NVIDIA Reveals SpaceX Holdings as Sixth-Largest Shareholder, Codex Unlocks Million-Token Context Window

Capital Moves: NVIDIA Bets on SpaceX, Stripe's Massive OpenRouter Acquisition

This week saw a flurry of capital activity in the AI space, with multiple blockbuster deals surfacing simultaneously. According to Bilibili's AI Daily Report, NVIDIA's latest holdings disclosure revealed for the first time that it holds approximately 122.8 million shares of SpaceX, making it SpaceX's sixth-largest shareholder. This revelation drew intense market attention — as the world's core supplier of AI computing power, NVIDIA's deep investment in a space technology company is widely interpreted as an attempt to build a broader frontier technology investment portfolio beyond AI infrastructure.

From a strategic perspective, NVIDIA's data center business revenue already exceeds 80% of total revenue, with its GPU product line supporting the vast majority of global AI training and inference tasks. The deeper logic behind investing in SpaceX lies in this: SpaceX's Starlink network is building a global low-Earth orbit satellite communications infrastructure, and the global deployment of future AI inference services — especially for remote areas and mobile scenarios — requires ubiquitous network connectivity. Additionally, frontier concepts like satellite edge computing and space data centers also require high-performance GPU support, creating deep potential synergy between the two companies at the "compute + connectivity" technology layer. More transaction details await disclosure in subsequent filings.

Meanwhile, payment giant Stripe also made major moves. Bloomberg reported that Stripe has finalized its acquisition of AI startup OpenRouter for over $7 billion, though neither party has issued an official announcement. OpenRouter, as a routing platform that aggregates APIs from multiple large model providers, has rapidly gained influence among developers in recent years. Its core business model provides developers with a unified API interface, aggregating dozens of model providers including OpenAI, Anthropic, Google, and Meta behind the scenes — developers don't need to integrate with each API individually but can call different models on demand through OpenRouter while enjoying automatic load balancing and cost optimization. This "model aggregation routing" positioning is highly consistent with Stripe's aggregated payment thinking in the payments domain. If the deal is confirmed, it means Stripe is extending its payment infrastructure aggregation logic to the AI model distribution layer, aiming to control the billing gateway for model calls and compete for dominance in the model invocation market.

Domestic Capital Markets Heating Up in Parallel

On the domestic front, Yushu Technology announced that its stock will debut on the Shanghai Stock Exchange's STAR Market on August 19, with an issue price of 150.80 yuan per share, marking another case of an AI-related company sprinting toward the capital market. The STAR Market, established in 2019, is positioned to serve "hard technology" companies, adopts a registration-based system, and allows unprofitable companies to list — representing China's institutional support for technological innovation. The 150.80 yuan issue price is in the mid-to-high range for STAR Market listings, reflecting the current market's valuation premium for AI sector companies while also meaning the company will face stricter information disclosure and performance delivery requirements.

Additionally, Alibaba reportedly plans to sell its gaming subsidiary Lingxi Interactive (gaming business) to accelerate its full pivot toward artificial intelligence. The existing management team is expected to remain stable in the short term after Xincheng Capital takes over — this move once again confirms the strategic trend of major tech companies "focusing on the AI main track and divesting non-core assets."

Model Capability Upgrades: OpenAI Codex Opens Million-Token Context Window

The most developer-anticipated news in this report is undoubtedly the capability update for OpenAI Codex. According to Tibor, ChatGPT subscribers can now use version 5.6 in Codex, gaining access to million-level (1M) context processing capability.

The context window is the maximum number of tokens a large language model can process in a single inference pass, directly determining how much information the model can "see." Early GPT-3.5 supported only 4K tokens (approximately 3,000 English words), GPT-4 extended this to 128K, and million-level context means the model can process approximately 750,000 English words at once — equivalent to over a dozen complete technical books, or the entire source code of a medium-to-large software project. This capability breakthrough relies on architectural innovations such as Sparse Attention and Ring Attention, as well as more efficient KV Cache compression techniques, freeing the model from information truncation and context loss problems under short windows. For handling large engineering projects, cross-file refactoring, and long-chain debugging, this represents a qualitative leap — developers can feed an entire code repository into the model at once and receive holistic understanding and suggestions.

However, capability upgrades come alongside tightening commercialization. Community users report that after exhausting Codex quotas, a paid reset option appears. Screenshots from Plus users show that paying $8 can restore quotas to 100%, though this mechanism has not been officially confirmed. This detail reflects that AI coding tools are gradually transitioning from "free trial" to "granular billing," and how to balance developer experience with commercial sustainability will become a long-term competitive focus for all players.

Domestic Large Models Continue to Push Forward: Zhipu GLM and Alibaba Qwen Audio 3.0

The open-source and domestic model camps are also active. Zhipu opened applications for the GLM series' cybersecurity line — security researchers and relevant institutions can gain greater usage freedom after approval, demonstrating model providers' openness in security research scenarios. This practice draws from the international "Red Teaming" approach — by opening model restrictions to security researchers, potential vulnerabilities and adversarial attack paths are proactively discovered, strengthening the model's security defenses before official release.

Alibaba released Qwen Audio 3.0 and launched it on the Qianwen platform, covering automatic speech recognition (ASR), text-to-speech (TTS), and real-time interaction, with multiple metrics reportedly ranking first globally. Continued breakthroughs in voice multimodal capabilities are opening new possibilities for on-device intelligent assistants and real-time interactive applications — from natural conversational human-computer interaction to real-time meeting transcription and multilingual simultaneous interpretation, the voice modality is becoming a key gateway for large models to enter everyday applications.

On-Device and Agent Deployment: Meituan's 90,000 Employee Coverage to Hardware Adaptation

On the AI application deployment front, several developments are particularly representative. Meituan has reportedly been advancing large-scale organizational Agent development — transitioning from earlier "spending tens of millions of yuan daily" in rough investment to systematic implementation, with its internal Agent platform now covering 90,000 employees and having built 30,000 Agents.

AI Agents differ from traditional conversational AI in their ability to autonomously plan, invoke tools, and execute multi-step tasks — not just answering questions, but decomposing tasks, calling APIs, operating systems, and completing end-to-end workflows. A deployment of this scale means the enterprise must solve a series of engineering challenges including Agent permission management, data security isolation, task orchestration reliability, and exception rollback mechanisms. The transition from "spending tens of millions daily" to organizational implementation reflects companies moving from rough trial-and-error in the exploration phase to production-grade system governance, including establishing Agent effectiveness evaluation systems, cost accounting mechanisms, and human-machine collaboration process standards. This figure also vividly demonstrates the aggressive pace of leading internet companies in enterprise-internal Agent deployment at scale, signaling that the "one AI assistant per person" office paradigm is moving from concept to reality.

Hardware On-Device Acceleration

On-device AI deployment is also picking up speed. MediaTek announced that its Dimensity automotive cockpit platform CX-1 and flagship mobile chips have completed adaptation for the Qwen 3.8B model, enabling faster on-device integration.

On-device inference refers to running AI models directly on local devices (phones, automotive chips, etc.), with core advantages including low latency (no network round-trip needed), privacy protection (data stays on device), and offline availability. The downside is limited device compute power, restricting operation to smaller parameter models. Qwen 3.8B (3.8 billion parameters) is Alibaba's lightweight model optimized specifically for on-device scenarios. MediaTek's Dimensity platform integrates an NPU (Neural Processing Unit) that can compress the model from approximately 7.6GB at FP16 to under 2GB for operation through INT4 quantization and other techniques, dramatically reducing memory and power requirements while maintaining core capabilities. Chip manufacturers proactively adapting mainstream large models signals that a hybrid inference architecture of "cloud for complex tasks, on-device for high-frequency simple tasks" is gradually becoming the industry standard, providing a hardware foundation that balances capability and speed for smart cockpit voice interaction, on-phone local assistants, and other scenarios.

On the consumer application front, Doubao added remote phone-to-computer control functionality, allowing users to handle cross-device tasks from mobile. Alibaba launched Happy Shrimp, where users simply describe an emotion, story, or memory, and the model can complete an end-to-end creation process covering lyrics, composition, arrangement, and vocals. AIGC is evolving from single-modality to full-pipeline creation where "one sentence generates a complete work." This end-to-end multimodal generation capability represents a milestone leap in AI creative tools from "assisting humans" to "autonomous completion."

Industry Concerns: Surge in AI Customer Service Complaints and Worker Rights

Amid the technology sprint, there are also warning signs worth noting. A report from the China Consumers Association shows that AI customer service-related complaints have increased significantly, with difficulty reaching human agents, false promises, and inaccurate generated content being the main points of contention. Most current AI customer service systems are based on large language model Retrieval-Augmented Generation (RAG) architectures that retrieve from enterprise knowledge bases to answer user questions. However, the model's inherent "hallucination" problem (generating seemingly plausible but actually incorrect information) is particularly fatal in customer service scenarios — incorrect return/exchange promises or warranty terms can directly cause consumers financial losses. This reminds the industry: while pursuing cost reduction and efficiency, excessive AI replacement of human agents may actually damage user experience and corporate reputation. "Human-machine collaboration" — where AI handles standardized questions while complex and emotionally charged scenarios are promptly transferred to humans — rather than "complete replacement," may be the more pragmatic path.

Additionally, one news story sparked widespread discussion: a 39-year-old programmer died suddenly in a company restroom, and the Shenzhen Human Resources and Social Security Bureau determined that although he had clocked in, he hadn't reached his workstation, and temporarily declined to classify it as a work-related injury. Against the backdrop of the AI industry running at high speed with practitioners generally enduring intense work pressures, this case once again brings the health and labor rights of tech workers to the forefront. The long-standing "996" and even more extreme work culture in China's internet industry, compounded by the urgency of the AI race, continues to amplify physical and mental health risks for practitioners — something the entire industry should deeply reflect upon.

Conclusion

From NVIDIA's space investments and Stripe's massive acquisition, to Codex's million-token context and domestic models' multi-front breakthroughs, this week's AI world has seen capital and technology racing forward on parallel tracks. Meanwhile, the emergence of AI customer service complaints and worker rights concerns provides a sobering mirror for this boom. The pace of technological progress is exhilarating, but how to make AI truly serve people and how to protect the practitioners behind the technology are equally unavoidable questions for the industry.