Era COMPUTATRUM: FEB.(2nd Half) 2026

February 15, 2026

Speaking of which, due to soaring DRAM prices, there are concerns that computer-related manufacturers may struggle with earnings this year. I also saw an article saying that even boxed CPUs are no longer selling well on Amazon’s online store. It seems that DIY PC users are paying the price for data centers absorbing massive amounts of memory. GPU cards have finally started to return in terms of product variety, but minimum prices have risen.
A little while ago, I mentioned in a “mini update” that installing the January Windows Update — “2026-01 Security Patch (KB5074109) (26200.7623)” — could cause frame-rate drops on NVIDIA graphics cards. Apparently, applying KB5077181, which will be distributed in the February Windows Update, resolves the issue. It also includes fixes for other bugs, including—surprisingly—a “Notepad vulnerability.” I’ll just wait for it to roll out automatically.
The day before yesterday, I wrote that AWS’s CEO commented that the market and industry are overreacting to AI writing code. Supporting that view, there was an article noting that about ten years ago, when AWS captured a large share of the software market, similar “doomsday” narratives emerged—but in the end, no software companies went bankrupt because of it. Thinking back, there were many transitions: freeware during the PC-communication era, GPL (now generally called OSS) in the UNIX world, Linux being released for free, and later the rise of SaaS on the web. Despite all these shifts, the software industry has not disappeared.

February 16, 2026

By the way, the India AI Impact Summit is being held this week. According to an interview with Intel’s Vice President for the India region, India generates 20% of the world’s data but owns only 2% of global server capacity. In terms of data center power capacity, the figures are said to be: the United States at 53 GW, China at 20 GW, Europe at 13 GW, and India at 1.6 GW (with an additional 1.7 GW planned by 2027). There is clearly significant room for growth.
There was also a press release stating that AMD will deliver Helios systems to India. The scale is said to be 200 MW. Since one Helios rack reportedly consumes about 0.18 MW, even rounding to 0.2 MW suggests roughly 1,000 racks will be deployed. With 72 GPUs per rack, that would amount to 72,000 GPUs across 1,000 racks. These will be introduced as an AI platform by HyperVault AI Data Center Limited, a subsidiary of Tata Consultancy Services (TCS). I recall reading an article around mid-last month about AMD forming a partnership with TCS.

Another AMD-related item: SoftBank issued a press release stating that it has jointly developed with AMD an orchestrator for partitioned use of AMD GPUs. AMD Instinct GPUs already support partitioning features. A demonstration is planned at MWC in Barcelona next month. In a similar vein, I’ve heard of Fujitsu’s AI Computing Broker, which I believe provides a mechanism to partition NVIDIA GPUs. NVIDIA offers features such as MIG (Multi-Instance GPU) and vGPU. While partitioning can help address GPU shortages, it also seems like a necessary mechanism to increase utilization of GPU servers and accelerate return on investment.

February 17, 2026

At the India AI Impact Summit 2026, which began yesterday, Qualcomm reportedly exhibited a humanoid robot. Apparently, Qualcomm has a robotics and automotive division and is building a general-purpose robotics architecture covering not only humanoids but also everything from home-use robots to industrial autonomous mobile robots. I don’t recall seeing much coverage of Qualcomm beyond smartphones, PCs, and communications chips.

Another Qualcomm-related topic: because Qualcomm holds essential patents related to the 5G communications standard, smartphone manufacturers must obtain patent licenses from Qualcomm. Naturally, these are not free, and the license fees are added to smartphone prices as royalties. In the UK, there had been a class-action lawsuit alleging that Qualcomm unfairly charged royalties and raised smartphone prices. However, since the royalties are not unreasonable per se, the lawsuit has now reportedly been withdrawn. The royalty for a 5G device is said to be around $16.
Speaking of smartphones, there are reports that Arm, which provides CPU architectures, plans to establish an AI semiconductor research facility at its Austin, Texas site, with support from the Texas state government. Until recently, Arm’s business centered on selling Neoverse N/V architectures and related IP, but it has shifted toward selling CSS (Compute Subsystems). Last year, I believe Arm also announced that it would develop its own chips. Given that Cortex and other designs have long been manufactured at TSMC and Samsung, producing chips in-house should be feasible. As semiconductor manufacturing moves away from being smartphone-centric and toward HPC and AI as the main drivers, it will be interesting to see what kinds of chips Arm brings to market.
On the AI front, NVIDIA issued press releases regarding GB200 and GB300. Compared to Hopper-based platforms, the cost per one million tokens is reportedly reduced to one-tenth with GB200 NVL72 and to one-thirty-fifth with GB300 NVL72. Furthermore, GB300 NVL72 is said to achieve up to 50× throughput per megawatt. GB300 MVL is being adopted by Microsoft, CoreWeave, and OCI (Oracle). CoreWeave had already announced at the end of last month that it would adopt Vera Rubin. With Vera Rubin, compared to Blackwell, the cost per million tokens for MoE (Mixture of Experts) inference is said to be one-tenth, and throughput per megawatt is 10× higher.

Another AI-related development: Fujitsu announced an “AI-driven software development platform.” It is said to address three challenges: labor shortages, replacement of maintenance work, and moving away from the traditional man-month calculation model. Understanding the Japanese role of SE (system engineer) may be helpful here. By AI-automating practical tasks such as system construction, coding, and maintenance in response to customer IT department requirements, the expectation is that SEs—previously overwhelmed with routine work—will be able to spend more time in in-depth discussions with customers. The core AI is reportedly Takane, a Japanese-language LLM developed jointly by Fujitsu and Cohere. As I wrote a few days ago, it seems to be Fujitsu’s proprietary LLM created in collaboration with Cohere. Customer-created specifications often contain local terminology and implicit assumptions; from the description, it appears that the system is designed to handle such nuances effectively in Japanese. This may represent one concrete way AI is entering real-world practice, and it is worth watching how it develops

February 18, 2026

By the way, it appears that NVIDIA has sold all of its shares in Arm. If I recall correctly, the discussion about NVIDIA acquiring Arm dates back to around 2020. That acquisition ultimately failed, but since Arm was relisted on NASDAQ in 2023, this complete divestment means that Arm’s future stock price movements will no longer have any direct impact on NVIDIA.

I don’t think this means that Arm-architecture CPUs such as Grace or Vera will disappear, but given NVIDIA’s investment relationship with Intel, how things unfold from here is something to watch closely.

Another NVIDIA-related item: it seems NVIDIA has signed a GPU supply agreement with Meta. The deliveries will start with the current GB300, followed by Vera Rubin. Meta plans to deploy these in its multi-gigawatt AI factory “Prometheus” in 2026, and later in the 5-gigawatt AI factory “Hyperion” planned a few years down the road. In terms of GPU count, this would amount to several million units—almost impossible to visualize. Jensen Huang’s efforts at the end of January to coordinate Taiwan-wide production of GPU racks may well have been in preparation for this.

February 20, 2026 Time of Arm Server (topic.1)

Speaking of which, the partnership between NVIDIA and Meta that was reported the day before yesterday appears to be having a broader impact. With NVIDIA making a major move into the data center space, there are concerns about the implications for Intel and AMD. Meta is already reportedly NVIDIA’s second-largest customer by volume, and beyond adopting GB300 and Vera Rubin for AI workloads, Meta is also introducing Arm CPUs for its data centers—raising concerns about the impact on x86. NVIDIA’s move has multiple dimensions.

I’d like to go into a bit more detail, since I took a break from blogging yesterday.

One aspect is NVIDIA’s positioning as a data center supplier. NVIDIA has been calling this an “AI factory,” but in reality it looks very much like a full-scale mega data center. What started as DGX SuperPODs for AI training has expanded to include inference, driving down per-token costs to promote wider adoption. By supplying systems at the rack level and even offering standalone Arm CPUs, NVIDIA has effectively become a supplier of mega-scale AI data centers.

Looking at CPUs, Grace has 72 cores and Vera has 88 cores, both Arm CPUs. In the Arm world, AWS’s Graviton5 has 196 cores, but Vera supports two threads per core, yielding 172 threads. On the x86 side, Xeon 6E offers 144 cores per socket, with two sockets totaling 288 cores. In a superchip configuration, Vera reportedly mounts two superchips per tray; assuming a dual-socket configuration, this would amount to 344 threads. AMD’s Turin (Zen 5c) reaches 768 threads in a dual-socket setup, so there’s always something larger to compare against—but Vera appears competitive as a data center CPU for the current generation.

As for GPUs, in supercomputers, GPUs already outnumber CPUs. When IBM and NVIDIA built ORNL’s Summit in 2018, the ratio was two IBM POWER9 CPUs to six Volta GPUs. This trend continues not only with GB300 and Vera Rubin but also with AMD’s Helios, where GPU counts exceed CPU counts. Seen this way, AI data centers are becoming structurally similar to supercomputers.

Perhaps the most significant point is the server ecosystem shift. Fujitsu demonstrated that Arm CPUs could be used in supercomputers by integrating SVE into the A64FX used in Fugaku, which led to SVE2 being included starting with the Armv9 architecture. This boosted Arm’s presence as a server CPU beyond its traditional role in smartphones. While Arm-based PCs such as Apple’s M series and Qualcomm’s Elite have gained visibility, Arm servers using Fujitsu’s A64FX or Ampere Computing’s Altra Max and AmpereOne never achieved major market presence. NVIDIA’s move this time likely means that a vendor—NVIDIA—is supplying Arm CPUs in large volumes specifically for AI data centers. This is clearly different from merely selling architecture licenses.

Meta (formerly Facebook) founded the Open Compute Project (OCP), a standardization body for data center hardware, together with Google and Microsoft. HPE and Supermicro are suppliers of OCP-compliant servers. In October last year, Arm, NVIDIA, and AMD all joined OCP as board members. AMD’s Helios rack is also believed to be OCP-compliant. In this context, Meta’s purchase of Arm CPUs from NVIDIA is likely intended to build OCP-based servers.

That concludes the discussion on NVIDIA and Meta.
Below are several investment-related topics. The fate of NVIDIA’s previously rumored $100 billion investment in OpenAI had become unclear, but it now appears that NVIDIA is participating in OpenAI’s funding round with a $30 billion investment. This is reportedly unrelated to the earlier $100 billion figure. In South Korea, there is also a plan to introduce 260,000 NVIDIA GPUs by 2030, following 13,000 units last year and an additional 15,000 units this year.

With the India AI Impact Summit underway, NVIDIA has also announced investments in India. Data centers totaling 70 MW are reportedly under construction in Mumbai and Chennai, with NVIDIA systems expected to be deployed there. A few days ago, AMD announced a partnership with Tata Consultancy Services (TCS), and now OpenAI is also said to have partnered with TCS, starting with a 100 MW AI data center and eventually expanding to 1 GW. Beyond investments, Qualcomm—who exhibited a humanoid robot—has reportedly partnered with Tata Electronics on automotive semiconductor manufacturing, with an OSAT facility located in Assam.

While not an investment or partnership, there was also an article about Fujitsu CEO Takahito Tokita giving a speech. He reportedly stated that excessive reliance on AI for specialized skills is risky, and that AI should instead support richer ideas and creativity by facilitating communication between experts and non-experts, including general employees and customers. With this philosophy at its core, industries can grow further and create more jobs. NVIDIA’s Jensen Huang and AWS’s Matt Garman appear to be expressing similar views. It increasingly feels like a picture is emerging in which AI creates more work, rather than eliminating it.

There was also an article stating that AMD has become a guarantor for financing raised by Crusoe, a data center operator in Ohio. The loan is reportedly from Goldman Sachs, with AMD’s chips themselves used as collateral. Neither AMD nor Goldman Sachs has confirmed this.

Google is reportedly offering a $100 million investment to Fluidstack, a cloud computing startup, likely in search of deployment opportunities for its TPUs. While Anthropic has adopted TPUs, NVIDIA still dominates overall AI compute power, and Google appears eager to expand its influence. Perhaps we are moving from an era of “build and sell” to one of “invest and get others to use.”
Finally, we get to chips. There are reports that Ryzen CPUs based on AMD’s Zen 6 architecture may be delayed until 2027. Zen 6 is expected to use TSMC’s 2 nm process. Internal testing reportedly achieved 6.5 GHz operation, suggesting that 2 nm prototype chips already exist. Zen 6–based EPYC Venice is slated for use in Helios and already has manufacturing and shipping plans, so AMD is likely prioritizing that. On the consumer side, there are also concerns about severe memory shortages, making it unlikely that conditions will be ready for a Zen 6 Ryzen debut before at least 2026.

Another AMD-related note: the Radeon RX 9060 XT reportedly reached 4.769 GHz under enhanced cooling conditions, apparently setting a world record for GPU overclocking.
A brief software-related item: there was an article noting that Linux kernel 7.0 will remove the driver for Intel’s 440BX chipset. On a personal note, I believe the ASUS motherboard in the PC I built in 2000 used the 440BX. That brings back memories. The CPU was a Pentium III (Coppermine, 0.18 µm), and despite having no real software to use it with, I built it as a dual-CPU system. Even now, I’m not entirely sure what I wanted to do with it.
Open-PGL, hosted on GitHub and rumored to have been discontinued by Intel, is reportedly being archived by the Academy Software Foundation (ASWF). ASWF is one of the projects hosted by the Linux Foundation.

February 21, 2026 Human & AI Error (topic.2)

There are also rumors that Intel’s Nova Lake-S will launch in 2027. It had been expected in late 2026, but appears to slip into the following year. I just read an article yesterday suggesting that AMD’s Zen 6–based Ryzen would also arrive in 2027, and it seems Intel may be on a similar timeline. In AMD’s case, the delay is likely due to prioritizing Zen 6 EPYC, but for Nova Lake, it may be that Intel’s 18A production lines are fully booked. They are probably being used for Xeon 6+E Clearwater Forest or Xeon 7 Diamond Rapids, both aimed at data center or HPC markets. It does seem that the consumer market will face a tough environment this year.

Another Intel-related topic: there have reportedly been layoffs of around 6,000 people at Intel’s Oregon sites, raising some concerns about corporate stability. Among researchers, Oregon (OR) is well known as Intel’s U.S. R&D base—many Intel research papers list “Intel Corp., Hillsboro, OR.” Seeing layoffs even at such a core R&D site is rather disheartening.
Over the past couple of days, many articles have appeared claiming that AWS’s coding AI caused service outages. AWS’s code-generation AI tool, called Kiro, apparently also performs operational tasks. The outages reportedly occurred twice last December, but I don’t recall hearing about them at the time—apparently because they occurred in China. One outage lasted 13 hours, while the other was not visible to customers. The information seems to have come from an internal AWS leak, while the official explanation cites human error. From what I’ve read, Kiro chose to delete and rebuild systems and executed this without human intervention. The users operating Kiro had overly broad privileges, leading to widespread impact. Whether this was Kiro’s fault or the users’ fault is debatable, but there are a few troubling points: execution without human involvement, and referring to humans with broad administrative authority simply as “users.” These “users” may be Kiro users, but within AWS they are likely employees with operational responsibility.

Executing actions without human oversight is something that should not be done, in order to prevent irreversible accidents. This is not because it “blurs responsibility,” but because automation without proper fail-safe and foolproof mechanisms has long been recognized as dangerous—ever since the days of Thomas Edison. This lesson is written in blood in the history books. AI may be intelligent, but it does not bleed, so caution is essential. AWS also announced layoffs of 16,000 employees last month, with internal emails reportedly sent before the public announcement. The fact that such internal stories are leaking may be related to the wave of workforce reductions.

One more AWS-related item, finally a technical one: AWS has released the Hpc8a instance. Compared to the previous Hpc7a generation, memory bandwidth is reportedly improved by 42%. The CPU is listed as a 192-core AMD Zen 5–based EPYC, which likely means Zen 5c (Turin), manufactured on TSMC’s 3 nm process. Both Zen 5 and Zen 5c support two threads per core, but SMT is disabled here, resulting in 192 threads. In HPC workloads, where data is processed sequentially and synchronously, single-thread performance is critical and SMT is often less valued. In data centers, concerns such as noisy neighbors and security also make SMT less desirable. While SMT is effective for increasing throughput with fewer cores, it may no longer be essential in the era of many-core CPUs with strong single-thread performance.
Finally, a political topic. A U.S. court has ruled that the Trump-era tariffs are illegal. It remains to be seen how this will affect TSMC’s massive planned investments in the United States, but the ruling does not address whether already-paid tariffs will be refunded or how such refunds would be handled. There also appear to be multiple alternative legal bases for tariffs, so the situation may not change much. Moreover, if a presidential order is issued and approved by Congress, it would become law, which seems quite possible. There is little point in reacting emotionally either way. At a press conference, former President Trump reportedly suggested that Taiwan had “stolen” semiconductor technology. In the DRAM era of the 1980s, Japan would almost certainly have been named instead. I can’t help but feel a bit sorry for Taiwan.

February 22, 2026

There was an article stating that NVIDIA has partnered with Idaho National Laboratory (INL) in Idaho on the development of nuclear AI applications. NVIDIA’s technology is expected to be applied to the design, manufacturing, construction, and operation of reactors developed at INL. This appears to be part of the Genesis project and a component of U.S. energy policy. Since digital twin technology is mentioned, NVIDIA Omniverse will likely be used. One wonders whether AI training models for nuclear reactions inside reactors might eventually be developed and even released as open source—though concerns about nuclear weapons applications make that unlikely.

Another NVIDIA-related note: for several years now, there have been occasional reports of GPU power connectors melting or burning. Possibly in response, there was an article noting that some Dell PCs use screw-fastened connectors for GPU auxiliary power. This likely helps prevent fluctuations in contact resistance caused by cable movement. High-end desktops can require power supplies exceeding 1,000 W. When large currents flow, even small increases in contact resistance can cause significant voltage drops and substantial heat generation via Joule heating. For example, at 400 W and 12 V, about 33 A flows; if contact resistance increases by just 10 mΩ due to heating, voltage drops by 0.33 V. If sensors detect this and raise voltage to compensate, current increases further, causing even greater voltage drop—leading to a runaway process that can result in burning. The trigger is variation in resistance.
There was also an article about a Toronto-based startup called Taalas developing an AI chip named HC1, manufactured on TSMC’s 6 nm process. It hardwires the generative AI model Llama 3.1 8B and reportedly achieves 17,000 tokens per second (TPS)—a figure that made me do a double take. Just last week I wrote that Llama 3.1 70B achieved 2,100 TPS on Cerebras; although the parameter counts differ (8B vs. 70B), this is an order of magnitude higher. According to the company, HC1 is two orders of magnitude faster than Blackwell, with one-twentieth the token cost, and can be air-cooled. While it’s common knowledge that specialized hardware can deliver higher performance and lower power consumption, the hardwired nature means it cannot support other LLMs. I also wonder whether reinforcement learning is possible on this chip alone.

The company says that changing the LLM can be handled by modifying wiring layers, but this reportedly takes two months. The number of layers modified and their position in the stack matter: the finest masks are those for the transistors and the metal layers immediately above them, while upper layers follow “reverse scaling,” with wider pitch and lower mask costs. However, wider pitch reduces routing density, potentially requiring more layers and higher mask costs, and there are limits to how many layers can be added. The chip size is reportedly 815 mm²—about 28 mm × 29 mm—very close to the practical manufacturing limit. Moving to 3 nm would reduce size somewhat, but wafer costs would then come into play. Still, as an exploration of how much faster LLMs can be when hardwired, this is a very interesting approach. A hybrid system—hardwiring a core LLM and supplementing reinforcement learning with GPUs—could be compelling in terms of both performance and power efficiency.
SoftBank Group CEO Masayoshi Son has reportedly proposed the “Arizona AI Mega Project,” a plan to build an AI and robotics industrial park in Arizona. The R&D hub would include semiconductor manufacturing units, housing for engineers, and a smart grid—effectively a technology development city. If robotics is included, this could serve as a hub for physical AI. SoftBank participated in the Stargate Project last year and invested in OpenAI, and through its ownership of Arm, it has decided to develop processors in-house, acquiring Ampere Computing. This may indicate that SoftBank is beginning to solidify a concrete exit strategy.

February 23, 2026

By the way, it appears that Intel may be designing a unified core that integrates its P-cores and E-cores. More precisely, Intel is reportedly planning to hire project members for a unified-core initiative, and it is not entirely clear whether this means a full integration of P and E cores.
In the past, there was a project called “Royal Core,” reportedly conceived by Jim Keller (with a CPU codename that may have been “Beast Lake”), and that design was said not to have separated P-cores and E-cores.

In reality, the decision to split P and E cores in the 12th-generation Core i series (Alder Lake) may have been a strategy to reduce power consumption while remaining on the 10 nm node, by bringing in Atom-based single-thread cores. Alder Lake’s Intel 7 process was essentially an enhanced version of Intel’s 10 nm SuperFin technology. Since the minimum feature size remained the same, packing in more circuitry would naturally increase power consumption, which may have led Intel to adopt separate P-core and E-core architectures.

As I wrote the other day, once single-thread performance improves and core counts increase, SMT may become less important. Starting with the second generation of Core Ultra (Lunar Lake / Arrow Lake, manufactured on TSMC 3 nm), the Lion Cove P-core has already moved to single-thread operation. Assuming process scaling continues smoothly, it may indeed become feasible to unify P-cores and E-cores again.
Another Intel-related item: there are reports about Intel’s 3000 W power supply units for servers and data centers. With a 12 V output, that corresponds to 250 A. These units are water-cooled rather than air-cooled. GaN and SiC are used as power devices, and the units have obtained 80 PLUS Platinum certification. In data centers, cooling water lines would typically be integrated into the infrastructure, but for standalone servers, would they attach an AIO (All-in-One) radiator unit? It is an interesting detail to consider.
On a slightly different note, some quantum computing news: a contract for an IBM quantum computer to be delivered to the Campania region of Italy has reportedly been suspended. The issue does not appear to lie with IBM’s proposal itself, but rather with questions raised about the bidding process. There is a project underway to establish a “Quantum Valley” in southern Italy, and this suspension may delay its progress.

February 24, 2026

It seems AMD has also entered into a GPU supply partnership with Meta. Just last week, Meta finalized a GPU supply agreement with NVIDIA, and now it appears to be AMD’s turn. Meta is reportedly planning to install 6 GW worth of Helios rack systems equipped with AMD’s Instinct MI400-series GPUs. There is also information suggesting that these GPUs will be tuned specifically for Meta, raising speculation that they may be customized versions of the MI400.
The CPU is EPYC, based on Zen 6 Venice, with Zen 7 Verano also under consideration. Assuming one Helios rack consumes about 0.2 MW, 6 GW would correspond to 30,000 racks. With 72 GPUs per rack, that comes out to a simple calculation of roughly 2.16 million GPUs. At that volume, producing customized GPUs would still make economic sense. Naturally, this cannot be completed in a single year, so the deal appears to be a multi-year contract.
As with the OpenAI deal, this agreement also includes warrants for 160 million shares of AMD common stock. The mechanism seems to be that as Helios deployments progress, Meta receives AMD shares accordingly. Issuing all 160 million shares at once would dilute AMD’s stock, so the allocation is apparently staged. There are also conditions tied not only to GPU volume but to AMD’s stock price at each stage. As Meta purchases GPUs and AMD’s performance improves, corresponding shares are allocated. During the OpenAI deal, this structure was criticized as potentially circular trading. AMD’s aim is presumably to catch up to NVIDIA in terms of market capitalization—though at present AMD’s market cap is still less than 10% of NVIDIA’s.
Several more NVIDIA-related topics. There are renewed reports suggesting that NVIDIA may be preparing to enter the notebook PC market, with announcements of N1/N1X possibly imminent. The N1/N1X CPU+GPU itself is already on the market as GB10. Despite the name, the CPU appears to be a hybrid of Arm Cortex-X925 and Cortex-A725, rather than Grace. The C1 used in Grace was likely a successor to the X925. The GPU uses the Blackwell architecture, though its performance is presumably scaled down.
NVIDIA’s target seems to be the market for CPUs with integrated GPUs (iGPUs). Rumors of a launch in the first half of 2026 have not yet disappeared, although there was no announcement at CES in January. Whether anything will be announced at the March GTC is still unclear. In any case, the consumer market environment this year looks unfavorable.
Another NVIDIA item: the company has partnered with Singtel to promote AI adoption in Singapore. Singtel is the country’s major integrated telecom operator. Together, they have established an applied AI Center of Excellence for both public and private sectors. At last week’s India AI Impact Summit, although CEO Jensen Huang was absent, NVIDIA demonstrated efforts such as building a digital twin of Tata Motors’ automotive factory using the Omniverse platform and launching initiatives for robot learning with logistics companies. NVIDIA appears to be actively pushing physical AI in India.
Next, an NPU-related topic. Intel reportedly considered acquiring SambaNova but has since abandoned the plan. However, Intel did participate in part of SambaNova’s funding, so the relationship has settled into an investment partnership. The two companies will also engage in technical collaboration over multiple years. They are expected to build systems combining SambaNova’s SN50 chip with Xeon processors. Intel’s consumer CPUs already include NPUs derived from its acquisition of Movidius, but SambaNova’s technology seems likely to be paired primarily with Xeon.
Finally, a software-related topic. Anthropic reportedly published a blog claiming that its Claude Code can rapidly rewrite COBOL code, traditionally used on mainframes. Following this, IBM’s stock price reportedly dropped by around 13%. Rewriting COBOL has been discussed for decades, and IBM itself has long pursued initiatives to migrate COBOL systems to Java. There have been many such efforts in the past, but I don’t recall stock prices being affected before.

This time, Anthropic’s blog seems to have had an impact because the software industry itself has become highly sensitive to AI coding developments. It may not be entirely realistic, but if AI can just as easily rewrite code into COBOL as out of it, one could imagine a future less constrained by programming languages. Whether that is truly the case remains an open question.

February 25, 2026

By the way, Apple reportedly plans to manufacture the Mac mini in the United States. The company will open an Advanced Manufacturing Center in Houston, Texas, intended as a domestic production base for AI servers. “Sovereignty” has become a key buzzword recently. This facility is likely envisioned as a massive complex where manufacturing lines, data centers, Apple Intelligence, and human training all coexist.

Arizona, neighboring Texas, is home to TSMC’s Fab 21, which is expected to supply the semiconductors for the Mac mini. There are also reports that Apple has signed a contract for the supply of 100 million chips. Within Texas itself, Austin hosts key companies such as Arm, TI, and Broadcom.
Speaking of Arm, its recent quarterly results were strong, with particularly strong royalty revenue from the data center segment. AWS released Graviton 5 in December, and instances have since become available. It seems likely that Arm began charging AWS royalties for these chips around that time. If royalties are charged per core, Graviton 5’s 196 cores would represent a substantial amount per chip. NVIDIA’s Vera, by comparison, has two threads per core, with 88 cores and 176 threads. Whether royalties are based strictly on core count is an interesting question.

Another Arm-related topic: Arm has partnered with the Indonesian government on engineer education, aiming to train 15,000 chip design engineers. There was recently news that Qualcomm completed the design of a 2 nm Snapdragon chip in India, and it appears that chip design engineers are in short supply globally. Designing chips requires IP macros such as basic logic gates, arithmetic units like adders and multipliers, on-chip memories like caches, and controllers for DDR and PCIe interfaces. Shortages of engineers capable of designing such IP have been discussed for several years. By training engineers familiar with the Arm architecture and Arm’s IP, this initiative aims both to alleviate talent shortages and to generate foreign revenue.
ASML in the Netherlands has reportedly succeeded in boosting the power of EUV lithography light sources, increasing output from the current 600 W to 1,000 W. Market introduction is targeted around 2030, with 1,500 W and even 2,000 W also in view. It seems the industry is entering the kilowatt era. The impact on semiconductor manufacturing would be significant: wafer throughput is expected to rise from 220 wafers per hour to 330, a 50% increase.

The key question is where these systems will be deployed. Likely candidates include TSMC, Intel, Samsung (and GlobalFoundries), SK Hynix, and Japan’s Rapidus. EUV tools are now being installed in an increasing number of regions: Taiwan (TSMC), the U.S. (Arizona for TSMC; Arizona, Oregon, and Ohio for Intel; Texas for Samsung; New York for IBM and others), Ireland (Intel), Korea (Samsung and SK Hynix), Hokkaido (Rapidus), and eventually Kumamoto for TSMC. Strengthened light sources imply applicability to advanced nodes, making Taiwan and Ohio particularly realistic destinations—though this is, of course, speculation.

February 26, 2026

By the way, today (February 25 in the U.S.) NVIDIA announced its earnings. Both fourth-quarter and full-year results were released. Full-year revenue reportedly reached $215.9 billion, which translates to approximately ¥32.4 trillion at an exchange rate of 150 JPUSD. This represents 65% year-on-year growth. Fourth-quarter revenue alone was $68.1 billion, up 20% quarter-on-quarter and 70% year-on-year, exceeding the annual growth rate. This suggests particularly strong growth in the latter half of the fiscal year. With 91% of revenue coming from data centers, these numbers underscore NVIDIA’s central role in the recent AI semiconductor boom.
Despite these stellar results, NVIDIA’s stock price has not risen as much as one might expect. This does not necessarily mean there is no room for growth. Massive shipments of GB300 and Vera Rubin for data centers are planned, and while memory supply constraints exist, demand for RTX 6090 GPU cards remains, and N1/N1X products for notebooks are also waiting in the wings.

Investor concerns seem to revolve around two opposing risks: caution toward a potential AI bubble, and fears that AI might fundamentally disrupt the software industry. Investors are caught between the possibility that AI investment becomes circular trading that eventually collapses, and the possibility that AI succeeds too well and destroys existing software business models. In either case—whether AI is “real” or “fake”—pessimism currently prevails.

Recently, CEO Jensen Huang and AWS CEO Matt Garman have repeatedly stated that the software industry will not be destroyed and that the future is bright. Yet when news spread that Anthropic could rewrite COBOL, even IBM’s stock saw drops of 13% or even 27%. Some are even predicting the extinction of SaaS, coining the term “SaaS-pocalypse.” Leaders like ServiceNow CEO Bill McDermott and Salesforce CEO Marc Benioff have pushed back against this narrative. While customer support staffing has been reduced, AI agents have made SaaS more useful, and the number of companies using SaaS is reportedly increasing. Jensen Huang has also suggested that the market is wrong.

AI agents will leverage SaaS and other existing software to support more companies than before. Existing software will still be needed, as will staff who support businesses in new ways. Executives likely already see a model in which AI actually increases the amount of work. When stock prices begin to rise again, that may be the point at which markets accept this view and concerns about circular trading subside.
Another NVIDIA-related note: last month, the U.S. government reportedly approved exports of H200 GPUs to China, and Chinese companies appeared open to importing them. However, it seems that H200s have not yet entered China. The export approval is likely conditional, with strict security monitoring, and Chinese firms may be hesitant to proceed under such constraints.
One CPU-related topic: Broadcom has reportedly shipped Fujitsu CPUs. The product is described as a “3.5D Face-to-Face Computing SoC,” which appears to refer to Fujitsu’s FUJITSU-MONAKA CPU under development. Beyond Fujitsu’s own materials, overviews have also been presented at TSMC symposia and Arm events.

“3.5D” likely refers to a configuration where some chiplets mounted on a silicon interposer (commonly called 2.5D integration) use 3D stacked LSIs. The computing die is manufactured using TSMC’s 2 nm process, while the cache memory die uses a 5 nm process. These are bonded face-to-face, with the cache underneath. In addition, an I/O die (5 nm) is included. AMD’s 3D V-Cache also uses face-to-face bonding, but in AMD’s case the CCD is on the bottom and the cache on top.

It feels as though Fujitsu is now the only company continuing to develop CPUs in Japan. Its previous CPU, A64FX for the Fugaku supercomputer, included HBM, but this new chip targets data centers and does not include HBM. Mass production is expected to begin in the latter half of this year. Broadcom appears eager to deploy this 3.5D SoC technology across a variety of chips.

February 27, 2026

By the way, in OpenAI’s recent funding round, NVIDIA was said to be investing $30 billion, but it now appears that SoftBank invested $30 billion and Amazon invested $50 billion, bringing the total raised to $110 billion. NVIDIA’s contribution reportedly takes the form of 3 GW of dedicated inference capacity and 2 GW of training capacity.
Amazon’s participation seems like a new development. As a result, OpenAI has secured 2 GW of capacity on AWS using Trainium 3 and, in the future, Trainium 4. On the AWS side, Amazon is set to become OpenAI’s exclusive third-party cloud distribution provider. This apparently means that when OpenAI Frontier—the OpenAI subsidiary that operates AI agent services—offers those services, they will be provided exclusively through AWS (Amazon Bedrock). The 2 GW of Trainium capacity is likely allocated for this purpose.

Microsoft remains an existing shareholder of OpenAI and continues to collaborate with it, but OpenAI’s AI agent services will not be offered through Azure. Existing OpenAI services delivered via Azure will continue. This suggests a division of roles: stateful APIs such as AI agents on AWS, and stateless APIs on Azure.
Another AWS-related item: there was news that AWS plans to build a data center in Louisiana. The investment is reported to be around $12 billion, though the power capacity was not specified. Infrastructure development is underway, including a power supply network incorporating 200 MW of solar generation, as well as water systems for cooling. About 87% of the cooling reportedly relies on outside air. While water cooling is used, heat exchange through radiators normally relies on ambient air. Construction is expected to create around 1,500 jobs, with approximately 540 full-time positions for data center operations. Supercomputer construction has long been considered a form of public works, and it seems data center construction has now entered the same category.
Turning to NVIDIA-related topics, the ripple effects of NVIDIA’s earnings continue to spread. NVIDIA’s contribution to TSMC’s revenue is reportedly $23 billion, accounting for 19% of the total—up 7 percentage points from 12% in 2024. By contrast, Apple’s share of TSMC revenue is $20.5 billion (17%), down 5 points from 22% last year. As a result, Apple has ceded its position as TSMC’s largest customer to NVIDIA. It had been said during TSMC’s earnings call last month that NVIDIA had overtaken Apple, but these figures make it clear.
A supercomputer using NVIDIA GPUs at Eli Lilly has reportedly gone into operation. Called “LillyPOD,” it is a DGX SuperPOD equipped with 1,016 Blackwell GPUs, construction of which began in October 2024. It is said to be the most powerful system in the pharmaceutical industry. With 72 GPUs per rack, 1,016 GPUs would amount to 14 racks plus 8 GPUs. Since each computing tray contains four GPUs, that leaves an extra two trays’ worth of GPUs—presumably installed somewhere.
Regarding Meta: after announcing that it would receive several million GPUs from NVIDIA, and then—just two days before NVIDIA’s earnings announcement—declaring that it would also receive 6 GW worth of GPUs from AMD, Meta has now reportedly announced that it will procure TPUs from Google as well. This may be a form of diversification.
Next, some AMD-related items. AMD CEO Lisa Su reportedly stated that the data center market will reach $1 trillion by 2030. According to Japan’s Ministry of Internal Affairs and Communications, the market size in 2024 was $416.1 billion. Reaching $1 trillion by 2030 would require annual growth of roughly 16%. There is no doubt that data centers are a growth industry. Terms like “mega data centers” and “hyperscalers” were already in use around 2020, but the wave of AI adoption has dramatically accelerated investment.
As I mentioned briefly during the India AI Impact Summit two weeks ago, current data center capacity by region is approximately 53 GW in the U.S., 20 GW in China, 13 GW in Europe, and 1.6 GW in India. Japan stood at 1.37 GW as of 2024.

AMD has also partnered with Nutanix. The deal includes a $150 million investment from AMD and technical collaboration, with $100 million allocated specifically for joint technical initiatives. Nutanix provides services that enable hybrid clouds spanning private and public cloud environments. This partnership is likely aimed at delivering a full-stack solution—including AI—on the private cloud side. Nutanix had previously supported NVIDIA GPUs, and with this partnership, AMD GPUs will now also be supported. From a confidentiality perspective, many companies avoid storing sensitive information on public clouds like AWS. It seems we are entering an era in which AI agents will also be deployed in private clouds. That said, Nutanix reportedly lowered its earnings outlook due to constraints in CPU and memory supply.
AMD also announced Sorano, a Zen 5–based processor in the 8xx5 series. It is the successor to Siena, with core counts increased from 64 to 84. It is not yet clear whether it uses Zen 5 or Zen 5c. If it is Zen 5, it would require twelve 7-core CCDs (with one core disabled per CCD). If Zen 5c, it would require six 14-core CCDs (with two cores disabled each). Using twelve Zen 5 CCDs would physically resemble the 96-core-class EPYC 96x5 configuration. If the processor is to fit into the smaller SP6 socket rather than SP5, a configuration using six Zen 5c CCDs seems more realistic—but that remains speculative.
One Intel-related item: a senior vice president at Intel Foundry has reportedly moved to Qualcomm. Just last month, it seemed that an engineer who had long led GPU development at Qualcomm moved to Intel. Such exchanges of personnel are not uncommon.
Finally, two political topics. Anthropic and the U.S. Department of Defense—commonly known as the Pentagon, whose current formal designation translates roughly as the Ministry of War—appear to be in some dispute over an AI usage contract. The Department of Defense reportedly wants unrestricted use of Anthropic’s Claude AI for any lawful purpose. The contract value is said to be $200 million, and Anthropic is reportedly hesitant.

DeepSeek is preparing to release its next v4 model, but NVIDIA and AMD have reportedly been excluded from the GPU vendors’ early access list. Chinese domestic vendors such as Huawei are said to be given priority for several weeks. This appears to be a move to reduce dependence on the U.S. and protect domestic vendors.

February 28, 2026 From GPU to NPU (topic.2)

There are reports regarding AMD’s Zen 7 “Grimlock Ridge.” Although this is a desktop CPU, its CCDs are expected to be used in EPYC as well. Zen 6 has not yet reached the market but uses TSMC’s 2 nm process, while Zen 7 is expected to use the A14 (1.4 nm) process. The CCD is said to have 16 cores with an area of 98 mm². This 16-core CCD is reportedly codenamed “Silverton.” Separately, there appears to be an 8-core CCD called “Silverking,” with an area of 56 mm². For reference, Zen 6 CCDs have 12 cores and an area of 76 mm².
According to TSMC materials, the density improvement from N2 (2 nm) to A14 is about 1.2×, implying an area reduction to roughly 83%. While moving from 2 nm to 1.4 nm represents a full-node generational shift, it does not result in halving the area for the same core count as in earlier eras. Keeping 12 cores would yield an area of roughly 63 mm² (76 × 0.83). A 24-core CCD would be around 124 mm², slightly larger than previous CCDs. It appears AMD may have opted for two variants: a 16-core die (1.5× the core count) and a smaller 8-core die at roughly half that size. Powers of two like 16 and 8 may also be more convenient from a design standpoint.
The article also mentioned APUs. In Grimlock Point/Halo, CCDs are reportedly stacked on top of the I/O die. In Zen 5’s Strix Point and Zen 6’s Medusa Point (not yet released), CCDs and I/O dies are connected using FOEB (Fan-Out Embedded Bridge). Moving to a 3D structure dramatically shortens wiring distances, improving both latency and power efficiency. Incidentally, Zen 6–based EPYC Venice also uses FOEB, and its appearance differs slightly from Zen 5–based Turin and earlier packages.
Regarding yesterday’s OpenAI funding round, it appears that the 3 GW of inference-only capacity provided by NVIDIA as part of its $30 billion investment will be based on Groq, with which NVIDIA partnered last year. The partnership with Groq is reportedly a non-exclusive licensing agreement, but NVIDIA has hired Groq’s founder and built an internal development team. The extent of internalization is unclear, but announcements are expected at GTC in March.
Groq’s LPU is expected to use 3D stacking with cache memory, and LPUs will be rack-mounted and connected using NVLink Fusion. The GPU used is reportedly “Feynman,” manufactured on TSMC’s A16 process using BSPD (backside power delivery), which TSMC refers to as Super Rail. Groq refers to its processor as an LPU—Language Processing Unit—adding yet another letter to the expanding “XPU” family.

NVIDIA has reportedly begun shipping samples of Vera Rubin, with full-scale shipments expected in the second half of 2026. This is just a personal impression, but both NVIDIA and AMD seem to be rushing GPU sales. It appears there is a growing expectation that as various NPUs like Groq’s emerge and CPU+NPU configurations become the mainstay for low-power inference, GPUs may no longer be indispensable.

With sufficiently trained AI models, systems may not necessarily require GPU-heavy configurations. For companies that do not perform reinforcement learning frequently, the need for GPUs may be intermittent. In such cases, GPU+CPU deployments could become cost-disadvantageous.
Personally, I find it hard to justify buying a high-end GPU costing several hundred thousand yen just for occasional reinforcement learning. Without a clear path to monetization, it is not an attractive investment. (That said, business models already exist—and will likely grow—in which companies offer reinforcement learning as a service.) GPUs will still be needed for supercomputing applications such as drug discovery and weather prediction, so demand will not disappear. However, the explosive growth in AI semiconductor demand since last year, and the massive multi-gigawatt data center contracts being signed over multi-year periods, may reflect an anticipation that a future dominated by inference will eventually lead to saturation in training demand. From that perspective, it is not unreasonable to think that vendors feel compelled to sell GPUs aggressively now.
Finally, some TSMC-related topics. TSMC’s 2 nm process capacity is reportedly fully booked for the next two years. Mass production began late last year. Publicly announced 2 nm users include AMD’s Zen 6 CCDs and Fujitsu’s FUJITSU-MONAKA computing die, both of which are likely already coming off the line. Ongoing production likely includes Apple silicon, and NVIDIA has announced plans to use 2 nm for Rubin Ultra. MediaTek is also expected to develop products on 2 nm.

Regarding fabs, TSMC reportedly plans to build a new 2 nm manufacturing site in Tainan. Procedures are expected to be completed within 2026, with production targeted to begin in 2028.

There was also news about TSMC’s withdrawal from the GaN business. TSMC announced last summer that it would exit the GaN foundry business for power devices by July 2027. After withdrawal, Rohm—TSMC’s partner in the GaN business—is expected to take over. Gallium is a semiconductor material known for quite some time and is considered one of the rare metals.

One more TSMC-related note, or rather Taiwan-related: reports say that the U.S. government, via the CIA, has warned U.S. tech companies about the risk of a Chinese invasion of Taiwan by 2027. There seems to be a growing number of articles expressing concern about the concentration of semiconductor supply in Taiwan. The February 24 edition of The New York Times reportedly ran a feature on risks surrounding Taiwan.

This Blog text was translated by AI from Japanese Source Blog.

Saturday, February 28, 2026

FEB.(2nd Half) 2026