Today, Amkor Technology's earnings report significantly exceeded expectations.

Revenue reached $1.69 billion, a substantial year-over-year increase, clearly surpassing market consensus; EPS also greatly outperformed expectations, driven by capacity utilization quickly rebounding from previous lows to the 70% range. More importantly, the company's guidance for the next quarter was further sharply raised.
However, the market showed no mercy, with the stock price dropping as much as 8% after hours.
Where exactly went wrong?
If we must find a reason, it’s only one: the company has directly increased its full-year capital expenditure from about $750 million to $2.5-3 billion, nearly tripling.
This looks like a replay of the market’s previous concerns about large tech companies’ capex growth and cash flow worries.
But do these concerns really make sense?
To answer this question, we need to break down the matter of advanced packaging.
Essentially, packaging addresses a question: how can computing power be efficiently “physically realized”?
Around this question, the industry chain has formed a clear division of labor: TSMC handles front-end manufacturing, etching circuits into silicon; Amkor handles back-end packaging and testing, turning bare dies into usable chips. Historically, there was almost no overlap between the two. But in the AI era, packaging has begun to directly impact bandwidth, power consumption, and system performance. Advanced packaging is gradually “front-endized,” with TSMC competing with CoWoS and Amkor, blurring boundaries. However, this shift mainly concentrates on the very top segment.
Amkor’s technological path happens to be on the other side. Its focus in advanced packaging is on Fan-Out systems, with HDFO (High-Density Fan-Out) being the current key growth driver, while also laying out 2.5D and 3D.
CoWoS is led by TSMC, based on silicon interposers, serving the extreme bandwidth needs of HBM and AI GPUs; HDFO, based on RDL (Redistribution Layer), does not rely on interposers, offering a simpler structure and lower cost but limited interconnect capability.
They are not competitors but layered solutions: one addresses performance limits, the other balances performance and cost.
From a technical hierarchy, the real ceiling lies in 2.5D and 3D packaging, especially in technologies like Hybrid Bonding that are approaching front-end process levels.
Meanwhile, OSATs like Amkor lead a different “engineering packaging” layer: Fan-Out, Flip Chip, and some 2.5D capabilities. The core of this layer is large-scale manufacturing capacity, yield control, and cost efficiency.
Looking at Amkor’s product structure,
The bottom layer consists of standardized packages like QFN and WLCSP, serving cost-sensitive markets such as automotive, analog, and power.
The middle layer includes FCBGA, fcCSP, and Fan-Out, used in data center CPUs, inference chips, and network switching chips—mid-to-high performance scenarios.
The top layer is extreme packaging like CoWoS, but this layer is not Amkor’s main battlefield; Amkor’s core business is in the first two layers.
FCBGA is essentially “high I/O, high power, high performance, but not aiming for extreme bandwidth.” It is widely used in server CPUs, non-HBM GPUs, cloud OEM ASICs, and switching chips.
Most compute chips do not require HBM. Only training-level chips like NVIDIA’s H100 and B100 must rely on CoWoS + HBM to overcome bandwidth bottlenecks.
Take Google as an example; its chip ecosystem is layered: training uses HBM and 2.5D packaging, while many inference, video processing, and networking ASICs are already using FCBGA or Fan-Out solutions.
Looking at AI development paths today, training is at the pyramid’s apex with limited quantities; inference is the main volume and is expanding exponentially.
From data centers to edge and terminals, the number of inference nodes far exceeds training nodes. In this process, the core constraint of packaging shifts from “performance ceiling” to “total cost of ownership.”
In most scenarios, “performance + cost control” beats “extreme performance,” which is the advantage of FCBGA and Fan-Out.
This is also why HDFO has entered a commercial volume phase and has become Amkor’s most important growth driver.
Returning to CapEx, we see it is laying the groundwork for “compute power diffusion in the inference era.”
From Arizona to Vietnam, from HDFO to high-performance testing platforms, it’s fundamentally about capacity positioning.
Business changes also confirm this. Traditional PCs and consumer electronics remain weak, but revenue from data centers and AI-related areas has hit new highs; HDFO is entering mass production, with customer numbers continuously increasing; meanwhile, the company is beginning to pass costs onto customers, and the long-term pricing power in the packaging and testing industry is marginally returning.
All these signals together indicate that the industry is not simply recovering but undergoing a restructuring.
In summary, advanced packaging is no longer a single path but coexists in two paradigms: “extreme performance” and “scale efficiency.”
The former is defined by TSMC and other manufacturers setting technological limits; the latter is determined by OSATs like Amkor shaping industry volume.
As AI moves from training to inference, from centralization to diffusion, the long-term value is often not in the very top segment but in the middle layer that serves the largest demand.
And this is precisely where Amkor is betting with its $3 billion capital expenditure.
Disclaimer: I hold the securities mentioned in this article. My views are biased and not investment advice. Stock investing carries huge risks; enter with extreme caution.