NVIDIA

Nvidia is simultaneously the supplier, the equity holder, the customer, and the backstop for CoreWeave. How is that not a closed financing loop?

Per CoreWeave's S-1 filed March 3, 2025, Nvidia holds a beneficial ownership stake of more than 5% at IPO. Through 2024, Nvidia paid CoreWeave approximately $320M as a customer. Nvidia is also contractually obligated as a buyer of last resort for unsold capacity through 2032, with an initial backstop obligation of approximately $6.3B. That is four simultaneous roles — supplier, shareholder, customer, and capacity backstop — with one counterparty. CoreWeave's total debt at December 31, 2024 was $8.0B, of which $7.6B was GPU-collateralized. Microsoft accounted for 62% of CoreWeave's 2024 revenue; the top two customers combined accounted for approximately 77%.

MIT says roughly 95% of enterprise GenAI pilots show zero measurable P&L impact. Nvidia's data-center revenue was $115.2 billion in FY2025. Those two facts exist simultaneously?

They do. MIT Project NANDA found approximately 95% of enterprise GenAI pilots showed no measurable P&L impact, with only approximately 5% showing measurable ROI — sourced via Fortune, August 18, 2025. Nvidia's FY2025 Data Center revenue was $115,186M, up 142% year over year, per the FY2025 10-K. The current buyer is not the enterprise end-user — hyperscalers drove the demand. The question is whether hyperscaler capex remains rational when the workloads they are building for show a 5% P&L conversion rate.

Director Mark Stevens sold $802 million in Nvidia shares. Is there a 10b5-1 plan on record for that?

Per EDGAR Form 4 XML filings covering 2025 through June 2026, director Mark Stevens executed approximately 4,669,912 code-S sales generating approximately $802M. No 10b5-1 plan was detected in the Form 4 footnotes for those transactions. Director Harvey Jones sold approximately $88M with no detected plan; directors Hudson, Burgess, Neal, and Drell also sold without detected plans. CEO Jensen Huang's $1,049M sale was on a 10b5-1 plan adopted March 20, 2025, with a defined ceiling of 6 million shares — a structurally different arrangement. Total all-insider code-S across all filers is approximately $2,501M. The CEO-only framing understates the full picture by roughly half.

Amazon shortened server useful lives from six to five years, citing AI obsolescence. Three other hyperscalers stretched their useful lives by billions of dollars. What does that combination say about the earnings funding Nvidia's $115 billion?

Amazon's FY2024 10-K (filed February 7, 2025) disclosed approximately $920M in accelerated depreciation in Q4 2024 plus roughly $0.6B in further 2025 operating-income impact from the server-life shortening, explicitly citing AI and ML obsolescence. Meta moved to 5.5 years in January 2025, reducing FY2025 depreciation by $2.9B. Google moved servers from 4 to 6 years in FY2023, reducing 2023 depreciation by $3.9B. Microsoft moved to 6 years in FY2023 with approximately $3.7B in favorable impact. An attributed estimate from Michael Burry dated November 11, 2025 puts the cumulative hyperscaler depreciation understatement at approximately $176B through 2028, with Oracle earnings overstated by approximately 26.9% and Meta by approximately 20.8% by 2028 on that estimate — attributed, not audited. Three of four hyperscalers stretched lives; Amazon is the first to reverse and name AI as the reaso…