〔Overnight Brief〕Nvidia's $96.2B Print (+106% YoY) Re-Anchors the 'Central Bank of AI' as 2M More AWS GPUs and a 25% AI-Lab Mix Stack Up — $500B+ Financing Piles and a 90%-Cheaper Qwen Test Capex Faith

0. Overnight Arc

The AI tape re-anchored on Nvidia's Q2 FY27 print — $96.2B revenue (+106% YoY), $89.0B data center (+117% YoY) — a clean beat against Mizuho's $92B consensus that re-confirmed the "central bank of AI" framing ^[1][2][3]. The reactive question is no longer demand, it's financing: the 2M-GPU AWS deal, the >$500B lender platform, the $105B OpenAI Ohio guarantee, and a $550B 2027 AI-revenue base that Mizuho calls conservative stack new debt on top of new demand ^{[4][5][6][7][3]}. China added the parallel track — MiniMax's $800M ARR, Qwen3.8-Flash at ~10% of prior training cost, an 8.64s humanoid 100m — keeping the supply race two-front ^{[8][9][10][11][12]}.

1. The Print: Nvidia Q2 FY27

• **[NEW] Revenue $96.2B, +106% YoY; data center $89.0B, +117% YoY** for Q2 fiscal 2027 ^[2]. Nvidia CFO separately guided that AI-lab demand will contribute ~25% of the overall business next year ^[13].
• **[NEW] Framing:** Jensen Huang: "Now, compute is revenue" ^[14]; Yahoo Finance framed Nvidia as "the central bank of AI" ^[1]; Reuters read the post-print bounce as evidence that "Wall Street's AI obsession is far from over" ^[15].
• **[NEW] IPO tell:** Huang separately hinted that OpenAI and Anthropic "will go public soon" ^[16] — same week the Wall Street Journal reported Google moving its AI-responsibility team out of the DeepMind lab ^[17], and the same week a Chinese-language long-form detailed Anthropic's talent raid of OpenAI/Google, with OpenAI's S-1 already filed in June and both startups eyeing mega-IPOs ^[18].

2. The Partnership: AWS Expansion and Rubin

• **[NEW] AWS-Nvidia stack expansion:** 2M additional NVIDIA GPUs across AWS global infrastructure, the NVIDIA Vera CPU on AWS for agentic AI, NVLink Fusion with NVHBM for faster, more power-efficient memory, and "AI factories" for the U.S. government including 100,000 GPUs on secure infrastructure ^[4][5][6].
• **[NEW] Rubin economics:** Morgan Stanley projects Rubin will contribute close to $9B in revenue in the fiscal quarter ending October, with Huang set to emphasise Rubin's improvement on Blackwell's AI-factory economics — the investor question is whether Rubin expands the pie or simply pulls Blackwell demand forward ^[19].
• **[NEW] Call setup:** the Q2 earnings call is the gating event for the next session's capex repricing ^[20].

3. The Sustainability Test: Financing and Capex Faith

• **[ESCALATED] Debt layer:** Nvidia, with six large financial institutions, has built a >$500B financing platform for AI infrastructure; separately, it is providing up to $105B in guarantees for OpenAI's Ohio data-center project ^[7]. Mizuho's Vijay Rakesh maintains a "Outperform" rating and a $300 target, expects $2.08 EPS on $92B revenue (the actual print came in at $96.2B), and argues the Street's $550B 2027 AI-revenue base is conservative — citing only the Nvidia-lender platform and SpaceX's gigawatt-scale Austin solar build as potentially adding another $400-500B ^[3]. A separate Jin10 read framed the consensus beat as a foregone conclusion, with capex guidance — not the headline — as the swing factor ^[21].
• **[ONGOING] Talent and corporate churn:** the Anthropic-vs-OpenAI talent war and OpenAI's June S-1 keep the IPO clock ticking; Google is moving its AI-responsibility team out of DeepMind per WSJ ^[18][17].
• **[NEW] Other capital conduits (single-source relay):** Softbank is weighing up to $20B in bonds for OpenAI financing, and OpenAI claims its in-house AI chip test performance surpasses the Nvidia GB300 ^[22].

4. The China Counter-Push

• **[NEW] MiniMax (per source transliteration):** August ARR exceeded $800M ^[12]; Q2 revenue +81.8% QoQ, with July token usage at 20x January's level ^[9]; the firm is adapting the M3 and H3 models to domestic chips, with a large-scale domestic compute cluster coming online to take real production traffic ^[8].
• **[NEW] Qwen3.8-Flash:** Alibaba released and open-sourced the model on August 26 — 1T total parameters activating only 6B, with claimed frontier performance beyond Claude Opus 4.6; training cost down ~90% vs Qwen3.7-Plus; input priced at 1 yuan per million tokens, output 3 yuan, as low as 1/3 of DeepSeek-V4-Flash ^[10]. First stop is the "Qianwen Office" product, with API access via the Qianwen AI platform.
• **[NEW] Robotics sprint:** the Tianjiao team reset the humanoid 100m record to 8.64s at the 2nd World Humanoid Robot Games, down from the opening 9.39s and the semifinal 8.85s ^[11][22].
• **[NEW] Other Chinese moves (single relay):** JD.com released "Super Brain" large-model 3.0; Tencent executives acknowledged Hunyuan Hy3 is competitive at equivalent parameter scale but overall compute shortage has slowed training and product development; Zhipu publicly claimed the "stealth model" Ox Alpha — 1M-token context, multimodal, free — and said usage on OpenRouter has run at 2x+ DeepSeek, with weights released the same evening; Tesla denied "FSD entry into China hopeless" rumours, said it has filed a police report, and that the data centre is operating normally with AD hiring accelerating ^[22][23][24].

5. Source Quality and Falsifiable Tests

• **Contradictions worth printing:** Mizuho's $550B 2027 AI-revenue base is called conservative, with a potential $400-500B add ^[3], yet the Jin10 piece frames near-term capex demand as strong but flags "financing sustainability worth attention" ^[21]. The two are not contradictory, but they bracket the trade.
• **Single-source items:** the Anthropic/IPO read comes from one Chinese-language long-form ^[18]; Huang's "Now, compute is revenue" quote is a single X wire post ^[14]; the 8.64s robot record traces to a single CCTV relay ^[11][22]; the OpenAI-vs-GB300 chip claim is a single relay ^[22]; the Softbank $20B bond item is a single relay ^[22]. Treat as data points, not consensus.
• **Falsifiable test for the next session:** Nvidia's capex commentary on the Q2 call ^[20] and any update on the $500B+ lender platform / SpaceX-Austin revenue bridge ^[7][3] are the swing items; on the China side, the print from Qwen3.8-Flash's first-night traffic on the Qianwen platform and any follow-on from the MiniMax domestic-cluster go-live will set the second-derivative ^[8][10].