Why Fabrinet fell 19%: Nvidia's share of it halved

Fabrinet builds optical components for other companies. You will not have used one of its products directly, but the data centres training large models are full of them, and the company’s fiscal 2026 was the best year it has had.

Revenue rose 35.7% to $4,641.1 million. Net income rose 42.3% to $473.0 million. The June quarter set a record and came in above the company’s own guidance range. Guidance for the next quarter was set higher again.

The stock fell 19.38% the next day, and 24.06% by the session after that.

That combination is worth taking apart, and the answer is not in the press release. It is in the annual report filed the following morning, in two places: who the revenue came from, and how much cash it produced.

The year, as reported

The fourth quarter alone was $1,315.8 million against $909.7 million, up 44.6%.

Growth was broad across what the company sells:

And the guidance was not a walk-back. For the September quarter Fabrinet guided revenue to $1.375–1.425 billion against the $1.3158 billion it had just reported.

Guidance midpoint        $1,400.0M
Q4 FY2026 actual         $1,315.8M
───────────
Implied sequential change    +6.4%

So: record year, record quarter, beat, raise. This is the same shape as Sandisk, which guided 17.7% above the quarter it had just reported and fell anyway. As there, the interesting part is what the headline does not carry.

The customer table

The 10-K discloses every customer at 10% or more of revenue. Here is that table, three years deep.

Two years ago Fabrinet was, to a first approximation, an NVIDIA supplier with other customers attached. It is not that any more.

NVIDIA’s own commitments have been moving in the opposite direction — outward. In August 2026 it guaranteed up to $105 billion of data-centre leases where OpenAI is the tenant, which is a supplier underwriting its customer rather than a supplier being depended on.

A falling share of a fast-growing number can still be a rising number. Here it is not.

NVIDIA   FY2024   35.1% × $2,883.0M  ≈  $1,012M
FY2026   16.3% × $4,641.1M  ≈    $757M
─────────
−25.2%
Cisco    FY2024   13.4% × $2,883.0M  ≈    $386M
FY2026   19.9% × $4,641.1M  ≈    $924M
─────────
+139.1%

Fabrinet’s largest customer of two years ago bought roughly a quarter less from it in dollars, during a period when the company as a whole grew 61%. Cisco more than doubled, and Nokia and Amazon arrived above 10% for the first time.

The receivables say the same thing

Customer concentration in revenue is an annual average, so it can lag. The balance sheet as of the year-end date is a snapshot, and it points the same way.

NVIDIA was a quarter of what Fabrinet was owed a year ago. At this year-end it does not clear the 10% disclosure threshold at all.

None of this is a problem the company hid. It is in the filing, in a table, in both places you would look. It is simply not in the earnings release, which is the document most coverage is built from.

The cash

The second thing the annual report shows is that a 42% increase in net income produced almost no free cash.

Operating cash flow fell 21.8% in a year when net income rose 42.3%. Capital expenditure doubled. What is left is $4.2 million, down 98.0%.

The June quarter on its own was worse:

Operating cash flow, Q4 FY2026     $55.0M
Capital expenditure, Q4 FY2026    −$91.9M
─────────
Free cash flow, Q4 FY2026         −$36.9M   (Q4 FY2025: +$4.7M)

This is the same gap that monday.com showed at a smaller scale — revenue up, operating cash flow down, one working-capital line explaining most of it. Here the line is inventory.

The gap itself is common. We screened 3,631 SEC filers and found 45 large companies where profit rose and free cash flow fell in the same year — and a third of them still converted more than 100% of net income into cash, which is why the ratio matters more than the direction.

Inventory, and the part that argues the other way

Inventories rose 75.8%, from $581.0 million to $1,021.2 million, against revenue growth of 35.7%. Trade receivables rose 34.1%. Trade payables rose too, which softens the cash effect but does not remove it.

A reader who stops there concludes that product is piling up unsold. The composition says otherwise, and it is worth showing because it cuts against the simple reading:

Finished goods fell. The build is at the front of the line — raw materials and work in progress — which is what buying ahead of production looks like, not what unsold output looks like. The company says so directly, attributing the increase to “an increase in inventories to support higher demand in the next quarter.”

That is a reasonable explanation and it is consistent with guiding the next quarter up 6.4%. It is also unverifiable from outside until the next quarter arrives. Both of those are true at once, and the September quarter is where they get settled.

Two one-off items that nearly cancel

GAAP net income of $473.0 million contains two large items pulling in opposite directions, and they almost exactly offset.

Gain on non-marketable equity securities   +$56.7M
Tax provision related to Pillar Two        −$57.4M
─────────
Net effect on GAAP net income               −$0.7M

The net effect on the headline is negligible, which is exactly why it is easy to miss that the number is built from two unrelated items of roughly $57 million each. Neither is operating. Neither recurs in the same form.

The margin that did not move

One more figure sits underneath the growth. Gross margin was 12.0% of revenue in fiscal 2026, against 12.1% in fiscal 2025. Cost of revenues was 88.0% of revenue, against 87.9%.

Revenue grew 35.7% and the gross margin did not move. For a contract manufacturer that is not surprising — the model is volume at a thin spread, and the company describes itself as providing manufacturing services rather than products of its own. But it does mean that scale here does not convert into margin, and a reader assuming operating leverage in an AI-linked supplier will not find it in this filing.

What would make this wrong

• The filing does not say why the stock fell, and neither do we. This article sets out what changed inside the numbers. Attributing the move to any one of those changes would require evidence we do not have.
• The dollar figures per customer are our calculation, not disclosure. The 10-K gives percentages rounded to one decimal place. Applying them to reported revenue produces approximate dollars, and small rounding differences move the result by several million.
• A customer falling below 10% is not the same as a customer leaving. Below the threshold the company discloses nothing, so NVIDIA’s fiscal 2026 receivable balance could sit anywhere under 10%. The direction is established; the level is not.
• Concentration falling is not automatically bad. A supplier whose largest customer drops from 35% to 16% while total revenue grows 61% has become less dependent, which is usually described as a strength. The observation here is narrower: the dollars from that customer went down.
• The inventory build has a stated, plausible explanation and the composition supports it. If the September quarter lands in the guided range, the working-capital swing was funding growth. If it does not, the same numbers read differently.
• These are audited annual figures, but the fourth-quarter columns in the earnings release are labelled unaudited, and prior-year figures are as originally filed.

Check it yourself

Both documents are linked below and were filed a day apart. The customer concentration table and the accounts receivable concentration table are in the notes to the financial statements in the 10-K; the inventory composition is in Note 10. The cash flow statement and the free cash flow reconciliation are in Exhibit 99.1 of the 8-K.

The one calculation we added is the dollar conversion of the disclosed customer percentages, and the arithmetic is in the article so you can redo it.

If a figure here does not match a filing, tell us and it will be corrected on the article and on the corrections log, with the date.