NightIndex Night navigation for signal-rich markets
August 25, 2026 8 signal desks
AI Industry 2026-08-25 中文

NIGHTLY INTELLIGENCE BRIEF

〔Overnight Brief〕Nvidia Bundles the AI Factory - Groq 3 LPX Live, Vera Rubin 30x/MW, SpaceXAI Starmind - Yet Pre-Earnings Selloff, Hyperscaler Defection and SEC Probe Converge

Nvidia put Groq 3 LPX into full production for agentic AI, claiming 3,400 output tokens/sec on Gemma 4 31B at 100K context and a 4x lead on agentic coding, while Vera Rubin NVL72 posted 30x throughput per MW and 35x lower per-token cost versus GB300 [3][13]. SpaceXAI will use Nvidia's Vera CPU and launch Starmind satellites on NVL72, and Nvidia made a strategic investment in power-infrastructure developer Lancium [4][5]. Yet the stock sold off pre-earnings after Trump and Bessent comments hit tech, and chip names fell on the Nvidia news [1][2]. The moat narrative is now openly contested: hyperscalers, AMD, Broadcom, AI startups, and even OpenAI and Anthropic are building chips [15]. The SEC has subpoenaed Wall Street banks over AI hedge fund Situational Awareness [7]. On the China side, Eoptolink's H1 revenue rose 100% to 20.91B yuan with optical gross margin at 48.46% [11], XPeng's He Xiaopeng targeted next-year no-safety-driver Robotaxi operations and took the robot CEO seat personally [8][9], and Tuosida took a 49% stake in optical-coupling maker Laile for 147M yuan [10]. Nvidia earnings decide the tape.

12 sources 5 min read

0. Overnight Arc

The overnight session split Nvidia into two stories running in parallel: a product blitz - Groq 3 LPX full production, Vera Rubin NVL72 throughput claims, SpaceXAI's Starmind satellites, a strategic Lancium investment - layered against a pre-earnings selloff as Trump and Bessent comments dragged the tape and chip stocks fell on the Nvidia news [1][2][3][4][5]. WSJ flagged "AI jitters and geopolitical threats" as the week's opening chord [6]. On the regulatory side, the SEC has subpoenaed major Wall Street banks for information on AI hedge fund Situational Awareness, a single-source relay from China's CLS wire - treat as unverified pending US-side confirmation [7]. In China, Eoptolink delivered a 100% revenue jump on optical interconnect, XPeng's He Xiaopeng personally took the robot CEO seat with a next-year no-safety-driver Robotaxi target, and Tuosida moved into optical-module coupling via a 147M-yuan stake [8][9][10][11].

1. The Nvidia Stack: Inference, Power, Orbit

  • **[NEW] Groq 3 LPX, full production:** Nvidia's interactive-AI inference accelerator is shipping, hitting 3,400 output tokens/sec on Gemma 4 31B at 100K-token context and up to 4x the response speed of the nearest competing platform on agentic coding workloads, per Nvidia and the Nvidia Newsroom X account [12][3][13].
  • **[NEW] Vera Rubin NVL72 benchmarks:** on SemiAnalysis's AgentX workload with DeepSeek V4 Pro, the platform delivers up to 30x throughput per MW and up to 35x lower per-token cost versus the prior-generation GB300 NVL72 [3].
  • **[NEW] SpaceXAI + Vera CPU:** SpaceXAI will use Nvidia's Vera CPU to accelerate agentic AI, with the Vera Rubin NVL72 system slated to launch Starmind AI satellites; Vera is positioned as the first CPU purpose-built for AI agents [5].
  • **[NEW] Power-stack integration:** Nvidia made a strategic investment in Lancium, a power-infrastructure developer, threading compute and grid together [4]. Musk separately described his design as "greatly simplified, lower cost, higher density and lighter than traditional racks" [14].

2. The Moat Under Siege

  • **[ESCALATED] Hyperscaler defection risk:** Google's and Amazon's in-house silicon efforts are extending to third-party sales, while AMD's and Broadcom's data-center businesses have reached tens of billions of dollars in revenue [15].
  • **[ESCALATED] Customer-chip collision:** OpenAI and Anthropic are pursuing their own silicon, reducing dependence on Nvidia [15].
  • **[ESCALATED] Startup capital wall:** AI chip startups are raising tens of billions from investors to chase the next architectural breakthrough [15].
  • **[NEW] Pre-earnings price action:** Investor's Business Daily said Trump and Bessent comments sparked tech losses, and Finimize flagged Nvidia news as the proximate trigger for chip-stock declines; WSJ tagged AI jitters and geopolitics as the week's backdrop [1][2][6].
  • **[NEW] IEA paradox:** the IEA surfaced the tension that AI is becoming dramatically more energy efficient even as data-center electricity demand surges [16].
  • Source quality: the "moat under siege" framing is single-source via a 财联社 relay republished on Tiger Sniff, and the SEC probe is single-source via a 财联社 wire; quote the band, not the point [7][15].

3. China: Robots, Optical, Compute

  • **[NEW] Eoptolink (XinYiSheng) H1:** revenue 20.91B yuan (+100.34% y/y), net profit 7.53B yuan (+90.98%), optical-interconnect product gross margin 48.46% (+0.99pp); capacity expanded from 15.2M to 28.36M units, sales from 6.95M to 11.19M units; product matrix spans 400G/800G/1.6T and LPO/LRO/XPO/NPO/CPO routes, with a 1.6T DR4, 6.4T NPO and 12.8T XPO family released in March 2026 [11].
  • **[NEW] XPeng:** He Xiaopeng disclosed the second-generation VLA pre-production Robotaxi has cleared 2,000 internal test orders in Guangzhou, with the cloud takeover platform complete and a target of next-year no-safety-driver passenger operations; he will personally serve as CEO of the robot business, integrating Turing AI chip, VLA model, Robotaxi and humanoid robot technology [8][9].
  • **[NEW] Leapmotor:** Vice President Li Tengfei confirmed robot-business plans on the H1 results call, with a formal announcement imminent; a July entity - Huzhou Lingsheng Precision Manufacturing, 210M-yuan registered capital - has been set up under subsidiary Zhejiang Lingsheng Power for industrial robot manufacturing and intelligent robot R&D [17].
  • **[NEW] Tuosida:** will pay 147M yuan for a 49% stake in Laile Optoelectronics, an optical-module precision-coupling equipment maker, with the seller committing 30M yuan to buy Tuosida A-shares within six months of first disclosure [10].
  • **[NEW] Standards track:** MIIT will seek public comment on the 2026 National Humanoid Robot Industry Standards Guideline (draft), targeting at least 100 key standards by 2028 across capability testing, key technologies, platforms/systems, scenarios and safety governance [18].
  • **[NEW] Nebius:** completed a $5.75B convertible-note issuance [19].
  • **[NEW] Yuejiang (2432.HK) H1:** revenue ~3.2B yuan (+107% y/y), gross margin ~47%, net loss ~110M yuan, adjusted net loss ~60M yuan; R&D ~100M yuan (+148% y/y), nearing 2025's full-year ~115M yuan; embodied-intelligence customer count reached 231 by end-June [20].
  • **[NEW] Robot conference dispatch:** a Caixin reporter's notebook piece - "applause and follow-up questions at the booths" - signals the industry remains in show-floor mode, not commercial scale [21].

4. Regulatory and Macro Cross-Currents

  • **[NEW] SEC probe:** subpoenas have gone to major Wall Street banks for information on AI hedge fund Situational Awareness's trading, per a 财联社 relay - single source, treat as unverified [7].
  • **[NEW] China agent-payment rules:** the Payment and Clearing Association issued the Agent Payment Application Self-Regulation Convention, building "Know Your Agent" (KYA) on top of KYC and requiring members to file before any live agent-initiated payment rollout [22].
  • **[NEW] ByteDance org:** TRAE and Coze teams are being absorbed into Doubao; "Doubao Work" is set to debut as the unified AI-office product this week, with TRAE IDE and CLI retained as a programming line under the Doubao brand; teams report to Doubao product head Zhao Qi [23].
  • **[NEW] Musk and the rack:** Musk's "simplified, lower-cost, higher-density, lighter" comment is the only direct read on the compute-rack economics debate in the packet [14].
  • **[ONGOING] AI-energy IEA framing:** AI efficiency gains coexist with surging data-center demand, a tension the IEA is now publicly unpacking [16].
  • **[ONGOING] Nvidia moat stress:** the competitive pressure piece catalogs customers, peers and startups all building alternatives, even as Nvidia's growth still outruns the field [15].

SOURCE TRAIL

Citations

23 records

  1. [1]
  2. [2]

    Google News — AI 产业Nvidia News Sends Chip Stocks Lower - Finimize ↗

  3. [3]
  4. [4]
  5. [5]
  6. [6]
  7. [7]
  8. [8]
  9. [9]
  10. [10]
  11. [11]
  12. [12]
  13. [13]
  14. [14]
  15. [15]

    虎嗅 · 全部资讯英伟达，“护城河”遭围攻 ↗

    relevance 0.64

  16. [16]
  17. [17]
  18. [18]
  19. [19]
  20. [20]
  21. [21]
  22. [22]
  23. [23]