@perceptron
But to give you a answer which is more accurate… and maybe you can tell us your results later? We would be thankful
Manufacturing “variable costs” for a silicon chip (e.g., BrainChip’s Akida AKD2500 in a fabless model) are not a single line item — you have to separate true per-unit variable costs from volume-driven/step costs and non-recurring costs that still have to be amortized if you want real unit economics.
1) True variable manufacturing costs (scale with each unit shipped)
A) Foundry / Wafer cost
- Price per wafer (depends on node, wafer size, foundry, volume tiers, contract terms)
- Wafer probe / E-test (often priced per wafer or per die)
To convert wafer cost into cost per good die, you need:
- Wafer price
- Gross dies per wafer (depends on die area + wafer geometry)
- Wafer yield (good dies / gross dies)
Rule of thumb formula:
Cost_per_good_die ≈ Wafer_cost / (Gross_dies_per_wafer × Yield)
B) Assembly / Packaging (OSAT)
- Package type (QFN/BGA/WLCSP, etc.), pin/ball count, substrate/leadframe, thermal/reliability requirements
- Typically quoted as cost per packaged unit, but strongly volume-tiered
C) Final test / QA
- ATE test time and coverage (test time is often the big driver)
- Burn-in / stress test if required
- Sampling, screening, and quality checks (some per-unit, some batch/step)
D) Backend handling & losses
- Marking (laser), tape-and-reel, handling
- Scrap / rework / backend yield loss (assembly yield is separate from wafer yield)
E) Logistics to make a deliverable unit
- Freight between foundry → OSAT → distribution
- Duties, insurance, warehousing/handling (often ignored in “manufacturing” discussions, but it’s still per-unit/volume-linked if you’re doing landed unit cost)
2) Other costs “associated with producing chips” (not clean per-unit, but volume matters)
A) NRE (non-recurring engineering) / mask costs
Not variable per chip, but you must amortize it to understand economics:
- Mask set, tapeout, sign-off, engineering runs
Amortized NRE per unit:
Amortized_NRE_per_unit = Total_NRE / Lifetime_units
B) EDA + IP licensing
- Tooling and IP block licenses (mostly fixed/periodic)
- Sometimes running royalties depending on the deal structure
C) Test program development, validation, customer enablement
- Test program creation, characterization, reliability quals
- Reference designs, support, and variant management (grows with customers/SKUs)
D) “Step costs” at higher volume
- Extra test capacity, QA/reliability lots, supply-chain headcount
These don’t scale perfectly linearly; they jump at volume thresholds.
3) Fabless reality check (relevant for Akida AKD2500)
BrainChip (fabless) typically pays:
• Foundry wafer costs
- OSAT assembly + test
- Logistics/distribution
- Plus NRE/EDA/IP and ongoing engineering/support
They do not carry the capex/opex of owning a wafer fab, but that doesn’t make “manufacturing cost” simple — it just shifts the cost structure into supplier pricing + yields + test time + volume tiers.
If you want to actually calculate a per-chip cost
You need (at minimum):
- Wafer cost (with volume tier)
- Die area → gross dies/wafer
- Wafer yield (%)
- Package/assembly cost per unit
- Final test cost per unit (ATE time, multisite, coverage)
- Backend yield (%)
- Freight/duty/handling per unit (or per batch)
- Total NRE + expected lifetime units (for amortization)
Without those inputs (especially die size, yields, package type, test time, and volume pricing), any single “cost per chip” number is just guessing.
Hope this helps!