폭스콘 CEO, 사상 최고 Q2 실적 이후 CoWoS 고급 패키징을 2027 AI 서버 천정으로 지명

For the first time in its history, Foxconn generated more revenue from AI servers than from iPhones and all other consumer electronics combined. Hon Hai Precision Industry—the Taiwanese manufacturer globally known as Foxconn—reported second-quarter 2026 net profit of NT$59.97 billion (approximately $1.86 billion), representing a 35% year-over-year increase from the NT$44.4 billion posted in the same quarter last year. The result surpassed the LSEG analyst consensus of NT$58.8 billion and established a new quarterly record for second-quarter net profit. The company that built its empire assembling iPhones in Shenzhen is now, by revenue, an AI infrastructure enterprise.

Foxconn’s cloud-and-networking segment, which houses its AI server operations, accounted for 51% of total Q2 2026 revenue, up from 48% in Q1 2026. This marks the first time the segment’s revenue share has exceeded 50% in a single quarter. Smart consumer electronics, including iPhone assembly, declined to 29% of quarterly revenue. The computing division contributed 15%, while electronic components accounted for 5%.

On a consolidated basis, Q2 2026 sales reached NT$2.53 trillion (approximately $78.5 billion), a 41% year-over-year increase and the company’s second-largest single-quarter revenue figure. Operating profit hit an all-time quarterly record of NT$94.8 billion (approximately $2.9 billion), up 68% year-over-year. The operating margin expanded to 3.75%, an improvement of 0.60 percentage points from a year earlier.

Rotating CEO Michael Chiang characterized the revenue shift as structural rather than seasonal. “AI-related business performance will continue to grow in the third quarter,” he stated in prepared remarks. “Combined with ICT products entering their peak season in the second half of the year, we expect significant quarter-on-quarter growth and strong year-on-year growth.” Chiang noted that cloud investment would become Hon Hai’s most critical growth driver over the coming years. The company reiterated its full-year guidance of “strong” growth—a stance maintained throughout 2026—while identifying geopolitical volatility as the primary risk to that outlook. Foxconn does not issue numeric forecasts.

The first-half 2026 results reinforce this trajectory. First-half net profit reached NT$109.89 billion (approximately $3.4 billion), up 27% year-over-year, on revenue of NT$4.65 trillion (approximately $144.2 billion), a 35% year-over-year increase.

Investors and supply-chain planners sought clarity on Foxconn’s involvement with Nvidia’s Vera Rubin AI server platform, and Chiang provided the most definitive timeline to date. Vera Rubin rack systems—Nvidia’s next-generation architecture succeeding the current Blackwell generation—entered mass-production preparation at Foxconn facilities in Q3 2026, with initial shipments slated for Q4 2026. “We expect production volumes to increase gradually over the next several quarters, and it will become our major product next year,” Chiang told Reuters on Wednesday.

In hardware terms, the Vera Rubin NVL72 platform integrates 72 Rubin GPUs and 36 Vera CPUs into a single liquid-cooled rack, linked by Nvidia’s sixth-generation NVLink fabric delivering 260 terabytes per second of total bandwidth—double the 130 terabytes per second of the prior Blackwell generation. This bandwidth advantage is critical for mixture-of-experts model architectures that dominate contemporary large language model design, where high-speed inter-GPU communication dictates performance. Vera Rubin also replaces Blackwell’s HBM3e memory with next-generation HBM4, enabling up to 50 petaFLOPS per GPU—five times the throughput of Blackwell—and reducing inference costs to roughly one-tenth per million tokens.

This development carries particular weight for Nvidia, which relies on Foxconn as its primary manufacturing partner for rack-scale systems. By its own accounting, Foxconn controls over 40% of the global AI server market, a position it has held since the Blackwell ramp. Chiang’s Q4 shipment confirmation provides the hyperscale cloud sector—including AWS, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure, all confirmed Vera Rubin launch customers—with its clearest production schedule to date. Four additional cloud partners, CoreWeave, Lambda, Nebius, and Nscale, are also confirmed for initial deployment this fall.

Despite uniformly strong demand signals and full order books, Chiang identified a single variable that could constrain how rapidly the AI server market scales in 2027. It is not manufacturing labor, factory floor space, or Nvidia’s chip design: it is CoWoS (Chip-on-Wafer-on-Substrate), an advanced packaging technology developed by Taiwan Semiconductor Manufacturing Company (TSMC) that only TSMC can produce at AI-chip scale.

Understanding CoWoS requires understanding its function. When TSMC fabricates a Rubin GPU on a 3-nanometer silicon wafer, the resulting die cannot ship as a functional AI accelerator independently. It must first be bonded to multiple stacks of High-Bandwidth Memory on a silicon interposer—a precision substrate with interconnect densities far beyond conventional PCB-based alternatives. CoWoS executes this bonding process and requires front-end-class cleanrooms that only TSMC operates at commercial scale. Without CoWoS, a finished GPU wafer remains idle. Consequently, the bottleneck for AI server shipments has shifted from silicon fabrication, where capacity is expanding, to advanced packaging, where TSMC’s CoWoS lines remain structurally oversubscribed.

The capacity metrics underscore the pressure. TSMC’s CoWoS output stood at approximately 35,000 wafers per month in late 2024. By the end of 2025, that figure rose to roughly 75,000. The target for the end of 2026 is 125,000 to 130,000 wafers per month—nearly a fourfold increase in under two years—yet still insufficient to meet demand. At TSMC’s June 2026 annual meeting, CEO C.C. Wei confirmed that CoWoS capacity “remains extremely tight and sold out through 2026.” Nvidia, which secures approximately 60% of TSMC’s total CoWoS output, has pre-committed the majority of available capacity through 2027, according to Morgan Stanley.

Chiang’s remarks on the earnings call were explicit: “The market currently expects CoWoS capacity to grow by more than 50% next year, but how much of that can ultimately be translated into shipments of next-generation AI server racks will depend on chip supply,” he said during the call. In practical terms, Foxconn’s Vera Rubin volume in 2027 will not be limited by its own factories. The constraint resides upstream at TSMC and is dictated by how much of TSMC’s 2027 CoWoS allocation Nvidia can secure relative to competing chip designers, including AMD, Broadcom, Google TPU, and Amazon Trainium. Foxconn assembles whatever packaged chips arrive. The limiting factor is not located in Houston or Mexico; it is governed by TSMC’s backend packaging facilities in Hsinchu.

For enterprise buyers planning 2027 AI infrastructure deployments, Chiang’s statement serves as direct guidance. The AI server market possesses sufficient demand to absorb any volume that ships; the uncertainty lies in actual delivery. Even if TSMC expands CoWoS capacity by the anticipated 50% and Nvidia maintains its ~60% allocation share, Nvidia-based AI rack availability could rise significantly. However, analysis from Silicon Analysts indicates TSMC’s backend packaging facilities remain sold out through 2027, with lead times stretching between 52 and 78 weeks. Alternative approaches, such as panel-level packaging and glass substrates, are not projected to reach commercial production scale before 2028.

Chiang also framed the AI buildout’s demand curve as being in its earliest stages. Cloud providers and AI laboratories currently comprise the buyer base, while government and corporate adoption remains nascent. This positioning frames the CoWoS constraint as a 2027 gating event within a broader, multi-year demand expansion—not a ceiling on market growth, but a governor on its velocity.

Foxconn anticipates its capital expenditure to increase by approximately 30% in 2026 compared to 2025, per the company’s May capex guidance. According to Reuters, Foxconn is currently constructing AI server manufacturing facilities in both Mexico and Texas dedicated to Nvidia platform production. While these sites expand Foxconn’s assembly capacity, they do not alter its position relative to TSMC’s CoWoS lines: Foxconn remains downstream of a constraint it cannot control.

Foxconn shares closed 2.7% higher in Taipei ahead of the earnings release.