December 27, 2025
Looking back, it really feels like 2025 was a year overflowing with topics related to computers and AI. The evolution of generative AI has been remarkable. At the same time, issues such as the copyright status of training data, fake videos, and ethical concerns have been growing steadily.
In the United States, it’s said that Generation Z can easily complete school assignments using AI, yet even after graduating from university, many struggle to find jobs—ironically because of AI itself. How AI should be used in education will clearly require much more discussion going forward.
On the hardware side as well, development of semiconductors specialized for AI processing has been accelerating, and architectures different from traditional CPUs and GPUs are drawing attention. These changes seem likely to affect not just technological innovation, but the structure of society itself.
December 28, 2025
With that in mind, I’d like to briefly look back at the industry this year, starting with January through June.
January:
NVIDIA released the RTX 50 series GPUs based on the Blackwell architecture. Jensen Huang delivered a keynote at CES, where the term “Physical AI” was mentioned.
February:
Intel announced additional models in the Xeon 6P series. The P-series is based on Granite Rapids and corresponds to so-called P-core products. Models positioned below last September’s top-end SKU (6900P)—namely the 6700P and 6500P—were introduced. I had expected higher-end models above last year’s mid-range 6700E for the E-series Sierra Forest, but they did not appear. (It also became clear that existing Xeon 6P products had been discounted toward the end of January.)
March:
At NVIDIA GTC 2025, Jensen Huang gave a keynote. NVIDIA announced plans to add Blackwell Ultra alongside the already shipping Blackwell, and also revealed that Rubin, announced last year, would have a Rubin Ultra variant. They also announced reduced power consumption for data center interconnects using MRM + COUPE.
April:
At the TSMC North America Technology Symposium 2025, the 1.4 nm process node (A14) was unveiled. In Japan, it was later introduced at the June Technology Symposium.
May:
NVIDIA announced NVLink Fusion, which standardizes the NVLink specification.
June:
At ISC, the TOP500 ranking placed El Capitan (LLNL) at number one. The previous champion, Frontier (ORNL), dropped to second place. Both systems use AMD EPYC CPUs and Instinct GPUs.
Looking back, the impression is simply that something happened every single month. I’ll look back at the second half tomorrow.
December 29, 2025
July:
Perhaps due to another hot summer, issues with Intel’s Raptor Lake CPUs made the news. My own PC uses an i7-13700K, also a Raptor Lake part, but since it’s liquid-cooled and not pushed particularly hard, I haven’t experienced any problems.
August:
It became public that Intel would receive support from SoftBank and the U.S. government. This raised questions about whether Intel could still be considered a private company if the government became a major shareholder. Lip-Bu Tan, who took office as CEO in March, indicated a policy of maintaining the foundry business.
September:
Reports emerged that Intel had entered into a partnership with NVIDIA and received investment from them. To me, this only made Intel’s difficult situation more apparent. In Japan, NHK aired a feature on Rapidus reaching the stage of prototyping a 2 nm process.
October:
Intel began manufacturing Panther Lake and Clearwater Forest using its 18A process. Clearwater Forest is reportedly built on 18A but categorized as Xeon 6+. It is seen as a replacement for the top-end Xeon 6E model that never appeared with Sierra Forest. In Japan, Fujitsu announced a collaboration with NVIDIA. Back in June, Fujitsu had already outlined a plan for Fugaku NEXT, connecting its next-generation Arm CPU with NVIDIA GPUs via NVLink, but this announcement described a much broader partnership.
November:
The SC TOP500 ranking remained unchanged from June, with El Capitan in first place and Frontier in second. At Arm UNLOCKED 2025 Tokyo, Arm reported that 50% of CPUs procured by cloud service providers are Arm CPUs. Fujitsu also showcased a prototype of the FUJITSU-MONAKA chip currently under development.
December:
At AWS re:Invent 2025, Graviton5 was unveiled and entered service. The Graviton series is also based on Arm CPUs. For AI training, Trainium3 also became available, and plans for Trainium4 were announced at the same time. From around October, AI-driven demand pushed up memory demand, leading to rising prices in memory-related markets. In December, this suddenly became a hot topic as people began talking about significant increases in PC prices next year.
Looking back, the first half of the year was dominated by NVIDIA, the middle by Intel, and the final two months featured many Arm-related developments.
One NVIDIA-related point that caught my attention is the growing narrative that GPUs were originally developed as if they were meant for AI (machine learning) from the start. I forget whether it was at CES or GTC, but Jensen Huang mentioned that NVIDIA had already recognized before 2012 that GPUs were well suited for machine learning, developed the DGX-1 AI cluster system, received no customer orders, and ultimately donated it to OpenAI.
The year 2012 is often cited as the trigger for the third AI boom, sparked by Google’s “cat” experiment, but in reality NVIDIA’s Bill Dally had already discussed GPUs being suitable for machine learning back in the 2000s. Former Intel executive Paolo Gargini also said in a webinar that NVIDIA understood around 2009 that GPUs could be used for machine learning. All of this is true, and it is also true that GPUs later became central to AI workloads—but I want to note that this particular narrative has been heard much more frequently starting this year.
December 30, 2025
Yesterday I looked back over the year, so today I’d like to summarize some recent developments. This industry is always in motion.
◆ NVIDIA passes on adopting Intel 18A
It was reported that NVIDIA decided not to adopt Intel’s cutting-edge 18A process, which it had been considering for next-generation GPU manufacturing. Intel itself is manufacturing Core Ultra 300 (Panther Lake) and Xeon 6+ (Clearwater Forest) using 18A.
◆ NVIDIA enters non-exclusive partnership with Groq for inference technology
NVIDIA licensed technology from U.S.-based Groq, known for its strength in AI inference, under a non-exclusive agreement. The move aims to complement NVIDIA’s GPU strategy, which has focused on training, with low-latency, high-efficiency inference capabilities. Groq engineers have joined NVIDIA to establish a cooperative development structure.
◆ AMD outlines direction for next-generation “Zen 6” CPUs
AMD released information about its next-generation CPU architecture, Zen 6. The CCDs are expected to be manufactured using TSMC’s 2 nm-class process, with a strong focus on AI processing and data center applications. FP16 and INT8 will be supported, and 3D V-Cache capacity is rumored to reach up to 288 MB.
◆ TSMC considers 2 nm production in Kumamoto
It was reported that TSMC is considering producing 2 nm-class chips at its Kumamoto facility in Japan, though TSMC has not commented. Rapidus is currently ramping up a national 2 nm initiative in Hokkaido, so the potential impact is worth watching.
◆ TSMC to raise prices for 3 nm and beyond
TSMC is reportedly considering phased price increases for advanced processes starting at 3 nm. Price hikes of 3–10% are expected in 2026, with rumors of annual increases continuing through 2029.
December 31, 2025
As 2025 draws to a close, it appears that TSMC began mass production of 2 nm chips starting yesterday. If production began by today, it can be said to have started in “Q4 2025.” The site is reportedly the Kaohsiung fab in Tainan, likely Fab 22. The initial products may be smartphone chips. There are rumors that Apple’s chips will be produced at Fab 20, so it’s possible that mass production has begun for non-Apple customers first. That said, next year’s iPhone 18 is widely expected to use the 2 nm A20, so Apple’s production lines should come online before long.
Intel is reportedly set to unveil Panther Lake, manufactured on 18A, at CES starting January 5. AMD plans to manufacture CCDs for EPYC Zen 6 on a 2 nm process as well, with shipments expected in the latter half of next year.
2026 will be the year when GAA transistor products truly begin operating in the real world.