24-Hour AI News: Ironwood Launches, Wuzhen Summit Opens, EU Assesses AI Bill Pace

In the past 24 hours (November 7), Google Cloud released its new generation TPU Ironwood and disclosed large-scale cluster metrics; in Europe, there were reports of a reassessment of the implementation pace of AI legislation; in China, the World Internet Conference Wuzhen Summit and the "Light of the Internet" Expo opened, iFlytek released the Spark X1.5 deep inference big data model, and XPeng released its second-generation VLA model and clarified its mass production and deployment schedule.

I. Google announces Ironwood TPU's inclusion in Google Analytics, raising the ceiling for training and inference computing power.

1. Announced on November 7: Ironwood offers approximately 10 times the peak performance of TPU v5p and over 4 times the single-chip performance of v6e in training and inference.
2. A single “SuperPod” can be scaled up to 9216 chips, forming approximately 1.77PB of shared HBM and 9.6Tbps of interconnect bandwidth, supporting a unified architecture from RL to massive online inference.
3. Google says Anthropic will "access up to 1 million TPUs" and collaborate with Ironwood to support large-scale services for cutting-edge models such as Claude, marking a further diversification of its AI computing power landscape.

II. The EU is considering a "targeted" suspension of the implementation of certain provisions of the AI Act.

1. On November 7, multiple media outlets cited sources saying that after communicating with the United States and major technology companies, the European Commission is discussing the implementation timetable for adjustments to a few provisions, and the final decision is expected to be made around November 19.
2. The EU reiterated its support for the goals and principles of the AI bill, focusing on optimizing the implementation sequence and pace, rather than "starting from scratch".
3. If the adjustments are implemented, they will affect the compliance roadmap for high-risk systems and the pace of enterprise technology transformation. Enterprises need to pay attention to the details and transition period arrangements.

III. Anthropic continues its expansion in Europe, opening offices in Paris and Munich.

1. On November 7, it was officially announced that it will set up offices in France and Germany, covering a complete team from R&D to sales and operations, to strengthen its local delivery capabilities in EMEA.
2. The company stated that its European business has seen "run-rate revenue increase more than ninefold" in the past year, and that it has established deep partnerships with regional clients (including automotive and retail giants).
3. Combined with its massive TPU resources and multi-cloud strategy, the availability and timeliness of model services in the European market are expected to be further improved.

IV. The US plans to block Nvidia from selling "scaled-down" AI chips to China.

1. A report on November 7th indicated that the US government plans to restrict the sale of a scaled-down version of an AI chip, codenamed B30A, to China. Samples of the chip had previously been demonstrated to some customers.
2. Nvidia responded that it currently has "zero market share in China's data center market and is not included in the guidelines."
3. On the Chinese side, there have also been recent guidelines that require government-funded projects to prioritize "domestic chips," and this two-way competition is accelerating the localization of the AI hardware supply chain.

V. OpenAI clarifies: Discussions with the government concerned loan guarantees for "chip factories," not data centers.

1. On November 6 (November 7 Beijing time), OpenAI stated that the federal loan guarantee under discussion focuses on "chip manufacturing" and does not seek to provide guarantees for self-built data centers.
2. According to the company's public statements, the capital commitment scale over the next eight years is approximately US$1.4 trillion, and the annualized revenue is expected to exceed US$20 billion. At the same time, the company is exploring the AI cloud model of "direct sales of computing power".
3. The White House has stated that it will not provide industry-wide relief, and the expansion of AI infrastructure will rely more on market mechanisms and multilateral capital structures.

VI. The World Internet Conference Wuzhen Summit opens, with the "Light of the Internet" Expo focusing on "Artificial Intelligence +".

1. The expo opened on November 6 with the theme of "AI Symbiosis, Intelligent Future". It set up seven themed exhibition areas, with 670 companies participating both online and offline, showcasing more than 1,000 AI technologies and products.
2. The main forum of the summit focused on topics such as cutting-edge big data models, data governance, trustworthy AI, and industrial implementation, serving as a window to observe the synergy between AI application and governance in China.
3. The conference will run until November 9th, and is expected to release a number of industry cooperation and technological achievements.

VII. iFlytek Releases Spark X1.5 Deep Inference Model

1. Released on November 6: According to the company, with a smaller parameter scale, the overall performance is "comparable to the mainstream international large models", and the overall performance is said to reach "more than 95% of GPT-5" (the statement needs further verification by third-party benchmarks).
2. The technical side claims that the "long mind chain reinforcement learning training efficiency" has increased from about 30% to over 84%, and the MoE (Expert Hybrid) full-link training efficiency has been significantly improved.
3. The deployment side emphasizes that "one Ascend standard server can be deployed", highlighting the domestic computing power route and the integrated hardware and software ecosystem.

8. XPeng Technology releases second-generation VLA large model, clarifies timeline for vehicle deployment.

1. Released on November 7: The second-generation VLA outputs end-to-end "vision → action", reducing the intermediate language link; the average takeover mileage of "small road NGP" for complex small roads has been increased by about 13 times.
2. Implementation schedule: The "Pioneer Co-creation" experience will be launched next month, and the full rollout to users will begin in the first quarter of next year along with the Ultra model.
3. Collaboration with Volkswagen is progressing in tandem, with Xpeng's Turing AI chip securing a Volkswagen contract, further streamlining the intelligent vehicle process from model to computing power.

9. Ant Group showcased its "Braille Large Model Family" for the first time in Wuzhen and won an award at the conference.

1. November 6-7: The entire Behring model family will be showcased for the first time. 18 models have been released, covering "language, thinking, and multimodal". It is claimed that two trillion-parameter models (Ling-1T and Ring-1T) have reached the leading level of open source (including the "silver medal" level performance of IMO2025, etc., all of which are based on the media and on-site technical personnel).
2. On the scene side, AI health manager AQ (which claims to be able to connect to resources from 5,000+ medical institutions nationwide), embodied robot Robbyant-R1, and trusted interaction gPass were showcased.
3. Ant Computing won the "Outstanding Contribution Award (Growth Potential Category)" at the World Internet Conference, emphasizing the industrial value of privacy computing and trusted data circulation.

10. DeepSeek researcher speaks publicly in Wuzhen: expressing caution regarding the long-term social impact of AI.

1. At the event on November 7, a senior researcher at DeepSeek stated at the "Six Little Dragons·Wuzhen Dialogue" that he holds a positive attitude towards AI's short-term efficiency improvement, but remains cautious about its medium- and long-term impact on employment and society.
2. Since DeepSeek gained popularity at the beginning of the year for its low-cost, high-performance models, it has made relatively few public appearances. This statement has sparked a renewed discussion about the relationship between "technological dividends and social resilience".
3. The incident also reflects the strategy and pressure of China's local AI ecosystem in pursuing three parallel paths: "open source models, adaptation of domestic computing power, and industrial application".

Frequently Asked Questions (Q&A)

Q: What are the key performance indicators and time to market for Ironwood TPU?

A: It was announced on November 7th that it has entered the GA window, and the official statement is that it will be rolled out to customers in the coming weeks. The performance is about 10 times higher than the peak performance of v5p, and more than 4 times higher than the single-chip training and inference performance of v6e; a single Pod can reach 9216 chips, with about 1.77PB of shared HBM and 9.6Tbps interconnect.

Q: Is the EU AI bill really going to be "suspended"? What will be the short-term impact?

A: The current delay is a "targeted postponement" within internal discussions and media reports, not a complete halt; the EU still supports the bill's objectives. The short-term impact is mainly on compliance scheduling and transitional provisions for high-risk systems and underlying models; companies need to pay attention to the official public statements in mid-to-late November.

Q: What exactly is the "$1.4 trillion" that OpenAI is referring to?

A: The company's public statement is "the scale of capital commitment over the next eight years," covering long-term investments in computing power and the upstream chip ecosystem; at the same time, it clarified that it has not sought federal loan guarantees for data centers, and the focus is on discussing chip manufacturing and supply chain strengthening plans with the government.

Q: When will the second-generation XPeng VLA be available? What are the improvements?

A: The official timeline is that the pioneer co-creation will start in December, and the full rollout will be rolled out with Ultra models in the first quarter of 2026; the average takeover mileage in complex road scenarios will be increased by about 13 times, and the new capabilities include "automatic assisted driving without navigation".

Q: What engineering modifications have been made to the iFlytek Starfire X1.5 "Deep Inference"?

A: The company claims that the efficiency of long-chain reinforcement learning has been significantly improved (approximately 30% → 84%+), the training efficiency of the MoE link has been increased, and emphasizes its deployability on the domestic Ascend platform; however, the specific ranking of the overall performance compared with mainstream international models still requires third-party benchmarks and public evaluations.