Weekly Digest 34

Topics we are tracking

OpenAI Pauses Model Development

Source: OpenAI — Pacing Model Development and Cyber Capabilities

In a blog post this week, OpenAI has announced that they are temporarily slowing the development of their latest model, Astra. This is the first time that any frontier lab has independently made the decision to slow the pace of developing one of their models. The lack of regulation and framework on a federal basis has prompted the labs to come up with their own internal frameworks and benchmarks to determine what is safe to release to the public. For OpenAI this is their “Preparedness Framework”, where Astra has potentially met a critical cybersecurity threshold. It’s good to see that frontier labs are taking these steps independently, with Anthropic’s Project Glasswing being a different type of intervention but having the same thought in mind, as well as Demis Hassabis advocating for stronger safety mechanisms. It’s our understanding that the Hugging Face incident has led to some internal turmoil and worry from employees at OpenAI, which prompted the company to react.

The updated safeguards introduced by OpenAI include:

• Better securing of research environments. This can be seen as a direct response to the Hugging Face hack, where the model was able to escape its sandbox.
• Expanding chain-of-thought monitoring. OpenAI was criticized for not being able to detect a hack while it was ongoing and has now implemented further classifiers and a more rigorous monitoring system.

Chain-of-thought monitoring is one of the most important parts of model safety evaluation today, but models are increasingly aware of evaluation settings, which raises the possibility of models “faking alignment”, where a model detects an evaluation setting and deceives the user with an aligned chain-of-thought reasoning trace while actually pursuing an alternative path.

Overall we are pleased to see frontier labs taking these steps, but are still in favour of a federal framework. Any such framework needs to work for both the frontier labs and smaller labs without disadvantaging either. Given the rapid pace of development and the expertise needed, this can only be set up in cooperation with the industry. The European AI Act is a great example of how not to do it, and it once again showed the complete technical illiteracy of European leaders.

Amazon Expands Drone Delivery

Source: The Wall Street Journal — Amazon Prime Air Drone Delivery Expansion

Amazon announced that it will expand its drone delivery program to over 500 cities in the U.S. by the end of the year, up from 11 currently. Jeff Bezos initially introduced delivery drones in 2013 and predicted the service would be available in about five years. The newest drones, the MK30, hover a foot above the ground and release the package. Packages are delivered within 30–60 minutes.

DoorDash has recently also announced an in-house drone delivery program after getting a certification from the U.S. Federal Aviation Administration (FAA). Zipline and Alphabet’s Wing are also expanding their operations as this technology is getting further integrated into our daily lives. Drones have faded from consumer, corporate, and venture-capital attention as AI has come to dominate the technology landscape. Discussion of drones has instead focused mainly on military applications, as the Ukrainian Armed Forces have shown how to use this technology in warfare effectively.

In China, drone delivery has been more widely adopted, with a reported 3 million parcels delivered via drones in 2025. (For context, domestic shipping volume in China was about 180 billion parcels in 2025.) It’s far from being profitable right now because of the high operating costs, but the government has made it a priority and the “low-altitude economy” has been included in official reports and frameworks. Companies like Meituan, JD and Alibaba are all investing in the space.

Opposition from local communities and politicians in the United States has so far been limited. We hope this continues, allowing the technology to spread more widely across the economy. All drones are equipped with cameras, which raise privacy concerns, but we think Amazon and other providers are already working to address them.

Unitree’s IPO and the Future of Humanoid Robotics

Source: TrendForce — Unitree Sees Embodied AI ChatGPT Moment in 2-3 Years at Earliest, Flags AI Model Limitations

The Unitree IPO has been the second blockbuster tech IPO after CXMT earlier this year. In its trading debut on the tech-focused STAR Market the stock rose fivefold, valuing the company at around $50 billion. The broader Chinese indexes dropped that day, with most Chinese tech names dropping in line with their American peers. This shows the unique demand from investors to have exposure to AI and embodied intelligence.

Unitree is among the leading companies in the Chinese-dominated humanoid robotics market. For the moment, most of its robots are sold to universities and research institutions, with only limited commercial deployment. The market is still in its infancy, but Goldman Sachs projects the global market to be $38 billion by 2035 with explosive growth afterwards. Morgan Stanley estimates the long-term TAM to be between $4.7 and $5 trillion by 2040.

Chairman Wang Xingxing cautioned against the optimism after the IPO, saying that the ChatGPT moment for embodied AI might still be two to three years away at the earliest. In his comments at the World Robot Conference he pointed to the lack of generalisation as the limiting factor. In fixed environments models can achieve close to a 100% success rate, but even small changes in the environment can lead to large drops in performance. AI models are now the largest investment for Unitree, which is pursuing multiple bets including world models and vision-language-action models (VLAs), as well as investing more in robotics data. Unitree so far has invested heavily in vertical integration and in-housing various components, which gives it a unique advantage even among its peers in China.

As the industry matures, it will also have to deal with its financial performance. Unitree’s latest numbers show H1 revenue growing between 36% and 45%, but net profit continuously declining. The US has contributed around 15% of revenue historically, but the FCC in July introduced restrictions on robotics equipment overseas. The current models by Unitree are allowed under the new FCC rules, but future models might be impacted by this as the tensions around AI and embodied intelligence are heating up between China and the United States.

Alibaba Lifts AI Capex 75% as Cloud Revenue Accelerates

Source: The Information — Alibaba CEO Expects AI-Related ARR to Reach $10 Billion in September

In its most recent earnings report, Alibaba reported that it increased AI capex by 75% to 67 billion yuan (roughly $10 billion) in the April to June period. CEO Eddie Wu expects to break even on AI-related capex within the next three years. In the earnings call he highlighted: “In order to be able to capture that future growth, we first need to make these capex investments to build out the necessary compute capacity.” AI cloud revenue rose by 45% in the quarter, a meaningful increase from previous growth numbers, an early sign that demand for AI services is strengthening.

A column chart of ten quarters from March 2024 to June 2026. Year-over-year growth rises without interruption from 3 per cent to 45 per cent, with the steepest gains through 2025.

Segment revenue for Cloud Intelligence as a whole, not the AI portion alone. The June 2026 quarter is the April to June period discussed above.

Source: Reuters · 2026-08-23

While this number is large for Alibaba, it pales in comparison to American hyperscalers: Amazon in the same period invested $54 billion, Alphabet $45 billion. Adjusted for revenue, however, the number is broadly comparable to Amazon but still below Alphabet.

Wu stated that Alibaba will continue to invest in developing frontier models, but also highlighted that Alibaba is focused more on monetisation: “AI model companies have their eyes on that ultimate end game [referring to AGI], where I think that the monetisation level will be significantly higher, much higher than what you see today.”

Alibaba occupies a unique place in China’s AI ecosystem and is perhaps the closest equivalent to US hyperscalers. They own frontier models via their Qwen family, are major investors in Moonshot (the developer of the Kimi model family), are the premier cloud provider in China and own significant distribution through their e-commerce platforms and fintech arm Ant Group. Chip export controls are also accelerating their own chip development work and allowing them to pursue closer hardware-software co-design.

Seen on X

Other interesting stories

• Financial Times: US buyers snap up Europe’s asset managers at fastest rate in decades
• The New York Times: How the A.I. Borrowing Binge Helps Drive Up Government Bond Yields
• The Wall Street Journal: How Saudi Arabia’s Neom Megacity Dream Collapsed
• Health Tech Builders: Be AI-Native and AI-Proof
• TrendForce: Marvell, AMD Reportedly Shake Up Google TPU Race, Putting Broadcom, MediaTek Under Pressure
• Chipstrat: The Hardest Computing Problem Ever, According to Arm’s Physical AI Chief