Wafer Fab Capacity Planning for AI Chips: From Die Demand to Wafer Requirements

    This article presents a reusable method for estimating foundry wafer capacity from AI chip demand. It explains how to convert die requirements into wafer starts using die-per-wafer geometry, how to model die yield with Poisson, Murphy, and negative-binomial models, and how to combine nominal capacity, effective capacity, utilization, and ramp-up. Two worked examples—a 400 mm² AI inference chip and an 800 mm² AI accelerator—show how yield, test and packaging loss, safety stock, and demand scale affect wafer demand. The article also covers common planning mistakes and a practical checklist.

    In the semiconductor industry, foundry capacity is often treated as a black box by outsiders. Fabs require large investments and have long delivery cycles, and it is not easy to understand intuitively why AI chip capacity ramps so slowly. In practice, capacity estimation follows a clear sequence of steps based on geometry, statistics, and production economics, and each step can be checked with numbers. This article follows the order of “from wafers to dies, from dies to wafers,” provides a reusable estimation method, and demonstrates the calculation with two examples.

    1. First, Convert Wafers into Dies

    A common unit for foundry capacity is wafer starts per month (WPM, wafers per month). How many dies a wafer can produce depends on three factors: wafer size, die area, and yield. The starting point for estimation is therefore to establish the conversion relationship between required die count and required wafer count.

    Most current mainstream lines use 12-inch (300 mm) wafers. The wafer area is approximately π × 150² ≈ 70,686 mm². The wafer edge cannot be fully used, and square dies cannot tile a circular wafer without gaps. In engineering practice, the following approximation is commonly used to estimate the number of whole dies per wafer, or DPW (Dies Per Wafer):

    DPW ≈ π × (d/2)² / S − π × d / √(2S)

    Here, d is the wafer diameter, taken as 300 mm, and S is the die area in mm². The first term is wafer area divided by die area, and the second term deducts edge loss. The formula assumes regular square dies and does not account for special layouts. Actual numbers should be based on layout tools and vendor data.

    Take a 600 mm² AI accelerator chip as an example. The first term is approximately 117.8, and the second term is approximately 27.2, so each wafer yields about 90 dies. If one simply divides the circular wafer area by die area, the result is about 118 dies, an overestimate of about 30%. This shows that edge loss cannot be ignored.

    2. Yield: The Core Variable Determining Real Output

    Yield needs to be defined carefully. There are three common categories. Wafer yield refers to the proportion of good wafers among input wafers, meaning whole wafers that are scrapped. Die yield refers to the proportion of good dies on a single wafer. Line yield combines all steps to calculate the final good dies as a proportion of input dies.

    Die yield decreases as area increases, mainly because defects land randomly on the wafer. The simplest model is the Poisson model:

    Y = e^(−A × D)

    Here, A is die area in cm² (1 cm² equals 100 mm²), and D is defect density in defects per cm². The Poisson model assumes defects are independent. When defects cluster in certain areas, it underestimates yield. In that case, the Murphy model can be used: Y = ((1 − e^(−A × D)) / (A × D))², or the negative-binomial model: Y = (1 + A × D / α)^(−α). Here, α is a clustering parameter. The smaller α is, the more pronounced defect clustering is, and the higher the yield is for the same A and D.

    Consider a set of assumed numbers. If defect density D is 0.1 per cm² and the die area is 100 mm² (1 cm²), the Poisson model gives a yield of about 90%. If area increases to 600 mm² (6 cm²), yield falls to e^(−0.6), or about 55%. A sixfold increase in area lowers yield from about 90% to about 55%. Large-area AI chips therefore place higher demands on process control. The D values above are assumptions for this article; actual values require fab data.

    3. Capacity Formula and Ramp-Up

    Let N_eff be the number of good dies per wafer. It is calculated as:

    N_eff = DPW × die yield × wafer yield

    If monthly market demand is Q dies, the required number of wafers is Q / N_eff. Considering rework, test loss, and safety stock, a factor of 1.1 to 1.2 is usually applied.

    Capacity planning must also distinguish between nominal capacity and effective capacity. Nominal capacity is the designed capability of the line, while effective capacity deducts the impact of utilization. The conversion is:

    Effective wafer capacity = nominal capacity × overall utilization
    Effective die capacity = effective wafer capacity × N_eff

    A new line usually does not reach full production immediately after launch; it goes through a ramp-up period. Equipment tuning, process stabilization, and yield improvement all take time. The yield curve is generally S-shaped: it may be only around 30% at first and improve gradually over several months. Ramp speed depends on process maturity and the experience of the engineering team. Ramp-up should be planned separately, rather than directly using steady-state numbers.

    4. Example: Back-Calculating Capacity for a 400 mm² AI Inference Chip

    Suppose a company plans to mass-produce an AI inference chip under the following conditions. Die area is 400 mm². Target monthly shipment is 1 million dies. The process uses 12-inch wafers and an advanced node. Die yield is assumed to be 65%, and wafer yield is counted as 100%, meaning no whole-wafer scrap is considered. The geometric formula gives about 143 whole dies per wafer; considering edge layout and design margin, this example conservatively uses 120 dies. Test and packaging loss is 5%. The safety factor is 1.15.

    The calculation steps are as follows.

    1. Accounting for test and packaging loss, the actual number of dies that must be started is 1,000,000 ÷ (1 − 5%) ≈ 1,052,600 dies/month.
    2. Good dies per wafer: N_eff = 120 × 65% = 78 dies.
    3. Required wafers: 1,052,600 ÷ 78 ≈ 13,495 wafers/month.
    4. With safety stock, required wafer capacity is 13,495 × 1.15 ≈ 15,520 wafers/month.

    This means the design company needs to reserve at least about 15,500 WPM of advanced-process capacity at its foundry. A 20,000 WPM advanced-process line would need about 78% utilization to meet this demand, before counting other customers’ orders.

    A sensitivity analysis shows how parameter changes affect the result. With die yield at 50%, 65%, and 80%, and safety factors of 1.1, 1.15, and 1.2, the required wafer counts are as follows.

    Die yield Good dies per wafer Wafers/month at safety factor 1.10 Wafers/month at safety factor 1.15 Wafers/month at safety factor 1.20
    50% 60 19,298 20,175 21,053
    65% 78 14,845 15,520 16,194
    80% 96 12,061 12,610 13,158

    If yield falls from 65% to 50%, required wafers increase by about 30%. This shows that the yield assumption has a greater impact on capacity planning than small adjustments to the safety factor.

    5. Where AI Chip Capacity Pressure Comes From

    Take an 800 mm² AI accelerator chip as an example. Using the approximation above, the number of whole dies per wafer is about 65. Assuming die yield of 50%, good dies per wafer are about 32.5.

    If a cloud service provider plans to deploy 100,000 such chips per year, that is about 8,333 chips per month. Required wafers are 8,333 ÷ 32.5 ≈ 256 wafers/month. On a 50,000 WPM line, this demand accounts for only about 0.5%.

    At the wafer level, demand from a single customer is therefore not large. The real pressure comes from scaling. If annual demand rises to 10 million chips, monthly demand is about 833,000 chips, and required wafers are about 26,000 per month—already exceeding the 20,000 WPM line assumed earlier.

    Why are AI chip lead times still long? There are three main reasons. First, advanced-process capacity is allocated among multiple customers, and leading customers order far more than a single customer. Second, advanced packaging capacity is limited; 2.5D or 3D packaging capacity is often tighter than wafer capacity. Third, supply of high-bandwidth memory and other supporting components may also become a bottleneck. Lead-time analysis therefore needs to cover the entire supply chain, not just the wafer segment.

    To secure capacity, the industry commonly uses capacity reservation agreements. Such agreements usually require long-term contracts, prepayments, or minimum purchase commitments. They can improve scheduling priority, but they do not equal guaranteed delivery. Actual priority depends on contract terms and the foundry’s scheduling arrangements. Before signing, check penalty clauses, capacity adjustment conditions, and delivery responsibilities; do not interpret such agreements as a universal guarantee.

    6. Common Misconceptions and Checklist

    Misconception 1: Looking only at theoretical layout counts and ignoring yield. Many teams estimate under ideal conditions early on, and mass-production yields then fall far below expectations. Use low, medium, and high yield cases for sensitivity analysis, and keep some capacity redundancy.

    Misconception 2: Confusing wafer capacity with chip capacity. Wafer capacity is measured in wafers, while chip capacity is measured in dies. The two are separated by N_eff. Without yield data, capacity planning cannot be completed.

    Misconception 3: Mixing die yield and line yield. Die yield reflects only the pass rate of dies on a single wafer, while line yield also includes packaging and testing. Calculations should clearly state which basis is being used and document it.

    Misconception 4: Reusing old parameters when migrating processes. Moving from 7 nm to 5 nm may reduce the area of a chip with the same function, but initial yield is often lower. The DPW increase from area shrinkage may be offset by lower yield, so the estimate must be redone.

    Misconception 5: Ignoring packaging and test capacity. Back-end steps must be included in planning; otherwise, wafers may be produced but cannot be scheduled for packaging.

    The checklist can be summarized as follows: confirm wafer size and die area; calculate DPW using the approximation formula or layout tools; clarify the yield basis and perform sensitivity analysis; distinguish nominal capacity, effective capacity, and ramp-up; include test, packaging loss, and safety stock; and verify back-end packaging and supporting component capacity.

    7. Conclusion

    Foundry capacity estimation is a calculation process that connects market demand, geometric layout, yield statistics, and supply chain arrangements. Each step can be checked numerically, and assumptions are easily exposed. For developers and technical managers, mastering this method helps assess the feasibility of mass production and view shortage and price-increase news more rationally. As AI models continue to scale, demand for advanced-process and packaging capacity may keep growing. Building a relatively accurate estimation method helps reduce blindness in supply chain decisions, but the results still need to be repeatedly validated against actual foundry data.