Nvidia: Record $235B Buyback, Morgan Stanley Top Pick, and Rubin at 20% of Data Center Revenue
Nvidia enters October 2026 at all-time highs with the strongest capital-return and demand signals of the cycle. On September 28, the board authorized a $150 billion increase to its repurchase program, bringing the total to $235 billion — "the largest single stock buyback authorization in U.S. corporate history" (Motley Fool, Oct 1). Morgan Stanley reinstated NVDA as its semiconductor top pick on Oct 2, betting "AI infrastructure cycle has room to run" (ROIC.ai); the stock sits at 98.2% of its 52-week range at a $5.65T market cap (/markets/NVDA/2026/10/05).
The demand side: fiscal Q2 2027 (ended July 31) revenue was $96.2B with data center revenue of $89.0B (+117% YoY, accelerating for a fourth straight quarter), and management guided FQ3 to roughly $108 billion ±2% — "notably without assuming any China data-center compute revenue." Rubin began shipping in August and is expected to account for 20% of data center revenue in FQ3 (reported in November); per the Fool: "Demand for Rubin is so massive that Nvidia is guiding for 70% year-over-year revenue growth in fiscal 2028.1" Higher memory costs "have yet to make a substantial dent in Nvidia's gross margins" (74.7% TTM) — a testament to pricing power.
China remains effectively foreclosed: Nvidia's fiscal 2026 filings describe the company as "effectively foreclosed" from China's data-center compute market and guidance assumes zero China data-center compute revenue; the only flow is small H200 shipments (ByteDance and Tencent each received ~10,000 chips per the FT).
The bear counterpoint resurfaced in the same week: Michael Burry revived his critique of Nvidia's GPU depreciation schedules, drawing a parallel to a 1960s market bubble ("We Have All Been Here Before") — the same circularity-and-financing debate tracked in The AI Slowdown Debate Becomes a Market Structure: Chips Sold, Hyperscalers Bought. And the memory shortage now reaches into Nvidia's roadmap: Micron is building "the industry's first custom-HBM4E implementation, NVHBM" for next-generation GPUs and NVLink Fusion (The Memory Shortage: Micron's Record FQ4 2026 Extends Tightness Through 2028), while Rubin Ultra's HBM capacity (192-256GB range still being tested) remains unresolved.
What it means: guidance and sell-side momentum remain decisively ahead of the slowdown narrative — but the marginal buyer now includes Nvidia itself at record scale, the depreciation/circularity critique is the live bear case heading into the November FQ3 print, and the ~$108B FQ3 guide ex-China is the number the Street will measure the cycle's durability against.
-
An instance of Every layer of the AI buildout is doubling at once. — The silicon layer keeps guiding growth ahead of the slowdown narrative, with inference demand compounding the entire compute stack. ↩︎