TLDR Hardware 2026-08-17

Nvidia's SEC filing revealed it holds nearly 123 million SpaceX shares, worth roughly $21 billion at the end of June (now closer to $17 billion after SpaceX's post-IPO share slide), a stake that originated through Nvidia's investment in xAI before Musk merged it into SpaceX. It comes right after Musk confirmed SpaceX will build its data centers exclusively on Nvidia hardware, citing the Vera Rubin architecture as "the best AI computer," locking in another major customer to Nvidia's silicon while Nvidia simultaneously holds equity in that same customer.

The bear case on AI infrastructure spending has always been simple: with Nvidia shipping a new architecture roughly every year, GPUs should lose economic value within two or three years, undermining the long-dated debt financing the whole buildout leans on. CoreWeave's Q2 earnings call complicated that story, disclosing a renewed contract for Nvidia's 2020-era A100 GPUs running all the way through 2029, at pricing CEO Mike Intrator said matches “full freight” from years ago.

Ukrainian military intelligence says Russia's S-71 “Monochrome” cruise missile, which offers reduced observability and autonomous targeting, uses an Nvidia Jetson Orin NX module, a consumer-grade system-on-module normally sold for robotics, drones, and edge AI research rather than military hardware. Nvidia confirmed the Orin NX isn't officially sold in Russia or export-controlled, and said it can't track products after resale but will act on confirmed export-control violations. In the missile, the chip likely powers a local electro-optical perception system, recognizing images in real time to assist terminal guidance, echoing an earlier Ukrainian claim that Russia used the same modules in Shahed drones.

Semiconductor equipment OEMs are increasingly adopting build-to-print contract manufacturing to scale production and meet soaring global chip demand without heavy capital expenditures. As expanding AI fabs drive severe equipment backlogs, toolmakers partner with electronics contract manufacturers to produce sub-assemblies from existing engineering drawings. This distributed model provides immediate access to cleanrooms and precision machining, cutting equipment lead times, preserving OEM capital, and improving supply chain resilience across regional fab expansions.

A 19-company industry coalition led by Lightmatter has formally launched the Open Silicon Photonics for AI Systems initiative as an official Open Compute Project workstream. Alongside a 300-page foundational architecture white paper, the initiative aims to establish an open, vendor-neutral blueprint for integrating co-packaged optics into datacenter infrastructure. Targeting OCP Modular Hardware System and Open Rack v3 standards, the framework scales clusters from 72 to over 1,024 nodes while overcoming copper reach and bandwidth bottlenecks.

The semiconductor and datacenter industries are actively debating the necessity of 1-megawatt server racks driven by continuous, power-intensive AI workloads. While hyperscalers explore ±400V DC power distribution, advanced liquid cooling, and 3D-IC packaging to support ultra-dense compute, critics argue this scaling vector is unsustainable. Instead of cramming accelerators into single dense enclosures to preserve short copper reach, alternative approaches advocate scaling horizontally across racks using optical interconnects and specialized silicon.

Intel VP Robert Hallock confirmed the company plans to keep Raptor Lake CPUs in production and stabilize supply of its LGA 1700-compatible chips “for years to come,” following a spike in demand for the four-year-old Alder Lake and Raptor Lake lineups. It's a direct response to the DDR5 price crunch: Raptor Lake supports both DDR4 and DDR5, letting buyers upgrade without paying today's inflated DDR5 prices, and rumors point to a possible third refresh, tentatively “Raptor Lake Next,” arriving alongside the DDR5-exclusive Nova Lake in early 2027.

Semiconductor Engineering reports that Co-Packaged Optics cannot scale to profitable high-volume manufacturing without industry-wide testing standardization. As datacenters adopt optical interconnects for AI workloads, test cells face severe bottlenecks from proprietary fiber connectors, non-standardized laser alignment schemes, and fragmented data formats.

United States utility-scale battery storage projects are stalling as aging electrical grids struggle to accommodate rapid interconnection demands.

The US Commerce Department has warned Apple against sourcing memory chips from Chinese suppliers CXMT and Yangtze Memory Technologies to alleviate ongoing supply shortages.

Nvidia has scaled back its planned financial guarantee for OpenAI's 10-gigawatt Ohio data center from $250 billion to under $120 billion, covering only the project's first phase, after investors raised concerns about the chipmaker's risk exposure.

The Missile Defense Agency is transferring mature, in-sustainment programs like THAAD and Aegis/SM-3 to the Army and Navy to free up bandwidth for newer priorities feeding into Golden Dome.