AI Accelerator Card Wafer Capacity Estimation: Definitions, Formulas, and Bottlenecks

    This article explains how to estimate AI accelerator card supply from wafer capacity without conflating design capacity, effective wafer starts, good-die output, and finished-card output. It provides a reusable calculation framework and pseudocode, explains why advanced-node wafer capacity cannot be treated as card capacity, and examines non-wafer bottlenecks such as advanced packaging, high-bandwidth memory, substrates, and test. It also covers demand allocation, sensitivity analysis, common pitfalls, and practical steps for building an auditable capacity model.

    From “Card Shortage” to “Wafer Gap”: Align the Definitions First

    When discussing AI accelerator card supply, the most common mistake is to treat “fab design capacity,” “advanced-node wafer starts,” “good-die output,” and “packaged finished-card output” as if they were the same thing. In reality, these four are separated by yield, scribe lines, edge loss, packaging and test, memory and substrate supply, and other steps. If any link is ignored, the estimate can be off by orders of magnitude.

    Therefore, the first step in capacity estimation is not to apply a formula, but to align definitions:

    • Design capacity: the wafer starts a fab can process under ideal conditions, usually expressed as wafers per month;
    • Effective capacity: actual wafer starts after accounting for maintenance, ramp, bottleneck processes, and product mix;
    • Die output: the number of good dies that can be cut from each wafer, affected by die area, scribe lines, edge loss, and yield;
    • Finished output: the final usable cards after good dies go through packaging, test, and module assembly.

    Only by converting “wafer starts” layer by layer into “finished cards” can you answer “how much advanced-node capacity is needed” and “why can’t we get supply.”

    Core Formula: From Wafer to Finished Card

    A reusable estimation framework is as follows:

    1. usable area per wafer = wafer area × usable area factor
    2. gross die = floor(usable area per wafer / die area)
    3. good die = gross die × wafer yield × edge loss factor
    4. monthly finished cards = monthly wafer starts × good die × package yield × test yield × module yield

    Here, die area should include scribe lines and guard bands; wafer yield is usually related to die area and process maturity and cannot simply be assigned a fixed value; the edge loss factor describes the impact of wafer edges that are not fully usable; package, test, and module yields correspond to losses in back-end steps.

    A pseudocode example is shown below:

    1. function estimate_monthly_cards(
    2. wafers_per_month,
    3. wafer_area,
    4. usable_area_ratio,
    5. die_area,
    6. die_yield,
    7. edge_loss_factor,
    8. package_yield,
    9. test_yield,
    10. module_yield
    11. ):
    12. usable_area = wafer_area * usable_area_ratio
    13. gross_die = floor(usable_area / die_area)
    14. good_die = gross_die * die_yield * edge_loss_factor
    15. cards = wafers_per_month * good_die
    16. cards = cards * package_yield * test_yield * module_yield
    17. return cards

    The value of this formula is that it is auditable: every parameter can be mapped to a specific business assumption. The example parameters are only used to demonstrate the calculation path and do not represent the real data of any vendor. In actual modeling, the assumptions should be calibrated jointly by process, packaging, test, procurement, and product teams.

    Why Advanced-Node Capacity Is Not Card Capacity

    The core output metric of an advanced-node fab is wafer starts, not the finished quantity of any one type of chip. The same line often carries multiple products, and capacity allocation is affected by long-term agreements, product priority, yield ramp, and process qualification progress. Even if demand for a certain AI accelerator card is strong, it does not mean the fab can instantly switch all effective capacity to it.

    In addition, advanced-node capacity has a clear ramp effect. New processes and new chip versions have lower yield in the early introduction stage and gradually improve as the process stabilizes and the design is optimized. If mature yield is used directly in the estimate, short-term supply will be overstated; if only early-stage yield is used, mid-term output will be understated. A more reasonable approach is to estimate a range and provide assumptions for the yield ramp curve.

    Cycle time also cannot be ignored. From wafer start to wafer out, from die to package, and from test to finished-goods inventory, every step has queue and turnover time. When demand rises suddenly, capacity release lags, and “monthly wafer starts” cannot be directly equated with “cards deliverable in the same month.”

    Non-Wafer Bottlenecks: The Other Half of the Barrel

    AI accelerator card supply does not depend only on advanced-node wafers. High-bandwidth memory (HBM), advanced packaging, substrates, power management, thermal modules, test equipment, and test time can all become bottlenecks. Advanced packaging in particular has a capacity buildout cycle and yield ramp pace that are not fully synchronized with wafer manufacturing.

    This means that even if wafer starts are sufficient, finished-card output still cannot rise if packaging or HBM supply is constrained. Conversely, if back-end capacity is ample but wafers are insufficient, the bottleneck returns to the front end. Capacity estimation should therefore use a “barrel model”: estimate available output separately for wafers, packaging, memory, substrates, and test, take the minimum as the system supply ceiling, and note the headroom for marginal improvement in each link.

    Demand Side and Allocation: No Supply Does Not Mean Zero Capacity

    When cloud service providers or enterprise customers feel that “cards are scarce,” the cause may be multifold: procurement plans that do not match capacity allocation, product qualification cycles, long-term agreement reservations, inventory strategy, product transition, budget approval timing, and so on. Explaining the situation only through wafer capacity can easily misclassify a commercial allocation problem as a manufacturing problem.

    In practice, demand should be split into levels such as confirmed demand, intended demand, framework agreement demand, and forecast demand, and then matched against supply-side “allocated capacity, obtainable capacity, and ramping capacity.” Only in this way can short-term gaps, structural gaps, and differences in communication be distinguished.

    Sensitivity Analysis and Common Pitfalls

    A capacity model should include sensitivity analysis rather than only a point estimate. Key variables to observe include changes in die area, wafer yield, package yield, and monthly wafer starts, and their impact on finished-card output. Usually, die area and yield have the most direct impact on die output, while back-end yield determines whether dies can be converted into deliverable finished products.

    Common pitfalls include:

    1. Using design capacity instead of effective capacity, ignoring maintenance, ramp, and product mix;
    2. Using gross die instead of good die, ignoring yield and edge loss;
    3. Calculating only wafers and not packaging and test, ignoring the barrel effect;
    4. Ignoring cycle time and treating wafer starts as same-month deliveries;
    5. Inferring capacity from revenue or unit price, which introduces price, inventory, and product-mix bias;
    6. Treating short-term allocation outcomes as long-term capacity conclusions.

    Implementation Recommendations

    It is advisable to build a reusable capacity estimation table that expands layer by layer from “wafer—die—package—test—module—finished card,” and to mark the source, update frequency, and owner for each parameter. For external communication, provide base, conservative, and optimistic scenarios at the same time to avoid a single-point number being misread. For internal decisions, prioritize identifying system bottlenecks rather than arguing only about wafer-level definitions.

    In short, AI accelerator card capacity estimation is a multi-step, multi-assumption, auditable engineering problem. Only by acknowledging definition differences, building a layered model, and continuously calibrating assumptions against actual output can questions such as “how much 7nm capacity is needed” move from slogans to verifiable engineering conclusions.