS-L and a 1–5% lane-repair yield figure could not be found and were removed, and burn-in conditions now match SemiEngineering's reporting.

CoWoS advanced packaging now represents 17–23% of AI accelerator manufacturing costs, with per-chip packaging ranging from $750 for H100-class designs to over $1,100 for NVIDIA's B200. Memory and packaging together account for 60–70% of total chip COGS, far surpassing the logic die itself. This guide synthesizes analyst estimates (Morgan Stanley, JPMorgan, Bernstein), TSMC earnings guidance, Epoch AI cost models, and TrendForce supply chain reporting into the definitive 2026 view of advanced packaging economics. Data as of April 2026.

The single most important chart in this entire article — where does the money actually go in an AI accelerator? Logic silicon is the smallest slice:

Memory and packaging dominate AI accelerator COGS — logic silicon is the smallest slice

Manufacturing cost breakdown per chip, 2026 estimates (USD)

Source: Silicon Analysts estimate aggregating Epoch AI, Raymond James, TrendForce, March 2026

CoWoS packaging cost by variant

CoWoS-S uses a monolithic silicon interposer fabricated on TSMC's N65-equivalent process with TSVs, multi-layer RDL, and deep-trench capacitors. For the H100 — an 814mm² die plus 5–6 HBM3 stacks on a ~2,500mm² interposer — total packaging cost runs approximately $750 per chip. A fully processed CoWoS wafer costs an estimated $10,000–$12,000 (Silicon Analysts estimate), with the interposer alone consuming 50–70% of packaging cost. With roughly 25–28 gross interposers per 300mm wafer and 60–70% interposer yields at this size, the math reconciles at $590–$800 interposer cost alone, before assembly and HBM attach. The practical size ceiling for CoWoS-S is ~2,700mm² (3.3× reticle).

CoWoS-L replaces the monolithic silicon interposer with an organic RDL substrate embedded with Local Silicon Interconnect (LSI) bridge dies — enabling packages far beyond CoWoS-S's reticle limit. NVIDIA's B200 uses CoWoS-L to integrate two ~800mm² compute dies plus 8 HBM3E stacks, a configuration physically impossible on CoWoS-S. Per-chip packaging cost runs $1,000–$1,100 (Silicon Analysts estimate; Epoch AI puts B200 CoWoS-L packaging at about $1,100 within a $1,000–$1,200 range, December 2025) — a ~47% premium over CoWoS-S. The premium comes from additional LSI bridge components, higher microbump counts, and initially lower assembly yields. Paradoxically, CoWoS-L is more cost-effective at very large package sizes because it sidesteps the catastrophic yield losses of fabricating monolithic silicon interposers beyond 2,700mm². Small LSI bridges yield at ~90% versus ~60% for large monolithic interposers.

CoWoS-R replaces the silicon interposer entirely with an organic thin-film interposer using InFO-based RDL — the cheapest CoWoS variant. AWS Trainium2 is the flagship user, employing CoWoS-R for a dual-chiplet configuration with 4 HBM stacks.

TSMC reportedly raised CoWoS prices 10–20% for 2025 (TweakTown, January 2025, reported 15–20%; Morgan Stanley had flagged a 10–20% rise for 2025 in November 2024, per Commercial Times via TrendForce). On TSMC's Q3 2025 earnings call, CEO C.C. Wei said AI-related "front-end and back-end capacity is very tight" and that TSMC was "working very hard to narrow the gap between the demand and supply." NVIDIA alone secured >70% of CoWoS-L capacity for 2025 (TrendForce, February 2025).

How alternatives compare. Intel EMIB embeds small silicon bridges only where die-to-die connections are needed, achieving 30–40% lower cost than CoWoS (TrendForce, November 2025). Bernstein estimates EMIB packaging at a few hundred dollars per chip versus $900–$1,000 for CoWoS on a Rubin-equivalent AI processor (Bernstein, via Investing.com, February 2026). Intel claims ~90% wafer utilization for small bridge dies versus ~60% for large interposers. TSMC InFO-PoP for iPhone A-series chips costs an estimated $10–$30 per chip — 3–10× cheaper than CoWoS — because it eliminates the silicon interposer entirely.

Correction, October 4, 2026: This sentence quoted Bernstein as estimating EMIB at "low hundreds of dollars per chip", and a later one tied its ~$900–$1,000 CoWoS figure to a configuration with 12+ HBM stacks. Neither the phrase nor the stack count is in the Investing.com report of Bernstein's note, which says "a few hundred $" per chip for EMIB and $900–$1,000 for CoWoS on a Rubin-equivalent processor. Both sentences now follow it, and the TweakTown citation points at its actual CoWoS story, which reported 15–20%.

TSMC is also outsourcing aggressively: an estimated 240,000–270,000 CoWoS wafers/year will move to OSATs in 2026, mainly Amkor (180,000–190,000) and SPIL (60,000–80,000) (Global Semi Research, December 2025). ASE's own CoWoS capacity is projected to more than triple to 20,000–25,000 WPM (equipment sources cited by Commercial Times, via TrendForce, December 2025), and ASE reportedly raised advanced packaging prices 5–20% for 2026 (January 2026).

Key takeaway: CoWoS-L costs 20–40% more than CoWoS-S per chip, but is actually more cost-effective for packages exceeding 2,700mm² because it avoids catastrophic silicon interposer yield losses. For everything larger than a single reticle + 8 HBM, CoWoS-L is the only viable option.

See a full per-chip cost breakdown in the Packaging Calculator.

When chiplets beat monolithic — and when they don't

The chiplet-versus-monolithic equation hinges on three variables: defect density, die size, and packaging cost. At TSMC N3's mature defect density of ~0.09 defects/cm², a monolithic 800mm² die yields approximately 48.7% under the Poisson model, producing ~32 good dies per wafer from 65 gross. At a wafer cost of ~$19,500 for N3, that translates to approximately $609 per good die.

Four 200mm² chiplets on the same N3 process yield 83.5% each, producing ~255 good chiplets per wafer at $76 each — or $306 for a complete set of four. The chiplet approach saves roughly $300 in silicon costs. But this is overwhelmed by advanced packaging: a CoWoS-L interposer plus assembly adds $700–$1,000, KGD testing adds $20–$40, and integration testing $50–$100. Total chiplet cost reaches $1,126–$1,546 versus ~$700 for the monolithic approach — monolithic is ~45% cheaper for 800mm² total silicon at mature defect densities.

The picture changes dramatically as defect density rises. The crossover is visible: monolithic stays cheaper until defect density passes ~0.17/cm², then chiplet wins.

Chiplets only beat monolithic when defect density exceeds ~0.17 defects/cm²

Total cost per good die at 800mm² silicon, TSMC N3, CoWoS-L packaging

Source: Silicon Analysts, Poisson yield model, TSMC N3 wafer at $19,500, CoWoS-L packaging at $800

The crossover at 800mm² occurs around D₀ ≈ 0.17–0.20 defects/cm² — typical of early production on a new node. The common rule of thumb that "chiplets beat monolithic above 400–500mm²" applies only when using cheap organic-substrate MCM packaging ($50–$200). AMD's EPYC Genoa proves this brilliantly: 12 small CCDs (~70mm² each) plus one IOD (~419mm²) on an inexpensive organic substrate delivers >40% cost reduction versus a hypothetical monolithic equivalent (AMD Chiplet Actuary paper, arXiv). But for AI accelerators requiring HBM — which mandates CoWoS packaging regardless — the comparison becomes less about cost and more about physical necessity (exceeding reticle limits) and risk management (yield insurance on immature nodes).

Yield versus die size at current defect densities (Poisson model):

Real-world chip cost comparisons illustrate the range:

The H100's cost breakdown is revealing: the 814mm² logic die costs only ~$300 (9% of COGS), while HBM3 memory is ~$1,350 (41%), CoWoS-S is ~$750 (23%), and test/assembly is ~$920 (28%). For B200, HBM3E rises to ~$2,900 (45% of COGS). NVIDIA moved to dual-die for Blackwell not for cost savings but because a single >1,600mm² die physically exceeds the reticle limit. Jensen Huang stated NVIDIA invested ~$10B in NV-HBI interconnect R&D to make dual-die work at 10 TB/s die-to-die bandwidth.

Key takeaway: The "chiplets are always cheaper above 400mm²" rule is wrong for HBM-bearing AI accelerators. On CoWoS, monolithic wins until defect density exceeds ~0.17/cm². Chiplets are chosen for physics (reticle limit) and yield insurance, not cost savings.

Compare cost structures across the full AI accelerator lineup in the Cost Bridge Chart.

Model this scenario yourselfOpen the Chip Price Calculator pre-loaded with an H100-class configuration — change the die size, defect density, or packaging type and watch the crossover move.

Test economics — the hidden cost multiplier

For a large AI accelerator like the H100, total test and assembly costs run ~$920 (28% of $3,320 COGS). Individual step costs scale sharply with die complexity:

ATE hourly rates run ~$100/hour for mainstream SoC testers; high-end configurations (Advantest V93000, Teradyne UltraFLEXplus) reach $100–$200/hour fully loaded (AnySilicon, 2024). Large AI accelerator dies require 30–120 seconds per die at wafer sort versus 0.5–2 seconds for simple IoT chips, with only 1–2 die per touchdown due to massive die area — collapsing parallelism and driving up per-die cost. Package-level burn-in typically runs 24–48 hours, longer for stricter quality targets, at 150–175°C under voltage stress (SemiEngineering, July 2025). Historically test was 2–3% of IC revenue; for advanced-node AI chips it is rising to 5–10%.

Chiplets insert a critical additional step: Known Good Die (KGD) testing. Once chiplets are bonded via microbumps or hybrid bonds, rework is essentially impossible — a single defective die scraps the entire package, including all other good chiplets, interposer, and HBM stacks (Cadence, 2024). The "Rule of Ten" applies: detecting defects costs 10× more at each subsequent manufacturing stage.

KGD testing costs $5–$15 per chiplet for thorough structural, parametric, and BIST testing. FormFactor notes that full KGD testing of every die is "often not economically feasible" — the industry compromise is "Good Enough Die" testing that balances cost against fallout risk (FormFactor/Amy Leong, 2022). Composite yield math makes quality paramount: for a 4-chiplet design at 95% KGD each, system yield before assembly losses is only 0.95⁴ = 81.5%. For AMD's MI300X with 12 chiplets, even 98% per-chiplet KGD produces only 0.98¹² = 78.5% composite yield.

The math is brutal: every chiplet you add punishes system yield multiplicatively. This is why MI300X-class designs demand ≥98% KGD per chiplet just to reach passable system yield:

Adding chiplets punishes system yield multiplicatively — MI300X needs ≥98% KGD just to reach passable yield

System yield as a function of per-chiplet Known Good Die yield, by chiplet count

Source: Silicon Analysts, composite yield = (KGD)^N

Total test cost for chiplet packages runs 15–30% higher than monolithic equivalents (Chiplet Actuary, arXiv). The B200's ~$2,400 in test and assembly dwarfs the H100's $920, reflecting dramatically higher multi-die testing complexity. UCIe lane repair provides partial mitigation — a defective lane can be swapped for a spare during test (SemiEngineering, August 2025).

Key takeaway: Once chiplets are bonded, rework is impossible — a single bad die scraps the whole package. That physics is why KGD testing adds 15–30% to chiplet test costs versus monolithic, and why "Good Enough Die" compromises are industry standard.

CoWoS capacity, allocation, and lead times

TSMC's CoWoS capacity has roughly doubled annually since 2023 — the kind of growth curve the semiconductor industry almost never sees outside of greenfield node ramps:

TSMC CoWoS capacity is expanding ~10× from 2023 to 2027

Total CoWoS wafer capacity, wafers per month (all variants combined)

Source: TrendForce, Global Semi Research, Morgan Stanley and investor estimates via 36Kr, Dec 2024–Jan 2026

CoWoS-L is the primary growth driver: of NVIDIA's ~595,000 CoWoS wafers for 2026, about 510,000 go to TSMC, mainly for Rubin-generation CoWoS-L (Morgan Stanley via 36Kr, December 2025). TSMC is boosting CoWoS-S capacity mainly through equipment reallocation rather than new builds. TSMC operates advanced packaging across AP3 (Longtan), AP5/AP5B (Taichung), AP6/AP6B (Zhunan), AP7 (Chiayi — opening ceremony December 4, 2025), and AP8 (Tainan — 96,000+ sqm, 9× AP6's size, equipment move-in began Q4 2025). A US packaging facility is planned to break ground in 2026 for completion by 2029.

Customer allocation is heavily concentrated. Global CoWoS demand in 2026 is projected at approximately 1 million wafers — up 40–50% YoY (Morgan Stanley, December 2025). NVIDIA alone takes more than the rest of the industry combined:

NVIDIA locks up ~60% of 2026 CoWoS capacity — more than all other customers combined

2026 CoWoS wafer allocation by customer, thousands of wafers

Source: Morgan Stanley, DigiTimes, December 2025

Broadcom is gaining share as hyperscaler ASIC demand accelerates. MediaTek is a new entrant booking ~20,000 wafers for Google's TPU project. The top customers lock in >85% of capacity, leaving <15% for second-tier players and startups.

No sign the booking window has compressed. Despite the capacity ramp and OSAT outsourcing, trade reports still describe CoWoS-S and CoWoS-L as fully booked, with booking windows for new orders of 52–78 weeks in 2026 — the window is how far in advance a customer must commit to secure a packaging slot, not production cycle time. TSMC publishes neither figure. What it has said is that its "packaging capacity is so tight that now it's limiting my customers' growth" (C.C. Wei, Q2 2026 earnings call, July 16, 2026), and its filings show customers paying to hold capacity well in advance: payments "to retain the Company's capacity" that fall due beyond twelve months rose from NT$43.3B at the end of 2025 to NT$92.4B at June 30, 2026 (TSMC Form 6-K, Q2 2026; company-wide, not CoWoS alone). An estimated 85%+ of 2026–2027 capacity is already locked. Separately, Wei said at an industry awards event on November 20, 2025 that TSMC's advanced-node capacity was "about three times short" of AI demand (Tom's Hardware, November 2025) — a remark about wafer capacity rather than CoWoS. Morgan Stanley's OCP Conference analysis and Jensen Huang's commentary suggest constraints are also building downstream — memory, power infrastructure, and rack assembly — but nothing on the public record shows packaging itself loosening. (A DigiTimes report from August 2025 indicated CoWoS utilization briefly dipped to ~60%.) For the full argument on why the booking window, not headline WPM, is the binding metric, see our CoWoS lead-times analysis.

Key takeaway: Capacity is estimated to nearly double, yet reported booking windows for new orders sit at 52–78 weeks, an estimated 85%+ of 2026–2027 capacity is pre-locked, and TSMC's customers are prepaying to hold capacity more than a year out. Watch the booking window, not the WPM headline — HBM and power are additional constraints stacking on top of packaging, not replacing it.

Track live allocation and queue status on the Allocation Dashboard.

AI chip packaging case studies

NVIDIA B200 integrates two ~800mm² GB100 dies (TSMC 4NP, 208B transistors total) connected via NV-HBI at 10 TB/s, with 8 HBM3E stacks delivering 192GB at 8.0 TB/s, all on a CoWoS-L organic interposer with embedded LSI bridges. Estimated total COGS is ~$6,400: packaging ~$1,100 (17%), HBM3E ~$2,900 (45%), logic dies ~$850 (13%), test/assembly ~$1,550 (24%). Packaging yield losses alone add roughly $1,000 in effective scrap cost. The GB200 NVL72 rack (72 GPUs + 36 Grace CPUs, 120kW liquid-cooled) implies GPU-only component COGS exceeding $460,000 per rack before system integration.

AMD MI300X is a 3.5D packaging tour de force: 8 XCD compute chiplets (N5, ~115mm² each) are 3D hybrid-bonded onto 4 IOD dies (N6, ~370mm² each) via SoIC at 9µm TSV pitch. The 12-chiplet stack sits on a CoWoS-S interposer at ~3.5× reticle alongside 8 HBM3 stacks. Total active silicon exceeds 2,400mm² across 153B transistors. Estimated COGS is ~$5,300, with packaging at $1,200–$1,800 reflecting the world's most complex commercial packaging — over 100 pieces of silicon per package including HBM layers.

Google TPU v7 (Ironwood) uses dual ~700mm² compute dies on TSMC N3P with 8 HBM3E stacks (192GB, 7.2 TB/s), delivering 4,614 FP8 TFLOPS — a dual-die + 8 HBM configuration that mirrors B200, a physics-driven convergence. Estimated packaging cost is $1,000–$1,300. Microsoft Maia 200 (launched January 2026) uses a monolithic 727mm² die on TSMC N3E with 216GB HBM3E across 6 stacks; estimated packaging is $900–$1,200. AWS Trainium2 uses CoWoS-R — the cost-conscious organic-interposer variant — for 2 chiplets + 4 HBM3 stacks. Meta MTIA v2 is the outlier: a 421mm² die on TSMC N5 using standard flip-chip BGA packaging with LPDDR5 — no HBM, no advanced packaging — at ~$50–$150 total COGS. Newer MTIA generations (300+) are transitioning to chiplet architectures with CoWoS-S and HBM.

Model any of these configurations yourself in the Chip Price Calculator.

Future packaging roadmap

CoWoS 9.5× reticle scales the interposer to ~8,100+ mm², accommodating 12+ HBM stacks alongside cutting-edge logic dies. Mass production is targeted for 2027 (TSMC North America Technology Symposium, May 2025; TrendForce, November 2024). Bernstein estimates CoWoS packaging for a Rubin-equivalent AI processor at ~$900–$1,000 per chip (Bernstein, via Investing.com, February 2026).

CoPoS (Chip-on-Panel-on-Substrate) replaces the round 300mm wafer with a 310mm × 310mm rectangular panel for interposer fabrication. Tool deliveries to TSMC subsidiary Xintec's R&D line began February 2026, with full pilot line completion targeted June 2026 (TrendForce/Commercial Times, April 2026). Mass production is expected late 2028 to early 2029 at AP7 Phase 4 in Chiayi. Panel utilization exceeds 95% versus ~85% for circular wafers, driving an expected 20–30% cost advantage over current CoWoS. The substrate-less CoWoP variant could achieve 30–50% cost reduction by eliminating ABF substrates, which currently account for ~40% of packaging cost.

SoIC capacity stands at ~10,000 WPM in 2025, targeting 15,000–20,000 WPM by end-2026 (TrendForce, March 2025). CapEx runs up to $7 billion per 10,000 WPM of SoIC capacity — among the most capital-intensive packaging technologies. Current customers include AMD (MI300 series); Apple, NVIDIA (Rubin), and Broadcom are confirmed future users.

Hybrid bonding is in production for AMD MI300 (SoIC), 3D NAND, and CMOS image sensors, but HBM4 will stick with microbumps, postponing hybrid bonding adoption to HBM4E around 2027 (Semiconductor Engineering, March 2026). The hybrid bonder market reached ~$152M in 2025 and is projected to reach $397M by 2030 at 21.1% CAGR (Yole Group, 2025). Thermocompression bonding remains simpler and cheaper — hybrid bonding only becomes economical when pad pitch drops below ~10µm.

UCIe 3.0 was released August 5, 2025 (UCIe Consortium), doubling data rates to 48 GT/s and 64 GT/s and extending sideband reach to 100mm. The consortium has grown to 150+ members, enabling multi-vendor chiplet interoperability and lowering barriers to entry for smaller design teams.

Industry economics

The total advanced packaging market reached approximately $43–50 billion in 2025 (Yole, IMARC, Acumen), growing at 9.5–10.6% CAGR through 2030. The high-performance segment (chiplets, AI/HPC) is the fastest-growing subsector at 23% CAGR, projected to reach $28.5B by 2030 (Yole/SEMI Summit, 2025). TSMC's advanced packaging revenue reached ~8% of total company revenue in 2025, targeting >10% in 2026. Industry-wide CapEx for advanced packaging exceeded $14 billion in 2025 (Yole), with ASE alone planning $7B for 2026.

Global CoWoS wafer demand: 370,000 wafers in 2024 → 670,000 in 2025 → ~1 million in 2026 (Morgan Stanley). Key equipment vendors span BESI and ASMPT (bonding), Canon and ASML (packaging lithography — ASML's TWINSCAN XT:260 launched in 2024 supports interposers up to 3,432mm² without stitching), and Advantest (~31% ATE share) and Teradyne (~23%). High-end ATE systems cost $1.5–$5 million per unit.

Ajinomoto controls >95% of the ABF film market for CPU/GPU substrates (Nikkei Asia), with supply constrained for high-layer-count packages. Ajinomoto plans a 50% boost in ABF output by 2030. Intel and NVIDIA have co-invested in substrate supplier expansions, covering ~50% of new production line costs. Glass substrates are emerging as a long-term alternative — AMD and AWS are reportedly accelerating glass substrate timelines.

CoWoS costs are rising, not falling: TSMC raised prices 10–20% for 2025. The progression from H100-class (~$750) to B200-class (~$1,100) to Rubin-class (~$900–$1,000) shows per-chip costs remain elevated even as TSMC scales. The most significant cost relief will come from panel-level packaging (CoPoS), which should deliver 20–50% cost reduction — but mass production timing of late 2028–early 2029 means meaningful relief is still 2–3 years away.

Frequently asked questions

How much does CoWoS packaging cost per chip?

CoWoS-S costs approximately $300–$800 per chip depending on die size and HBM count. CoWoS-L costs $800–$2,000 for multi-die configurations. For an H100-class chip, packaging is about $750 (23% of total COGS). For NVIDIA's B200, packaging is ~$1,100 (17%).

Is chiplet packaging cheaper than monolithic?

It depends on packaging technology. On organic substrates (like AMD EPYC), chiplets save >40% above ~400mm². On CoWoS (required for HBM), chiplets only break even above ~800mm² on immature processes (D₀ > 0.17 defects/cm²). On mature nodes, a monolithic design is ~45% cheaper at 800mm² total silicon because CoWoS packaging costs ($700–$1,000) overwhelm the silicon yield advantage.

What is Known Good Die (KGD) testing?

KGD testing verifies each chiplet works before assembly. Once bonded via microbumps or hybrid bonds, rework is impossible — a single defective die scraps the entire package. KGD costs $5–$15 per chiplet and is mandatory for cost-effective multi-die packaging.

How long is the CoWoS lead time in 2026?

Trade reports put the booking window for new CoWoS orders at roughly 52–78 weeks in 2026; TSMC does not publish lead times. This is how far in advance a customer must commit to secure a packaging slot — not production cycle time. TSMC's CEO said in July 2026 that packaging capacity is "so tight that now it's limiting my customers' growth," and an estimated 85%+ of 2026–2027 capacity is already locked, so there is no sign the window has compressed.

What is CoWoS-L and how is it different from CoWoS-S?

CoWoS-L uses an organic RDL interposer with embedded Local Silicon Interconnect (LSI) bridges instead of a monolithic silicon interposer. It supports packages up to 5,000mm²+ (versus 2,700mm² for CoWoS-S) and is used for NVIDIA's B200 and future Rubin GPUs. It costs 20–40% more per chip than CoWoS-S but is actually more economical at very large package sizes because it avoids catastrophic interposer yield losses.

Who gets the most CoWoS capacity?

NVIDIA secures ~60% of total CoWoS allocation in 2026 (~595,000 wafers). Broadcom gets ~15% (with Google TPU, Meta ASIC, and OpenAI projects), AMD ~11%, Marvell ~5.5%, and Amazon/Alchip ~5%. The top customers lock in >85% of capacity, squeezing out smaller AI chip companies.

References & Sources

<!-- Auto-populated from frontmatter sources by the rendering pipeline -->