CoreWeave Deploys Multi-Rack NVIDIA Vera Rubin NVL72 Clusters

Amazon Web Services Chief AI and Technology Officer Matt Wood said the cost of delivering a given level of AI intelligence is falling by one or two orders of magnitude every three to six months outside the most advanced frontier models, a decline he attributed to improving efficiency as companies learn to operationalize frontier models at lower cost. Speaking to CNBC on Thursday, Wood rejected the idea that more capable AI agents must become less safe, saying developers can surround models with controls, guardrails, security and privacy protections while still improving the underlying technology, and pushed back on the notion that companies must choose between development speed and safety. Wood also defended AWS's $220 billion infrastructure investment, pointing to customer demand and internal business metrics, and said he expects additional computing capacity and greater operating efficiency to bring AI to more businesses and users. Accelerating AWS growth helped push Amazon above $3 trillion in market value for the first time in August 2026, after second-quarter results showed AWS revenue growing at its fastest pace since 2021; the stock surged more than 15% in one session, adding nearly $400 billion in market value, before gaining as much as another 5.3% as it crossed the threshold. TD Cowen analyst John Blackledge said AWS's AI business exited a quarter at roughly a $25 billion annual run rate, its fifth consecutive quarter of accelerating revenue growth, and pointed to CEO Andy Jassy's view that AWS could eventually generate $1 trillion in revenue, compared with the $600 billion long-term opportunity Jassy discussed in his April shareholder letter. Amazon expects AWS to remain capacity constrained through 2026 and 2027 because of AI demand, even as it plans to double AWS capacity by the end of 2027.

Reactor, the developer platform for real-time generative video and world models, announced that NVIDIA and Sapphire Ventures have joined its Series A round, which is led by Lightspeed Venture Partners. The company said the investment comes as demand grows for infrastructure capable of running increasingly powerful world models in real time. Reactor's proprietary inference engine, global GPU cloud and low-latency streaming layer let developers and enterprises run models at 60+ FPS and sub-40ms latency through a unified SDK and API. The platform is already used across media and entertainment, where major Hollywood studios, advertising platforms and video streaming services have active projects, and by leading world model labs in the U.S. and internationally. Reactor was co-founded by CEO Alberto Taiuti and CTO Bryce Schmidtchen, both former technical leads on Apple Vision Pro, and its team includes engineers and researchers from Apple, Meta, Google, Adobe, Replicate and Microsoft.

Nebius Group N.V. has acquired Inferize, an inference optimization company, to strengthen the production inference capabilities of its Nebius Token Factory managed inference platform, with the terms of the transaction not disclosed. The acquisition brings Inferize's technology and systems engineering team into Nebius Token Factory, and the technology is designed to reduce the time required to launch and scale large AI models and make inference workloads more elastic. Inferize's technology focuses on addressing cold starts, the time required for AI models to load before serving a request, which can leave assigned GPUs idle when demand rises and new instances are launched or when model weights are updated during workloads such as reinforcement learning. By reducing this idle GPU time, the technology is designed to let capacity scale more closely with actual usage while supporting higher capacity utilization and improved token economics. Management said Inferize brings technology aimed at accelerating how quickly systems respond to changing customer demand, along with a team experienced in GPU systems, and that the team's contribution is expected to extend beyond the initial integration. The deal adds another layer to Nebius' production inference efforts, following Eigen AI's model, kernel- and system-level optimization work and the addition of Clarifai's core team and licensed technology for system-level inference and compute orchestration.

Firmus Grid plans to allocate about half of the shares in its initial public offering to existing shareholders, potentially allowing Nvidia and Blackstone to increase their stakes, Bloomberg News reported on Monday, citing people familiar with the matter. The Australian data centre operator has received investor indications well above the offer size and is seeking to raise as much as A$5.5 billion, or $3.8 billion, including a greenshoe option. The bookbuilding deadline has been brought forward to Thursday from Friday. Firmus priced the IPO at A$11 a share, implying a valuation of about A$43.7 billion, or $30.3 billion, a listing that could rank among Australia's largest IPOs, comparable with Medibank's 2014 offering, which raised just under $5 billion. Firmus has secured commitments from investors including Nvidia and Blackstone as it expands data-centre capacity to meet demand for artificial intelligence computing, with proceeds helping fund GPUs for its first data centre in Batam, Indonesia.

Jim Cramer said on the September 28 episode of Mad Money that Space Exploration Technologies Corp. is developing into a broader technology and infrastructure company, with Starlink, AI computing, and Starship as multiple potential growth engines. Cramer pointed to SpaceX's GPU infrastructure business, saying Elon Musk is renting GPUs bought from NVIDIA to Google at an $11 billion a year run rate and has a similar deal with Anthropic at a $15 billion run rate starting in October, roughly $2.17 billion a month in compute rentals. The company's second-quarter results showed AI revenue of $2.56 billion, up 247% year over year, alongside $14.1 billion in contracted Cloud Services Agreements, though that AI revenue came with a $1.26 billion operating loss. Cramer also highlighted the September 28 Starship flight, which reached orbit and deployed 26 Starlink V3 satellites, but a Raptor Vacuum upper-stage engine shut down prematurely, cutting the planned 10-hour mission to about three hours. SpaceX generated $7.81 billion of second-quarter revenue while spending $18.37 billion on capital expenditures, $15.83 billion of it on AI, and Cramer separately warned that a large amount of restricted stock will become eligible to enter the market over the next year. Insider Monkey reported that 119 hedge fund portfolios held SpaceX at the end of the second quarter of 2026, and the stock trades at a price-to-sales multiple of 85.5x.

Amazon has established an $8 billion special purpose vehicle to buy and lease back Nvidia Grace Blackwell chips to AWS clients, an off balance sheet financing tool for AI hardware as cloud capital demands rise. Morgan Stanley has reinstated Nvidia as a Top Pick, citing its view on AI data center demand and capacity expansion. The Amazon vehicle and the Morgan Stanley call both sit inside a wider Nvidia AI infrastructure story, alongside Nvidia's own US$500 billion ecosystem financing plans. The structure lets AWS keep building AI capacity without stacking all the hardware on its own balance sheet, pointing to new funding structures that can support large orders as AI factories become more capital intensive across cloud and sovereign projects. The clearest early signal on whether the model scales would be other hyperscalers or sovereign AI buyers setting up similar chip vehicles that explicitly name Nvidia hardware, with disclosed sizes and timelines.