January 1, 2026
There seems to have been a great deal of movement in semiconductor- and computer-related news even toward the end of the year.
Following the U.S. decision to lift export restrictions on the H200 for China, NVIDIA has reportedly already placed orders with TSMC to manufacture the H200. Until now, only a downgraded version, the H20, was permitted for export, but with this decision the H200 has also been approved. For the Blackwell architecture, NVIDIA has prepared the B40, but the B40 uses GDDR7 memory, which differs from the HBM3 (3e) used in the H20/H200. With the H200, it should be possible to carry out larger-scale training than with the B40. Incidentally, both Hopper and Blackwell are manufactured using TSMC’s N4-series process.
TSMC has reportedly obtained permission to export manufacturing equipment purchased from the United States to its fab in Nanjing on mainland China. TSMC operates several fabs in mainland China; the Nanjing site is likely Fab 16, which, as the number suggests, dates back quite some time. The transistor process nodes used at Fab 16 are said to be 28/22 nm and 16/12 nm. The 28 nm node is at roughly the same level as JASM in Kumamoto (TSMC Fab 23). TSMC’s 28 nm process uses a conventional planar structure, nodes from 16 nm onward use FinFETs, and from 2 nm onward the company will transition to GAA-type transistors.
January 2, 2026
By the way, over the year-end holidays I watched NHK Plus’s program “The Century of Video: Butterfly Effect.” Nostalgic scenes of Tokyo were picked up from fragments of various films. Since I myself came to Tokyo only after the start of the Heisei era, footage from the Showa period came back to me as fond memories of things I had seen on television as a child. In contrast, footage from after the collapse of the Heisei bubble reminded me, “Yes, that’s what it felt like back then,” as something I had actually experienced.
In the early part of the Heisei era, the internet had not yet reached society at large. There were no smartphones, not even mobile phones, and people younger than us were using pagers. Now, I can hardly even remember how we used to arrange to meet people back then. It almost seems mysterious that we could meet up without carrying smartphones. Were we all somehow psychic back then?
January 3, 2026
Looking at the news ahead of CES 2026, which starts in Las Vegas on the 5th, it seems that gaming PC CPUs will be particularly prominent this time.
AMD will have a keynote by Lisa Su. There are rumors that the Ryzen 7 9850X3D, the Ryzen 9 9950X3D2, and a Ryzen 9 APU may be announced. (Apparently, Ryzen 8 is reserved for mobile, and the number is skipped for desktop products.)
Intel has already announced that it will unveil Panther Lake (Core Ultra 300). There is no deep reason for this, but I am personally hoping to see photos of 18A wafers for the computing tile. The 4Xe GPU tile is manufactured on Intel 3, but the 12Xe tile is reportedly made by TSMC. Since NPU5 is included, AI performance will also be worth watching.
Qualcomm will announce the Snapdragon X2 at a media event. It is an Arm CPU supporting SVE2 and SME (apparently not SME2). It uses Nuvia’s Oryon cores and emphasizes better power efficiency than Intel/AMD x86 CPUs. New Windows on Arm laptops are expected.
NVIDIA is said to be holding a media briefing by Jensen Huang. There will be no new GPUs, but laptops equipped with the RTX 50 series are rumored. It may be better to look forward to the March GTC rather than CES.
IBM is also holding several sessions related to AI and quantum computing. Within CES, an international quantum conference, A World Quantum Congress, will be held.
January 4, 2026
With CES approaching, there is little other industry news, but an article reported that wafer costs at TSMC’s U.S. fabs are likely to be high. The main reason seems to be that depreciation costs in the U.S. are about four times those in Taiwan. Overall, profit margins are said to fall to about one-eighth. For geopolitical risk mitigation, operating fabs in the U.S. is unavoidable. The chip war will likely continue in 2026, especially with rising memory prices.
Another piece of news is that NVIDIA is reportedly investing $1 billion in Nokia. Perhaps this is aimed at smartphones and base stations as the endpoint for inference-oriented edge AI. SoftBank has been promoting AI-RAN since last year, and if NVIDIA joins in, deployment could accelerate.
January 5, 2026
With that, the first day of work is done. CES starts late tonight. It runs from January 5 to 8 local time in Las Vegas, with pre- and post-events on either side. I’ll be waiting for media reports and information released by participating companies.
January 6, 2026
As expected, news sites are filled with CES headlines. CPU-related announcements are largely in line with prior expectations. New laptops were also announced as rumored, with HP unveiling models equipped with new CPUs from Intel, AMD, and Qualcomm.
All three new CPU chips include NPUs, signaling the arrival of the AI PC era. Perhaps because Microsoft Copilot requires 40 TOPS, a minimum of 50 TOPS is being targeted.
Intel Panther Lake Core Ultra 300 series (NPU5): 50 TOPS
AMD Ryzen AI 400 series (XDNA2): 50–60 TOPS
Qualcomm Snapdragon X2 series (Hexagon NPU6): 80 TOPS
NVIDIA did not appear to make any new announcements, but it did reveal that production of Vera Rubin has ramped up in earnest. Since it uses TSMC’s N3 process, mass production should be feasible. (N2 also entered mass production toward the end of the year.)
With CPU/GPU announcements seemingly complete, I’d like to read through everything in detail over the weekend.
January 7, 2026
The news about NVIDIA’s Vera Rubin at CES was particularly interesting. NVIDIA announced that it has prepared six chips as the next-generation AI platform.
Rubin (GPU): 50 PF + HBM4
Vera (CPU): 88 cores + LPDDR5
NVLink 6 Switch: 3,600 GB/s
ConnectX-9: 800 Gb/s
BlueField-4 DPU: Grace (CPU) + CX-9 800 Gb/s
Spectrum-6 Ethernet Switch: 102.4 Tb/s + CPO
Applying HBM4 to a GPU product may be something new.
As announced in advance, NVIDIA has incorporated CPO (Co-Packaged Optics) into its Ethernet switches. This is expected to reduce communication power consumption in AI data centers to about one-third per package while maintaining the same speed (1.6 Tb/s). The MRM technology mentioned in the presentation is likely being used.
January 8, 2026
This may also be CES-related news, but it seems that comments by NVIDIA’s Jensen Huang suggesting that data center cooling may not need to be so low-temperature caused water-cooling-related stocks to fall. I haven’t looked closely at the details, but as semiconductor technology advances, the next generation is expected to adopt backside power delivery networks (BSPDN). BSPDN requires thinning the silicon layer that forms the transistors. Current wafers are roughly 700 µm thick for silicon, with about 20 µm for wiring layers, but with BSPDN the silicon thickness is expected to be under 100 µm, perhaps around 50 µm. Signal wiring layers are built on the front side and power wiring on the back side, so the total thickness from front to back may be under 200 µm. In that case, heat generated by transistor operation can be dissipated more efficiently than with conventional single-sided silicon. In other words, the total thermal resistance from transistor to cooling pad is reduced, meaning water cooling may not need to be as cold as it is now. Cooling power consumption is significant, so reducing cooling requirements would greatly help operating costs.
Unrelated to CES, there is information that TSMC has halted development of new 3 nm projects. Customers requesting 3 nm production may be encouraged to move to 2 nm instead. In other words, 3 nm capacity may be sold out. Although there is still demand and supply is tight at 3 nm, mass production of 2 nm has begun, so TSMC likely wants to focus on that line and recover its investments more quickly.
It also seems that AWS has raised GPU instance prices by around 15%. Rising prices are now reaching cloud service fees as well.
January 9, 2026
Intel announced Panther Lake at CES, but it seems that a lower-cost version called Wildcat Lake has also appeared. Even the code names—Panther and Wildcat—give the impression of a budget variant. I have a feeling that performance is being intentionally limited for the Chinese market, likely targeting the mini PC segment.
When it comes to wildcats, Lynx is the most familiar name (from browsers and constellations), but bobcats, pumas, and cougars are also wildcats. I thought Panther referred to a leopard, but leopards are actually “leopard,” while panther refers to large wildcats such as leopards or jaguars. Speaking of jaguars, I occasionally see “Chiba no Jaguar-san” on X.
January 10, 2026
It seems that after completing the Panther Lake announcement, Intel has stated that it will focus on 14A. The 18A node is expected to be used only for Intel’s own products, but industry observers believe there may be external customers for 14A. Logically, NVIDIA—having given up on 18A—could be a candidate.
AMD announced its rack-scale computing platform Helios at CES. It is said to deliver 3 exaflops per rack. However, this is likely not in double precision (FP64); viewed as half precision (FP16), it would be roughly equivalent to about 1.5 Fugaku systems. For AI performance, it may be FP8. Since racks can be scaled out, it is said to support yotta-scale computing. Because AMD develops both CPUs and GPUs in-house, it can vertically integrate AI support. This is seen as a competitor to NVIDIA’s GB200 NVL72 and its successor, the VR200 NVL72.
January 11, 2026
By the way, one of the announcements made at CES 2026 was that Jensen Huang, founder and CEO of NVIDIA, has been selected as the recipient of the 2026 IEEE Medal of Honor. His consistent R&D efforts since the development of the GPU in 1999, as well as his leadership in industry and impact on society, were highly recognized. The award ceremony will be held in New York in April. For reference, AMD’s Lisa Su received the IEEE Robert N. Noyce Medal in 2021.
Three days ago, on January 8, I wrote that TSMC appeared to have stopped accepting new orders for its 3nm process. The original source was Taiwan’s Industrial and Commercial Times on January 8, and the article was reposted on SemiWiki on January 9.
January 12, 2026 CES2026
I finally watched all three keynote speeches from NVIDIA, AMD, and Intel at CES 2026. As expected, the amount of information was overwhelming. Below are my notes on keywords, guests, and major topics, in the order I watched them. NVIDIA and AMD in particular feel like they are in a head-on competition, given how closely their business domains now overlap.
◆ NVIDIA: ~1.5-hour keynote by Jensen Huang
Keywords: Autonomous driving platform Alpamayo, DGX10, Physical AI platform, five-layer AI stack, robotics, collaboration with three EDA vendors, partnership with SISW, Vera Rubin POD
Five-layer AI stack (example: autonomous driving):
Top layer: Applications – Mercedes-Benz
Layer 2: Infrastructure models – Alpamayo (open source)
Layer 3: Training and model generation – Omniverse, Cosmos
Layer 4: System configuration – GPU, CPU, networking
Bottom layer: Physical infrastructure – vehicle body
The Vera Rubin NVL72 rack was introduced:
Weight: 2.5 tons (including cooling water)
NVFP4 Inference: 3.6 ExaFLOPS
NVFP4 Training: 2.5 ExaFLOPS
Memory: 54TB (LPDDR5), 20.7TB (HBM4)
HBM4 bandwidth: 1.6 PB/s
Scale-up bandwidth: 260 TB/s
Components:
CPU: Vera (88 cores + LPDDR5)
GPU: Rubin (HBM4)
NVLink 6 Switch: 3,600 GB/s
NIC: ConnectX-9 (800 Gb/s)
DPU: BlueField-4 DPU (Grace CPU + CX-9 800 Gb/s)
Spectrum-6 Ethernet Switch: 102.4 Tb/s + CPO
The Vera Rubin compute tray has no wiring—only slots.
Cooling water temperature is 45°C, requiring no dedicated chiller.
The idea that a server tray can be assembled simply by plugging VR superchip cards into slots, with almost no wiring or piping, is very well thought out. It clearly shows NVIDIA’s intention to mass-produce at scale. The rack weighs 2 tons dry, but apparently someone forgot to drain the water, and it was shipped at 2.5 tons. Even Jensen Huang joked, “We shipped a lot of water from California to Las Vegas.” Granted, even a desert city like Las Vegas probably has cooling water available—and for an exhibition, cooling water wasn’t even necessary in the first place.
◆ AMD: ~2-hour keynote by Lisa Su and guests
Keywords: Cloud, Helios (GPU/CPU), PC, Ryzen AI 400, Ryzen AI Halo, gaming, healthcare, physical AI/robotics, space, HPC roadmap, science and education
Guests: OpenAI, Luma AI, Liquid AI, World Labs, Absci, Illumina, AstraZeneca, Generative Bionic, Blue Origin, White House (OSTP)
The Helios rack was introduced:
Weight: 7,000 pounds (just over 3 tons)
2.9 ExaFLOPS (AI compute)
31TB HBM4
43 TB/s bandwidth
2nm/3nm process, 4,600 cores (Zen 6)
18,000 GPU units
Components:
CPU: EPYC Venice (Zen 6), 256 cores
GPU: Instinct MI455X
DPU: Pensando Salina 400
NIC: Pensando Vulcano 800 AI NIC
The presence of a White House guest likely reflects AMD’s selection as a participant in the Genesis initiative. Lisa Su herself also served on the President’s Council of Advisors on Science and Technology (PCAST) under the Biden administration, where she was involved in reports related to engineering education. The final hackathon award ceremony likely followed from that context.
Incidentally, EPYC Zen 6 Venice appears to use a silicon interposer rather than an organic substrate, suggesting a significant design change.
◆ Intel: Opening by Lip-Bu Tan, followed by a 45-minute keynote by Jim Johnson
Keywords: Panther Lake Core Ultra 300 series, 18A process, RibbonFET, PowerVIA, 12Xe GPU, Arc GPU, handheld devices, AI PC, XMX (matrix engine in GPU), ByteDance, Perplexity, Comet, AI agents, VLM
Guests: Electronic Arts, Perplexity
Panther Lake appeared largely as expected. Intel’s GAA transistor, branded as RibbonFET, and BSPDN branded as PowerVIA, are finally entering the market. GPU performance, in particular, has been receiving high praise.
Compared with NVIDIA and AMD, Intel did not emphasize rack-scale or data center scale-out solutions. However, since CES is fundamentally a consumer electronics show, focusing on gaming and comfortable laptop experiences may be entirely appropriate.
January 13, 2026
It appears that Intel has provided a 12-qubit quantum device to Argonne National Laboratory (ANL). Quantum devices leverage semiconductor manufacturing technologies. Since 2023, Intel has been offering a research-oriented quantum device codenamed Tunnel Falls.
TSMC’s earnings announcement is approaching on January 15. U.S. Trump-era tariffs are expected to settle at around 15%, but in exchange, TSMC is reportedly planning to build four additional fabs in the U.S. At Fab 21 in Arizona, 4nm is already in operation, while the 3nm building has been completed and is expected to begin tool installation this year and production next year. Since TSMC expands fabs in “phases,” Phase 1 corresponds to 4nm and Phase 2 to 3nm. If four more phases are added, that would imply expansion up to Phase 6. In addition to wafer fabs, packaging facilities (AP fabs) are also likely to be added. Fab 21 may expand to P3 (2nm) and P4 (A14), alongside AP9 or AP10 and another packaging fab (tentatively AP11) in the U.S.
I hadn’t mentioned this before, but Apple reportedly unveiled a crease-free display at CES, likely intended for the next iPhone. Apple also seems to have partnered with Google’s Gemini as part of its AI infrastructure. I thought Apple Intelligence was based on ChatGPT.
January 14, 2026
Intel has announced a workstation-oriented Xeon 6 CPU: the Xeon 6 698X, featuring 86 cores / 172 threads, 4.6 GHz, and 336 MB of L3 cache. This is Granite Rapids, likely built on the Intel 3 process. It wasn’t shown at CES, probably to keep the spotlight on 18A. Meanwhile, AMD announced Zen 5 (4nm)-based Ryzen CPUs at CES, so while Intel took the lead with 18A Panther Lake, it may be reinforcing its workstation lineup with Intel 3-based Granite Rapids. Intel’s 18A Xeon lineup will start with Clearwater Forest (E-cores) as Xeon 6+, while performance-focused Diamond Rapids (P-cores) is expected later as Xeon 7.
Although NVIDIA’s H200 has reportedly received U.S. export approval to China, the Chinese government has not yet approved imports. This may indicate moves to protect domestic semiconductor industries.
January 15, 2026
Intel reportedly has nearly sold out its data center server CPUs for the entire year, according to financial media. “The entire year” still means another 11.5 months, but it’s unclear whether this refers to Xeon 6 P-cores, E-cores, or both. My guess is both.
Xeon 6+ E-core (Clearwater Forest) is expected in the first half of 2026, and Xeon 7 P-core (Diamond Rapids) in the second half. This likely means all remaining Intel 3-based inventory and planned production has been sold. From here on, Intel can focus on selling 18A-based Xeon 6+ and Xeon 7.
AMD is also reportedly planning to raise server CPU prices by around 15%. AMD’s Zen 6 Venice (TSMC 2nm) is scheduled for late 2026. However, about 50% of CPUs procured by hyperscalers in 2025 were Arm-based, so it remains to be seen whether these price increases will further benefit the Arm ecosystem.
On the NPU front, OpenAI’s inference chip Titan, developed on TSMC 3nm, has reportedly entered production and is scheduled for release by the end of the year. Inference can be handled by CPUs, but NPUs are generally more power-efficient. Training will likely continue to rely on GPUs for some time, but inference may increasingly see competition between CPU inference engines and dedicated NPUs—or perhaps a coexistence with effective offloading, depending on compilers and software.
January 16, 2026
With TSMC’s earnings announcement yesterday, semiconductor news has been dominated by that topic. As expected, the results were strong, yet it has become clear that TSMC’s current capacity is still insufficient to meet AI demand. Semiconductor demand shows no signs of slowing.
Supply shortages inevitably affect product lineups. NVIDIA is reportedly starting to limit shipments of the RTX 50 series. While the GPUs themselves are Blackwell-based, the constraint appears to be related more to LPDDR7 memory availability.
AMD has announced a partnership with Tata Consultancy Services (TCS) of India, with AMD solutions supporting TCS’s business initiatives. The results of this collaboration will likely be presented at a future CES or similar event.
OpenAI has also reportedly partnered with Cerebras, known for its wafer-scale engine (WSE), for a three-year period. Cerebras produces what is essentially an entire 12-inch wafer cut into a square—something that initially left many in the industry thinking, “Sorry, I don’t quite understand what this is.” The WSE is a dedicated AI engine, and this partnership means OpenAI has secured additional AI compute capacity. For Cerebras, this may finally be their moment.
January 17, 2026
Intel has apparently released a new CPU called Bartlett Lake. It’s not for general retail, but for embedded applications, meaning it’s only available through specialized vendors. It’s a P-core-only product using the Raptor Cove architecture. Benchmark results for the Core 7 253PE (10 P-cores / 20 threads) show about 20% higher multi-threaded performance than the Raptor Lake Refresh Core i5-14400 (6P + 4E / 16 threads), and about 2% higher than the i5-14500 (6P + 8E / 20 threads). On a per-thread basis, the results seem fair.
While NVIDIA’s H200 has received U.S. export approval, Chinese import approval is still pending. By contrast, AMD’s MI325X has reportedly received U.S. export approval. It’s possible China is reluctant to pay Trump-era tariffs.
AMD is well known for stacking L3 cache on CCDs using 3D V-Cache in EPYC and Ryzen CPUs. Now there are reports—based on patent filings—that L2 cache may also be stacked. This reportedly reduces access latency from 14 cycles to 12 cycles (a 14% improvement). It should also reduce CCD area, and suggests that L2 cache may not need to use the same process as the cores. It feels like the industry has moved forward another step.
January 18, 2026
It appears that SiFive, a fabless RISC-V company, plans to support NVLink. With NVIDIA investing in Intel, Arm already supporting NVLink, and now RISC-V joining, x86, Arm, and RISC-V are all in the game.
Apple is reportedly preparing the A20 Pro chip, manufactured on TSMC 2nm, for the iPhone 18 Pro—possibly a foldable model. 2026 will likely be the year GAA reaches consumer products. TSMC’s 3nm mass production began on December 29, 2022, making 2022 effectively the “3nm year.” Similarly, 2nm mass production began on December 30, 2025, making 2025 the “2nm year.” The roughly three-year cadence per node is being maintained, although the scaling factor no longer seems to be 0.7. Even if transistor scaling slows, 3D integration can compensate—though power remains an issue. The iPhone 18 Pro is also rumored to use WMCM (wafer-level multi-chip module) packaging.
TSMC’s Kumamoto fab (JASM) was thought to only support planar 28/22nm processes, but it now appears to include 16/12nm FinFET processes as well. At 12nm, EUV is not required. A second fab planned for 6nm is reportedly on hold due to process re-evaluation. Whether leading-edge nodes are feasible depends on the NA of available EUV tools, and TSMC may be consulting ASML. Moving to 2nm could put it in competition with Rapidus, but if they serve as mutual backups, it could strengthen BCP and attract orders.
An older article noted that AWS has begun offering sovereign cloud services in Europe. Incidentally, the Davos conference begins next week.
January 19, 2026 Time of AI Semiconductor
AWS has reportedly launched a new Xeon 6-based instance, Amazon EC2 X8i. Given references to large memory capacity, all-core turbo, and in-memory databases (SAP/HANA), this is likely based on Xeon 6 P-cores. Xeon 6E 6900E has not yet appeared, with Xeon 6+ expected in the first half of the year. AWS is likely an early user of 6900E for testing, but E-core-based data center CPUs may have limited market presence. This is something I’m watching closely.
Two more AWS-related items:
Amazon announced a partnership with mining company Rio Tinto to support copper mining. The initiative aims to reduce environmental impact through IT-driven optimization of low-carbon copper mining and refining in the U.S. The copper will of course be used for data centers and electronic circuits.
AWS also announced a partnership with SUDO Consultants in the Middle East. At re:Invent last December, AWS announced a partnership with Saudi Arabia’s HUMAIN to promote AI cloud services. Following its recent sovereign AI announcements, AWS has been in the spotlight repeatedly.
In summary, TSMC’s earnings clearly highlighted how the AI boom is driving semiconductor demand. Before AI, it was smartphones and EVs (since ~2010), before that servers and PCs (since ~1995), and before that DRAM (since ~1980). From 2025 onward, AI will likely be the main driver. This suggests a roughly 15-year cycle for dominant technologies—implying AI could last until around 2040, though the pace feels so fast that the next shift may come sooner.
What’s truly striking about the AI semiconductor boom is that even the entirety of existing semiconductor manufacturing capacity is already insufficient to meet demand. China’s entry into smartphones and EVs since 2010 has massively increased demand, but within China, shortages of advanced semiconductors at 7nm and below are estimated to amount to as much as one million wafers. U.S. export controls on manufacturing equipment and TSMC’s concentration of leading-edge technology both contribute. TSMC follows an “N-2 rule,” exporting only technologies two generations behind the leading edge. This underpins the so-called “silicon shield” that protects TSMC geopolitically.
Computers have long been considered strategic goods subject to export controls, but their importance has grown even further as the AI boom accelerates this strategic dimension. I plan to continue watching where semiconductor technology heads next.
The Davos conference has begun in Switzerland. This year reportedly has the highest participation of heads of state in history.
January 20, 2026
Intel’s earnings announcement is scheduled for January 22 (January 23 Japan time). With the release of 18A-based Panther Lake and reports that yields have exceeded 60%, market sentiment appears positive. The CEO’s commitment to focusing on 14A has also been well received. With TSMC’s capacity fully booked by AI demand, there is speculation that customers may shift toward Intel. If Intel Foundry can secure external customers at 14A, this could mark a true turnaround. Intel’s Ohio fab, whose construction has been delayed, is currently targeting operations by 2030—but if that timeline moves forward, it would be a strong signal. Two years ago, when Pat Gelsinger stepped down and Intel was removed from the Dow (replaced by NVIDIA), it was quite a shock. Given the amount of external investment Intel has received, a recovery is essential—and it’s encouraging that the external environment seems to be aligning in Intel’s favor.
NVIDIA is reportedly planning to bring Arm-based laptops to market. The CPU (CPU+GPU) will be N1X, the same as the GB10 used in DGX Spark. The CPU consists of 20 Grace Arm cores, meaning Neoverse V2. The GPU is Blackwell—at full power equivalent to an RTX 5070, though power consumption remains a question.
Interestingly, both GB10 and N1X are reportedly manufactured on TSMC 3nm. Grace and Blackwell were thought to be N4-based, so this may be for power efficiency.
Windows 11 laptops based on N1X are expected in the first half of the year. Qualcomm’s Snapdragon X2 was announced at CES, and NVIDIA likely wanted to announce N1X there as well, but Windows readiness may have lagged. The growth of Windows on Arm PCs is an exciting development for AI PCs. Compared to x86, Arm seems to offer advantages in tighter CPU–GPU–NPU integration. Personally, I had expected RISC-V to take that role, but perhaps not yet.
Tesla is reportedly restarting its DOJO project for autonomous driving AI training—this time as DOJO 3. DOJO 2, intended for training, was canceled last August. Tesla’s AI5 chip, capable of both training and inference, has reportedly completed design, following the current AI4 inference chip. Manufacturing will be handled by TSMC and Samsung, and the same chip is expected to be used in DOJO 3. AI6 samples are expected within 2026, with mass production next year, and Samsung will be the sole manufacturer.
January 21, 2026
With that said, things are busy, so I’d like to take a break from semiconductor and computer industry topics today and tomorrow. The discussions coming out of the Davos conference also look interesting, and I’d like to catch up on them over the weekend. Trump-style rhetoric is, of course, something to watch, but I’m also curious about how topics such as AI, energy issues, climate change countermeasures, and the widening inequality often described as a “K-shaped economy” are being discussed.
Just one note. It appears that TSMC has approached Apple with a significant price increase. As seen in TSMC’s earnings, AI has become its main profit driver, meaning smartphones are no longer the primary source of revenue. The A20 Pro, expected to be used in the iPhone 18 Pro, is likely being manufactured on the 2nm process, and iPhones may become even more expensive than before. Given the weak yen, prices in Japan could rise further.
January 22, 2026
I said I’d be taking a break from semiconductor watching through today, but just one item.
Over the past couple of days, there’s been growing discussion about TSMC’s packaging plant AP7, which appears to have been built in Chiayi, Taiwan. This is an advanced packaging facility, and technologies such as CoWoS and WMCM—previously concentrated at AP3—are expected to shift there. It is also believed to support SOW-X, frequently mentioned in NVIDIA presentations, as well as the more recently introduced CoPoS (Chip-on-Panel-on-Substrate).
Part of the attention AP7 is receiving comes from articles emphasizing its enhanced safety measures. There were reportedly fatal accidents last year, and safety protocols have since been strengthened. While wafer fabs are largely automated, packaging still seems to leave room for manual handling—especially compared to printed circuit boards. Whenever human operation or transport is involved, accidents are hard to avoid. That said, efforts to reduce accident rates do matter. Stay safe.
January 23, 2026
Intel’s earnings were announced. Judging from the news coverage, many articles are pessimistic, noting that the stock price fell. This likely reflects a pullback after buying pressure through yesterday—essentially, results were not quite as good as expected.
While Intel did return to profitability at the consolidated level, Intel Foundry remained in the red, even though losses narrowed. CEO Lip-Bu Tan summarized the situation as “still being on a turnaround journey,” which seems like an appropriate assessment.
Several articles mention that the reason Xeon 6 sold out early this year was a sudden surge in demand from hyperscalers (cloud providers). Intel is apparently even using PC production lines to manufacture them, running close to hand-to-mouth. Xeon 6 uses Intel 3 for the core tiles and Intel 7 for I/O, which overlaps with Intel 7 lines used for Raptor Lake and Intel 3 lines used for Panther Lake’s Xe3. Intel clearly wants to prioritize Xeon, but can’t entirely stop producing PC CPUs.
Looking ahead, Xeon 6+/7 cores and Panther Lake cores will both move to 18A, and Nova Lake is also expected on 18A by the end of 2026.
In short, Intel Foundry appears to be capacity-constrained. Given that TSMC itself says it cannot meet AI-driven demand, it’s not surprising that Intel’s lines are also full. The bottleneck may well be the supply of ASML’s EUV lithography tools—without EUV, anything beyond 5nm is essentially impossible. Intel is a company that once led the industry by fundamentally changing CMOS transistor fabrication at 45nm and 22nm. Seeing Intel Foundry described as a financial burden is, frankly, a little sad.
January 24, 2026
There are reports that NVIDIA has halted shipments of the RTX 50 series for six months, and that it has ended its open price program. Both appear to stem from shortages of GDDR7 memory chips. Looking at domestic prices on Kakaku.com, even lower-end RTX 30-series cards seem to have risen by about ¥10,000 since the end of last year. More than that, the number of available models feels noticeably reduced.
On X, I occasionally see posts about people pulling DDR4 DIMMs from junk systems and selling them for decent prices. There was a time when DRAM was called “the rice of industry.” After last year’s surge in the price of actual rice in Japan, will industrial “rice” also become expensive this year?
January 25, 2026 DAVOS2026
I watched several Davos conference videos—though it’s impossible to watch everything. I chose four: NVIDIA’s Jensen Huang, Tesla’s Elon Musk, Microsoft’s Satya Nadella, and Palantir Technologies’ Alex Karp. There are others I’d like to watch eventually.
I’ll write down some impressions. Since it’s Sunday, this may run a bit long.
◆ Tesla & SpaceX: Elon Musk
Toward the end, he said something like, “It’s more fun to be optimistically wrong than pessimistically right.” His direct engagement with energy issues is genuinely impressive. He’s thinking seriously about space-based solar power. It’s an idea anyone can imagine, but he’s probably the only one who can actually execute it. (There were reports some time ago about experimental success in Japan.)
He mentioned that on Earth, only about 20–25% of solar generation capacity can be relied on as stable power, whereas moving it into space increases efficiency by roughly five times—almost 100%. About 12–13 years ago, when the global population was around 7 billion, he reportedly said that if 70% of people lived in cities and everyone lived like those in developed countries, Earth would need to be 1.8 times its current size. Now the population is about 8 billion, projected to exceed 10 billion before stabilizing around 9 billion later this century.
Energy demand will only grow, fossil fuels will eventually run out, and a successor energy source is necessary. Musk’s perspective seems to view the solar system itself as an energy system, with the sun as the only true source. That seems reasonable. Oil, of course, remains indispensable—not just as energy, but as material. Just look inside any computer. Solar energy may indeed be the only path that reconciles population growth with sustainability.
◆ NVIDIA: Jensen Huang
He reiterated the concept of Physical AI and the five-layer “AI cake” he discussed at CES 2026. Few people are better at industrializing AI than he is. From the bottom up: physical infrastructure; hardware (GPU, CPU, networking); cloud and edge services; AI models; and finally, the application layer. Anyone with a background in communications engineering will recall the OSI seven-layer model.
Physical AI doesn’t just mean robots—it refers to AI that understands physical laws well enough to move limbs correctly and collaborate with humans. I first heard this term a year ago, but I feel like I finally understand it now.
Using the example of radiology technicians, he explained that AI doesn’t take jobs; it automates tasks, allowing professionals to focus on their core roles, improve service quality, attract more customers, and ultimately create more jobs. Perhaps skilled professions will be reevaluated in this way. In Japan, the distinction between engineers and technicians is often blurred, but internationally, engineers design and define specifications, while technicians execute them—though one person can be both. According to international engineering standards, engineers are expected to have university-level education.
Work isn’t just about earning money; contributing quality labor to society also matters. AI, then, is a tool to automate tasks and let humans focus on jobs. If services improve, businesses grow, and more people are needed.
◆ Microsoft: Satya Nadella
Compared with Tesla or NVIDIA, Microsoft is a long-established company, and that historical perspective gave weight to the discussion. Microsoft today is a cloud service provider offering Copilot. While NVIDIA talks about “AI factories,” Microsoft speaks of “token factories”—spreading tokens worldwide and improving tokens per dollar or per watt. He mentioned that token prices halve roughly every three months.
Given the immense compute demand of AI and the current inability of semiconductor manufacturers to keep up, this makes sense. Token generation is skyrocketing, and prices should fall. For white-collar workers, this is a major threat. From a management perspective, traditional white-collar labor is expensive, making layoffs inevitable unless workers defensively adopt AI themselves. This finally helped me logically understand why we’re told to “use AI” at work. Productivity has risen with DX, but wages haven’t kept up—unit labor costs are falling.
When asked how individuals should engage with AI in a future of multiple models, he suggested distilling multiple models into a personal one, feeding it personal data, and using it for individualized feedback. It was an enlightening discussion overall.
He also referred to AI as a “cognitive amplifier,” which reminded me of IBM’s old term “cognitive computing” during the early Watson days.
◆ Palantir Technologies: Alex Karp
I don’t know Palantir well, but I recall seeing posts on X about U.S. federal directives in 2025 to consolidate state data through Palantir, sparking resistance from some states. They seem to be a data analytics company.
In the latter half, he talked about identifying aptitudes that traditional entrance exams cannot detect, and placing people in roles that truly suit them. Like Jensen Huang, he emphasized that skills will become more important. White-collar work is rooted in university education, but that doesn’t guarantee it aligns with one’s true aptitude. Palantir claims it can analyze this more accurately, and apparently operates this way internally.
He also addressed concerns—raised earlier with Satya Nadella—about how to spread AI globally and evenly. In developing countries, AI access is often limited to the highly educated, potentially worsening economic inequality and accelerating the K-shaped economy. This concern predates COVID and remains unresolved.
January 26, 2026
Regarding NVIDIA’s H200 not being importable into China, it appears that the Chinese side has begun moving toward approval. Since the CEO recently visited China, perhaps the winds have shifted, though I haven’t read the details.
Another NVIDIA item: the Earth-2 family of AI-based Earth models for weather prediction has been released. “Family” refers to multiple models optimized for different purposes, with larger models coming later this year. These sit in the second layer from the top of NVIDIA’s five-layer AI stack. At CES, NVIDIA introduced Alpamayo for autonomous driving; Earth-2 seems to be the equivalent for climate and weather modeling. NVIDIA is providing these AI models openly.
The DAWN supercomputer at the University of Cambridge has reportedly been upgraded with new CPUs and GPUs. Originally launched in 2023 with Intel Xeon 4 (Sapphire Rapids) and Data Center GPU Max (Ponte Vecchio), it will reportedly switch to AMD MI355X GPUs by spring 2026. The CPUs will likely change as well, though details weren’t provided. A typical pairing would be MI355X with Zen 5c (Turin), both on 3nm.
January 27, 2026
Cloud provider CoreWeave has reportedly decided to adopt NVIDIA’s Vera Rubin. This suggests NVIDIA’s vision of AI factories is becoming reality. Power consumption is expected to reach 5 GW. With CPO-based optical cables enabled by Spectrum-X’s COUPE technology, this marks a major step forward. AMD’s Helios rack uses Broadcom’s Tomahawk 6, also with CPO. Which becomes operational first remains to be seen, but either way, the industry is advancing.
Microsoft has released its AI chip Maia 200, manufactured on TSMC 3nm. It uses CoWoS packaging with six HBM3e stacks (1.18 Tb/s each), totaling 7 TB/s and 216 GB. On-chip SRAM is 272 MB. It delivers 10 PFLOPS at FP4, with power consumption of 750 W. From photos, the die appears roughly the size of six HBM stacks—around 480–500 mm².
At present, most NPUs and AI chips appear to remain monolithic rather than chiplet-based. For now, scaling through large monolithic dies seems to be the trend—Cerebras being the extreme example. Large chips are expensive at advanced nodes, which is why CPUs moved to chiplets. AI chips may remain large and confined to the cloud rather than consumer devices.
January 28, 2026
There were reports late last year that NVIDIA abandoned Intel’s 18A process, but new articles suggest that the next-generation GPU, Feynman, may use Intel Foundry for I/O dies and packaging, while GPU dies remain with TSMC. Rubin was likely monolithic (or dual-die superchips), so Feynman may move to a chiplet architecture. Feynman is expected to be manufactured on TSMC A16 at Fab 18 around 2028. Fab 18 appears to be evolving into a central advanced fab in southern Taiwan, supporting 5nm, 4nm, 3nm, A16, and advanced packaging.
China has also reportedly approved imports of H200. After initially pausing for domestic protection, authorities likely judged that competitiveness without H200 would suffer.
The Vera Rubin VR200 NVL72 rack announced at CES 2026 reportedly costs $6 million—about ¥900 million at ¥150 per dollar. For comparison, GB200 NVL72 was $2.9 million and GB300 was $4 million. Inflation plays a role, but VR pricing is still striking. More than anything, it’s notable that there is a listed price.
In unrelated news, Amazon AWS reportedly accidentally emailed layoff notices to employees before officially announcing workforce reductions—around 16,000 people.
January 29, 2026
Following yesterday’s NVIDIA news, there are reports that Apple may also outsource part of its M-series production to Intel Foundry—possibly the base (non-Pro, non-Max) models.
January 30, 2026
This didn’t get much attention during Intel’s earnings, but its custom x86 business appears to be doing well. CEO Lip-Bu Tan comes from Cadence, an EDA vendor, so the view is that x86-based ASIC development could be a strong growth area. Given NVIDIA’s investment, it’s also speculated that NVLink Fusion–compatible Xeons may be offered as custom x86 products.
On Xeon, there are reports that Rakuten Mobile has begun using Xeon 6E—likely the 144-core Xeon 6 6700E—for 5G core network validation.
Intel also released an “AI Chip Test Vehicle,” essentially an implementation sample with chiplets mounted on a substrate. It likely includes basic serial interconnect wiring to verify assembly integrity via resistance testing. While that may sound trivial, photos show four large main chips (~450 mm² each), twelve HBM stacks, and two elongated I/O chips (~250 mm² each). These main chips are assumed to represent AI accelerators. I wrote a few days ago that AI chips are growing larger and may be hard to chipletize, but Intel seems to be assuming chiplets from the outset. Since inter-chip wiring exits the die and runs through the package, spacing inevitably widens, and how tightly this can be packed may determine success.
NVIDIA also announced that robotaxi services using Mercedes-Benz vehicles have begun. This collaboration was mentioned at CES 2026, but it has now entered real-world deployment. The service reportedly operates through Uber, and the vehicles use Thor based on Blackwell.
For AMD’s Zen 6, information on CCD size, core count, and L3 cache has emerged. Zen 5 (4nm) had 8 cores and 32 MB L3; Zen 6 (2nm) is expected to have 12 cores and 48 MB. Both cores and cache increase by 1.5×, maintaining L3 per core. CCD size reportedly grows from 71 mm² to 76 mm² (+7%). Assuming a 0.7 scaling factor from 4nm to 2nm (two generations), this effectively looks like only one generation of shrink. Of course, anyone familiar with semiconductor internals knows that “4nm” and “2nm” are largely nominal labels.
January 31, 2026
With January coming to an end, this month’s main topics were CES 2026, TSMC’s earnings, and the Davos conference. It’s been a busy month.
Jensen Huang is reportedly visiting Taiwan and has met with TSMC founder Morris Chang. In a meeting with Chairman C.C. Wei, reports say TSMC expects to double its manufacturing capacity over the next ten years. Other major Taiwanese companies were also invited, likely to strengthen the supply chain for Vera Rubin production. As AI factories spread, power and cooling are expected to become bottlenecks, with “800V HVDC” emerging as a key term. NVIDIA is no longer just a GPU company—its reach now spans autonomous driving, robotics, data centers/HPC, and PCs.
Another item: NVIDIA and MediaTek appear to have partnered on AI PC development. I recall seeing articles last week about NVIDIA entering the PC market with N1X (= GB10). The Vera Rubin superchip uses SOCAMM for LPDDR—thin, secured by three screws, and replaceable. It’s possible the same could be used in N1X PCs. SOCAMM stands for System-on-Chip Attached Memory Module and differs slightly from SODIMM.
In power devices, reports say TSMC has partnered with VIS (Vanguard International Semiconductor Corporation) to license GaN technologies (650V, 80V). This includes GaN-on-QST, where GaN is formed on QST substrates. QST, developed by Qromis, reportedly matches GaN’s thermal expansion coefficient, enabling larger wafers (up to 8 inches).
On architecture, Arm has reportedly signed a long-term agreement with Apple extending beyond 2040, and also announced collaboration with Meta. This reinforces Arm’s importance as a customizable AI chip architecture for edge use. Arm already has strong track records in servers, cloud, and supercomputers. Hyperscalers such as AWS (Graviton), Google (Axion), and Microsoft Azure (Cobalt) all use Arm-based CPUs.
Finally, it was reported that TSMC’s C.C. Wei and Broadcom’s Hock Tan received the Robert N. Noyce Award last November, and that Broadcom raised prices for TSMC at the award ceremony. The article noted that Broadcom—not TSMC—raised prices, suggesting that TSMC may be using Broadcom chips in its systems. Whether internal IT or fab systems, TSMC surely operates substantial infrastructure. It made me realize I’d never really thought about what those systems actually are.
This Blog text was translated by AI from Japanese Source Blog.