24-Hour AI News: EU AI Bill May "Slow Down," Wuzhen Awards to Share the Stage with Google Ironwood

In the past 24 hours (November 7-8), news emerged from Europe that the pace of AI regulation enforcement may slow down; major US tech companies continued to increase investment in large-scale AI infrastructure and accelerate their international expansion; meanwhile, the Wuzhen Summit announced a series of awards and industry progress, and XPeng publicly "disassembled" its humanoid robot to clarify controversies. The dynamics in China and North America/Europe are basically symmetrical.

I. Google releases Ironwood TPU, fully launching for the "inference era"

1. Google announced that its seventh-generation TPU Ironwood will be widely available within weeks, claiming a 10x peak performance improvement over the TPU v5p and a 4x single-core performance improvement over the v6e, targeting large-scale training and low-latency inference.
2. A single Pod can scale to 9216 accelerators, providing approximately 1.77PB of shared HBM and 9.6Tb/s of interconnect for “hyperscale inference and services”.
3. Simultaneously launch Arm architecture Axion instances, N4A preview and C4A metal will be available soon, focusing on cost and energy efficiency in conjunction with inference clusters.

II. Meta has invested $600 billion in the US over three years, with a core focus on AI data centers.

1. Meta says it will invest at least $600 billion in the U.S. over the next three years in infrastructure and jobs, including several mega-AI data centers.
2. The company is simultaneously promoting financing for ultra-large projects and building distributed industrial parks to address the AI computing power demand curve through "pre-emptive expansion".
3. Management stated that the goals are aligned with the long-term roadmap of "General Intelligence," and capital expenditures will "significantly increase" starting in 2026.

III. The EU plans to "slow down" its AI legislation, introducing a transition period and exemptions.

1. According to a draft European Commission document disclosed by the media, a grace period of up to August 2027 is proposed for the registration and penalties of high-risk AI systems.
2. Exemptions from database registration will be granted for high-risk AI scenarios that only perform "procedural/limited" tasks; AI content labeling obligations will be implemented in stages.
3. The final proposal is expected to be submitted on November 19, but may still be adjusted during the process, involving the demands of large platforms such as Apple and Meta, as well as European industrial companies.

IV. Anthropic announces new offices in Paris and Munich, further strengthening its presence in EMEA.

1. Official confirmation of the addition of offices in France and Germany, continuing the expansion of the European footprint after London, Dublin and Zurich.
2. Over the past year, EMEA's revenue has grown at an annualized rate of more than 9 times, and the number of major regional customers has increased more than 10 times, covering manufacturing, pharmaceuticals and finance.
3. The company stated that it will further localize its research, engineering, business and policy cooperation functions, and undertake the TPU supply expansion previously agreed upon with Google Cloud.

V. Wuzhen: Tsinghua University and Zhipu's "Key Technologies and Large-Scale Application of GLM Large Model" won the Leading Technology Award

1. The project provides a systematic solution in terms of pre-training architecture, chip adaptation and multimodal alignment, and claims to be compatible with more than 40 AI chips.
2. The daily API call volume of the relevant platforms reaches the level of "trillions of words", covering more than 20 industries such as Internet and finance, and more than 40 open source models.
3. This achievement was selected as one of the 17 annual projects for the Leading Technology Award at the 2025 World Internet Conference.

VI. Baidu wins Wuzhen Leading Technology Award again for its "Script-Driven Multi-Modal Collaborative Highly Realistic Digital Human Technology"

1. This is the third consecutive year that an AI company has won the award, with key technologies covering multimodal planning, real-time interaction, and long video synthesis.
2. According to official sources, the technology has produced over 100,000 digital humans, and has been implemented in e-commerce, education, law, and other scenarios, significantly reducing the cost of live streaming.
3. The continuous awards have strengthened its industrialization and content production capabilities in the "Wenxin" technology stack.

VII. XPeng's humanoid robot IRON sparks heated discussion; the company responds to doubts twice with "self-proof".

1. From November 5th to 7th, XPeng publicly demonstrated the internal structure of IRON twice at its Technology Day and press conference, including cutting open the leg cover while it was powered on.
2. The official statement emphasizes the use of designs such as a humanoid spine, bionic muscles, and fully covered flexible skin, with the goal of achieving mass production by 2026.
3. At the same time, the second-generation VLA "physical AI" capability was released, claiming that the average takeover mileage on complex roads has been significantly improved, forming a collaborative narrative of vehicle/machine/aircraft.

8. China Mobile releases 2026 version of the 5G mobile phone white paper, setting a mandatory requirement of "BeiDou SMS".

1. It is proposed that from July 1, 2026, new mobile phones priced at 4,000 yuan and above must support native Beidou SMS; each brand must have at least one model that supports Beidou SMS + positioning fusion and be shipped on a large scale.
2. It was disclosed that as of now, more than 2.65 million 5G base stations have been opened, covering 96% of the population, and there are more than 599 million 5G users.
3. With the advancement of eSIM, satellite communication and 5G-A, terminal-network collaboration will impose "hardware constraints" on the connectivity and positioning perception capabilities of AI terminals.

Frequently Asked Questions (Q&A)

Q: What is the performance relationship between Ironwood and Nvidia Blackwell?

A: Publicly available metrics show that Ironwood's single-core FP8 computing power is approximately 4.6 PFLOPS, with a maximum of 9216 core interconnects per Pod and 1.77PB of shared HBM, indicating "ultra-large-scale inference" and high-concurrency services; it is competitive with Blackwell's high-end SKUs at the same precision level, focusing on large-scale clusters and system-level collaboration.

Q: Where did Meta spend its $600 billion, and over what period of time?

A: The plan covers data center parks, power supply and cooling facilities, network and computing equipment, and jobs within the United States, with a timeframe of the next three years, in addition to previously signed large-scale project financing and state-level investment commitments.

Q: Why is the EU considering slowing down the implementation of its AI law? What does this mean for businesses?

A: Large enterprises and European and American industry organizations believe that the implementation costs are high, the execution is complex and the cycle is tight. The committee plans to reduce compliance friction through a transition period and registration exemption. In the short term, this will benefit existing systems in high-risk scenarios and AIGC content labeling, but in the long term, formal compliance requirements still need to be met.

Q: Wuzhen's "Leading Technology Award" is being awarded to large models and digital humans in quick succession. What is the industrial implications of this?

A: The concentration of awards reflects a shift from "paper metrics" to "replicable engineering and service capabilities." GLM emphasizes cross-chip training and multimodal robustness, while Digital Human emphasizes long-term interaction and content production efficiency, both of which drive the large-scale supply of AI applications.