TFME is the second-largest mainland OSAT and the one most closely tied to high-performance compute. Its H1 2026 report is unusually informative because it breaks results out by base, which lets you see exactly where the money is made.
It also contains a number that needs reading carefully: attributable net profit grew 317%, but most of that growth is not operating.
The headline, and what sits behind it
H1 2026 revenue was 16.04 billion RMB, up 23.0% year-on-year. Attributable net profit was 1.72 billion RMB, up 316.8%. Ex-items net profit was 747 million RMB, up 77.8%. Second-quarter revenue was 8.56 billion RMB, up 23.2% year-on-year and 14.4% sequentially.
The gap between those two profit figures is the thing to notice. The company states that 970 million RMB of the 1.72 billion came from non-recurring items, mainly fair-value gains of about 1.15 billion RMB on shareholdings in several listed semiconductor companies.
Ex-items growth of 77.8% is still strong and is the honest operating number. But a programme decision based on the 317% figure would be based on an investment result, not on the assembly line.
Four bases, four different stories
The segment breakdown is where the report earns its keep:
- TF-AMD Penang: revenue 5.395 billion RMB, up 23.5%; net profit 233 million RMB, up 29.4%
- TF-AMD Suzhou: revenue 4.266 billion RMB, up 8.4%; net profit 395 million RMB, down 27.5%
- Chongchuan (parent): revenue 4.222 billion RMB, up 8.1%; net profit 847 million RMB, up 553.6%
- Nantong: revenue 1.635 billion RMB, up 38.3%; net profit 51 million RMB, against a 228 million RMB loss in the prior-year period
What the base split shows
Two of the four bases are the AMD-associated joint-venture sites carrying FCBGA and chiplet work. Penang is growing faster and more profitably than Suzhou, whose profit fell despite revenue growth. Nantong turned positive. Chongchuan's profit surge is the one to treat with caution, given how much of the group's non-operating income sits in the same period.
The AMD concentration
TFME's own report states that its largest customer, AMD, posted record revenue and profit in the period, with second-quarter data-centre revenue of 6.7 billion USD, up 107% year-on-year, driven by EPYC processor and Instinct GPU demand, and client revenue of 3.1 billion USD, up 23%.
That is a genuine advantage and a genuine risk. The advantage is process maturity: a line running one customer's hardest parts at volume learns faster than a line running many customers' easier ones. The risk is allocation. When that customer's demand surges, capacity follows it.
For a smaller programme the practical question is not whether TFME can build your part — it is what queue position and support model applies at your volume when the anchor customer is running at record levels.
Technology and other segments
The company reports cumulative patent applications above 1,800 and granted patents above 800. First-half R&D spend was 825 million RMB, up 9.25%. Beyond compute, it cites high double-digit growth in automotive packaging revenue supported by Nantong and Hefei capacity, strong demand for PMIC wafer-level packaging, SiP and bumping from domestic power management customers, and double-digit growth in display driver output following two major customer wins.
A 4.22 billion RMB private placement completed during the period at 55.38 RMB per share, with 18 institutions participating.
Which programmes fit TFME
Large-body FC-BGA for server and AI compute is the obvious fit, and it is where the line's maturity is real rather than claimed. Automotive, PMIC wafer-level packaging and display driver are the secondary fits with visible momentum.
It is a weaker fit when your programme needs a route the line holds only in validation, or when you need schedule flexibility during a period when the anchor customer is running at record volumes. In both cases the answer is the same: ask what is qualified today, and ask what queue position your volume band actually gets.
