Amkor: revenue is up 26%, but capital spending is tripling and free cash flow is negative

In the first half of 2026 Amkor grew revenue 26%, with data center computing at a record. It also spent $688 million on equipment, three times as much as a year earlier, and free cash flow was negative $307 million. The company guides to $2.5 to $3.0 billion of capital spending for the full year, against $905 million in 2025. The shares have fallen about 26% in October alone and are 60% below their 52-week high of $96.68. The question is whether a business that turns about a third of its EBITDA into free cash flow deserves a premium for the AI build-out.

At $39.05 on October 22, 2026, the shares trade 23% above my triangulated Base value of $31.66, a downside of 18.9% to Base. The Bull value of $66.98 is 71.5% above the price and the Bear value of $15.23 is 61.0% below it. The probability-weighted value of $36.38 is 6.8% below the price. The required margin of safety of 50% puts the maximum buy price at $15.83. The framework verdict is hold, do not add. Third-quarter results arrive on Monday, October 26, after the close.

Valuation route: Technology hardware and semiconductor services, applied to an outsourced semiconductor assembly and test company. Economic classification: Standard Company (4/12), because the return on capital is about 10.7%, below 15%, and free cash flow is negative this year. Two methods carry equal weight: mid-cycle EV to normalized EBITDA, built up from a sector base with adjustments, and a ten-year cash flow model. Reference checks: historical and peer multiples, and a reverse calculation of what the price assumes. Values are present fair-value estimates, not 12-month price targets.

What Amkor is

Amkor packages and tests semiconductor chips for other companies: it takes finished wafers from chip designers and foundries, puts the chips into the cases and modules that go onto circuit boards, and tests them. Advanced products (flip chip, memory, wafer-level packaging and the related testing) were 82% of revenue in the June quarter, mainstream wirebond products 18%. Packaging services are 88% of sales and testing 12%.

The customer base is concentrated: the ten largest customers were 66% of second-quarter sales, down from 72% a year earlier. In 2025 Apple was 29.8% of net sales and Qualcomm 11.1%. By end market, communications (smartphones and tablets) was 42% in the second quarter, computing 22%, automotive and industrial 22% and consumer 14%. Amkor had about 30,800 employees at the end of 2025, roughly 95% in Asia-Pacific, and is building a large advanced packaging campus in Arizona, where it has a ten-year agreement with TSMC and, since July 24, a $1.5 billion prepayment agreement with NVIDIA.

The latest quarter

Computing revenue rose 20% from the first quarter to a record, driven by data center demand, and automotive and industrial rose 17%. Communications rose 6%. For the third quarter, management guided computing up nearly 30% from the second quarter and communications down by a high single digit percentage. Call coverage says about half of the communications shortfall against normal seasonality comes from market factors, including memory supply constraints, and half from moving a module product line to Vietnam. For the full year, only capital spending was guided: $2.5 to $3.0 billion, 65% to 70% of it for buildings and facilities including Arizona.

Why the current earnings are not the whole story

The first reason is cash. Revenue over the last three years has barely moved: $6.50 billion in 2023, $6.32 billion in 2024 and $6.71 billion in 2025. Free cash flow was $359 million in 2024 and $308 million in 2025, 5.7% and 4.6% of revenue, and negative $307 million in the first half of 2026. This year the company is spending about a third of its revenue on equipment to produce this growth, so profit and cash have moved far apart.

The second reason is one-offs. The second quarter included the $21 million real estate gain in operating expenses and a 14% tax rate helped by $14 million of discrete tax benefits. Together they added roughly $0.13 to the quarter's $0.70 of EPS. The company's own third-quarter guide implies EPS of $0.72 to $0.82 without those items.

The third reason is the base. Earnings are rising from a low starting point: operating margin was 6.9% in 2024 and 7.0% in 2025, and has reached 10.5% in one quarter. Gross margin of 16.8% is the best in the data I have, and management guides to 18.5% to 19.5% in the third quarter. Utilization, product mix and fixed-cost leverage can all reverse. And on the first-quarter call, according to call coverage, management said the Arizona ramp would reduce operating margin by 1 to 2 points from 2027, before the site scales.

Balance sheet, capital spending and shares

Cash and short-term investments are $2.51 billion against total debt of $2.49 billion ($150.8 million due within a year), so net cash is only $26 million, down from $546 million at the end of 2025. In May the company issued $1.15 billion of 0% convertible notes with a capped call; I could not verify the conversion price, so I treat the notes as debt at face value and do not add any dilution. Restricted cash is $188.5 million. Debt to EBITDA is 1.8x gross. Property, plant and equipment grew 16% in six months to $4.48 billion, and capital spending bills not yet paid rose from $244 million to $621 million.

The shares are stable: 247.9 million basic and 250.0 million diluted in the second quarter. I use 250.0 million for a market value of $9.76 billion. Minority interests are $36.4 million, which I treat as a claim at book value, and I hold back $250 million of cash as an operating reserve. Net financial claims in the model are $260 million and enterprise value is about $10.02 billion, 8.6 times 2025 EBITDA of $1.162 billion and 6.4 times my estimate of 2026 EBITDA of about $1.56 billion (first half $685 million, third and fourth quarters about $425 and $450 million).

On July 24 the company and NVIDIA announced a $1.5 billion prepayment agreement to expand advanced packaging in Arizona. Call coverage says the prepayment is expected in 2027 and will be returned over five to ten years through services delivered in the United States. I treat it as financing for capacity that is already in my capital spending, not as value for shareholders, and I do not add it to cash flow. The campus has a first phase that is fully committed, and a second phase has been announced; the company says total planned investment is $12 billion. I could not find the amount of any CHIPS Act grant.

The cash flow levels check was not triggered on the numbers I have: net debt is near zero, and interest expense is $17 million a quarter against first-half operating cash flow of $382 million. In the Base case net debt rises to about $1.0 billion at the end of 2026 and about $1.8 to $1.9 billion at the end of 2027, about 1.0 times 2027 EBITDA, before any NVIDIA prepayment. The scheduled debt maturities were not in the sources I could read.

Valuation: two methods, equal weight

The first method is mid-cycle EV to normalized EBITDA. The second is a ten-year cash flow model on cash flow after stock compensation, capital spending and working capital. Each runs through Bear, Base and Bull and each gets half the weight.

The Base normalized EBITDA of $1.6 billion is 38% above 2025 ($1.162 billion) and about in line with my estimate for 2026 ($1.56 billion). I take the current margin and capacity as roughly sustainable at mid-cycle, but not the best quarters: a 20% EBITDA margin on about $8 billion of revenue.

Why 7x. The multiple is built up: a sector base of 6.0x, which is where outsourced assembly and test has historically traded, plus 1.0x for growth tied to data center computing, plus 0.5x for the balance sheet, minus 0.5x for customer concentration and the Arizona ramp. I add nothing for return on capital, which is about 10.7%, and nothing for dilution, which is small. Provider data put the current trailing multiple at 7.7x. A multiple has to be justified by what it implies: at 7x, the enterprise value of $11.2 billion on normalized free cash flow of about $520 million (a third of EBITDA, the share the cash flow model produces in the long run) is a 4.7% yield, and at a 10% discount rate that pays for perpetual growth of about 5.1%. The Base assumes 2.5%, so 7x is a generous multiple, and I say so. At 5x the implied growth is about 3.3%.

The cash flow model starts with a second-half 2026 outflow of $1.0 billion (capital spending of about $1.7 billion against EBITDA of about $0.9 billion). Revenue grows 12%, 10%, 8%, 6%, 5%, 4.5%, 4%, 3.5%, 3% and 3% from $7.63 billion in 2026 to $13.5 billion in 2036. EBITDA, which is already after stock compensation under GAAP, is 20.5% of revenue and fades to 19.5%. Capital spending falls from 25% of revenue in 2027 to 20%, 16%, 14%, 13% and then 11% to 12%, against depreciation of 9%. Working capital takes 12 cents of each new revenue dollar. It uses a 10% discount rate, reflecting a beta above 2, a 17% cash tax rate, and 2.5% terminal growth. Free cash flow is negative in 2027 and in 2028, so terminal value is 87% of the Base enterprise value.

Common basis and why the two methods differ

The gap is $6.05 billion of enterprise value, or $24.21 a share, and it has one source: how much of EBITDA this business can keep. A multiple of EBITDA treats depreciation as if it were free. Amkor spends 11% to 13% of revenue on equipment even at maturity, and depreciation is 9% of revenue, so the cash flow model keeps only about 33 cents of each EBITDA dollar. The divergence is 76.5% in the Base, above the framework's 30% flag. I handle it in the confidence score (robustness) and not in the margin of safety, so it is not counted twice.

Historical and peer cross-check

On trailing earnings the stock trades at about 17.7 times, and forward price to earnings is 14 to 17 times depending on the data provider (an implied forward EPS of about $2.78 at the lowest). A provider PEG ratio is not available, and free cash flow growth cannot be measured because free cash flow is negative this year. A provider shows operating margin of 8.9%, return on invested capital of 10.7% and a trailing EV to EBITDA of 7.7x (6.8x forward). A ten-year median multiple was not available to me.

Figures are from third-party data providers in October 2026. ASE, the largest competitor, trades at a higher multiple with a similar margin, but its EBITDA includes a larger share of non-packaging business and I did not have a clean basis to compare, so this is context and not a basis for the value. The stock fell sharply after the late-July results, when third-quarter communications guidance came in weak, and about 26% in October as of this morning. I did not find one confirmed cause for the October fall; coverage points to selling across semiconductor stocks and to concern about the cost of the Arizona ramp. Beta is 2.2.

Valuation triangulation

Divergence is 88.5% in the Bear, 76.5% in the Base and 33.4% in the Bull. All three are above the 30% flag, so read each scenario as a range between the two methods: the Base is $19.55 to $43.76. Terminal value is 78.9%, 87.0% and 75.2% of enterprise value in the Bear, Base and Bull cash flow models, because cash flows are negative for the first two years and small afterwards. The conclusion depends on what the business earns after the build-out ends, which no one can see yet.

What is the market already pricing in?

Run the cash flow model backwards and the price needs revenue to grow about 20% a year for four years from 2027 and fade after that, reaching roughly $18.4 billion in 2031, a compound rate of about 19% a year from 2026, against $11.3 billion in the Base (8.2% a year). Put differently, Base cash flows would have to be about two times higher every year. If the Base cash flows play out and the shares are bought at $39.05, the return on the enterprise is about 3.7% a year, far short of the 10% required.

On multiples the picture looks different. At 7x, the price implies mid-cycle EBITDA of about $1.43 billion, below the $1.56 billion I estimate for this year. At 6x it implies $1.67 billion and at 8x $1.25 billion. The market is paying for roughly this year's profit at a normal multiple. The disagreement with the cash flow model is about whether capital spending of a third of revenue is temporary.

Bear, base and bull scenarios

Bear: the AI build-out slows and smartphone demand stays weak, revenue falls 8% in 2027, margins settle near 16%, and the new capacity sits partly idle. Capital spending is cut to 11% of revenue in 2028 and 9% after that. Base: computing keeps growing, communications recovers slowly, growth fades from 12% toward the pace of the economy, and the Arizona ramp cost is absorbed by mix and scale. Bull: the data center and NVIDIA programs run well ahead of the Base, revenue passes $17 billion by 2036, EBITDA margins hold above 21% and the market pays 9 times EBITDA.

The probability-weighted value is $36.38, 6.8% below the price.

Sensitivity: multiple method, value per share

On this method alone the price sits inside the range, which is why the multiple method looks fair.

Sensitivity: cash flow model, value per share

No cell reaches the price. Even an 8% discount rate with 3% terminal growth gives $35.18.

How the buy price is set

Confidence C starts at a 35% margin of safety. Five points are added for the cyclicality of semiconductor demand, five for customer concentration and five for the dependence of the cash flow model on terminal value, for 50%. A 50% discount to the $31.66 Base gives a maximum buy price of $15.83. Applied to each method separately, the same margin gives $9.78 for the cash flow model and $21.88 for the multiple method. Each risk is priced in one layer:

Dividend

The quarterly dividend is $0.08352 a share ($0.33 a year, a 0.85% yield), raised 1% in December 2025, and costs about $83 million a year. In the first half of 2026 the company paid $41.4 million of dividends while free cash flow was negative $307 million, so it was funded from cash and the new convertible notes. The dividend is small and safe in size, but it is not a reason to own the stock.

Reasons to own AMKR

• Growth that is showing up. Revenue up 25.6% in the quarter, computing at a record and guided up nearly 30% next quarter.
• Margins are expanding. Gross margin from 12.0% to 16.8% in a year, with 18.5% to 19.5% guided for the third quarter.
• Strategic position. A ten-year TSMC agreement, a $1.5 billion NVIDIA prepayment agreement and a large U.S. advanced packaging campus under construction.
• A clean balance sheet so far. Debt to EBITDA of 1.8x gross, net cash near zero and $3.6 billion of liquidity.
• Less dependence on one customer. The top ten customers were 66% of sales, down from 72%.
• A lower price. The shares are 60% below their high and trade at 14 to 17 times forward earnings.

What could go wrong?

• Cash burn. Free cash flow was negative $307 million in the first half and capital spending of $2.5 to $3.0 billion implies a large outflow in the second half.
• Capital intensity. The business keeps about a third of EBITDA as free cash flow, so a margin or utilization miss hits equity value hard.
• Customer concentration. Apple was 29.8% of 2025 sales and Qualcomm 11.1%; the communications end market, 42% of sales, is guided down next quarter.
• Arizona ramp. Start-up costs and margin dilution of 1 to 2 points from 2027, according to call coverage, before the site fills.
• Cyclicality. Revenue was flat over 2023 to 2025 and the stock has a beta above 2.
• Supply chain. Management cited tight advanced materials, memory supply and export controls.
• Earnings quality in the latest quarter. About $0.13 of the $0.70 came from a property gain and tax items.
• Convertible notes. $1.15 billion of 0% notes with an unverified conversion price could dilute shareholders.
• Missing data. Debt by instrument, maturities and stock compensation were not in the sources I could read.

Management execution

Management has delivered on its guidance for several quarters, lifted gross margin by more than four points in a year, diversified the customer base and won agreements with TSMC and NVIDIA. It has also chosen to raise capital spending to three times last year's level and to issue convertible debt in the same year, while keeping the dividend, and it has not guided revenue or margins for the full year. I have not seen a call transcript, only call coverage. Management and capital allocation scores 6/10.

Stock Analyza scorecard

The valuation score is 38/100: the price is 23% above the Base and below the Bull, with the cash flow model far below it.

Confidence score

A grade of C is the lowest this analysis can earn without a no-buy conclusion: one point lower would make it D. Accounting is straightforward under GAAP; the lower scores reflect a cyclical business, a 76% gap between the two methods and the missing debt detail.

What I would watch from here

The next review point is third-quarter results on Monday, October 26, 2026, after the close, with the call at 5:00 p.m. Eastern. Guidance is revenue of $1.95 to $2.05 billion, gross margin of 18.5% to 19.5% and EPS of $0.72 to $0.82.

The three most important thesis breakers

1. Capital spending that keeps rising while revenue growth fades, leaving idle capacity and negative free cash flow into 2028.
2. Gross margin falling back as the Arizona and Korean capacity starts, before the volume arrives.
3. Losing share at a large smartphone customer, or a pause in data center spending by the NVIDIA and TSMC ecosystem.

Stress test and independent check

The whole model was recomputed separately in high-precision arithmetic and matched to the cent. Three judgments were tested because they move the answer. Capital spending matters most: two points of revenue in every year moves the cash flow value from $19.55 to $30.41 or $8.69. Margins matter nearly as much: two points either way gives $10.54 or $28.57. The second-half 2026 outflow is not material: at negative $1.3 billion the value is $18.37 and at negative $0.7 billion $20.74. Minority equity and the $250 million cash reserve each move the Base by less than $2.

The strongest counter-argument is that capital spending of 25% of revenue is growth capital for a data center business that will earn much more per square meter of cleanroom than phone packaging, that customer prepayments reduce the funding burden, and that a 10% discount rate is too high for a company with little net debt. At an 8% discount rate and the Bull operating path the cash flow value is $70.66, well above the price. That is a legitimate view, and it needs the Bull scenario and a low required return together. On the multiple method alone the price is already inside the Base range.

Only the combination of the Bull operating path and a discount rate of 8% clears the price on the cash flow model. Main data limits: debt by instrument and maturities, stock compensation, the conversion terms of the new notes, any CHIPS Act grant, a standalone second-quarter cash flow statement and a full-year revenue outlook were not available to me, and the third quarter has not been reported.

Final verdict: hold at $39.05, do not add

Amkor is a better business than it was a year ago, and the shares have fallen 60% from their high. But the cash flow model, which counts what the business must spend to grow, does not support the current price, and the multiple method needs a generous 7x to reach it. The distance between those two answers is the main reason for the hold and for a margin of safety of 50%.

Verdict: HOLD at $39.05, do not add. Triangulated Base value $31.66, downside 18.9%, Confidence C (13/25). Required margin of safety 50%, maximum buy price $15.83.

For an existing holder, the probability-weighted value of $36.38 is close to the price, so there is no strong case to sell. New money waits for $15.83 or below, or for evidence that capital spending is peaking and free cash flow is turning positive while gross margin holds above 18.5%.

The main thesis breaker is capital spending that keeps rising after demand stops accelerating.

Sources

Primary sources include Amkor's second-quarter 2026 results release, the fourth-quarter and full-year 2025 results release and the fiscal 2025 Form 10-K. Earnings call coverage, the NVIDIA agreement, peer and historical multiples and the market price come from third-party sources and data providers.

The reference market price is $39.05, the October 22, 2026 regular session close. An intraday price of about $39.4 on October 23 is within 5% of it. Balance sheet figures are as of June 30, 2026.