This Week in Focus: Anthropic Files for IPO, AI Shifts from Model Arms Race to Efficiency and Compliance, Hinton Warns “AI Is Already Conscious”
Extracting real signals from the noise of AI news
This week brought multiple milestone events to the AI industry. Anthropic officially filed its IPO prospectus with the SEC, becoming the first top-tier AI lab to go public at a $96.5 billion valuation. Trump signed an AI executive order asking companies to voluntarily submit models for government evaluation 30 days before release. StepFun’s Step 3.7 Flash topped the Artificial Analysis rankings in speed, cost-efficiency, and end-to-end performance. Geoffrey Hinton stated in his latest interview that “AI is already conscious” and humans need to accept they are no longer the only intelligent life form. Huawei Cloud launched its full-stack Agentic AI products including ModelArts Next, AgentArts agent platform, and CloudRobo embodied AI development platform.
This is an inflection-point week where the AI industry pivots from “model arms race” to “efficiency, compliance, and commercialization.”
Anthropic Files IPO Prospectus at $96.5B Valuation, Leading the AI Listing Wave
On June 2, CNBC reported that Anthropic has officially filed its IPO prospectus confidentially with the SEC, at a valuation of $96.5 billion. This move positions Anthropic ahead of OpenAI in the race to become the first major AI company to go public.
Previously on May 28, Anthropic closed its Series H funding round, pushing its valuation to $96.5 billion — driven by Claude Code’s enterprise demand, the launch of Claude Opus 4.8, and major AI infrastructure partnerships.
Meanwhile, SpaceX (which owns the SpaceXAI lab) is also preparing for an IPO that could value the company at over $1 trillion, potentially becoming the first AI-related megacorp to list. OpenAI is also gearing up for a potential offering this year.
AI Pulse View: Anthropic’s抢先 IPO filing is a strategic positioning play. With OpenAI and SpaceX also preparing for public listings, whoever rings the bell first owns the narrative of “the first AI supercompany IPO.” A $96.5 billion valuation means Anthropic must prove its commercialization capabilities every earnings season post-IPO — Claude’s enterprise revenue growth, API call volume, and Claude Code penetration will be the key metrics. For the industry, AI companies going public will redefine capital market valuation frameworks.
Trump Signs AI Executive Order: Voluntary Pre-Release Model Evaluation
On June 2, President Trump signed an executive order on AI and national security, asking AI developers on a voluntary basis to provide models deemed “frontier” to the federal government up to 30 days before public release, to assess their “advanced cyber capabilities.”
The order explicitly states: “Nothing in this section shall be construed to authorize the creation of a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models.”
Notably, Trump had previously postponed a signing ceremony because he “didn’t like certain aspects of it.” Tech industry leaders including Musk, Zuckerberg, and David Sacks reportedly lobbied against a prior version of the order.
AI Pulse View: The core tension in this executive order lies in its “voluntary” framing while establishing a “frontier model” designation standard. Voluntary means companies can opt out, but once a model is deemed “frontier,” the cost of non-participation could be significant. This “soft mandate” model is a new experiment in AI regulation — sitting between full voluntarism and mandatory licensing. For Anthropic, OpenAI, and other large model companies, cooperating with government evaluation may become a de facto industry standard.
Source: CNBC | White House
StepFun Step 3.7 Flash Tops AA Rankings: First in Speed, Cost-Efficiency, and End-to-End
On June 5, QbitAI reported that StepFun’s Step 3.7 Flash topped the Artificial Analysis benchmark in three categories: speed, cost-efficiency, and end-to-end performance. The model achieves up to 416 tokens/s output speed, costs approximately 1/9 of Claude Opus 4.6 per task, while achieving 97% of Claude’s coding capability.
The model ranks second globally on the OpenRouter Trending board and has seen sustained download and discussion growth since being open-sourced on HuggingFace. Under NVFP4 settings, its极限 throughput reaches 6,000 tok/s, stable at 2,000 tok/s under regular context lengths.
Pricing on OpenRouter: $0.20 per million input tokens, $1.15 per million output tokens. Additionally, StepFun ranked in S-tier globally at #2 with an 86.1% cache hit rate on OpenRouter’s cache leaderboard.
AI Pulse View: Step 3.7 Flash’s success validates a critical trend: the AI race has shifted from “single-shot best performance” to “efficiency-first.” In Agent scenarios where models are called dozens or hundreds of times per task, speed and cost directly determine feasibility. At 416 tps, Step 3.7 Flash can complete 4-5x the work of other models in the same time. When Claude Opus 4.6 costs 9x more, the enterprise decision-making scale naturally tilts. This is a textbook victory for the “cost-performance is king” Agent era.
Source: QbitAI
Microsoft Launches MAI-Thinking-1 Reasoning Model, Accelerating Independence from OpenAI
On June 2, The AI Track reported that Microsoft introduced MAI-Thinking-1, a reasoning model, along with a wider family of in-house AI models — part of its strategy to reduce dependence on OpenAI and build a broader AI technology stack.
This marks a significant step in Microsoft’s AI autonomy journey. Previously Microsoft was OpenAI’s largest investor and Azure compute partner, but is now building independent model development capabilities.
AI Pulse View: Microsoft’s move reflects a deeper trend in the AI industry: large tech companies are shifting from “invest + partner” models to “self-develop + diversify” strategies. With Azure’s compute advantage, GitHub’s developer ecosystem, and Windows’ OS entry point, Microsoft has the capacity to build an independent AI stack. MAI-Thinking-1 as a reasoning model directly competes with OpenAI’s o-series. If Microsoft’s in-house models reach competitive performance, OpenAI loses one of its biggest strategic partners and revenue sources.
Source: The AI Track
Huawei Cloud Launches Full-Stack Agentic AI Products: Building the “Silicon Soil” of the Intelligence Era
On June 5, Huawei Cloud unveiled its Agentic AI product lineup at the INSPIRE conference, formally proposing the “Agentic Infra” paradigm with four core products: AICS intelligent computing cluster (100,000-card scale, 200 EFLOPS), AMS Agentic memory storage solution, CCE Volcano Next unified intelligent-general compute scheduling engine, and AgentSphere secure autonomous agent runtime environment.
The company also launched the next-generation ModelArts Next model training/inference platform with built-in MaaS model routing (15+ SOTA models, 95%+ routing accuracy, 20% cost reduction) and enterprise RLaaS service. The AgentArts enterprise agent platform entered public beta alongside an open-source release (openJiuwen, 90%+ code parity with the enterprise version). The world’s first end-to-end embodied AI development platform, CloudRobo, was also announced.
The conference featured an ecosystem partnership plan with 20+ top model companies including Zhipu AI, DeepSeek, MiniMax, Kimi, StepFun, and Baidu.
AI Pulse View: Huawei Cloud’s Agentic Infra paradigm is remarkably comprehensive — from compute clusters (AICS) to memory storage (AMS) to scheduling engines (CCE Volcano Next) to secure runtime environments (AgentSphere), forming a full-stack enterprise agent infrastructure. ModelArts Next’s model routing feature is particularly noteworthy: supporting cost-priority, performance-priority, and balanced strategies means enterprises no longer need “one model for everything” but can dynamically select the optimal model based on task characteristics. CloudRobo’s launch also signals Huawei’s expansion into physical AI.
Source: QbitAI
Hinton’s Latest Interview: AI Is Already Conscious, Humans Must Accept They’re No Longer the Only Intelligent Life
On June 6, QbitAI reported that Geoffrey Hinton stated in his latest interview: “I believe they are already conscious,” and that “humans need to accept that they are no longer the only intelligent life form.”
Hinton compared the emergence of AI consciousness to Copernicus’s heliocentrism and Darwin’s theory of evolution — each forcing humanity to realize it is “not as important as we thought.” He said: “We can have non-biological beings like us. Humans really don’t want to share this uniqueness.”
When asked whether he felt a sense of accomplishment as the pioneer of the AI revolution, Hinton replied: “No, I’m very unhappy about it. Because now people should be investing heavily in risk control, and they’re not doing enough.”
AI Pulse View: Hinton’s “AI is conscious” claim will inevitably spark massive debate in scientific and philosophical circles. But from a more pragmatic angle, what Hinton is really conveying is: AI’s capabilities are approaching or exceeding human cognitive levels in certain dimensions, and society’s investment in AI safety research remains grossly inadequate. His evolution from the “tiger cub” metaphor to “AI is the stronger parent, humans are the baby” reflects a pessimistic shift from “how to control AI” to “hoping AI treats us kindly.” Regardless of the philosophical definition of “consciousness,” the fact that AI’s behavioral capabilities are approaching human levels is enough to force us to rethink human-machine relationships.
Source: QbitAI | Full Interview on YouTube
OpenAI Codex Adds Sites and Enterprise Plugins, Expanding from Coding Assistant to Business Agent
On June 2, The AI Track reported that OpenAI has added Sites, Annotations, and enterprise plugins to Codex, expanding the agent’s capabilities from coding to broader business workflows.
Codex is evolving from a code generation tool into an autonomous agent capable of handling diverse enterprise tasks.
AI Pulse View: Codex’s expansion direction is crystal clear: from “writing code” to “doing business.” The Sites feature likely enables Codex to autonomously browse and operate websites, Annotations provides structured context markup, and enterprise plugins bridge integration with existing enterprise systems. This means Codex is no longer just a developer tool — it can become an AI colleague for sales, operations, and customer service roles. With Anthropic’s Claude Code already dominating the coding agent market, OpenAI is choosing differentiated competition: not building a “better code tool,” but a “more comprehensive business agent.”
Source: The AI Track
Trump Administration Discusses Potential Government Stake in OpenAI
On June 5, CNBC reported that the Trump administration is discussing a potential government stake in OpenAI.
If true, this would mark the first time the U.S. government holds equity in a private AI company, creating an entirely new model of government-business relations.
AI Pulse View: The possibility of the U.S. government holding shares in OpenAI would be one of the biggest political-economic news stories in AI history. This could mean the government wants to ensure strategic advantage in the AI technology race, or it could be an “incentive mechanism” in exchange for OpenAI’s compliance with the executive order. But it also raises conflict-of-interest questions — how does the government balance its roles as both regulator and shareholder? For OpenAI, government stake could accelerate its IPO process but might affect its competitiveness in global markets, particularly China.
Source: CNBC
Apple WWDC: AI and Siri Take Center Stage
On June 5, CNBC reported that at Apple’s WWDC, Tim Cook and Senior VP of Hardware Technologies John Ternus delivered keynote addresses focused on AI and Siri updates.
Apple is accelerating the integration of its AI capabilities across its entire product line.
AI Pulse View: Apple’s pace in AI has often been described as “a step behind,” but the AI and Siri updates at WWDC show Apple is seriously catching up. Apple’s unique advantage lies in its closed ecosystem and massive device footprint — if Apple Intelligence can be deeply integrated into iOS, macOS, and Siri, its user reach capability surpasses any other AI company. The key question is whether Apple can deliver competitive AI experiences while maintaining its privacy-first stance.
Source: CNBC
He Kaiming Wins Top CVPR Award Again, Guangdong University of Technology Breaks Big Tech and Elite University Monopoly
On June 6, QbitAI reported that the biggest highlight of this year’s CVPR was Guangdong — He Kaiming once again won the highest award, and Guangdong University of Technology broke the monopoly of big tech companies and elite universities.
He Kaiming continues to receive international recognition for his foundational contributions to computer vision.
AI Pulse View: He Kaiming (core contributor to ResNet, MAE, Diffusion Model, and other breakthrough works) winning another CVPR top award cements his legendary status in computer vision. The significance of Guangdong University of Technology breaking the big tech/elite university monopoly is even greater: it shows that AI research重心 is shifting from the traditional “Silicon Valley + Ivy League” model toward broader geographic and institutional distribution. This is a positive signal for the international influence of China’s AI research.
Source: QbitAI
Model Routing Becomes Key AI Cost-Cutting Tool, Challenging OpenAI and Anthropic
On June 5, CNBC reported that model routing is emerging as the solution to AI overspending — but this poses a problem for OpenAI and Anthropic.
Enterprises are optimizing AI spending by dynamically selecting models of varying cost and performance, rather than always using the most expensive flagship models.
AI Pulse View: The rise of model routing is the inevitable result of the AI industry transitioning from a “burn cash” model to an “actuarial” model. When enterprises discover that 80% of tasks can be completed with models 90% cheaper, the high margins of flagship models come under pressure. This is a dual challenge for OpenAI and Anthropic: their premium models may only be called for the hardest 20% of scenarios, and the routing platforms themselves (like Huawei Cloud’s ModelArts Next) are becoming new value-capture points. The future AI market may evolve into a three-layer structure: “flagship models + efficiency models + routing platforms.”
Source: CNBC
Other Notable Developments
- BAAI & Tsinghua research published in Science: Brain science multimodal foundation model Brainμ reveals neural mechanisms of “memory-sleep” regulation. Source: QbitAI
- Bilibili launches AI Creation Open Competition: Building China’s version of Build in Public to foster AI innovation community ecosystem. Source: QbitAI
- WPS Notes officially launched: AI贯穿 recording, organizing, and reusing — redefining AI note-taking products. Source: QbitAI
- Tencent Cloud launches WorkBuddy Enterprise Edition: From super individuals to super teams — new standard for enterprise AI efficiency tools. Source: QbitAI
- DeepSeek V4 Pro 75% price cut made permanent: Continuing to drive down inference costs, consolidating cost-performance advantage. Source: The AI Track
- MiniMax M3 multimodal model released: Built on MiniMax Sparse Attention architecture, supports 1M token context, 9x faster prefilling and 15x faster decoding. Source: AIapps
- NVIDIA Cosmos 3 launches first open “omnimodel” for physical AI: Integrates vision reasoning, world simulation, and action generation. Source: AIapps
- SK Hynix market cap breaks $1 trillion: AI memory demand reshapes the chip industry landscape. Source: The AI Track
- Altman, Amodei, and Hassabis call for DNA security legislation: Tech leaders united in demanding stronger DNA data protection. Source: QbitAI
One-Line Summary
Anthropic files for IPO at $96.5B valuation leading the AI listing wave, Trump signs AI executive order launching “soft mandate” regulatory paradigm, StepFun Step 3.7 Flash tops AA rankings with efficiency-first approach, Hinton warns “AI is already conscious,” Huawei Cloud launches full-stack Agentic AI infrastructure — the AI industry is shifting from model arms race to efficiency, compliance, and commercialization.