NVIDIA
Oct 15, 2026

NVIDIA’s OCP booth spotlighted NVLink PTFE suppliers as analyst flags tightening M6-M7 CCL supply

NVIDIA drew market attention at the OCP conference after displaying an "NVLink PTFE Board" at its booth alongside a list of seven companies tied to the exhibit: Flexium, Kinwong, SCC, SYTECH, Taiflex, TTM and VGT. The list spans materials suppliers and PCB makers across Taiwan, mainland China and the United States, though the booth photo did not indicate supply shares or confirm whether any of the companies had moved into mass production. At the same booth, NVIDIA also showed an "Ultra Dense Interconnect Cable Cartridge" and named six other companies: Amphenol, Foxconn, Lead Wealth, LOTES, Luxshare and TE. That second list points to connector and cable-module suppliers rather than the PTFE board supply chain. Taiwan-based veteran PCB industry analyst kokycpcb said market focus has been fixed on M8 and PTFE, but M6-M7 grade CCL, or copper clad laminate, may be entering its own "memory moment." He cited worsening shortages of E-glass and capacity crowding from top-tier manufacturers as factors tightening supply. On the demand side, he linked the shift to the rise of AI Agent workloads, arguing that CPU servers and the PCB materials used in them are facing mounting pressure as PCIe 5.0 and PCIe 6.0 platform demand grows.

1180
NVIDIA’s OCP booth spotlighted NVLink PTFE suppliers as analyst flags tightening M6-M7 CCL supply
Microsoft
Oct 14, 2026

Microsoft unveils Maia 300 with TSMC 2 nm N2P compute die for 2027 rollout

Microsoft introduced its third in-house AI accelerator, Maia 300, at the Open Compute Project Global Summit on Oct. 12 U.S. time, with the chip’s main compute die built on Taiwan Semiconductor Manufacturing Co.’s 2 nm-class N2P process, according to a report cited by ABMedia. The company expects large-scale deployment to begin in 2027. Maia 300 is Microsoft’s first AI chip to use a chiplet design. Microsoft said the processor splits compute, memory control, and I/O into separate dies. The chip is paired with eight stacks of HBM4 memory for a total of 288GB, up from 216GB in the previous Maia 200 generation. Power consumption is listed at 950 watts per chip, and each rack can hold 96 accelerators, double the Maia 200 configuration. ABMedia also cited The Information as saying Microsoft is in talks for capacity exceeding 300,000 Maia 300 units, with a long-term target of as many as 1 million units for delivery in 2027. The report added that TSMC is expanding 2 nm output across five fabs in Hsinchu and Kaohsiung, while the company expects N2P mass production in the second half of this year.

990
Microsoft unveils Maia 300 with TSMC 2 nm N2P compute die for 2027 rollout
DG Matrix
Oct 12, 2026

DG Matrix unveils Interport 360 Flex power platform for NVIDIA MGX-based AI factories

DG Matrix introduced its Interport 360 Flex core technology at the Open Compute Project global summit, targeting NVIDIA MGX modular systems and NVIDIA’s 800VDC architecture. The offering combines the Interport SST Cell hardware module with the Interport OS real-time software layer to help data center operators shift from conventional AC distribution to higher-efficiency DC power without scrapping existing facilities. According to ABMedia’s summary, the SST Cell uses silicon carbide power semiconductors and liquid cooling, accepts 34.5 kVAC medium-voltage input, and can output 800 VDC directly from existing site power. Interport OS manages power routing across units, ports, and storage devices on a sub-millisecond timescale while integrating UPS, STATCOM, and load-smoothing functions. DG Matrix positioned the system as a skid-mounted, end-to-end power solution for high-density AI racks, aimed at operators dealing with fast-changing AI workloads and rising compute density. The company said the modular design allows deployment of native 800 VDC power in existing spaces, which can shorten AI compute center buildout timelines and help manage energy costs while preserving prior investment in land and electrical infrastructure.

860
DG Matrix unveils Interport 360 Flex power platform for NVIDIA MGX-based AI factories
AMD
Oct 12, 2026

AMD unveils Helion agentic AI rack system with support for up to 32,768 CPU cores

AMD said it has introduced the AMD Helion agentic AI rack system, an open infrastructure design built on Open Compute Project, or OCP, standards for large-scale AI agent workloads. The company outlined two primary configurations. One is a CPU-only rack using 6th Gen AMD EPYC processors and supporting as many as 32,768 CPU cores for hyperscale deployments, where it can provide dedicated compute for agent execution while linking to separate model inference infrastructure. The other is an integrated rack that combines AMD EPYC CPUs, AMD Instinct MI350P GPUs and general-purpose compute nodes for enterprise agentic AI use cases. AMD said Helion supports both liquid and air cooling, allowing customers to adjust CPU and GPU resources based on workload, performance, power and cost needs. The system also includes open hardware and software components plus an integrated token router that can direct model calls across frontier and open-source models. AMD plans to submit the Helion specification to OCP for broader community development and adoption. The company said the product is being developed with ecosystem partners and is expected to launch in November 2026.

1130
AMD unveils Helion agentic AI rack system with support for up to 32,768 CPU cores
SK hynix
Aug 4, 2026

SK hynix and Sandisk unveil first HBF standard at FMS 2026, with Google and Tenstorrent joining alliance

SK hynix said at FMS 2026 that it has jointly released the first High Bandwidth Flash, or HBF, standard specification with Sandisk, roughly six months after the two companies formed the alliance in February this year. The specification has been disclosed through the Open Compute Project and sets out key parameters including capacities of up to 512GB, three bandwidth tiers topping out at about 3.0TB/s, and use of the UCIe interconnect interface. Google and Tenstorrent have confirmed they are joining the HBF alliance. The technology is positioned as a new storage tier between high-bandwidth memory and SSDs, combining HBM-class transfer speeds with NAND’s larger capacity scaling. During the event, SK hynix also publicly showed its 10th-generation V10 375-layer 4D NAND wafer and related product for the first time, saying energy efficiency is 2.5 times better than the previous generation and that mass production is expected early next year. In a conference presentation, an SK hynix executive also outlined a tiered memory architecture as a way to address data-processing demands in the AI inference era.

2950
SK hynix and Sandisk unveil first HBF standard at FMS 2026, with Google and Tenstorrent joining alliance
SanDisk
Jul 20, 2026

SanDisk and SK hynix push High Bandwidth Flash as a new memory layer for AI inference

SanDisk and SK hynix are promoting a new architecture called High Bandwidth Flash, or HBF, that applies HBM-style advanced packaging and die stacking to NAND Flash. The idea is not to replace High Bandwidth Memory, but to insert a new layer into AI server memory hierarchies for large, read-heavy data sets used during inference. According to public material cited in IEEE Spectrum, first-generation HBF is targeting up to 16 stacked NAND chips, as much as 512GB per stack, and read bandwidth of up to 1.6TB/s. SanDisk’s roadmap then points to 2TB/s and 3.2TB/s in later generations. SK hynix executive Hoshik Kim said the approach could deliver bandwidth well beyond standard NVMe storage and help relieve the severe capacity pressure faced by HBM. The concept is aimed at inference rather than training. In training, models require constant reads and writes, which is a poor fit for Flash. In inference, model weights are largely fixed, making static, read-intensive data such as billion-parameter weights and precomputed KV cache more suitable for HBF. The companies have already launched standardization work under the Open Compute Project, though IEEE Spectrum reported that shipping is still at least a year away and broad deployment may take several more years.

2950
SanDisk and SK hynix push High Bandwidth Flash as a new memory layer for AI inference