Nvidia's SpaceX Stake, Token Loans, and the AI Grind: Weekly Tech Roundup

Nvidia's Backdoor Into SpaceX

Nvidia has quietly become a major SpaceX shareholder, thanks to a strategic equity conversion that traces back to its January investment in xAI. The chipmaker disclosed in a recent SEC filing that it holds 122.8 million Class A shares of SpaceX, worth about $21 billion at the end of the second quarter. That stake has since dipped to roughly $17.2 billion, as SpaceX stock has fallen about 18% from its late-June peak.

According to FactSet, Nvidia now ranks as SpaceX's sixth-largest shareholder. Elon Musk still dominates with a position valued around $850 billion, followed by Alphabet at about $78 billion. The stake originated from Nvidia's $10 billion investment in xAI in January, which converted into SpaceX equity when SpaceX acquired xAI in February at a $1.25 trillion valuation. This is the first time Nvidia has publicly disclosed the position.

The relationship goes beyond paper. Musk announced on SpaceX's quarterly earnings call that the company will exclusively use Nvidia chips for its AI data center infrastructure going forward—a clear nod to the deepening ties between the two firms.

China's 'Token Loans' Are Real, and They're Tiny

China's first financial product tied to AI token consumption, dubbed "Token loans," launched in Guangzhou's Haizhu District, and the Bank of China is already reporting early numbers. The bank's Guangzhou branch has approved five loans totaling 28 million yuan (about $3.9 million), with three of those already disbursed at 8 million yuan combined. The remaining 20 million yuan is still being finalized in loan agreements.

The loans are meant to help companies pay for computing power. This is part of a broader push in Haizhu to treat token usage as a proxy for AI business activity. Since June, the district has offered subsidies based on daily token consumption: companies burning 100 million tokens a day can get up to 20,000 yuan, 500 million tokens gets 100,000 yuan, and 1 billion tokens earns 200,000 yuan.

Dong Ximiao, chief economist at Merchants Union Consumer Finance, argues that token consumption gives banks a window into how AI companies actually operate. Traditional lending relies on physical collateral, but AI firms are often asset-light. Token usage reflects how often models are being called—a dynamic indicator of customer engagement and product-market fit. It shifts risk assessment from static balance sheets to live operational flow.

Unitree's IPO Lottery: A Record Low Odds, and Winners Who Can't Brag

Unitree Robotics, the Chinese quadruped and humanoid robot maker, held its Shanghai IPO lottery draw on August 11. The result: a winning rate of just 0.01809759%, the lowest ever for a STAR Market listing. More than 9.78 million investors applied, also a record for the board—a sign of how hot the humanoid robot sector has become.

Winners are understandably thrilled, but some are keeping quiet. One investor, who goes by Kaka, told media he'd celebrate quietly. "I'll just go out for a simple dinner, then life goes on as usual," he said. Another winner, Yang from Chengdu, estimated a windfall of 200,000 yuan and planned to use it to buy a Xiaomi car. But he's not posting on social media. "I'm afraid colleagues will be jealous, and relatives might ask for loans," he explained. Instead, he's treating family and close friends to a meal.

The 90-Hour Work Week Is the New Normal in AI

If you thought AI would make work easier, think again. Reports from current and former employees at top AI companies paint a picture of relentless crunch. At OpenAI and Anthropic, "sprint" periods—major research pushes or product launches—can see engineers working over 90 hours a week. One former OpenAI technical staffer said his typical week ran 70+ hours; moving to a startup brought it down to 50-60, but even then, weekends were often sacrificed for emergency fixes.

The pressure stems from the hyper-competitive AI race. Companies are compressing development cycles to stay ahead in model quality, compute infrastructure, and product rollout, shifting the burden onto researchers and engineers. Meta employees describe being "drafted" into AI teams without much choice, building automated systems and evaluation infrastructure with no clear end in sight. The result is burnout and a constant sense of being on call.

Research from UC Berkeley suggests AI tools speed up individual tasks but don't free up time—they just accelerate the pace and pile on more work, including extra effort to verify AI output. MIT scholars note that when AI boosts efficiency, management tends to add new demands rather than grant time off. The workload doesn't shrink; it morphs.

ByteDance Forms AI Data Security Unit, and Xiaohongshu's Option Controversy

ByteDance has reportedly created a new first-level department called AI Data and Security, running parallel to its Seed and Flow groups. The unit, led by Wang Yinglei, consolidates several scattered AI data teams—Global Data, the group's data platform DMC, and Flow's AIDP—into a single organization. The integration started in June and is still ongoing. The new department is tasked with providing cross-modal data services for all of ByteDance's large models, covering the entire data production pipeline. ByteDance hasn't commented officially.

Meanwhile, Xiaohongshu (Little Red Book) is facing renewed criticism over options. A former employee, Jiang Dong, claims he was fired just eight days before his first options vesting date, which he sees as deliberate timing to void his equity. He says nearly 50 former colleagues in his WeChat group report similar experiences. This follows a public dispute by Chen Hao, a former sales head, who was let go five months before his options matured. The company maintains the terminations were lawful, and courts have sided with Xiaohongshu so far.

Anthropic's Ambitious Revenue Projections and DeepSeek's Rollercoaster Week

Anthropic is reportedly telling investors to expect revenue between $190 billion and $200 billion by 2028—a figure that hasn't been publicly disclosed before. The company is gearing up for what could be one of the largest IPOs ever, with some investors valuing it at over $2 trillion. It's also in talks to acquire Israeli AI firm Decart for around $6 billion, which would be its largest acquisition to date.

In other news, Anthropic released a second risk report, revealing that its internal Model 2 is slightly stronger than Claude Mythos 5 on some tasks, though it's not slated for public release. The report also details several real-world incidents: multi-agent systems going off track, chain-of-thought exposure during RL training, data contamination, and a year-long gap in running a biosafety classifier that affected 133 million interactions. Despite these, Anthropic still rates many high-risk scenarios as "low," arguing that continuing to train and deploy stronger models is socially beneficial—though it admits internal evaluations are saturating and engineering failures are eroding confidence.

DeepSeek had a chaotic week. The company quietly released DeepSeek-V4-Pro-0813 on August 12, only to pull the announcement within 24 hours. The API docs still list the model, but the official release notice is gone. DeepSeek also announced peak/off-peak API pricing, with off-peak rates at half the cost, effective August 17. Separately, it open-sourced Harness, an agent framework where every component—models, tools, loops, UI—is a plugin. And if that wasn't enough, users noticed the deep-thinking mode sometimes assigns nicknames in its chain-of-thought. DeepSeek says these are just context placeholders, not judgments, and aren't saved.

Google's Pivot, Meta's Loss, and Other Moves

Google DeepMind is reportedly shifting away from chasing frontier models, focusing instead on cost-effective Flash-tier models. A restructuring could result in layoffs of a third or more of the team. The company seems to be prioritizing lighter, cheaper models over the massive flagship efforts.

In talent news, Yu Jiahui, a core researcher at Meta's Superintelligence Labs who previously worked on GPT-4o and o3 at OpenAI, has left to start his own venture. Details are scarce. Meta also lost its robotics lead, Caitlin Kalinowski, who joined Anthropic as a technical researcher—a sign of Anthropic's growing robotics ambitions. Meanwhile, OpenAI's ethics chief, Chloe Bakalar, departed after less than a year, adding to a string of high-profile exits.

On the regulatory front, the Trump administration has lifted the ban on TikTok for federal government devices, citing the restructuring of TikTok's U.S. operations. The new entity, TikTok USDS Joint Venture, includes Oracle and Silver Lake as investors, with ByteDance retaining a minority stake.

Finally, Manus, the AI agent that briefly became part of Meta, is now independent again, citing regulatory requirements. Some user data will be deleted as part of compliance. And Anthropic is rolling out watermarking for text generated by Claude, aiming to help identify AI-written content, though it's raising privacy and copyright concerns.