The TSMC manufacturing pivot and the Tensor G5 yield risks

TSMC manufacturing and the yield gap

Google uses TSMC 3nm for the Tensor G5. This shift follows reports that Samsung Foundry’s 3nm yield for the Exynos 2500 dropped below 20%. Samsung’s 2nm yield for the Exynos 2600 also stays below 50%, while TSMC maintains 70% to 80% yields for its 2nm process. While Google moved to TSMC 3nm to bypass Samsung’s low 3nm yields for the Exynos 2500, the company must still battle massive demand for TSMC capacity from Apple and Qualcomm. This competition forces Google to fight for manufacturing space against companies like NVIDIA and AMD. Samsung faces a 1 trillion won loss in its foundry division and struggles to secure large orders for advanced nodes. Most customers currently give TSMC a 95% share of orders, leaving Samsung with only 5% for advanced processes. Samsung’s recruitment efforts struggle against Tesla’s "Terafab" project, which targets senior engineers in South Korea and Taiwan to build sub-3nm capacity. Samsung reported a 60% yield for 2nm chips for clients like Canaan and MicroBT in March 2026, but internal Exynos production lags. Samsung’s 3nm process also shows 10% to 20% lower power efficiency compared to TSMC’s 3nm. TSMC’s production capacity for 3nm is expected to remain sold out through 2026 because of the heavy demand for AI-capable hardware from Apple and NVIDIA.

Thermal management and 8K video failures

I see the Pixel 10 Pro handles 15 minutes of 8K video in the sun after the latest thermal management patch. This software update optimizes power draw and manages background tasks more aggressively when the camera app runs. Without this patch, the device suffered a "Zoom Freeze" bug where the frame locked up during lens transitions. The base Pixel 10 performs worse because it lacks the vapor chamber used to cool the Pro models. This hardware omission causes the base model to struggle with stability during sustained heavy tasks. You know the Tensor history involves constant thermal struggles. Google’s gChips division internal documents show the Tensor program is not meeting financial goals. Google previously addressed consumer complaints regarding devices that overheat and battery life that fails to last a single day, noting that 86% of older Pixel owners struggled with daily usage. While older Pixel owners wanted 36-hour battery life, the G5 focuses on better efficiency through TSMC’s 3nm process. The G5 delivers 34% faster CPU performance than the G4, but software still dictates thermal success.

Can the Tensor G5 beat the Snapdragon rivalry?

The Tensor G5 falls short of the Qualcomm Snapdragon 8 Elite. The Snapdragon 8 Elite hits 38.27 FPS in 3DMark Wild Life Extreme, while the Tensor G5 Pro only hits 19.49 FPS. The G5 uses an Imagination Technologies DXT-48-1536 GPU, which provides ray tracing but fails to compete in gaming. The GPU performance remains a massive disappointment. The Exynos 2500 actually beats the Tensor G5 in multi-core tests. The base Pixel 10 uses 12GB of RAM and UFS 3.1 storage, while the Pro models use 16GB of RAM and UFS 4.0. In Geekbench 6 single-core tests, the G5 scores 2,285 compared to the G4’s 1,929. In multi-core, the G5 hits 6,191 against the G4’s 4,586. The Honor Magic 7 Pro hits 2,965 in single-core, and the Samsung Galaxy S25 Ultra reaches 3,001. The G5 TPU provides a 40% increase in TOPS, yet real-world AI tasks only see a 14% improvement. Gemini Nano runs 2.6x faster on the G5, whereas the G4 was slower. The G5 uses a 3.78 GHz Cortex-X4 core, while the G4 relied on 3.10 GHz. Can Google bridge the performance gap against Qualcomm with just better manufacturing partners? I find the Tensor G5 a frustrating upgrade for anyone prioritizing gaming or heavy multitasking.