The $19 Million Tapeout: What a Fabless Miss Costs

Archi4 min read

semiconductorsfinancestartupsmanufacturing

Contents
  1. The invoice arrives before the silicon
  2. Where the money went
  3. The second spin is the expensive decision
  4. What to change in the budget
  5. Cash discipline beats design heroics

The invoice arrives before the silicon#

A recent fabless program missed first silicon. The debug log got attention. The finance team got the invoice. That order matters because a chip program can survive an engineering miss if the cash plan has room for a second spin. It can fail even with a working part if the budget treats tapeout as the finish line.

The case here is a specialty process, not a leading-edge phone chip. The first spin cost about $15 million before a customer saw a working part. Masks were $7.2 million. Wafer starts were $2.4 million. Package tooling and test development were $1.5 million. EDA seats, IP licenses and verification hardware added $1.8 million. The rest went to assembly NRE, expedite fees and the people costs that payroll sees every month. The second spin added $4.2 million. The final bill landed near $19 million, against a plan that had reserved $11 million.

Where the money went#

Mask costs are visible and easy to argue about. They are not the whole fixed cost. A new package can require a substrate tool, a socket, a handler kit and a reliability qualification. Test development can consume several months of engineering time and a load board that no one budgets until the first probe card fails. IP licenses often start at tapeout. A license for a PHY or a controller can invoice whether or not the chip boots.

The wafer commitment is the quiet risk. Foundries sell wafer starts in lots. A hot lot costs more. A cancelled lot may still cost something. At mature nodes the numbers are smaller, but the shape is the same: money leaves before yield is known. A team that orders a full production lot to 'save time' turns an engineering problem into an inventory problem.

The second spin is the expensive decision#

First silicon failures are normal. The money question is whether the second spin is planned or improvised. In this case, the team had no reserve and no customer milestone tied to bring-up. The board approved the respin in a single meeting because the alternative was writing off the first $15 million. That is a rational choice and a dangerous one. It keeps the program alive while consuming the cash that should fund qualification and production.

A respin budget should be set before the first mask order. For a complex part, 30 to 50 percent of the first silicon cost is a reasonable planning range. That money does not sit idle. It is a reserve, a line the team can draw on without renegotiating the year. If the part works, the reserve pays for yield learning and customer samples. If it fails, the reserve pays for the fix.

What to change in the budget#

Start with a stage gate that has a dollar value. A design review that ends with 'looks good' is not a gate. A gate says the team can spend the next $2 million only after the test plan, package choice and IP license list are frozen. That sounds bureaucratic until a missed license shows up on a purchase order.

Price the shuttle. A multi-project wafer run can cost a few hundred thousand dollars and delay the full mask set. It is often the right trade for a small team. The shuttle does not remove the need for a respin reserve, but it can turn a $7 million mistake into a $300,000 lesson.

Tie customer money to real milestones. NRE should cover engineering work, with the design win treated as a separate event. Samples are not revenue. Qualification is not production. A purchase order that starts at mass production may arrive 18 months after the first wafer. The cash plan has to survive that gap.

Track the cost of delay in the same spreadsheet as the mask bill. Every month of slip has payroll, EDA seats, lab time and the opportunity cost of the team. Those numbers are less visible than a mask invoice, and they compound.

Cash discipline beats design heroics#

The failure here was not a single bad block. It was a budget that stopped at tapeout. The team knew the mask cost. It did not know the cost of a second mask set, a new package, a test program rewrite and six more months of payroll. The design engineers did their jobs. The finance plan did not do its job.

A chip program needs a cash owner who sits in the same reviews as the architecture lead. That person should ask three questions at every gate: what has been committed, what can be cancelled, and what happens if the next silicon does not boot. The answers belong in writing before the purchase order goes out. The alternative is a postmortem that reads like an invoice.

More articles