Amkor’s computing revenue rose 37% and is guided up more than 50% in the fourth quarter, but gross margin missed the company’s range, and only about 3% of a $12bn Arizona campus had been spent by June. Free cash flow stays negative through 2030 on our numbers. On the company’s own targets to 2030, with the second Arizona phase filling from 2031, the shares are worth $26.17 to $49.49, most likely $35.45. At $37.60 the price is the plan, delivered: it needs a 10.26% discount rate instead of 10.5%, or revenue 2.8% higher than our path, and nothing for the faster ramps of our bull case. We rate the shares fairly valued.
The mix is doing its job; the margin guide still missed. Computing revenue was about $518m in the third quarter, up 37% on a year earlier and 26% of sales against 19%; communications fell to 37% of sales from 51%. Materials, the largest cost, fell to 50.7% of sales from 57.5%. Gross margin rose to 18.3%, but the company had guided 18.5–19.5%; management attributes about 0.6 points to currency and a July earthquake at its Japanese plant. Fourth-quarter guidance is 18.0–19.0%.
The bill comes before the revenue. The Arizona campus is now planned at about $12bn across two phases; $333.6m had been incurred by 30 June. Facilities alone take more than $2bn in 2027, Arizona spending peaks in 2028–29 and its anchor customer ramps from late 2028. In our base case capital spending from October 2026 to 2030 is $15.3bn and free cash flow over the period is negative $3.4bn. Without the outside money — customer advances, the 35% investment credit on phase 1 and the CHIPS award — the base case would be worth $26.48 instead of $35.45.
The price is the plan. Our base case follows the company’s own 2028 and 2030 targets and lets the second Arizona phase fill from 2031: $35.45 a share, band $26.17–$49.49, probability-weighted $38.91. At $37.60 the price needs a 10.26% discount rate, or revenue 2.8% higher than our path. Nothing is priced for a digestion year and nothing for the bull case’s faster ramp ($56.79); the case for owning the shares rests on the upside tail, and the case against on how long the cash takes to come back.
At $37.60 Amkor’s market value is about $9.30bn on 247.45m shares, and its enterprise value about $9.65bn once debt at face, leases, pensions and minority interests are added and cash is taken away. That is 1.3 times trailing revenue, 6.5 times trailing EBITDA and 15 times trailing earnings: cheap against the semiconductor index, and cheap for a reason, because the earnings are being reinvested. Our base case gives management its own Investor Day framework — revenue of $9.3bn in 2028 against a target of $9.0bn ± $0.5bn, $11.3bn in 2030 against “about $11bn and above”, gross margin of 17.5% in 2028 as Arizona starts depreciating and 21% by 2030 — and then lets the second Arizona phase and the larger Korean footprint fill from 2031, to $16.3bn in 2036 at a 21.5% gross margin. That path is worth $35.45 a share, 6% below the price.
Run the model in Section 02 backwards and the price needs any one of three things, each on its own with everything else held at the base case. A discount rate of 10.26% instead of 10.5%. Or revenue 2.8% higher than our path in every year from 2028: $9.6bn in 2028, $11.6bn in 2030 and $16.8bn in 2036. Or a gross margin 0.4 points higher every year, 21.9% in the long run. Put plainly, the market is paying for the company’s plan, roughly delivered: no premium for the bull case’s faster ramp, and no discount for a digestion year. The price sits inside our $26.17–$49.49 band.
The signature measure for Amkor is the build-out against the money that pays for it. The first bar is the Arizona campus and what has been spent on it; the second is the money from outside Amkor identified for it so far; the last two are Amkor’s own cash and debt. All four are drawn to one scale.
Campus cost and phase timing from the company’s announcements of 6 October 2025 and 8 September 2026 and the third-quarter call; $333.6m incurred from the Form 10-Q for the quarter to 30 June. The split of the $12bn between the phases is not disclosed: phase 1 (33,000 square metres of clean room) carries most of the site and infrastructure, and our estimate is about $5.5bn. The NVIDIA Arizona agreement is valued at $1.5bn and includes a prepayment whose size has not been disclosed; a second NVIDIA agreement, $1.3bn, is for Korea. The credit estimate assumes 75% of phase 1 cost qualifies; phase 2 starts construction after the deadline and is excluded in the base case. Cash excludes $201.7m of restricted cash. Identified outside money covers a little more than a quarter of the campus; the rest is operating cash flow and, possibly, new debt from 2028.
The fourth-quarter results in February bring the 2027 capital-spending guide, the single number that sets how deep the free-cash-flow trough goes; facilities alone are above $2bn before equipment. Our base case has $3.4bn in 2027. The second test is the cash itself: the size and timing of the NVIDIA prepayments, which our base case takes at the agreement values. At half those values the base case is worth $33.79.
The rate backdrop is not friendly to a long-duration build. The 10-year Treasury yield was 5.00% on 26 October, against about 4.6% in late July. Amkor’s five-year beta is about 2.2 and its thirty-day realised volatility about 60% annualised. We use a 10.5% discount rate in every case: Amkor is profitable, but it is a high-beta, cyclical manufacturer funding a construction programme larger than its annual revenue. Demand, concentration and execution risks are carried in the scenarios and their weights rather than added to the rate. The band runs the rate from 9.5% to 11.5%.
2026–36 revenue CAGR · 2036 operating margin · 2036 free cash flow · credits, grants and net customer advances, 2027–36 · terminal value is of enterprise value · shares, the 0% notes .
This is the model that produced every figure in this report, reduced to its nine moving parts. It is an unlevered discounted cash flow on a ten-year path, 2027 to 2036, with the fourth quarter of 2026 as a stub. Free cash flow to the firm is operating income (revenue times gross margin less operating expenses, with stock pay already inside them) after cash tax, plus depreciation, less capital spending, less working capital at 10% of each year’s added revenue, plus the outside money: investment credits, the CHIPS award and customer advance payments, less the advances later delivered as revenue without new cash. The valuation date is 27 October 2026; each year is discounted at its mid-point, and the Gordon terminal value at 3% on 2037 cash flow is discounted from the end of 2036. On base settings the terminal value is 92% of enterprise value, because free cash flow is negative through 2030. The scenario cards load whole published paths, and the sliders bend the loaded path rather than replace it, so each card reproduces its published value exactly.
Why these inputs. 2027 revenue of $8.4bn (assumed). Computing up about 35%, automotive about 12%, communications and consumer flat; the consensus before the results was $8.66bn, and there is no company outlook for 2027 yet. 2028 of $9.3bn (assumed). The Investor Day target is $9.0bn ± $0.5bn. 2030 of $11.3bn and 2036 of $16.3bn (assumed). The 2030 target is about $11bn and above; growth of 10% in 2030 and 9% in 2031 as phase 2 fills, fading to 3% by 2036. Gross margin of 21.5% in the long run (assumed). 18.0% in 2027, 17.5% in 2028 as Arizona and the Korean buildings start depreciating, 21.0% in 2030 against a target of about 22%; ASE Technology earns 27.3% in packaging and test. Capital spending at 100% of plan (assumed): $3.4bn, $4.0bn and $3.9bn in 2027–29, $3.0bn in 2030 and $2.1bn in 2031, then 12.5% of revenue falling to 11%. Outside funding at 100% of plan (assumed), detailed below. Discount rate 10.5% and terminal growth 3.0% (assumed), the same in every case.
What the controls hold fixed. Operating expenses fall from 6.9% of revenue in 2027 to 6.3%; 2026 is about 7.1%. The cash tax rate is 18% in 2027 and the company’s 20% planning rate after it. Depreciation rises from 10.5% of revenue in 2027 to 12.5% in 2029–30 as the new buildings open, then falls to 10.5%. The outside money runs on its own schedule: about $1.45bn of credit on phase 1 in 2028–31, received after the property is placed in service; the $407m CHIPS award in 2027–29; customer advances of $3.0bn received in 2027–28 (NVIDIA $1.5bn for Arizona and $1.3bn for Korea, about $0.2bn from others), each delivered back as revenue without cash over seven years. Net, that is $1.9bn over 2027–36 in the base case. The fair-value band in Section 03 is this model run across the discount rate and terminal growth, and every cell of that grid is reproducible from the two controls.
Derivations, base case. Fourth quarter of 2026: revenue $1,950m and gross margin 18.5%, the guidance midpoints; operating expenses about $140m; tax 17%; depreciation $195m; capital spending $1,010m, the company’s figure of about $2.5bn for the year less the $1,489m accrued in nine months; free cash flow an outflow of $632m, discounted 0.09 years. Shares 247.45m: 247.99m at 30 June less about 0.6m bought back in the third quarter (estimated; the third-quarter 10-Q had not been filed). The $1.15bn of 0% notes due 2031 convert at $106.37, and a capped call offsets dilution up to $139.50; they are counted as debt at face, and only value above the cap would add shares. Unvested awards are excluded because stock pay is charged as a cost. Net debt and debt-like items of $829m: cash and short-term investments of $2,376.2m, less debt at face of $2,495m, finance leases of about $120m, pension and severance of $73.9m, minority interests of $36.5m and about $480m of customer advances already received, which Amkor will repay in packaging rather than cash.
The number above is not a fact about Amkor Technology. It is what these assumptions produce, and the two that decide it are not in any filing: how fast the new buildings fill after 2028, and the gross margin they earn when they do. The rating is a statement about the band, not this cell: it changes to undervalued at a price below $26.17 and to overvalued above $49.49, and the price is 6% above our base-case value. Move the 2030 and 2036 revenue settings and the margin until they are yours, and use the number that falls out.
Each card loads its published case into the model above and prints its value: $18.01, $35.45, $56.79 and $85.25. All four use the same 10.5% discount rate and 3.0% terminal growth, so they differ only in how much Amkor sells, at what margin, how much it builds and how much of the build others pay for; demand, margin and execution risk sit in the cases and the weights, not in the rate. Weighting the four at 25 / 45 / 25 / 5 gives $38.91.
| Input | Bear (25%) | Base (45%) | Bull (25%) | Super-bull (5%) |
|---|---|---|---|---|
| 2027 revenue | $7.9bn | $8.4bn | $8.8bn | $9.2bn |
| 2028 revenue | $7.8bn | $9.3bn | $10.2bn | $11.2bn |
| 2030 revenue | $8.8bn | $11.3bn | $13.1bn | $15.3bn |
| 2033 revenue | $9.9bn | $14.5bn | $16.5bn | $20.4bn |
| 2036 revenue | $10.8bn | $16.3bn | $18.7bn | $23.6bn |
| Gross margin, 2027 / 2028 / 2030 | 16.5 / 15.0 / 17.0% | 18.0 / 17.5 / 21.0% | 18.5 / 18.5 / 22.5% | 19.0 / 19.5 / 24.0% |
| Long-run gross margin | 17.5% | 21.5% | 23.0% | 25.0% |
| Operating expenses, % of revenue | 7.2% → 7.0% | 6.9% → 6.3% | 6.7% → 6.0% | 6.6% → 5.7% |
| Capital spending, 2027 / 2028 / 2029, $bn | 3.0 / 2.4 / 1.8 | 3.4 / 4.0 / 3.9 | 3.6 / 4.3 / 4.2 | 3.9 / 4.8 / 4.8 |
| Capital spending from 2032, % of revenue | 11.5% → 11% | 12.5% → 11% | 12.5% → 11% | 13% → 11% |
| Investment credit | Phase 1, ≈$1.45bn | Phase 1, ≈$1.45bn | + phase 2, ≈$1.7bn | + phase 2, ≈$1.7bn |
| Common to all | Discount rate 10.5% and terminal growth 3.0%; cash tax 18% in 2027 and 20% after; working capital 10% of added revenue; CHIPS award $0.41bn in 2027–29; customer advances of $3.0bn received in 2027–28 and delivered over seven years; valued at 27 October 2026 from the 30 September 2026 balance sheet; stock pay inside operating costs | |||
| Value per share | $18.01 | $35.45 | $56.79 | $85.25 |
Company targets from the Investor Day presentation of 21 May 2026: 2028 revenue $9.0bn ± $0.5bn and gross margin 17.5% ± 1 point; 2030 revenue about $11bn and above and gross margin about 22% and above. The consensus for 2027 before the results was $8.66bn.
| $ millions, base case | Q4 2026 | 2027 | 2028 | 2029 | 2030 | 2031 | 2033 | 2036 |
|---|---|---|---|---|---|---|---|---|
| Revenue | 1,950 | 8,400 | 9,300 | 10,304 | 11,304 | 12,434 | 14,502 | 16,311 |
| Growth | — | 12% | 11% | 11% | 10% | 10% | 7% | 3% |
| Gross margin | 18.5% | 18.0% | 17.5% | 19.0% | 21.0% | 21.5% | 21.5% | 21.5% |
| Operating income | 221 | 932 | 995 | 1,278 | 1,650 | 1,890 | 2,204 | 2,479 |
| Less cash tax | (38) | (168) | (199) | (256) | (330) | (378) | (441) | (496) |
| Plus depreciation | 195 | 882 | 1,116 | 1,288 | 1,413 | 1,492 | 1,595 | 1,713 |
| Less capital spending | (1,010) | (3,400) | (4,000) | (3,900) | (3,000) | (2,100) | (1,740) | (1,794) |
| Less increase in working capital | 0 | (88) | (90) | (100) | (100) | (113) | (95) | (48) |
| Plus credits, grants and net advances | 0 | 2,236 | 806 | 248 | 31 | (179) | (429) | 0 |
| Free cash flow to the firm | (632) | 395 | (1,372) | (1,442) | (335) | 613 | 1,095 | 1,854 |
Sonde estimates. The fourth quarter of 2026 is the stub to 31 December; 2027 growth is on full-year 2026 revenue of $7.52bn. Free cash flow from the fourth quarter of 2026 to 2030 sums to negative $3.4bn. Credits and grants: the 35% investment credit on phase 1 and the CHIPS award; advances: receipts less revenue delivered from them.
| Base case | $m | Note |
|---|---|---|
| Present value of cash flows, Q4 2026–2036 | 788 | Mid-year convention, the fourth quarter as a stub |
| Present value of terminal value | 8,813 | 92% of enterprise value; 2037 free cash flow $1,826m at 10.5% less 3.0% |
| Enterprise value | 9,600 | |
| + Cash and short-term investments | 2,376 | 30 Sep 2026 |
| − Debt at face and finance leases | (2,615) | $2,495m and about $120m |
| − Pension and severance, minority interests | (110) | $73.9m and $36.5m |
| − Customer advances already received | (480) | Revenue Amkor will deliver without new cash; estimated |
| Equity value | 8,771 | |
| Shares, millions | 247.45 | The notes are debt below the $139.50 cap of the capped call |
| Value per share | $35.45 |
The grid holds the base case fixed and varies the discount rate against terminal growth. Its corners are the published fair-value band, $26.17–$49.49: a WACC of 9.5–11.5% by terminal growth of 2.5–3.5%.
| WACC ↓ / terminal growth → | 2.5% | 3.0% | 3.5% |
|---|---|---|---|
| 9.5% | $42.47 | $45.71 | $49.49 |
| 10.0% | $37.50 | $40.18 | $43.28 |
| 10.5% base | $33.20 | $35.45 | $38.01 |
| 11.0% | $29.46 | $31.35 | $33.49 |
| 11.5% | $26.17 | $27.77 | $29.58 |
The $37.60 price lies inside the grid, 6% above the centre cell: fairly valued. The rating would change to overvalued above $49.49 and to undervalued below $26.17. Run across the same grid, the probability-weighted value spans $29.71–$52.78. The second table holds the rates at the base case and moves one input at a time:
| Change to the base case, all else held | Value | Change |
|---|---|---|
| Revenue 10% higher from 2028 | $43.03 | +21% |
| Revenue 10% lower from 2028 | $27.86 | −21% |
| Gross margin 1 point higher every year | $40.47 | +14% |
| Gross margin 1 point lower every year | $30.42 | −14% |
| Long-run gross margin 20% instead of 21.5% | $29.43 | −17% |
| Capital spending 10% higher | $28.57 | −19% |
| Capital spending 10% lower | $42.32 | +19% |
| Customer advances at half the agreement values | $33.79 | −5% |
| Phase 2 qualifies for the 35% credit | $39.81 | +12% |
| No credits, grants or customer advances | $26.48 | −25% |
Modal base case $35.45, 6% below the price. Fair-value band $26.17–$49.49, the WACC 9.5–11.5% by terminal growth 2.5–3.5% grid, which anchors the rating; the price is inside it: fairly valued. Probability-weighted value $38.91 across the four cases at our own weights, 3% above the price. The skew is to the upside: the bear sits $17.44 below the base and the bull $21.35 above it, because in the bear case Amkor also slows the build and keeps more cash.
| Single input changed, all else at base | Price implies | Our base |
|---|---|---|
| Revenue from 2028, every year | 2.8% higher | The company’s targets, then phase 2 |
| 2030 revenue on that path | $11.6bn | $11.3bn |
| 2036 revenue on that path | $16.8bn | $16.3bn |
| Gross margin, every year | +0.4 points; 21.9% long-run | 21.5% long-run |
| Discount rate | 10.26% | 10.5% |
Each reading is a small move from our base case. The discount-rate reading is the most natural way to see the price: the market is applying almost exactly our rate to almost exactly the company’s plan.
| Scenario | Value | Against price | Weight | Contribution |
|---|---|---|---|---|
| Bear | $18.01 | −52% | 25% | $4.50 |
| Base | $35.45 | −6% | 45% | $15.95 |
| Bull | $56.79 | +51% | 25% | $14.20 |
| Super-bull | $85.25 | +127% | 5% | $4.26 |
| Probability-weighted | $38.91 | +3% | 100% | $38.91 |
We weight the bear at 25% because computing is now the growth engine and the main destination of the new capacity, so a pause in data-centre spending would hit revenue and the build at once, and because ten customers are two-thirds of sales. We weight it no higher because a dozen customers are paying for capacity in advance and the mix shift is already in the reported numbers. Weights are our judgement, not a model output.
| Risk | Priced in | Not also in |
|---|---|---|
| AI packaging demand pauses | The bear revenue path (2028 lower than 2027) and its 25% weight | The discount rate |
| Communications and Apple share | Bear revenue and margin; base communications slightly down in 2027 | The terminal growth rate |
| Arizona cost and ramp | Base gross margin dips to 17.5% in 2028; capital spending at the company’s framework | A separate overrun allowance |
| Capital-spending overrun | The capital-spending control and the 10% sensitivity | The base cash flows |
| Advance-payment terms | Advances as interest-free financing at the agreement values; a sensitivity at half | A price concession |
| Investment-credit eligibility | Phase 1 only in the base case; phase 2 in bull and super-bull | — |
| Customer concentration | The scenario weights | The discount rate |
| Dilution | Stock pay inside operating costs; the notes as debt below $139.50 | The share count (unvested awards are not added) |
| $ millions | Value | Note |
|---|---|---|
| Market value | 9,304 | 247.45m shares × $37.60: 247.99m at 30 June less third-quarter repurchases (estimated) |
| + Debt at face | 2,495 | Carrying value $2,467m; includes the $1.15bn of 0% notes due 2031, which convert at $106.37 |
| + Finance leases, pension and severance, minority interests | 230 | About $120m, $73.9m and $36.5m |
| − Cash and short-term investments | (2,376) | 30 Sep 2026; $201.7m of restricted cash excluded |
| Enterprise value | 9,653 | Customer advances already received (about $480m) are excluded here and deducted in the DCF |
Source: Q3 2026 results release; Form 10-Q for the quarter to 30 June 2026 for debt at face, finance leases and contract liabilities. Reference price the last Nasdaq pre-market trade, at 4:40 a.m. Eastern on 27 October 2026.
| Measure | AMKR | Basis |
|---|---|---|
| EV / trailing revenue | 1.29× | $7,461m, Q4 2025 to Q3 2026 |
| EV / 2027E revenue | 1.11× | $8.66bn consensus (11 analysts, before the results) |
| EV / trailing EBITDA | 6.5× | $1,489m |
| P/E, trailing | 15.2× | $2.48 |
| P/E, 2026E | 15.1× | $2.49: nine months plus the fourth-quarter guidance midpoint |
| P/E, 2027E | 13.4× | $2.81 consensus (9 analysts, before the results) |
| PEG, 2027 | 1.04 | 2027 consensus EPS growth 13% |
| Price / trailing free cash flow | n/m | Twelve-month free cash flow an outflow of about $8m |
| Price / book | 1.94× | Amkor equity $4,800m, $19.40 a share |
| Dividend yield | 0.89% | $0.08352 a quarter |
Consensus: Yahoo Finance, as at 26 October 2026, before the third-quarter results. Free cash flow is operating cash flow less payments for equipment plus proceeds from equipment sales (company definition), derived.
| Company | Mkt value | EV | EV / trailing revenue | EV / trailing EBITDA | P/E, trailing | Gross margin, latest quarter | Growth, latest quarter |
|---|---|---|---|---|---|---|---|
| Amkor Technology $37.60 | $9.3bn | $9.7bn | 1.3× | 6.5× | 15.2× | 18.3% | +0.2% |
| ASE Technology $40.10 per ADS | $88.0bn | $94.8bn | 4.2× | 19.3× | 45.7× | 21.0% 27.3% packaging and test | +26.7% |
| SOXX aggregate 26 Oct 2026 close | — | — | P/S 10.6× | — | 37.7× | — | — |
Amkor at the last Nasdaq pre-market trade, at 4:40 a.m. Eastern on 27 October 2026; ASE Technology at its last NYSE pre-market trade, at 4:23 a.m. Eastern. ASE: 2Q26 results release (30 July 2026) and the 4Q25 and 1Q26 releases; 4,386.7m basic weighted shares in 2Q26, two shares per ADS; debt NT$296.4bn, cash and current financial assets NT$107.4bn and minority interests NT$30.0bn at 30 June, converted at NT$31.85; trailing revenue (NT$711.2bn), EBITDA (NT$154.9bn) and net income (NT$60.8bn) converted at NT$31.59. Its third-quarter results were not yet published. SOXX: Sonde aggregate multiples built from the fund holdings at the 26 October 2026 close. JCET, Powertech and other Asian OSATs do not file with the SEC and are not included.
Cheap, fair or expensive? Cheap against peers, and the discount is mostly earned. Amkor trades at 1.3 times trailing revenue and 15 times trailing earnings, against 4.2 times and 46 times for ASE and 38 times earnings for the semiconductor index. ASE earns nine more points of gross margin in packaging and test, grew revenue 27% in its latest quarter against Amkor’s flat year, and is not funding a greenfield campus in a high-cost country. On next year’s consensus Amkor trades at 13 times earnings with a PEG close to 1. But its trailing free cash flow is roughly zero and will be negative through the Arizona peak, so a mid-teens P/E is a multiple on profit that shareholders will not receive as cash for several years. The peer comparison says what the market would pay for Amkor if it already had ASE’s margins and no build to fund; the DCF says what the build is worth. Because the discount to ASE reflects lower margins and the build-out, the DCF does not penalise those risks again through the discount rate.
| What would expand the multiple | What would compress it |
|---|---|
| A 2027 capital-spending guide that implies a shallower free-cash-flow trough | Another gross-margin miss as new space depreciates |
| Disclosure of the prepayment amounts and their timing | A weaker 2027 communications year |
| The bridge programme and further 2.5D wins | A pause in data-centre spending |
| Arizona phase 2 qualifying for the investment credit | New debt or equity, or TSMC keeping more AI packaging in-house |
The base case is close to the company’s own targets and inherits their optimism; the 2027 consensus ($8.66bn) and earnings predate the third-quarter results and the fourth-quarter guide. The split of the $12bn between the phases is not disclosed: our phase 1 estimate (about $5.5bn) drives the investment-credit figure, and the 75% qualifying share is assumed. The NVIDIA prepayments are taken at the agreement values ($2.8bn); at half, the base case is $33.79. The customer advances already on the balance sheet ($480m), the share count and finance leases are estimated from the second-quarter Form 10-Q and the third-quarter release, pending the third-quarter 10-Q. Fourth-quarter capital spending ($1.01bn) is derived from the full-year figure. Terminal value is 92% of enterprise value, because free cash flow is negative through 2030: small changes in long-run margin or growth move the value a great deal.
Amkor reported its third quarter, the three months to 30 September 2026, on 26 October after the close, and held its call at 5.00 p.m. Eastern. Kevin Engel, president and chief executive, and Megan Faust, chief financial officer, took questions. The shares closed at $36.81 before the release; the reference price in this report is $37.60, the last Nasdaq pre-market trade, at 4:40 a.m. Eastern on 27 October 2026. The Form 10-Q for the quarter had not been filed when this report was published.
| Q3 2026 | Value | Change | Why it matters |
|---|---|---|---|
| Net sales | $1,991m | +4.9% q/q; +0.2% y/y | Low half of the $1.95–2.05bn range |
| Computing revenue | ≈$518m | +25% q/q; +37% y/y | Guided at nearly +30% sequentially |
| Communications revenue | ≈$737m | (8)% q/q; ≈(27)% y/y | Down again in the fourth quarter |
| Automotive and industrial | ≈$438m | +7% q/q; +37% y/y | A record |
| Gross margin | 18.3% | +1.5 pts q/q; +4.0 pts y/y | Below the 18.5% bottom of the guided range |
| Materials, % of sales | 50.7% | from 57.5% a year earlier | The mix shift in one line |
| Operating margin | 11.3% | from 8.0% a year earlier | Operating expenses $139m, flat on the quarter |
| Diluted EPS | $0.76 | from $0.51 a year earlier | Inside the $0.72–0.82 range |
| EBITDA | $435m | from $400m in Q2 | 21.8% margin |
| Free cash flow | $(62)m | from $(199)m in Q2 | Nine months: $(333)m |
| Cash and short-term investments | $2,376m | $(136)m in the quarter | Against $2,467m of debt |
| Top ten customers | 66% of sales | from 73% a year earlier | Concentration falling, still high |
Sources: Q3 and Q2 2026 results releases. End-market dollars are derived from the company’s rounded percentages; second-quarter free cash flow is the six-month statement less the first quarter (derived).
One: computing kept compounding, a little slower than promised. Revenue rose 25% sequentially against a guided rise of nearly 30%. Management said nothing was short of supply: a high-density fan-out part passes through more than 300 process steps, and filling that line with work in progress took longer than planned. Computing is guided up by a low-double-digit percentage in the fourth quarter, more than 50% on a year earlier, with growth of more than 30% for 2026 and a similar or higher rate expected in 2027.
Two: margin rose 1.5 points but missed the range. Gross margin was 18.3% against 18.5–19.5% guided. The chief financial officer put about 0.6 points of drag on currency and on the July earthquake at the Kumamoto plant in Japan, which was back to full operation within the quarter. The fourth-quarter range of 18.0–19.0% implies only about 0.2 points of further gain at the midpoint, with mix helping and currency still hurting.
Three: communications falls again, in both phone ecosystems. High memory prices hit low-end Android hardest; component shortages also held back premium phones and wearables; and moving SiP production from Korea to Vietnam leaves a gap while lines are re-qualified. Management expects Apple’s new launch cadence to change the usual seasonal pattern and said it was too early to call 2027 for the segment.
Four: the spending steps up again. 2026 capital spending was narrowed to about $2.5bn, from $2.5–3.0bn, on the timing of construction payments. For 2027, facilities alone exceed $2bn: Arizona up to 15% of its $12bn cost, Korea $400–450m and Vietnam about $50m, with equipment on top. Arizona spending peaks in 2028–29, when phase 2 is built while phase 1 is equipped. Full 2027 guidance comes in February.
Five: customers are paying for capacity in advance. A dozen customers have signed advance-payment arrangements since 2025 and more are under discussion. Management said the agreements contain terms that protect Amkor, that the equipment can be redeployed across advanced formats, and that the balance sheet is solid through 2027; whether 2028 needs new debt is too early to say.
Fourth quarter: revenue $1.9–2.0bn, gross margin 18.0–19.0%, operating expenses about $140m, net income $160–185m and earnings per share $0.65–0.75; the consensus before the release was about $2.04bn and $0.78. By end market: computing up by a low-double-digit percentage, automotive and industrial up by a low-single-digit percentage, communications down by a high-single-digit percentage and consumer down into the mid-teens. There is no 2027 outlook; capital-spending guidance comes in February.
The tone was confident and measured. The chief executive framed the quarter around three themes — stronger earnings power, advanced packaging as the growth engine and customers committing earlier — and was specific about buildings, dates and programme counts. The chief financial officer was precise on the causes of the margin miss and on the 2027 facilities budget, and declined to put numbers on 2027 margins, total capital spending or long-run capital intensity. Neither executive gave a revenue outlook for 2027 beyond computing.
| Firm | Question | Answer, in short |
|---|---|---|
| Bank of America | What drives the small fourth-quarter margin gain? How do Korea and Arizona differ in margin and return, and how fungible is the equipment? | Mix helps, currency hurts. Arizona costs more in depreciation and labour but will run fuller and more automated, so long-run margins should be close. Equipment is usable across fan-out, 2.5D and bumping. |
| Morgan Stanley | Why +25% in computing rather than about 30%? | Not supply: loading a 300-step line took longer than planned. |
| Melius Research | Can computing keep its pace in 2027? What about phone seasonality? | 2026 above 30%, 2027 similar or higher; phone seasonality is changing and 2027 is too early to call. |
| UBS | How is the fan-out bridge progressing? Do material costs and tight capacity support pricing? | The bridge is a late-2027 into 2028 opportunity. Costs are being passed through and the pricing environment is attractive; SiP should rebound after the transfer. |
| Fox Advisors | How will the new capacity affect 2027 margins? | Buildings are capitalised until complete; depreciation rises in 2027; mix should help. Too early to quantify. |
| KeyBanc | Can the build be funded without new debt? What if consumer stays soft? | Solid through 2027, too early for 2028. Computing growth cushions weak phones and consumer. |
| Needham | Memory prices and phones and PCs; long-run capital intensity? | Low-end Android hit hardest; PCs affected, offset by the move to Arm. Capital spending stays elevated through the Arizona peak in 2028–29, then tapers; no long-run figure. |
Paraphrased from a published transcript of the 26 October call (The Motley Fool); every figure checked against the results release.
| Positive surprises | Negative surprises |
|---|---|
| Computing guided more than 50% above a year earlier in the fourth quarter, with a second hyperscaler CPU programme and two 2.5D programmes starting | Gross margin below the guided range |
| Automotive and industrial up 37% on a year earlier, with mainstream automotive packages recovering too | Computing growth short of the guide |
| 2026 capital spending narrowed to the bottom of the range | Fourth-quarter revenue and earnings guidance both below consensus |
| First share repurchases of the year ($32m) | Communications and consumer down again, on both iOS and Android |
| A dozen customers paying in advance, with more in discussion | 2027 facilities spending above $2bn before equipment, with no full guide until February |
What to watch next. The February results and the 2027 capital-spending guide, the single number that sets how deep the free-cash-flow trough goes; whether gross margin holds 18–19% as Korean space starts depreciating; the size and timing of the NVIDIA prepayments, which the third-quarter Form 10-Q may describe; communications in the first half of 2027 as the Vietnam SiP lines qualify; and whether the two new 2.5D programmes ramp in the fourth quarter as scheduled.
Amkor is an outsourced semiconductor assembly and test company, an OSAT. Chip designers and foundries hand it finished silicon wafers; Amkor bumps and dices them, mounts the chips in packages that connect them to a circuit board, and tests them. It is the second-largest company in the industry and the largest headquartered in the United States, with factories in Korea, Vietnam, the Philippines, Taiwan, China, Japan and Portugal, and a campus under construction in Arizona.
For decades packaging was the cheap last step. That changed when shrinking transistors stopped delivering reliable cost gains: the industry now builds large processors from several smaller chips joined inside one package, and the joining (2.5D interposers, high-density fan-out, silicon bridges) has become a performance bottleneck. Amkor charges per unit packaged and per hour of test time; about 88% of revenue is packaging and 12% test. It is a fixed-cost manufacturing business, so profit follows factory utilisation and product mix.
| End market | Q3 2026 | Share | Q3 2025 | Change y/y | Q4 2026 guide, q/q |
|---|---|---|---|---|---|
| CommunicationsSmartphones, tablets | ≈$737m | 37% | ≈$1,013m | ≈(27)% | High-single-digit decline |
| ComputingData centre, PC, storage | ≈$518m | 26% | ≈$378m | +37% | Low-double-digit growth |
| Automotive, industrial and other | ≈$438m | 22% | ≈$318m | +37% | Low-single-digit growth |
| ConsumerIoT, wearables, home | ≈$299m | 15% | ≈$278m | ≈+7% | Decline into the mid-teens |
| Total | $1,991m | 100% | $1,987m | +0.2% | $1.9–2.0bn |
Source: Q3 2026 results release, end-market distribution; fourth-quarter guidance from the call. Dollar figures derived from rounded percentages.
Who is paying. Concentration is falling but remains high. The ten largest customers were 66% of third-quarter sales, against 73% a year earlier. The 2025 Form 10-K named Apple at 29.8% of sales and Qualcomm at 11.1%. Customers named publicly in 2026 include NVIDIA (agreements in Arizona for $1.5bn and in Korea for $1.3bn) and TSMC, which has made Amkor the turnkey back end for wafers from its Arizona fabs. The new computing revenue is concentrated too: management describes a flagship high-density fan-out CPU programme as the largest single driver, with a second hyperscaler CPU programme ramping since the third quarter and two 2.5D programmes in networking and data centre starting in the fourth.
| Cost of sales, % of net sales | Q3 2025 | Q1 2026 | Q2 2026 | Q3 2026 |
|---|---|---|---|---|
| Materials | 57.5% | 53.5% | 52.6% | 50.7% |
| Labour | 9.2% | 10.7% | 10.0% | 10.1% |
| Depreciation | 7.5% | 9.3% | 8.6% | 8.6% |
| Other manufacturing | 11.5% | 12.3% | 12.0% | 12.3% |
| Gross margin | 14.3% | 14.2% | 16.8% | 18.3% |
Materials pass through at a low mark-up, so the fall from 57.5% of sales to 50.7% is the mix shift: the dollars that disappeared were phone packaging and the dollars that arrived were computing and automotive. Labour and other manufacturing costs rose slightly as a share of sales, and depreciation has held at 8.6% while capital spending trebled, because most of the new money is still in construction and not yet depreciating. That changes as Korea’s new space opens at the end of 2026, the Gwangju buildings in late 2027 and Arizona phase 1 at the end of 2027. The chief financial officer expects depreciation to rise by a mid-teens percentage in 2026 and to increase again in 2027.
Arizona. In September Amkor added a second phase, taking the plan to about $12bn and about 93,000 square metres of clean room (33,000 in phase 1, 60,000 in phase 2), with more than 3,500 jobs. Phase 1 construction finishes at the end of 2027 and its anchor customer ramps from late 2028; phase 2 starts construction in late 2027 and should be complete by the end of 2029. Management’s framework for 2027 is up to 15% of the $12bn, which is $1.8bn. Korea. Two buildings at Gwangju, broken ground this quarter and complete in late 2027, cost $400–450m in 2027, and the new Songdo test building completes by year-end; the Korean footprint grows about 40% from 2025 to 2028. Vietnam. About $50m in 2027 for the SiP centre. Facilities alone therefore exceed $2bn in 2027 before equipment.
The funding has three outside sources. A dozen customers have signed advance-payment arrangements since 2025, with more under discussion; NVIDIA’s two agreements total $2.8bn, and none of that money had arrived by September. The US 35% advanced manufacturing investment credit applies to qualifying property whose construction began before 2027, which covers phase 1. The $407m CHIPS award is milestone-based and nothing has been received. Management says the balance sheet is solid through 2027 and that it is too early to say whether 2028 needs new debt.
| $ millions | Q3 2026 | Nine months 2026 | Nine months 2025 |
|---|---|---|---|
| Operating cash flow | 487 | 868 | 451 |
| Payments for property, plant and equipment | (548) | (1,237) | (473) |
| Proceeds from sale of equipment | 0 | 36 | 5 |
| Free cash flow company definition | (62) | (333) | (16) |
| Rise in capital expenditures payable | (125) | 253 | — |
| Accrued capital spending (derived) | ≈423 | ≈1,489 | — |
| Share repurchases | (32) | (32) | — |
| Dividends | (21) | (62) | (61) |
Source: Q3 and Q2 2026 results releases, statements of cash flows. Third-quarter figures are nine months less six months (derived). Accrued capital spending is payments plus the change in capital expenditures payable (derived).
Operating cash flow was $487m in the quarter, enough to fund almost all of the $548m paid for equipment and buildings, and Amkor bought back $32m of shares, its first repurchases this year. The nine months still show $333m of negative free cash flow. Cash and short-term investments were $2.38bn against $2.47bn of debt, so the company is now marginally net indebted; liquidity including the undrawn revolving facility was $3.4bn and debt was 1.7 times EBITDA.
ASE Technology, the largest OSAT, is about three times Amkor’s size including its electronics-manufacturing arm and earns a 27% gross margin in packaging and test. TSMC is both partner and competitor: it keeps the most valuable AI packaging (CoWoS and SoIC) inside its own fabs, while sending Arizona wafers to Amkor. Chinese OSATs led by JCET compete in mainstream and mid-range packaging but are cut off from the tools for the most advanced work by US export controls. Integrated manufacturers such as Intel and Samsung package in-house. Capacity for the advanced formats is short across the industry: on the call the chief executive said packaging “is becoming one of the key bottlenecks”, and that customers are securing capacity into 2028, which is why they are paying in advance.
| $ millions | Q3 2025 | Q4 2025 | Q1 2026 | Q2 2026 | Q3 2026 |
|---|---|---|---|---|---|
| Net sales | 1,987 | 1,888 | 1,685 | 1,898 | 1,991 |
| Advanced products | 1,684 | 1,580 | 1,372 | 1,557 | 1,632 |
| Mainstream (wirebond) | 303 | 308 | 313 | 341 | 359 |
| Gross margin | 14.3% | 16.7%* | 14.2% | 16.8% | 18.3% |
| Operating income | 159 | 185 | 100 | 200 | 225 |
| Operating margin | 8.0% | 9.8% | 6.0% | 10.5% | 11.3% |
| Net income to Amkor | 127 | 173 | 84 | 174 | 191 |
| Diluted EPS | $0.51 | $0.69 | $0.33 | $0.70 | $0.76 |
| EBITDA | 340 | 369 | 285 | 400 | 435 |
| Depreciation and amortisation | 164 | 166 | 171 | 180 | 187 |
| Top ten customers, % of sales | 73% | 72% | 68% | 66% | 66% |
Sources: Q3 2026, Q2 2026, Q1 2026 and Q4 2025 results releases. *Fourth-quarter 2025 gross margin derived from full-year and nine-month figures; it includes a $30m asset-sale benefit.
Disclosed drivers: computing, with more than 30% growth in 2026 and a similar or higher rate expected in 2027 as Korean space opens and new data-centre programmes ramp; automotive and industrial, from advanced in-car computing and a recovery in mainstream packages; a communications recovery once the SiP lines are re-qualified in Vietnam; pricing, with costs passed through and capacity tight; and US capacity, with Apple, NVIDIA and TSMC’s US wafers as anchor demand. Assumed in our base case: the company’s 2028 and 2030 targets, then phase 2 filling from 2031.
Risks, in order of weight: capital intensity, with about $17.4bn of capital spending from October 2026 to 2031 in our base case against $7.5bn of 2026 revenue; customer concentration, with ten customers 66% of sales; AI capital spending, which is both the growth engine and the main destination of the new capacity; the communications transition; Arizona execution, in a higher-cost location; TSMC keeping the most advanced AI packaging in-house; memory and component shortages; and trade and geography, with factories across seven countries and a campus under construction in an eighth.
| Source | Score / 5 | Assessment |
|---|---|---|
| Brand | 2.5 | Known to engineers as the US-headquartered OSAT; little value outside procurement |
| Network effects | 1 | None |
| Switching costs | 4 | Requalifying a package is slow and risky, so programmes stay for years; it protects existing work, not new awards |
| Cost advantage | 2.5 | A competitive Asian cost base and automation in Vietnam and Arizona; ASE has the scale edge and Arizona costs more to run |
| Scale | 3.5 | Second-largest OSAT, $7.5bn of sales; about a third of ASE |
| Intellectual property | 3.5 | High-density fan-out, 2.5D, wafer-level, a bridge in development; TSMC leads the most advanced formats |
| Distribution | 3.5 | Turnkey back end for TSMC’s Arizona wafers and the NVIDIA agreements; the channel is partly rented from a competitor |
| Regulation and geography | 4.5 | First high-volume US OSAT site; 35% credit; CHIPS award; export controls on Chinese rivals. Policy can change |
| Data | 1.5 | Yield learning across programmes; not defensible on its own |
| Customer loyalty | 3.5 | Apple for more than a decade; prepayments from a dozen customers; large customers dual-source by habit |
Scores are our judgement. Composite about 3.0 out of 5: narrow, widening at the advanced and US end.
| Amkor | ASE Technology | TSMC (packaging) | JCET | |
|---|---|---|---|---|
| Position | No. 2 OSAT, $7.5bn sales | No. 1 OSAT, ≈$22.5bn including EMS | Foundry with in-house CoWoS and SoIC | No. 1 in China |
| Advanced packaging | Fan-out, 2.5D, SiP; bridge from late 2027 | Full range; leading-edge capacity at scale | The most advanced AI formats | Mid-range; limited by export controls |
| US capacity | Arizona from late 2028 | Limited | Arizona fabs, packaging partners | None |
| Packaging margin | 18.3% gross (Q3 2026) | 27.3% gross, packaging and test (Q2 2026) | Not disclosed separately | Not covered here |
| Main vulnerability | Capital spending and concentration | Cyclicality; EMS dilution | — (it is the threat) | Export controls |
Sources: Amkor Q3 2026 release; ASE 2Q26 release (30 July 2026). ASE revenue is trailing twelve months converted at NT$31.59 per US dollar.
Strongest advantages. Switching costs and geography. Once a chip is qualified in a package on a given line, moving it means new reliability testing, new yield learning and schedule risk on a product with a short life, so programmes stay where they start. And Amkor is the only large OSAT building high-volume capacity in the United States, with a 35% investment credit, a CHIPS award and US export controls that keep Chinese competitors away from the tools for the most advanced work. NVIDIA’s and TSMC’s choices of Amkor for US capacity reflect that position, not only its technology.
Biggest threats. TSMC first: every generation in which the foundry keeps more of the AI package inside CoWoS and SoIC is a generation in which the most valuable work never reaches an OSAT. ASE second: it is three times the size, earns about nine points more gross margin in packaging and test, and spent US$1.7bn on equipment in its second quarter alone. Third, the US cost base: Arizona will have higher labour and building costs than Korea or Vietnam, and management’s answer, running it fuller and with more automation, still has to be proved.
Direction: widening at the advanced and US end, where prepayments and policy reinforce each other; stable in automotive and mainstream wirebond, protected by qualification; shrinking in commodity SiP and phone packaging, where Chinese capacity and customer bargaining power grind margins. Our base case reflects the net by taking gross margin to 21.5%, still six points below ASE’s packaging and test today.
| Concern | Severity | Evidence and why it matters |
|---|---|---|
| Capital intensity | HIGH | Nine-month free cash flow $(333)m; 2027 facilities above $2bn before equipment; Arizona peaks in 2028–29. In our base case free cash flow is negative from the fourth quarter of 2026 to 2030 |
| Customer concentration | HIGH | Ten customers 66% of third-quarter sales; Apple 29.8% and Qualcomm 11.1% of 2025 sales. The computing ramp leans on a few large programmes |
| Scope growth at Arizona | MED-HIGH | The plan was $2bn when announced in 2023, $7bn in October 2025 and $12bn in September 2026. Each step added capacity, but the bill grew by $5bn in eleven months |
| Advance-payment disclosure | MEDIUM | The prepayments within the NVIDIA agreements are not disclosed separately; terms (volumes, minimums, refunds) are not public. Contract liabilities were $393m at 30 June |
| Controlled ownership | MEDIUM | The founding Kim family holds a controlling interest; an entity linked to it sold 10.0m shares for $485m on 12 February 2026. The risk factors say certain holders can determine shareholder votes |
| Guidance delivery | MEDIUM | Third-quarter gross margin below the range and computing short of the guide, after two quarters of large beats |
| Leverage trend | MEDIUM | Debt $2.47bn from $1.45bn in December; net cash of $546m at year-end became net debt of about $91m. The 0% notes ($1.15bn, 2031) convert at $106.37 |
| Inventory build | LOW | Inventories $604m from $438m in December (+38%) on revenue about 5% above the fourth quarter of 2025; days of inventory from about 26 to about 34 |
| Insider selling | LOW | About $3.9m sold by directors and officers from 30 July to 13 October, including $2.8m by the former chief executive, now a director; about $1.0m under 10b5-1 plans. No open-market purchases |
| Earnings quality | LOW | Other income $23m, about 10% of third-quarter pre-tax profit; a 17% tax rate below the 20% planning rate |
Severity is our judgement. Sources: Q3 and Q2 2026 results releases, Q2 2026 Form 10-Q, Forms 4 filed 31 July–15 October 2026 and 17 February 2026, company announcements of 2023, 6 October 2025 and 8 September 2026.
Nothing here points to the accounts being unreliable. Goodwill is $18m, receivables are about 70 days of sales, revenue is recognised on delivery, and the cash is real. The concerns are the size of the build relative to the business and how much of the plan rests on a handful of customers.
| Strength | Weight | Evidence |
|---|---|---|
| Customer money up front | HIGH | A dozen advance-payment arrangements since 2025; NVIDIA agreements of $1.5bn (Arizona) and $1.3bn (Korea). Customers fund capacity they expect to need |
| Mix shift | MED-HIGH | Computing 26% and automotive 22% of sales, from 19% and 16% a year ago; materials down to 50.7% of sales from 57.5% |
| Policy position | MED-HIGH | First high-volume US OSAT site; 35% investment credit on phase 1; CHIPS award up to $407m; export controls keep Chinese rivals out of the leading-edge tools |
| Operating leverage | MED-HIGH | Operating income up 41% on a year earlier on flat revenue |
| Pricing | MEDIUM | Management says cost increases are passed through and capacity constraints keep pricing attractive among a small number of qualified suppliers |
| Financing | MEDIUM | Liquidity $3.4bn; the 2031 notes carry no coupon and a capped call to $139.50 limits dilution |
| Korean redeployment | MEDIUM | Space released by the SiP move to Vietnam is being refilled with advanced-packaging equipment already on order |
The balance of the evidence. The strongest signal is that customers are paying for capacity before it exists. NVIDIA alone has agreements worth $2.8bn, and twelve customers have done the same in some form. Prepayments do not prove the capacity will earn its cost of capital, but they make an empty Arizona factory much less likely. The concerns are about scale and timing, not about the quality of what is reported.
| Dimension | Score / 5 | Assessment |
|---|---|---|
| CEO track record | 3.5 | Kevin Engel, chief executive since 1 January 2026 after running operations. Utilisation and mix improved sharply under him; the strategy (Arizona, advanced packaging, the TSMC link) was set largely before |
| CFO credibility | 4 | Megan Faust gave a quantified bridge for the margin miss and the 2027 facilities budget, and said plainly that 2028 debt needs are not yet known |
| Guidance accuracy | 3.5 | EPS above the range in the first and second quarters of 2026 and inside it in the third; revenue above the midpoint in the first two, below it in the third; gross margin above the range in the first two, below it in the third |
| Transparency | 4 | Discloses cost of sales by type, end markets and top-ten share every quarter; prepayment amounts and contract terms are the gap |
| Capital allocation | 2.5 | A $12bn US campus, up from $7bn a year ago, plus Korea and Vietnam, in a business whose filings cite no backlog; partly customer-funded, but the returns will not be visible until 2029–30 |
| Acquisitions | 3 | None of note; goodwill $18m |
| Buybacks | 2.5 | First repurchases of 2026 ($32m) in the third quarter, under a $300m authorisation, while capital spending is at a record |
| Dilution | 4 | Diluted shares 249.5m, up 0.5% on a year earlier; the 2031 notes carry a capped call to $139.50 |
| Insider ownership and dealing | 3 | Controlled by the founding Kim family, which sold 10.0m shares in February; directors and officers sold about $3.9m since July and bought none |
| Compensation | 3 | Standard restricted-stock structure; nothing unusual in the filings reviewed |
| Board quality | 3 | Experienced directors, but minority holders have limited practical influence in a controlled company |
| Communication style | 4 | Plain and specific about buildings, dates and programme counts; reticent on 2027 numbers until February |
Composite about 3.3 out of 5. Sources: results releases and calls for 2026, Forms 4, Q2 2026 Form 10-Q.
Do they act like long-term owners? As operators, yes. Moving SiP to Vietnam at the cost of a year of communications revenue, committing to a second Arizona phase before the first is finished, and asking customers to prepay for capacity are all decisions that trade near-term reported numbers for position. As stewards of shareholder capital the record is still open. The Arizona budget has grown sixfold since 2023, the payback depends on programmes that ramp from late 2028, and the first buyback of the year coincided with the largest spending programme in the company’s history. Insiders have been sellers, not buyers.
| Window | Event | What it decides | Impact · source |
|---|---|---|---|
| Late Oct 2026 | Third-quarter Form 10-Q | Prepayment amounts and terms, contract liabilities, the share count | Medium · filing calendar |
| Late Oct 2026 – Jan 2027 | Hyperscaler and memory-maker results | 2027 data-centre budgets; memory prices | High · company calendars |
| Q4 2026 | Two new 2.5D programmes (networking, data centre) | Fourth-quarter computing | Medium · 26 Oct call |
| End-2026 | Songdo test building completes | Automotive and data-centre test capacity | Low · 26 Oct call |
| 30 Nov – 3 Dec 2026 | NVIDIA GTC Washington, Amkor exhibiting | Programme visibility | Low · company, 21 Oct 2026 |
| Early Nov 2026 | Quarterly dividend declaration | Whether cash is conserved | Low · timing typical |
| Feb 2027 | Fourth-quarter results and the 2027 capital-spending guide | The depth of the free-cash-flow trough; facilities alone above $2bn | Very high · CFO, 26 Oct call |
| Feb 2027 | 2026 Form 10-K | Customer concentration, credit and grant receivables, purchase obligations | Medium · filing calendar |
| 2027 | First NVIDIA Arizona prepayment | The cash that funds the 2027 step-up | High · company |
| H1 2027 | Communications as the Vietnam SiP lines qualify | Whether the segment recovers in 2027 | High · 26 Oct call |
| Late 2027 | Gwangju buildings complete | Space for the fan-out ramp; interposer capacity up more than 300% by end-2028 | High · company |
| End-2027 | Arizona phase 1 construction complete; phase 2 breaks ground | Depreciation starts; the anchor customer ramps from late 2028 | High · company, 8 Sep 2026 |
| Late 2027 into 2028 | Fan-out bridge programmes | The next advanced format after 2.5D | Medium · 26 Oct call |
| Rolling | Trade and export rules | Tariffs on packaged parts; export controls on China | Medium · policy |
Dated where the company or a counterparty has given a date; “timing typical” where we infer it from past reporting. Upside in each case is the event arriving as described or better; downside is a delay or a weaker figure. The single most informative item is the February capital-spending guide, because it sets the depth of the trough the base case already assumes. Impact reflects how much the event can move the value, not the direction of its outcome.
BULLThe mix. The shift is not a forecast any more. Computing and automotive were 35% of sales a year ago and are 48% now. Operating income rose 41% on flat revenue. Materials fell seven points as a share of sales. That is a better business, visible in one quarter’s numbers.
BEARAnd gross margin still missed its own range, computing came in five points short of its guide, and the fourth quarter is guided below consensus. The improvement is real; the guidance is not as conservative as it used to be.
BULLDemand. Sixty basis points of the miss was currency and an earthquake. Customers are telling you what demand looks like: a dozen of them are paying in advance, NVIDIA for $2.8bn across two agreements. Nobody prepays for packaging capacity they think they will not need.
BEARPrepayments are loans repaid in kind, not profits. Look at what they are funding: a $12bn Arizona campus that was $7bn a year ago, more than $2bn of buildings in 2027 alone, and free cash flow negative through 2030. A cyclical toll-packager is betting the balance sheet on a five-year build.
BULLFunding. Not the balance sheet: 35% of phase 1 comes back as a tax credit, CHIPS adds $0.4bn, and management says no new debt is needed through 2027. Arizona is the only high-volume US OSAT site, and export controls keep the Chinese rivals away from the high end.
BEARThe best AI packaging stays inside TSMC. ASE is three times bigger, nine points more profitable in packaging and test, and spent $1.7bn on equipment in one quarter. Amkor’s slice is the overflow plus what customers want in America, and America costs more.
BULLValuation. Then why does ASE trade at 46 times earnings and Amkor at 15? On next year’s consensus Amkor is 13 times earnings with a PEG of about 1, for a company whose own plan doubles earnings by 2030.
BEARBecause Amkor’s earnings will not be cash for years, and because the plan is the price. The reverse DCF says $37.60 already assumes the 2028 and 2030 targets are met. You are paying for execution, not getting it free.
BULLPaying for the plan, with the bull case thrown in for free: $56.79 if 2.5D and the bridge ramp faster.
BEARAnd $18.01 if 2027 is a digestion year. A quarter of the probability sits there.
The bull has the better of the argument on the business: the mix shift, the margin expansion and customers paying in advance are facts, not forecasts, and Amkor’s US position is a real advantage that did not exist five years ago. The bear has the better of it on cash and timing: the build has grown faster than the business, the payback starts in late 2028 at the earliest, and the third quarter was the first in 2026 in which a guided figure was missed. On the price it is a draw: our base case, which grants management its own targets, is $35.45; the price is $37.60. Fairly valued is the evidence-based conclusion. Most uncertain: the size and terms of the prepayments, the phase split of the $12bn and how much qualifies for the credit, and whether gross margin holds as new buildings start depreciating. What would change it: a price above $49.49 or below $26.17; or a 2027 capital-spending guide and prepayment schedule that move the trough materially. Verify next: the third-quarter Form 10-Q, the February capital-spending guide, first-half 2027 communications revenue and gross margin through 2027.
Companies like Apple, NVIDIA and Qualcomm design chips, and factories like TSMC print them onto silicon wafers. Amkor does the next step: it cuts the wafer into chips, puts each chip in a protective package that connects it to the rest of a device, and tests that it works. It is the finishing shop of the chip industry, and the second-biggest in the world.
It charges for each chip it packages and for the time spent testing. Its factories cost roughly the same to run whether they are full or half empty, so profits rise quickly when they are busy and fall quickly when they are not.
The newest AI chips are built from several smaller chips joined together inside one package, and joining them is hard. That makes the finishing step a bottleneck. Amkor’s computing business, which serves data centres, grew 37% in a year, and NVIDIA has agreed deals worth $2.8 billion, paying part up front.
Amkor is paying for its new factories years before they earn anything. Customers and the US government are paying part of the bill, through advance payments and tax credits. How much of the bill they cover, and how fast the factories fill, decides what the shares are worth.
If the company meets its own plan, we think the shares are worth about $35, with a range of about $26 to $49. At $37.60 we call them fairly valued: you are paying for the plan to work, and not for anything better. The shares move a lot: they are down about 60% from their June high.
Right: AI packaging keeps growing faster than expected, the new factories in Korea and Arizona fill quickly, and margins reach the company’s own target of about 22% by 2030.
Wrong: AI spending pauses and Amkor’s new space sits half empty; a handful of customers, led by Apple, provide most of its sales; or TSMC, its partner, keeps more of the most advanced packaging for itself.
The 2027 spending plan, in February. Whether phone revenue recovers in the first half of 2027. And whether gross margin holds as the new buildings start to cost money in depreciation.
| Status | What it covers |
|---|---|
| Verified from Amkor filings and releases | Revenue by end market and by product group, cost of sales by type, margins, operating income, EBITDA, cash flow, capital spending and capital expenditures payable, cash, debt, the convertible notes and their capped call, finance leases, pension and severance, minority interests, contract liabilities at 30 June, customer concentration, the Investor Day targets, the Arizona, Korea and Vietnam plans, the NVIDIA agreements, the CHIPS award, insider transactions (Forms 4) and the fourth-quarter outlook |
| Verified from market data and peer filings | Amkor at the last Nasdaq pre-market trade, at 4:40 a.m. Eastern on 27 October 2026; ASE Technology at its last NYSE pre-market trade, at 4:23 a.m. Eastern; ASE’s share count, debt, cash and results from its 2Q26, 1Q26 and 4Q25 releases; the 10-year Treasury yield from the US Treasury daily par yield curve |
| Third-party, attributed | Call remarks and analysts’ questions from a published transcript of the 26 October 2026 call (The Motley Fool); consensus revenue and EPS (Yahoo Finance); the NVIDIA Korea agreement (Benzinga); SOXX aggregate multiples (Sonde) |
| Our estimate or judgement | The revenue, margin, operating-expense and capital-spending paths in every scenario; the split of the $12bn between phases; the investment credit and its timing; the timing of the CHIPS award and the customer advances; contract liabilities and the share count at 30 September; fourth-quarter capital spending; the scenario weights |
| Still open | The prepayment amounts and terms inside the NVIDIA and other agreements; the phase split of the Arizona budget; how much of phase 1 qualifies for the credit; the 2027 capital-spending guide; whether 2028 needs new debt |
Method. Unlevered discounted cash flow on a ten-year explicit path, methodology version 1.3. The fourth quarter of 2026 is a stub to 31 December and each later year is discounted at its mid-point; the Gordon terminal value at 3.0% is discounted from the end of 2036; everything is valued at 27 October 2026 from the 30 September 2026 balance sheet. One discount rate (10.5%) and one terminal growth rate are used across all scenarios, and each material risk is priced once. Stock-based compensation is inside operating costs and unvested awards are excluded; the 0% notes are debt at face, and value above the $139.50 cap of the capped call would add shares. Customer advance payments are cash when received and revenue without cash when delivered; advances already received are deducted as debt-like. The investment credit and the CHIPS award are cash on their own schedule. Scenario weights are our judgement, the fair-value band is the corners of the WACC 9.5–11.5% by terminal growth 2.5–3.5% grid, and the reverse valuation is run against the market price. The rating changes to overvalued if the price rises above $49.49 with no change in the evidence, and to undervalued below $26.17; the conclusion is most sensitive to the long-run gross margin, revenue after 2028 and the discount rate. Ratings are undervalued, fairly valued or overvalued according to where the price sits relative to the band; Sonde does not issue buy, sell or hold recommendations or price targets. The reference price used throughout is $37.60, the last Nasdaq pre-market trade, at 4:40 a.m. Eastern on 27 October 2026, and every multiple in this report scales directly with it. Ratings are reviewed each quarter on results and immediately on a material contract, a financing event or a share-price move of more than 25%.