Today's News
Nvidia discloses $530B in off-balance-sheet guarantees, up 188% in one quarter; backstops cover only ~20% of planned global AI infrastructure growth
SemiAnalysis
Nvidia's 2Q F1/27 10-Q revealed $530B in gross off-balance-sheet guarantees—nearly six times its on-balance-sheet liabilities of $91B—up from $184B last quarter. The jump was driven by three items: supply commitments rising to $279B (96% due by F1/29), a new $108.5B guarantee on SB Energy's PORTS-Pike campus in Ohio (4.25 GW leased to OpenAI for 20 years with Nvidia guaranteeing land, power, and shell), and $36B in AI Cloud Partner program commitments ($2.35/hr/GPU floor pricing vs. $4.50–$4.60 market rates). Nvidia also disclosed $20B in datacenter leases it has signed as tenant but expects to reassign.
SemiAnalysis models that Nvidia's maximum potential gross backstop of 46.6 GW represents less than 20% of the ~240 GW of net AI IT capacity additions forecast between end-2025 and end-2030, suggesting the industry is too large for Nvidia to fund alone. The asymmetric risk-reward structure favors Nvidia: it wins on GPU sales at full margin on day one, then again on revenue-share floors or lease guarantees if demand holds; if demand falls, the backstopped Neoclouds remain solvent but Nvidia's upside is capped. The real tail risk emerges only if a downcycle steep enough to overwhelm Nvidia's ability to meet obligations coincides with its own business slowdown—a correlated event, since a Neocloud that cannot pay its lease has stopped buying GPUs. With consensus estimates of ~$441B EBITDA in F1/28 and projected cash balances reaching $1.4T by F1/31, Nvidia has substantial firepower, yet the $11T cumulative capex forecast for CY24–CY29 underscores that Gigascalers' implicit backstops remain the dominant force.
Jensen Huang (NVIDIA) via Liberty's Highlights
Jensen Huang confirmed that OpenAI's GPT-6 Astra model was trained on approximately 100,000+ Nvidia Grace Blackwell GPUs with NVLink72 interconnect, with 400,000 additional GPUs coming online next. The disclosure signals that frontier labs are rapidly scaling training clusters beyond current capacity, making electricity availability and data center power delivery a critical constraint. The question of whether 400K GPUs can work together on single training runs remains unanswered, but the infrastructure race is accelerating.
Reuters via Techmeme
Anthropic is in advanced IPO talks seeking to raise as much as $100 billion at a ~$2 trillion valuation, with Nvidia considering investing up to $10 billion as an anchor investor. The potential involvement would give Anthropic an early strategic backer while further deepening ties between the chipmaker and one of its major customers. Anthropic's annualized revenue run rate had climbed above $65 billion by end of July 2026, up from ~$9 billion at end of 2025, with company projections of $190–$200 billion in revenue by 2028. The listing, expected to complete before November's U.S. midterm elections, would mark another jump in valuation from the $965 billion post-money valuation in May 2026. Anthropic has been diversifying chip suppliers as demand for Claude has strained available computing capacity, including commitments to Amazon AWS, Google TPU capacity, and in-house custom chip design.
DeepSeek via AINews
DeepSeek released v4.1-Flash with a fundamental architectural shift—separate 8B prefill and 16B decode parameters—achieving an Artificial Analysis Index score of 40 (above v4 Pro, below GLM-5.3-Flash) while demonstrating exceptional cost efficiency at $0.27 per task, roughly 7x cheaper than competing open models. The model runs successfully on consumer hardware via SSD offloading at 300+ tokens/second with KV cache compression down to ~890 bytes/token and sliding-window attention enabling long-context agentic workloads at frontier capability levels. Sebastian Raschka argues the architectural innovation should have been labeled v5 rather than a minor point release.
OpenAI via AINews
OpenAI released GPT-Live-1 full-duplex voice API achieving 0.798s response latency and 83.6% first-attempt task completion when paired with GPT-6 Astra, with integrations from LiveKit, HeyGen, Telnyx, and Cognition's Devin Voice signaling rapid ecosystem adoption. Combined with ChatGPT Work's Data agent for enterprise data access and Box's integration framing files as "the file system for AI," the market is consolidating around persistent, governed agent environments rather than stateless model endpoints. The public-beta Agents API with hosted sandboxes represents a shift toward stateful, organization-aware infrastructure.
Bloomberg via Techmeme
Chinese AI developer Moonshot AI told investors its annual recurring revenue topped $1 billion in August 2026, up from $300 million in June, fueled by the July release of Kimi K3—a model that leapt to the top of performance benchmarks while costing drastically less than top U.S. contenders. The company is targeting $2 billion annualized revenue by end of 2026, demonstrating rapid scaling in the competitive frontier model market and significant traction outside Western labs.
Cognition via AINews
Cognition released SWE-2 claiming frontier parity at 70% lower cost while scaling reinforcement learning to multiple trillions of parameters. Cognition's acquisition of Dioxus Labs for VM and computer-use infrastructure signals vendors are building integrated stacks rather than relying on model improvements alone. Cursor's new Projects feature with shared memory across subagents and Claude Code's managed-agent session viewer reflect industry-wide shift from single-task chats toward long-lived software project substrates with explicit orchestration controls and inspectable sessions.