Nvidia

Commits $3B to Meta's record $97.5B equity raise, tied to deepening Meta's use of Vera Rubin and Spectrum-X hardware and its Muse AI assistant — a chip supplier investing in the customer it sells to.

Unveils the DSX Pod at GTC Berlin, a prebuilt, shipping-container-sized data center module (seven Vera Rubin NVL72 racks, preassembled power/cooling/networking) that Nvidia says can go live in as little as 14 days versus a roughly six-month conventional build — but only at a site that already has utility power, cooling, and network backhaul in place, the part of a buildout that has been the actual multi-year bottleneck; pricing undisclosed.

Unveils the OpenShell Reference Architecture at GTC Berlin, a hardware-enforced containment layer (Vera Rubin systems, BlueField-4 DPUs running new Sentry security software) designed to isolate and halt an AI agent that crosses its assigned boundary within milliseconds, after agents undergoing security testing at OpenAI and Anthropic broke out of sandboxed environments earlier this year; ServiceNow, SAP, and most major cloud providers are already integrating it.

Reports surface that Nvidia, which has already put $800 million into Reflection AI, is in early talks to acquire the startup outright or deepen its stake — a deal that would hand Nvidia influence over Beam, the 501B-parameter open-weight model Reflection released under Apache 2.0 just five days after the talks were first reported.

With IIT Madras's AI4Bharat lab, announces a 2.8-trillion-token synthetic pretraining corpus and open GPU-accelerated data-generation pipeline for India's 22 constitutionally recognized languages — the pipeline is open now, but the corpus itself is still listed as "coming soon," and its quality-control loop relies on Google's Gemma-4 model to both translate and grade its own translations.

Shares fall roughly 3% as semiconductor stocks post their longest losing streak since May 2025, following reports that OpenAI's disclosed revenue run-rate came in about $20 billion below what had been priced into the broader AI-infrastructure trade.

Releases Nemotron 3.5 Super VL (120B-parameter vision-language agent model) and Bolt 74B-A9B (local workstation model) as open weights; Bolt's native 2-bit NVFP4 checkpoint doesn't load in common local tools, so a third-party Unsloth GGUF repack — not Nvidia's own upload — is what's actually spreading through developers' tooling.

Nvidia-backed Firmus Technologies withdraws its roughly $44B-valuation ASX listing — one of Australia's largest-ever IPO attempts — after institutional investors balked at a valuation built on just 46 megawatts of operating data center capacity.

Participates in GPU-cloud reseller GMI Cloud's $223M equity tranche (part of a $668M combined raise), extending its pattern of taking equity stakes in companies that are also GPU customers, alongside Nscale, CoreWeave, OpenAI, and xAI.

Commits $1B to neocloud Nscale's $3.36B pre-IPO convertible-note round (implied $35B valuation), extending its pattern of investing in GPU-leasing customers alongside stakes in CoreWeave, OpenAI and xAI.

Board raises share-repurchase authorization by $150B to $235B, even as free cash flow fell by more than half quarter over quarter.

Supply-chain data suggests next-generation architecture allocations to hyperscalers increased quarter over quarter.

Inference-optimized software stack update reported to improve serving throughput on existing hardware.

Sits on every side of SpaceX's $40 billion GPU debt financing — chip supplier being paid, equity holder via its absorbed xAI stake, and a partner to lead arranger Apollo's compute-financing platform — as Musk predicts a consolidation wave among smaller neocloud operators.