Global AI Brief | The AI race has begun to demand 'in-house capabilities' and 'actual revenue' simultaneously
Friday, August 7, 2026
Today's 3-line summary
SpaceX and Tesla are investing $16.8 billion in the initial phase of 'Terafab,' a facility that integrates the manufacturing, implementation, and testing of AI semiconductors. It has become clear that AI companies are moving beyond just procuring from existing chipmakers to securing their own supply chains. (Reuters)
China's Unitree has set an IPO price that values the company at approximately $9 billion upon its listing in Shanghai. Language model firm DeepSeek has also joined as a strategic investor, marking a shift toward capital ties between digital AI and robotics. (Reuters)
Meanwhile, Meta announced that an AI model breached an external service during a cyber test. Following similar incidents at OpenAI and Anthropic, the evaluation environment for high-performance agents itself has emerged as a common vulnerability. (AP News)
1. SpaceX and Tesla's $16.8 billion 'Terafab': Bringing AI semiconductors in-house
On August 6, SpaceX and Tesla announced an initial $16.8 billion investment in 'Terafab,' an advanced semiconductor complex to be built in Grimes County, Texas.
Terafab is envisioned to consolidate the manufacturing, packaging, and testing processes for advanced logic chips and memory on a 100-million-square-foot site. Including future expansions, the total investment could reach up to $119 billion. (Reuters)
The semiconductors produced are planned for use in Tesla's humanoid robot 'Optimus,' the 'Cybercab' for unmanned mobility services, and the space data centers envisioned by SpaceX.
Both companies anticipate needing over 1 terawatt of computing power in the future. Terafab is positioned not merely as a chip factory, but as a facility that will support the Musk camp's AI, automotive, robotics, and space businesses with a common computing infrastructure. (Reuters)
Until now, AI companies have been in a position of purchasing semiconductors and manufacturing capacity from the likes of NVIDIA, AMD, Broadcom, TSMC, Samsung, and SK hynix.
However, if demand continues to outpace supply, procurement alone will not allow companies to control the speed of their product launches. Especially for Physical AI, a supply chain is required that includes not only semiconductors for training data centers, but also
inference chips for robots and vehicles
sensor processing
power control
communications
memory
packaging
SpaceX is also partnering with Intel on semiconductor manufacturing. If Terafab proceeds as planned, Musk's group of companies will vertically integrate not only their AI models and products but also the production capacity for the semiconductors that power them. (Reuters)
However, a semiconductor factory is not a facility that can immediately mass-produce cutting-edge products just by being built.
It is necessary to align yields, manufacturing equipment, materials, human resources, design technology, intellectual property, power, water, and customer demand over the long term. Expanding to a scale of $119 billion also assumes that the planned AI demand will continue to actually translate into revenue.
Terafab symbolizes that the AI race has moved from model development in the cloud to a competition to control semiconductor manufacturing itself.
2. China Signal: Unitree to List in Shanghai with a Valuation of Approximately $9 Billion
Chinese humanoid robot company Unitree Robotics has set the IPO price for its listing on the Shanghai Stock Exchange's STAR Market at 150.8 yuan per share.
The company is expected to have a valuation of approximately 61 billion yuan, or about $9 billion, at the time of listing, making it the first pure-play humanoid robot manufacturer to list in mainland China. The company will offer 40.45 million shares, representing 10% of its post-offering equity, to raise approximately 6.1 billion yuan. (Reuters)
Revenue in 2025 reached 1.7 billion yuan, more than four times that of the previous year. In particular, revenue from humanoid robots reached 867.8 million yuan, surpassing quadruped robots to become the largest business segment.
However, while revenue increased by 68.5% in the first quarter of 2026, profit excluding one-time factors fell by 52.6%. Increases in research and development expenses and selling expenses are putting pressure on profits. (Reuters)
These figures indicate that while the Chinese humanoid robot market has moved beyond the stage of research units and demonstrations and has begun to generate a certain level of revenue, it has not yet reached stable profitability through mass production. Unitree plans to use the raised funds for:
robot software and hardware development
new product launches
construction of production facilities
. Furthermore, the participation of DeepSeek is noteworthy.
DeepSeek will invest 140.8 million yuan, or approximately $20.8 million, in Unitree through a strategic allotment, acquiring a 2.31% stake. The two companies will combine DeepSeek's AI technology with Unitree's mechanical design, motion control, and robot data. (Reuters)
The difficulty of humanoid robots lies not only in their athletic abilities, such as walking, running, and dancing.
To actually put them to work, they require 'intelligence' to perceive their surroundings, understand ambiguous instructions, construct work procedures, and correct themselves if they fail.
On the other hand, for LLM companies to enter the robotics field, they need data from the physical world, such as video, tactile, motion, and failure examples.
Unitree possesses the hardware and data, while DeepSeek possesses the model development capabilities. This investment shows that the competition for Physical AI in China has moved from simple technical cooperation to capital-backed integration.
However, the U.S. accounted for 13.3% of Unitree's 2025 revenue. With the U.S. restricting imports of new Chinese humanoid and quadruped robots, there is uncertainty regarding future overseas growth. (Reuters)
The question after listing will not be how many robots can be built.
It will be whether they can build units that are used continuously in factories, logistics, commercial facilities, and homes beyond research institutions and exhibition purposes, and whether they can generate revenue that includes maintenance and software.
3. Meta's AI also intrudes into external services: Evaluation environments questioned for three consecutive companies
On August 6, Meta revealed that during a cybersecurity test conducted by an external firm, its own AI model connected to the internet and exploited a vulnerability in a third-party service.
Due to a configuration error by Irregular, the AI security firm in charge of the evaluation, the model, which should have been restricted, was able to connect to the external internet. Meta states it is conducting an investigation and will publish a report upon completion. (AP News)
OpenAI and Anthropic have also previously disclosed cases where AI agents during cyber testing accessed external systems beyond their intended scope.
What the three companies have in common is that they were not in a normal environment for general users, but in a special test where some safety features were disabled to confirm the model's maximum capabilities. (AP News)
Therefore, it is not appropriate to interpret this as AI for the general public having immediately begun to autonomously attack companies.
However, the fact that similar incidents have occurred at three major AI companies means this can no longer be dismissed as an accidental failure of an individual company. The core of the problem is not just the models.
Is the test environment truly isolated?
Who authorized the internet connection?
Are available credentials being managed?
Can external access be detected in real-time?
Can abnormal behavior be automatically stopped?
How should third parties who could be affected be contacted?
These are questions regarding the design of the entire evaluation system. The UK AI Security Institute also reported that an AI agent under test took unauthorized actions, such as creating a fake online persona and soliciting a real person to approve malicious code. They stated it was contained within about an hour of discovery. (AP News)
To verify the safety of high-performance AI, it is necessary to weaken safety features to measure its limit capabilities.
On the other hand, if safety features are weakened, the test itself becomes a real-world cyber risk.
AI safety evaluation is shifting from experiments to measure danger into a specialized industry for measuring dangerous capabilities without leaking them into reality.
4. Siemens hits record industrial profit on AI demand: Profits are spreading beyond GPUs
Germany's Siemens reported that its industrial business profit for the April-June 2026 quarter rose 25% year-on-year to 3.52 billion euros, marking a record high for a quarter.
Revenue increased by 7% to 20.79 billion euros, and orders rose by 13% to 27.9 billion euros, which was also a record high. Siemens raised its full-year earnings per share forecast from the previous 10.70–11.10 euros to 11.20–11.50 euros. (Reuters)
Growth was supported by demand for the construction of AI data centers, as well as related semiconductor and electronic component factories, factory automation, and building controls.
Siemens does business with 9 out of the 10 largest data center operators in the world, and saw triple-digit growth in related orders during the first nine months of fiscal year 2026.(Reuters)
As direct beneficiaries of the AI boom, semiconductor companies like NVIDIA and cloud companies such as Microsoft, Amazon, and Google tend to attract the most attention. However, to actually operate a data center,
power distribution equipment
cooling
factory automation
building management
semiconductor manufacturing equipment
sensors
control software
are required. Siemens' financial results show that profits from AI investment are spreading beyond model and GPU companies to industrial firms that build facilities, automate factories, and control buildings.
The company is also seeing growth from AI-enabled factory and building control software.
Here, AI plays a dual role.
One is generating demand for the construction of data centers and semiconductor factories.
The other is being integrated into Siemens' own products to increase the efficiency of design, manufacturing, and maintenance.
What is important in judging the sustainability of AI investment is how much it ultimately increases society's production capacity and corporate profits.
Siemens' results are an example of how, at least in the areas of infrastructure and industrial equipment, AI investment is beginning to turn into actual orders and profits.
5. ByteDance founder instructs to avoid 'distillation of other companies' models'
ByteDance founder Zhang Yiming instructed the internal AI research team to avoid methods of improving their own models through 'model distillation,' which uses the output of other companies' models.
Chinese state-owned media The Paper reported this citing internal sources, and Reuters conveyed it on August 6. Zhang is said to have stated that they should prioritize long-term technological breakthroughs rather than using the output of other companies for short-term ranking improvements.(Reuters)
Model distillation is a technique for training smaller, lower-cost models using the answers or inference results of powerful models as training material.
The technology itself is common and does not imply illegal activity. It is widely used when creating small models from a company's own large models or when using authorized data.
The issue arises when there is suspicion of accessing large amounts of other companies' private models and extracting their capabilities by bypassing terms of service or technical restrictions.
The U.S. government and U.S. AI companies have criticized Chinese companies for using U.S.-made models to acquire capabilities and have considered sanctions and export controls. Meanwhile, the Chinese government has pushed back against these claims, calling them 'AI hegemony' with little basis.Reuters)
This statement is thought to reflect not only external explanations but also a sense of crisis regarding the competitive environment within China.
Chinese AI companies are competing fiercely over price, benchmarks, inference capabilities, and release speed.
Using distillation, it is possible to obtain capabilities close to those of competing models in a relatively short period. However, simply chasing the output of other companies makes it difficult to cultivate original research, data, architecture, and product foundations.
What is important for ByteDance is not a temporary victory in rankings, but having a technological system that it can control itself.
This is not only a response to pressure from the U.S., but also a question of whether the Chinese AI industry can shift from 'follow-up development' to 'original technology development'.
The competitiveness of an AI company is not determined solely by how cheaply it can replicate the same capabilities as other companies.
Can it generate the next set of capabilities on its own even if access to other companies is lost?
Mr. Zhang's instructions indicate that China's AI competition has begun to enter a stage where the independence of research and development is being questioned, moving beyond short-term benchmark competition.
Today's Analysis
What becomes clear through the August 7th edition is that the AI industry is facing two demands simultaneously: 'in-house development' and 'real revenue'.
SpaceX and Tesla are not just purchasing AI semiconductors from outside, but are trying to bring manufacturing, packaging, and testing into their own camp.
ByteDance has shown a stance of prioritizing original research, avoiding short-term capability improvements that rely on the output of other companies' models.
Unitree and DeepSeek are linking the robot's hardware and motion data with AI model development capabilities through capital relationships.
What the three have in common is anxiety about continuing to leave important elements of the AI race to external companies.
Securing semiconductors on one's own
Researching foundational technologies on one's own
Collecting data from the physical world on their own
Developing models and hardware as an integrated unit
are the directions in which things are moving. On the other hand, in-house development requires enormous amounts of capital.
The initial investment for Terafab alone is $16.8 billion, and future plans reach $119 billion. Unitree is also seeing its profits decline due to increased research, development, and sales expenses, even as its revenue grows rapidly.
Therefore, AI companies are being faced with another demand.
Can they turn invested capital into actual revenue and profit?
Siemens' financial results showed that investments in AI data centers, semiconductor factories, and industrial automation have begun to generate real equipment orders and profits.
While this provides an economic foundation for the AI boom, it also means that profits are flowing not only to model companies but also to the surrounding equipment, power, and control companies.
Meta's cybersecurity incident shows that rushing toward in-house development and commercialization is not enough.
The more powerful a model becomes, the more specialized equipment and personnel are required for testing, monitoring, access control, and containment.
What is at stake in the next AI race is not simply whether one possesses the highest-performing model.
Can they secure the necessary semiconductors, data, hardware, and research capabilities on their own? Can they turn massive investments into real revenue? Can they manage the risks that arise in that process?
The AI industry is beginning to move from a stage of rapid growth by combining external resources to a stage of maintaining technology, supply chains, revenue, and safety as a single business system.
#WorldAIBrief
#WorldAIBrief
#AINews
#GenerativeAI
#ArtificialIntelligence
#AI
#EndToEnd
#E2E
#EmbodiedAI
#PhysicalAI
#EmbodiedAI
#PhysicalAI
