AI Chip Supply-Demand Gap Calculation: From Wafer Starts to Finished Accelerator Cards
This article explains how to estimate the supply-demand gap for AI chips by tracing capacity from wafer starts to finished accelerator cards. It covers die-per-wafer math, Poisson yield, backend and packaging losses, a reproducible Python script, and how to compare deliverable card output with demand over the same time window. It also outlines what procurement and architecture teams should verify before relying on nominal capacity.
1. More Wafers Do Not Necessarily Mean More Cards
Over the past year, many cloud service providers and algorithm teams have faced the same problem: chip shipments are growing, but purchase orders are still delayed. Outside observers often attribute this to fabs not expanding capacity or capacity being taken by other products. Those explanations have some merit, but they do not show how wafer capacity translates into card output.
There are three common estimation mistakes. First, treating wafer count directly as chip count. How many dies can be cut from one wafer depends on die area and edge waste; how many usable dies result also depends on yield. Second, ignoring process ramp-up. A new node typically takes several quarters to move from trial production to stable yield, and during ramp-up, even high wafer starts can produce relatively low qualified output. Third, looking only at front-end manufacturing. AI chips are usually packaged together with high-bandwidth memory (HBM), and packaging capacity is often tighter than wafer capacity.
Below is an estimation method that runs from wafer to finished card and can be calculated by hand or reproduced in code. All numbers in this article are illustrative assumptions. In practice, replace them with data for the specific vendor and production line.
2. Step 1: Calculate Dies per Wafer
Take a 300 mm (12-inch) wafer as an example. Its area is:
Wafer area = π × (150 mm)² ≈ 70,686 mm²
Assume an AI accelerator chip uses a 7nm-class process and has a die area of 400 mm². Dividing wafer area directly by die area would ignore two losses: the wafer edge is round, so square dies near the edge cannot be arranged completely; and there are scribe lines between dies. A commonly used approximation is:
Theoretical dies ≈ π × (d/2)² / A − π × d / √(2A)
Here, d is the wafer diameter (300 mm) and A is the die area (400 mm²). The first term is wafer area divided by die area, and the second term approximately subtracts the unusable edge region. Substituting the values gives:
70,686 / 400 − 942.5 / 28.28 ≈ 176.7 − 33.3 ≈ 143 dies
This is the theoretical cut count before yield is considered. Different vendors use different layouts and scribe-line widths, so actual layout efficiency will vary.
3. Step 2: Convert to Good Dies Using Yield
Yield is the proportion of good dies among the theoretical cut count. The industry often uses the Poisson model as an approximation:
Y ≈ e^(−D0 × A)
Here, D0 is defect density and A is die area. The units must be consistent. In process data, defect density is usually expressed as defects per square centimeter, written as defects/cm². If die area is measured in mm², D0 must be converted to defects per square millimeter. Because 1 cm² equals 100 mm², 0.1/cm² equals 0.001/mm². Both forms give the same result when substituted consistently.
Take D0 = 0.1/cm² and die area A = 400 mm² = 4 cm². Then:
Y ≈ e^(−0.1 × 4) = e^(−0.4) ≈ 0.670, or about 67%
The number of good dies per wafer is:
143 × 0.670 ≈ 96 dies
The impact of die area is nonlinear. If area doubles to 800 mm², the theoretical cut count falls to about 65, yield falls to about 45%, and good dies per wafer fall to about 29. Doubling the area reduces the good die count by about 70%. This is one reason chiplet designs have gained attention: splitting a large chip into multiple smaller chiplets can significantly improve yield.
4. Step 3: From Good Dies to Finished Cards
Good dies cannot be delivered directly. The backend flow usually includes the following stages:
- Chip probing (CP) test: Screen out dies with normal functionality.
- Advanced packaging: Integrate compute dies with high-bandwidth memory and other components on the same substrate. 2.5D packaging commonly uses a silicon interposer, whose capacity is limited by equipment and process constraints and is a scarce step worldwide.
- Card-level testing and flashing: Includes full-load testing, firmware writing, and retesting.
- Assembly and logistics.
Each stage has its own losses. For simplicity, this article assumes a combined backend yield of 85% and that each card uses only one compute die.
The table below lists all assumptions and results for this example.
| Parameter | Value | Notes |
|---|---|---|
| Wafer diameter | 300 mm | Mainstream advanced process |
| Die area | 400 mm² | Illustrative assumption |
| Defect density D0 | 0.1/cm² | Illustrative assumption, not public data |
| Monthly wafer starts | 3,000 wafers | Allocation to this product, not the line’s nominal capacity |
| Combined backend yield | 85% | Illustrative assumption |
With these parameters, theoretical dies per wafer are about 143, die yield is about 67%, good dies per wafer are about 96, monthly good dies are about 288,000, and monthly deliverable finished cards are about 245,000.
Note that the 3,000 wafers in the table are the allocation actually available to this product. The same line’s nominal capacity may be higher, but only the portion allocated to AI chips determines supply.
5. Reproducing the Calculation in Code
The Python script below connects the steps above. Change the parameters and rerun it directly.
import mathWAFER_DIAMETER_MM = 300DIE_AREA_MM2 = 400D0_PER_CM2 = 0.10 # Defect density in defects/cm²WAFERS_PER_MONTH = 3000 # Monthly wafer starts allocated to this productBACKEND_YIELD = 0.85 # Combined yield for packaging, test, and assemblydef gross_dies_per_wafer(wafer_d_mm: float, die_area_mm2: float) -> float:"""Theoretical die count: wafer area divided by die area, minus edge loss."""wafer_area = math.pi * (wafer_d_mm / 2) ** 2edge_loss = math.pi * wafer_d_mm / math.sqrt(2 * die_area_mm2)return wafer_area / die_area_mm2 - edge_lossdef poisson_yield(d0_per_cm2: float, die_area_mm2: float) -> float:"""Poisson yield model. Area is converted to cm² internally."""die_area_cm2 = die_area_mm2 / 100return math.exp(-d0_per_cm2 * die_area_cm2)gross = gross_dies_per_wafer(WAFER_DIAMETER_MM, DIE_AREA_MM2)yield_rate = poisson_yield(D0_PER_CM2, DIE_AREA_MM2)good_per_wafer = gross * yield_rategood_per_month = good_per_wafer * WAFERS_PER_MONTHcards_per_month = good_per_month * BACKEND_YIELDprint(f"Theoretical dies/wafer: {gross:.1f}")print(f"Die yield: {yield_rate:.1%}")print(f"Good dies/wafer: {good_per_wafer:.1f}")print(f"Monthly good dies: {good_per_month:,.0f}")print(f"Monthly deliverable cards: {cards_per_month:,.0f}")
Run result:
Theoretical dies/wafer: 143.4Die yield: 67.0%Good dies/wafer: 96.1Monthly good dies: 288,358Monthly deliverable cards: 245,104
6. How to Calculate the Supply-Demand Gap
To judge the supply-demand gap, compare demand and supply within the same time window. Deliverable card count is the minimum of the following three values:
- Good dies allocated to the product multiplied by backend yield;
- Packaging capacity;
- Test capacity.
The gap equals demand in the same time window minus deliverable card count.
The choice of time window matters. Wafer start to shipment usually takes months, so when judging supply for a given quarter, use wafer-start data from several months earlier, not the current month’s wafer starts. If current-month wafer starts rise sharply, actual shipments may not appear until later quarters.
In addition, packaging capacity and wafer capacity often do not grow in sync. Even with sufficient wafer supply, packaging can become the actual ceiling.
7. Frequently Asked Questions
How should yield units be converted?
1 cm² equals 100 mm², so 0.1/cm² equals 0.001/mm². The area and defect density in the formula must use the same unit system.
How can CoWoS-class packaging capacity be estimated?
First determine the packaging plant’s monthly capacity ceiling, measured in package bodies or interposer wafers. Then compare it with the deliverable die count and take the smaller value. Packaging yield must be counted separately. Packaging and wafer ceilings often do not move in sync; the actual bottleneck is the tighter of the two.
How is the AI chip supply-demand gap calculated?
Compare demand with deliverable card count over the same time window. Deliverable card count is the minimum of good dies multiplied by backend yield, packaging capacity, and test capacity.
8. Procurement and Architecture Considerations
Procurement lead time depends on manufacturing cycle time, packaging queues, and process ramp-up progress, and is often measured in months. Before purchasing, teams should confirm die size, yield level, and packaging allocation with suppliers rather than relying only on nominal capacity.
There are usually three ways to obtain compute capacity, each with different costs and risks:
| Approach | Suitable Case | Main Costs | Main Risks |
|---|---|---|---|
| Self-built cluster | Long-term stable demand, large scale, operations team available | Upfront capital expenditure, data center, power | Hardware delivery delays, equipment depreciation, operations labor |
| Public cloud GPU | Demand fluctuates, rapid scaling needed | Usage-based or annual billing | Price volatility, vendor lock-in, quota limits |
| Hybrid deployment | Base load self-built, peak load in cloud | Both cost types, unified scheduling required | Higher architecture complexity |
Whichever approach is chosen, two architecture measures can help. First, keep a hardware abstraction layer so inference services can adapt to multiple accelerators and avoid a single supply source. Second, use containerized deployment and resource monitoring to improve utilization per card. These measures are independent of where the cards come from.
9. Summary
Chip capacity may seem distant from application teams, but it determines procurement timing and delivery schedules. By understanding the calculation chain from wafer to finished card, teams can distinguish real supply constraints from market sentiment and plan budgets and resources accordingly.
