AMD's Madhu Rangarajan, corporate VP of Compute and Enterprise AI, says the company never set out to build its Zen 6 EPYC 9006 "Venice" chips around agentic AI, a not-so-subtle dig at NVIDIA and Arm, which both pitch their latest server CPUs as made for that exact job. Rangarajan still called EPYC 9006 the best CPU for agentic workloads during a media session covered by Tom's Hardware. Chief Zen architect Mike Clark said much the same in a separate talk, admitting AMD didn't see agentic computing coming, though he added that the company always expected the CPU to keep a role in AI software. Here's what Rangarajan said:

We think that the EPYC 9006 CPU is the best CPU for the agentic era, and spoiler alert, we didn't magically know four years ago that agentic was going to be a thing and purpose-built something for agentic. The reality is an agentic workflow is a diverse set of tasks… and in order to execute those tasks well, you need compute that is able to handle a diverse set of workloads with a full portfolio.

ASRock Rack is promising up to 4 TB of DDR5 on a single socket with its DMRD16-2L+, one of the first motherboards built for Intel's upcoming Xeon 7 "Diamond Rapids" processors. The board was shown at the OCP Global Summit 2026 and has since been detailed by ServeTheHome. It has 16 DIMM slots in a one-DIMM-per-channel layout, matching the 16-channel memory setup Intel detailed at HOT CHIPS. Regular RDIMMs top out at 128 GB per module and 8000 MT/s, while MRDIMMs run at up to 12,800 MT/s, with taller 2U MRDIMMs reaching 256 GB each. Populate all 16 slots with those, and you hit 4 TB, along with a theoretical bandwidth of roughly 1.6 TB/s.

Expansion is unusual, with no standard PCIe slots and no OCP NIC 3.0 connector. All I/O goes through up to 16 MCIO x8 connectors, each rated for PCIe 6.0 or CXL 3.0, for 128 lanes in total. The rest is fairly basic: one PCIe 5.0 x4 M.2 slot, two Intel I350 Gigabit Ethernet ports, an ASPEED AST2700 BMC with a dedicated management port, and VGA output. ASRock Rack fits it all on a 12 x 13-inch EEB PCB, about the size of a large E-ATX board, so it could squeeze into some bigger workstation cases. Pricing and availability have not been shared yet, but that is hardly surprising, since Intel has the Xeon 7 family slated for a 2027 debut on the new LGA9324 socket. The board is aimed at servers, though anyone planning a high-memory workstation for local AI will probably keep an eye on it.

Maxsun has shared full specifications for the SK-PTLNAS and SK-WCLAIO, two NAS-focused MoDT (Mobile on Desktop) motherboards built around Intel's Panther Lake and Wildcat Lake chips. Both were first shown at Computex 2026 and are expected to launch this month, though pricing is still unknown. The SK-PTLNAS now tops out at the Core Ultra 9 386H, a 16-core chip with four Xe3 GPU cores, a 25 W base power and up to 80 W of turbo power. The Computex listing mentioned options up to the Core Ultra X9 388H, which no longer seems to be on the table. It supports up to 128 GB of DDR5-7200 memory across two SO-DIMM slots, with five M.2 slots and three Mini SAS ports for storage. The board also gets a PCIe 5.0 x8 slot, 10GbE, 2.5GbE and two 1GbE ports, plus two Thunderbolt 4 ports. Maxsun lists Windows 11 and Ubuntu support.

The SK-WCLAIO measures just 90 x 165 mm, a big change from the mATX design shown in June. It runs the six-core Core 7 350 (2P + 4LPE) with two Xe3 cores, a single DDR5-6400 SO-DIMM slot for up to 48 GB, 1 TB of onboard UFS 3.1 storage and four M.2 slots. Wildcat Lake has very few PCIe lanes, so those drives probably won't each get full x4 bandwidth. Maxsun rates the system at 45 W when active and 16 W with drives on standby, which should suit an always-on home NAS. The power input is a bit unclear for now, since the listing mentions both a 12 V DC jack and USB-C PD.

PCs with 32 GB and 64 GB of RAM are becoming less common, at least in the latest CPU-Z Validator statistics. According to Doc TB, one of the developers behind CPU-Z, this is the first time RAM capacity on PCs has declined year over year. 32 GB is still the most popular configuration at 37.2% of validations, though its share is down 2.5% compared to a year ago. 64 GB took a much bigger hit, falling 18% to 7.3%. Lower capacities are moving the other way, with 16 GB up 12.9% to 33.0% and 8 GB up 5.5% to 11.0%, while 96 GB and 128 GB setups dropped by more than 26% each. Memory type tells a similar story, with DDR4 now making up 54.5% of validations, up 6.0%, while DDR5 slipped 2.9% to 32.9%.

Of course, these figures only cover systems that users actually submit to the validator over the last three months, so they aren't a full picture of the PC market. Steam's September survey, for example, showed 32 GB overtaking 16 GB as the most common configuration among gamers, but that promptly flipped with the October survey. Still, the timing makes sense. Memory prices have climbed sharply this year, and Team Group expects the worst of the shortage to land between Q4 2026 and Q2 2027. DDR4 platforms may also get a second wind, with DDR4 motherboards getting released for unannounced LGA1700 processors expected in early 2027. It looks like plenty of people are sticking with older DDR4 systems or settling for 16 GB instead of paying current prices.

We might not be past the worst part of the global memory shortages, as the most severe supply chain squeeze could happen soon. During Team Group's earnings conference on November 2, the company stated that the worst period in memory demand is projected to be between Q4 2026 and Q2 2027, which is just a few months away. Major memory makers such as SK hynix, Samsung, and Micron have sold out their 2027 supply and entered into long-term agreements with customers, making it increasingly difficult to source memory for any third-party purposes. Team Group states that even if you have the funds, it might be impossible to purchase memory products. Currently, it is possible to buy memory, albeit at a high price, but soon enough, that may no longer be an option, noted the company.

Even though CXMT and YMTC are supplying close to 10% of the memory market share with their products, it is still impossible to satisfy demand. A supply-demand balance may not emerge until the second half of 2028 or even later. Currently, Team Group sells most of its memory modules to businesses in the B2B sector, which commands 64.6%, while the B2C sector, selling directly to consumers, accounts for 35.4%. This is a sharp reversal compared to the 2025 Q3 figures, where B2B was 24.6% and B2C was 75.4%. Most of the revenue now comes from the business sector developing AI servers, rather than traditional consumer purchases.

Tuxedo Computers has just announced its Nano Pro Gen15, a new mini PC powered by Intel's Panther Lake Core Ultra X7 358H CPU and its capable Arc B390 iGPU. The mini PC is a workstation targeting developers and home lab enthusiasts who want to run "local, secure, sovereign" LLMs, but Tuxedo also fully acknowledges that it can be used as a no-fuss mini gaming PC, thanks to Intel's graphics integrated into the Linux kernel of the pre-installed TuxedoOS. Physically, the Nano Pro Gen15 measures in at just 110 × 117.5 × 49 mm, and it features a plastic chassis, although Tuxedo has maintained a repair-friendly design, with just four screws required to access the internals, where the upgradeable RAM, storage, and wireless cards are easily accessible.

Tuxedo supplies a 120 W external power supply that connects via a barrel adaptor, but it makes no claims about the actual maximum TDP of the Intel CPU inside. Connectivity on the Nano Pro Gen15 is reasonable, if not stellar, with Wi-Fi 7 and dual 2.5 GbE for networking, alongside Bluetooth 6 for peripheral connectivity. On the back of the mini PC, there is a DC-in power port, the aforementioned 2.5 GbE ports, 2× USB 3.2 Gen 2 Type-A ports, 2× HDMI ports with HDMI 2.1 support on Windows and 2.0b support on Linux (for now). On the front of the mini PC, there is a USB4 port, 1× USB 3.2 Gen 2×2 Type-C port with DP 1.4a Alt Mode, and a USB 3.2 Gen 2 Type-A port alongside a 3.5 mm audio combo jack. Tuxedo says that, between the USB Type-C ports and the dedicated HDMI ports, the Nano Pro Gen15 can support up to four simultaneous displays.

A lot of information relating to Intel Nova Lake-S has emerged leading up to the official launch in early 2027, including official information from Intel detailing the Nova Lake architecture and a recent benchmark leak that promised significant single-thread performance improvements on a 28-core Nova Lake-S part. However, a recent leak out of @Haze2K1 on X, shared by VideoCardz, purports that the real-world performance improvements for Intel's Nova Lake desktop CPUs are in the double digit range.

While the leaker has been around for a while, there is no guarantee that this information is accurate, since the only cited source is "circulating chatter," however, the claim is that Intel's Nova Lake will outperform comparable last-generation Arrow Lake CPUs by 8-17% in raw single-thread performance, 15% in general-purpose workloads, and up to 20% in gaming. There is some nuance to these claims, with the 17% improvement figure exclusive to the higher-end CPUs that feature bLLC, while the 17% boost allegedly comes from the implementation of Intel's new Advanced Performance Extensions.

Samsung Electronics today reported financial results for the third quarter ended Sept. 30, 2026. The Company posted KRW 195.3 trillion in consolidated revenue, another all-time quarterly high, representing a quarter-on-quarter (QoQ) increase of 14%. Operating profit reached KRW 107.4 trillion, making Samsung Electronics the world's first tech company to exceed KRW 100 trillion in quarterly operating profit.

The Device Solutions (DS) Division posted a QoQ sales increase of 15%, with the Memory Business setting an all-time high for quarterly revenue and operating profit, as semiconductor demand is set to remain robust. The Device eXperience (DX) Division posted a QoQ revenue increase of 8% and recorded an operating loss of KRW 0.8 trillion.

SK hynix reported its latest Q3 2026 quarterly results with a record $78.129 billion in net revenue. However, the company also shared a market outlook, highlighting its long-term supply agreements and plans for expansion with new fab capacity and facilities under construction. For the consumer market, the outlook is grim as SK hynix expects memory demand to continue outpacing supply through 2030. The main supply constraint is the complexity of leading-edge semiconductor nodes and the lengthy lead times for fab construction, which take years to translate into regular mass production volumes.

Despite sustained demand, pricing is also on an upward trajectory, with the average selling price of memory modules in Q3 increasing by over 20% compared to Q2. This is concerning, as it suggests that the prices of consumer electronics devices might jump by more than 20% once memory inventory is depleted. Although SK hynix expects a decrease in PC and smartphone shipments, this will be offset by a strong rollout of AI data centers. About 14 major customers, including most hyperscalers, have signed 5-year long-term agreements, which will account for over 50% of total capacity sold for 2027 and 2028.

AMD's Zen 6 just got one of its first rough per-clock comparisons against Zen 5, and the early numbers look good. Leaker HXL spotted a Geekbench 7.1.0 listing for the 256-core EPYC 9996, the flagship "Venice" chip AMD launched in July, scoring 2,847 points in single-core on a Dell PowerEdge R9815 server running Linux. User three twenty six on X then locked a Ryzen 7 9700X to 4.1 GHz, matching the EPYC 9996's maximum boost, and got 2,444 points on the same Geekbench version. That works out to roughly 16.5% higher single-core performance for Zen 6 at the same frequency. Apparently, Zen 6 came out ahead in all 15 single-core subtests, with gains ranging from 8.2% in File Compression to 25.7% in HTML5 Browser.

Of course, these scores only give us a general idea about the gains. The EPYC 9996 uses the dense Zen 6c cores, not the full Zen 6 cores expected in desktop Ryzen chips, so it's hard to say how closely these numbers will carry over. The two systems are also wildly different, with a server platform on one side and a 9700X running DDR5-6000 CL34 under CachyOS Linux on the other, and the EPYC listing has since been pulled from the Geekbench database. Still, a double-digit per-clock gain from the dense core variant is quite encouraging. AMD has already confirmed that "Olympic Ridge" desktop CPUs will stay on AM5, though they reportedly won't arrive until 2027, so we can expect more comprehensive performance leaks well before then.

YMTC, known as a NAND Flash manufacturer, is reportedly preparing to start mass production of LPDDR5X memory. While it has been known for some time that YMTC plans to enter DRAM production alongside its main NAND Flash business, it was not widely known that the company had already begun low-scale manufacturing. According to Counterpoint Research, YMTC has been producing about 5,000 wafers per month of 16 Gb LPDDR5X DRAM running at 7,500 MT/s at its Fab 2 since October 2025. The company is now preparing to begin volume production of LPDDR5X memory with speeds of 9,600 MT/s at its Fab 3, using Vertical Channel Transistor (VCT) 4F² cell structure technology.

YMTC has a significant advantage with 3D silicon, which represents the next phase of memory evolution. With its Xtacking 4.0 technology, YMTC can separate memory cell wafers from peripheral logic circuits and combine them using hybrid bonding wafer-to-wafer stacking. Hybrid bonding creates a permanent bond between two silicon chips using embedded metal bumps that are fused together to form a single chip. Multiple dies or entire silicon wafers can be connected by stacking the back ends of two chips together.

Micron has decided not to bring 4F² DRAM to market, even though it says it invented the cell design and has been researching it for around two decades. Scott DeBoer, Micron's President and Chief Technology and Products Officer, shared the decision during the company's technology webinar on October 21. On paper, 4F² uses vertical channel transistors to shrink each memory cell's footprint by about a third compared with the 6F² layout used in today's DRAM. However, DeBoer argued that real-world density is limited by how closely the capacitors can be packed, not by the cell's footprint, calling the density advantage over 6F² "a bit of a myth." He also said 4F² can be built, but at a higher cost and complexity, and that it runs into fundamental scaling problems at leading-edge nodes, which points toward 3D DRAM instead.

That puts Micron at odds with its rivals. Samsung and SK hynix are both developing 4F² designs, and China's CXMT says its first 4F² product will be a server DDR5 RDIMM due by the end of this year. Micron will keep scaling 6F² instead, with its 1-delta node set to ramp in Boise, Idaho, in the second half of 2027 and the 1-epsilon node after it moving to wafer bonding. 3D DRAM is being developed as a parallel path, but there's no timeline yet, and Micron says it will only switch once the economics make sense. It's a risky bet if rivals manage to make 4F² cheap at scale, especially with memory supply expected to get even tighter in 2027 and 2028, per Micron's own estimates.

Xiaomi has officially spilled the beans on the 2027 update to its Redmi Book laptop series, detailing both a launch date and what to expect from the laptops in terms of performance and battery life. The information comes courtesy of the Xiaomi Redmi account on Weibo, and the announcement reveals an October 27 release date for the Redmi Book 14 and likely the 16, alongside it. It's so far unclear if and when the new Redmi Book laptops will launch internationally or what the pricing will be, but the outgoing Redmi Book 14 starts at 6,499 CNY with 16 GB of RAM and 512 GB of storage.

With the post, Xiaomi also confirms that the Redmi Book 14 and 16 will be using Intel's Wildcat Lake Core Series 3 processors, with previous leaks having tipped the Core 5 320 for the Redmi Book 16, and the fine print on the marketing images now confirming the Core 5 320 for the Redmi Book 14. The Redmi Book 14 2027 has also been confirmed to ship with a 72 Wh battery, which Xiaomi claims will deliver over 20 hours of office use, with a claimed 1.26 days of local video playback. It will also use LPDDR5X-7467 RAM, although capacities are unclear at the time of writing, aside from the confirmation that there will be a 16 GB version available. The fine print also confirms that the display will be 2.8K, however the panel tech isn't mentioned.

Axiomtek, a world-renowned leader relentlessly devoted to the research, development, and manufacture of innovative, highly efficient, and reliable industrial computer products, proudly introduces the MANO330, a compact Thin Mini-ITX motherboard powered by the Intel Processor N97 or N150 and DDR5 memory. Designed for space-constrained systems that require flexible connectivity and expansion, the MANO330 combines multi-display capability, dual LAN connectivity, versatile M.2 expansion, and industrial I/O in a compact platform. With a wide 12-24 VDC power input, the MANO330 provides system developers with a flexible foundation for smart manufacturing and industrial automation, digital signage, smart retail, and POS applications.

Compact Performance for Space-Constrained Systems
The MANO330 combines a compact Thin Mini-ITX design with Intel Processor N97 or N150 and DDR5-4800 memory support. The N97 supports up to 32 GB of memory, while the N150 supports up to 16 GB. Measuring just 170 x 170 mm, the MANO330 provides the essential computing performance and I/O capabilities needed for embedded systems, helping developers build space-efficient solutions for a wide range of industrial and commercial applications.

Team Group Inc., a global leader in memory and storage solutions, today announced that its T-CREATE EXPERT Ai CQDIMM DDR5 Desktop Memory has won the GOOD DESIGN AWARD 2026. Designed for creators running generative AI locally, the EXPERT Ai CQDIMM DDR5 is Team Group's first DDR5 CQDIMM desktop memory to offer a capacity of 128 GB per module. Combining an independent clock driver with patented IC grading verification technology, it delivers a default data rate of 8,000 MT/s at the JEDEC-standard voltage of 1.1 V without overclocking. A two-slot configuration provides a total capacity of 256 GB, overcoming memory capacity limitations and enabling creators to reliably handle larger AI workloads for a smoother, more flexible creative experience.

Established in 1957, the GOOD DESIGN AWARD employs a rigorous, multistage evaluation process. Entries are assessed from four perspectives—human, industry, society, and time—with consideration given to design philosophy, user experience, and long-term value. For multiple consecutive years, Team Group has received this award recognizing outstanding designs that improve people's lives. The EXPERT Ai CQDIMM DDR5 continues T-CREATE's minimalist design approach, balancing aesthetics with practical functionality. By integrating creators' needs for high performance, low power consumption, and large capacity into its design, it offers a desktop memory solution tailored to local AI computing. Recognized by an international panel of design experts, the EXPERT Ai CQDIMM DDR5 demonstrates Team Group's expertise in addressing creators' needs through technological innovation.

The ever-growing list of mini PCs powered by AMD's Gorgon Halo APUs has just received its latest addition in the TOPC TXA2, a new workstation built around AMD's Ryzen AI Max+ Pro 495 APU and its powerful Radeon 8065S iGPU. The TOPC TXA2 is available in two configurations, with either 2 TB of storage for $6,399 or 4 TB of storage for $6,699 on TOPC's online store—even though the price is high, it's still a solid $900 cheaper than AMD's in-house Ryzen AI Halo Developer platform.

The TXA2 also packs 192 GB of LPDDR5X-8533 unified memory, with up to 160 GB shared with the iGPU for workstation use. Despite the 180 ×180 × 78 mm, or ~2.5 L chassis, TOPC claims that the TXA2 is capable of a 120 W TDP. It's also no slouch in the I/O department, with Wi-Fi 7 and Bluetooth 5.4 for wireless networking and connectivity, and it has a healthy port selection, to boot.

Intel's upcoming "Nova Lake" desktop platform is drawing nearer, and today we have some intriguing details thanks to the Chinese electronics and motherboard maker Decenta. In leaked information about the motherboard SKU "M26-NVLS-ATX09," we see that it uses Intel's new LGA1954 socket and is listed with Q970 and B960 chipset options. The most notable difference between the two chipsets seems to be their expansion capabilities. Previously leaked specifications suggest that the Q970 provides a total of 44 PCIe lanes, including eight PCIe 5.0 lanes supplied by the chipset. In contrast, the B960 reportedly offers 34 lanes, with no additional PCIe 5.0 lanes from the chipset itself.

Both chipsets feature PCIe 4.0 connectivity, although the Q970 offers 12 chipset lanes compared to the B960's 14. Decenta's ATX motherboard supports four DDR5-6400 memory modules, with dual-channel operation and a maximum capacity of 256 GB. Its expansion layout includes PCIe 5.0 slots, two M.2 storage connectors, and a legacy PCI slot for older industrial hardware. For wireless connectivity, we see that Decanta's platform supports expansion through an M.2 E-Key socket that supports Wi-Fi 7 and Bluetooth 6.x modules. VGA, HDMI, and DisplayPort outputs allow up to three displays to operate simultaneously.

Marvell Technology, Inc. , a leader in data infrastructure semiconductor solutions, today announced the second-generation Marvell Structera X 3406 and Marvell Structera X 3506 CXL memory expanders, designed to help cloud, enterprise and AI infrastructure providers build more efficient, scalable and higher-performance systems for the next era of AI and computing. Designed to address the evolving needs of leading hyperscale datacenters, Structera X addresses the most demanding data center workload requirements.

AI and ML workloads are rewriting the rules of data center design. Growing LLM sizes, expanding context windows and growing KV-cache requirements are driving unprecedented memory demands that traditional server architectures simply cannot meet. Memory is a critical constraint of AI performance and infrastructure efficiency.

Microsoft today launched its biggest in-house compute processor for deployment on the Azure Cloud, the Microsoft Azure Maia 300. Announcing the silicon, Microsoft CEO Satya Nadella said that the company would be scaling its Maia family of chips to multi-gigawatt deployments, making it the company's primary AI inferencing system for its own products and services. The company is currently deploying its predecessor, the Maia 200, across the US, Europe, and Australia regions, where it is delivering 70% more tokens per Watt across Copilot and Teams, than AI GPUs. The new Maia 300 will offer a further 40% uplift in tokens per Watt over GPU-based solutions.

The Microsoft Azure Maia 300 is a multi-chiplet tile-based processor consisting of a main compute die built on the latest TSMC N2P (2 nm) foundry node, flanked by two I/O dies built on an older node, four DDR RDIMM memory tiles that serve as system-level main-memory; and 288 GB of 9 GT/s HBM4 memory for the compute accelerator across eight stacks. The HBM4 bandwidth on offer is a whopping 18 TB/s. The entire SoC has a maximum TDP of 950 W.

GIGABYTE has released BIOS updates that prepare its entire Intel B760 and H610 motherboard lineup for upcoming LGA1700 processors, which the company says are "expected to launch in early 2027." As reported by Videocardz, the update covers DDR4 and DDR5 boards alike, so owners only need to flash the new firmware and drop in the new CPU once it arrives. GIGABYTE doesn't name the processors in its press release, and Intel has yet to announce anything new for the socket. However, the timing lines up what we've been reporting for months. Intel is said to be readying "Raptor Lake Next" chips for LGA1700 in 2027, based on the original Raptor Lake silicon.

That makes GIGABYTE's BIOS the closest thing to a confirmation of new DDR4-compatible Intel processors for 2027, and it comes before Intel has announced them. It looks like the firmware was rolling out quietly even before the press release. The B760 Gaming X AX, for example, received its F18 BIOS on September 29, with a changelog that lists a CPU microcode update and "preparation for upcoming processor support." GIGABYTE also claims to be the first vendor to cover every B760 and H610 model. The press release makes no mention of Z790, Z690, or B660 boards, and there's no word on whether those will follow.

HPE (NYSE: HPE) today announced the first four of its HPE ProLiant Gen13 Servers powered by 6th Generation AMD EPYC Server processors. The new systems expand compute, memory, and I/O capabilities to help enterprises accelerate agentic AI and inference, analytics, virtualization, simulation, and modeling while strengthening security, simplifying operations, and providing reliability across distributed IT environments.

As AI systems evolve from generating insights to taking autonomous action, trust becomes a prerequisite for scale. McKinsey's 2026 AI Trust Maturity Survey found that nearly two-thirds of respondents cited security and risk concerns as the top barrier to fully scaling agentic AI. Within this evolving landscape, HPE ProLiant Gen13 helps customers address security and operational barriers with a trusted infrastructure foundation that combines HPE iLO 8 innovations for automatic protection, infrastructure trust, and post-quantum readiness with enhanced visibility and control from HPE Compute Ops Management.

AMD's upcoming EPYC "Verano" processors will apparently use a new socket called SB1. The name comes from server cooler maker Dynatron, which now states on its website that its cooling solutions support "AMD EPYC Verano on Socket SB1." Dynatron also has a product page up for a 4U active cooler listed as "SB1-4U-ACTIVE (TBD)," so the final product name is still pending. The "B" in SB1 could point to a BGA package soldered to the board, although that is a guess based on the name and nothing official. The listing doesn't help much either. Maximum power is marked as TBD, and the 128.05 x 106.6 x 130.7 mm dimensions are identical to Dynatron's J26 cooler for SP7, so the specs could simply be placeholders for now.

Verano is the codename for the EPYC 9006 LP, a Zen 6 part that AMD announced alongside the rest of its 6th Gen EPYC lineup. It tops out at 72 cores with boost clocks of up to 5 GHz, and it runs on 24 channels of LPDDR5X memory through replaceable SOCAMM2 modules. AMD is aiming it at AI racks as a host processor for GPU clusters, with a launch planned for the second half of 2027. The regular "Venice" chips use different sockets, namely SP7 for up to 256 cores and SP8 for up to 128. For "Verano," AMD itself hasn't revealed the socket name yet.

Valve's September 2026 Steam Hardware & Software Survey shows that 32 GB of RAM is now the most common memory capacity among participating gamers. The 32 GB tier gained 4.77 points to reach 42.22%, while 16 GB fell 3.38 points to 37.82%. Last month, 16 GB still led with 41.20% against 37.45%, so that is a big swing for a single month. The idea had been around for a while, with 32 GB already being floated as the next standard back in August 2025. Steam only reports capacity, so the survey says nothing about speed or whether those kits are DDR4 or DDR5. It makes sense, though, with several recent games asking for 32 GB RAM in the recommended specs, with 16 GB now often listed in the minimum requirements.

Elsewhere, AMD gained 1.89 points on Windows to reach 47.41% CPU share, and the GeForce RTX 5070 is the most common GPU at 5.86%, up 2.30 points. 1440p primary displays climbed 5.18 points to 27.04%, and Windows 11 keeps growing at 71.34%. Of course, one month doesn't make a trend, so it's reasonable to wait for October's numbers to see whether 32 GB remains at the top. The survey is also voluntary, and the sample changed a lot this month. For example, Simplified Chinese jumped 7.42 points to 31.39% of surveyed systems. That change in sampled systems could explain part of the RAM swing, though it could also simply be 32 GB RAM gaining popularity.

NVIDIA's supply commitments climbed to $279 billion in the quarter ended July 26, up from $119 billion the quarter before. The company says the increase is primarily related to memory procurement. Morgan Stanley estimates that NVIDIA has claimed about 37.3% of global high-bandwidth memory (HBM) supply for 2027, with Alphabet at 36% and AMD at 12.1%. That adds up to roughly 85%, leaving everyone else to split the rest. These are Morgan Stanley's estimates, as reported by The Inference.

About $267 billion of the commitments come due by the end of fiscal 2029, according to the filing. NVIDIA reported record revenue of $96.2 billion for the quarter, but memory costs are weighing on margins. Gross margin was 75% and is guided to 74% this quarter, with a trough of 71% to 72% expected in the January quarter. CFO Colette Kress tied that pressure to memory pricing. Operating cash flow also fell to $24.1 billion from $50.3 billion the quarter before. Jensen Huang said on September 17 that he expects NVIDIA to sell twice as many chips next year, and the company's formal guidance is about 70% revenue growth.

Gigabyte, a global leader in computer hardware, today announced the Gigabyte AI TOP 100 B850, a high-performance desktop engineered to bring workstation-class AI compute directly to the deskside. Designed to reduce cloud dependence and lower the barriers to local deployment, the system enables businesses, developers, and researchers to run advanced AI workloads locally, giving them greater control over sensitive data and AI infrastructure while reducing reliance on cloud and token-based services.

Built for Next-Generation AI Development
At the heart of the Gigabyte AI TOP 100 B850 is the 16-core, 32-thread AMD Ryzen 9 9950X processor, delivering the performance required for modern AI workloads ranging from local LLM fine-tuning and generative AI to deep learning research and agentic AI. The system comes equipped with 128 GB of DDR5 memory and a 2 TB PCIe Gen 4 SSD, providing ample capacity and high-speed access for large datasets, local model storage, and complex development environments.