title: "AI Weekly | ByteDance AI Four Major Propositions, MaaS Target 150B, Claude Mythos Deployment in 15 Countries' Infrastructure"

description: "This week’s AI highlights: ByteDance discloses four major AI propositions for 2026; Volcano Engine raises its MaaS target to 150 billion; Claude Mythos deploys critical infrastructure in 15 countries; Meta builds tent data centers; Nvidia acquires Kumo AI; Anthropic calls for a pause on AI self‑improvement training."

pubDate: 2026-06-05

category: '周刊'

tags: ['AI周刊', '字节跳动', 'Anthropic', '火山引擎', 'Claude', '具身智能', '融资']

featured: true

readTime: "13分钟"

views: 1

I. Major Players' Moves: From Valuation Competition to Real Competition in Computing Power and Products

ByteDance Discloses Four Key AI Propositions for 2026, Targeting Google Genie 3

36Kr exclusively reveals ByteDance's key AI propositions for 2026:

This marks the first time ByteDance has explicitly stated "targeting Genie 3" in its internal positioning, signifying that world models have officially become the next-generation flagship competitive focus for domestic tech giants.

Huoshan Engine Raises MaaS Full-Year Revenue Target to 15 Billion, Seedance 2.0 Breaks 1 Billion in Single Month

36Kr exclusively learned that Huoshan Engine raised its MaaS business revenue target to 150 billion yuan in April, with near-monthly adjustments—another step up from the 100 billion target at the end of 2025. Video model Seedance 2.0 has broken 1 billion yuan in monthly revenue, becoming the fastest monetization product in the AIGC video track.

> A single MaaS business with a target increase of 50% within half a year, and one video model alone generating 1 billion per month—this is the first time domestic large model commercialization has provided a "visible cash" sample.

II. Models and Products: Agents Move Toward the "Workbench"

Claude Mythos Deployed in 15 Countries' Critical Infrastructure, Anthropic Officially Files for IPO

Anthropic officially filed its IPO prospectus to sprint toward listing, and simultaneously elevated Claude Mythos to critical infrastructure scenarios in 15+ countries (electricity, transportation, communication, etc.). This is a landmark event for Anthropic in the enterprise market—"selling AI into the systems that cannot afford to make mistakes."

Claude Cowork Desktop Agent: Bringing Claude Code Capabilities to Non-Developers

Anthropic launched Claude Cowork—a desktop AI Agent on macOS that can directly read, edit, and create local files. Paired with Claude Opus 4.8's focus on "honesty" (proactively marking uncertainty when errors occur), Anthropic's product strategy has expanded from "selling to engineers" to "selling to everyone who uses a computer."

Alibaba Qwen3.7-Plus Launched, New Foundation for Multi-modal Agents

Alibaba's Tongyi Qianwen Qwen3.7-Plus was officially released, featuring multi-modal agent capabilities—capable of one-click replication of professional desktop software workflows (Photoshop, IDE, Office). This is the first time the Qwen series has explicitly positioned "multi-modal + Agent" as its flagship label, meaning Alibaba Cloud aims to upgrade Agent from a Chatbot to an "operating system's co-pilot."

StepUp(step) Star Step 3.7 Flash Open-Sourced, Domestic Agent Model Efficiency Route

Step 3.7 Flash (198B MoE) ranked first in ClawEval and SimpleVQA Search benchmarks, focusing on agent workflow efficiency. Paired with Alibaba Cloud Bailian CLI and ByteDance Bernini framework open-sourcing—domestic models are shifting from "catching up on general capabilities" to "being optimal in specific scenarios."

Microsoft Build 2026: MAI-Thinking-1 + Scout Personal Assistant + Project Solara

Microsoft unveiled its self-developed reasoning model MAI-Thinking-1, the Scout personal assistant based on the OpenClaw concept, and Project Solara, an operating system designed specifically for AI Agent hardware. Microsoft's strategy is clear—it doesn't just want to do models and applications; it also wants to build the "AI hardware operating system" layer.

3. Embodied Intelligence and Robotics: Mass Production Race Kicks Off

CVPR 2026 On-Site: NVIDIA, Tesla, and Waymo Share the Stage to Hear Chinese Companies Discuss Physical AI

The CVPR 2026 Physical AI track was packed with teams from NVIDIA, Tesla, and Waymo, with Chinese embodied intelligence and autonomous driving companies as the main speakers. Chinese manufacturers have taken the lead in implementing the "data collection → world model → closed-loop training" flywheel, seizing the right to define standards on the physical AI track.

Tesla Optimus Robot Factory Breaks Ground, Planning Annual Capacity of 10 Million Units

The dedicated Optimus humanoid robot factory officially broke ground inside Tesla's Texas Gigafactory, planning a peak annual capacity of 10 million units, with large-scale production scheduled for summer 2027. In parallel:

The World Model Direction LeCun Bet $1 Billion On, Domestic Vision Large Model Teams Already Had in Place

LeCun's latest venture JEPA 2 bets on latent space world models. Several leading domestic vision large model teams started laying out this path as early as 2024, with multiple CVPR 2026 oral reports coming from Chinese teams. The competition for discourse power in world models has entered a "China vs USA" bipolar pattern earlier than general LLMs.

Four. Computing Power and Hardware: Hardware Becomes an Engineering Problem

Meta Follows Tesla: Building AI Data Centers in Tents

TechCrunch reports that Meta replicated Tesla's early approach of using tents to quickly ramp up production capacity, building some newly constructed AI data centers directly in tent structures, compressing the time to get computing power online from several years to several months.

> The next step in the computing power arms race is not "making better chips," but "getting chips online faster."

Nvidia Acquires Kumo AI, Doubling Down on Enterprise Generative AI Inference

Nvidia officially acquired Kumo AI, a platform focused on generative AI inference and prediction on enterprise data. Kumo's capabilities will be integrated into Nvidia NIM and AI Enterprise suites—further strengthening Nvidia's full-stack presence in the enterprise GenAI market.

Groq Raises $650 Million, AI Inference Chips Become a New Hot Spot

Following Nvidia's $20 billion "non-acquisition hiring," AI inference chip company Groq raised $650 million. During the same period, XCENA raised $135 million at a $570 million valuation, betting that "memory is the real bottleneck for AI."

BYD Releases China's First 4nm Intelligent Driving Chip Xuanji A3

BYD released China's first 4nm process intelligent driving chip, which has already begun mass production, supporting L3/L4 autonomous driving, with three chips working in coordination to achieve a total computing power exceeding 2100 TOPS.

OpenAI CFO: AI Hardware Will Officially Launch Before the End of This Year

OpenAI CFO Sarah Friar revealed that she has personally experienced the company's AI device, confirming it will officially launch "before the end of this year"—significantly earlier than the previously internal document-projected mass production timeline of February 2027.

Five, Safety and Governance: Anthropic Rarely Calls for Pause on AI Self-Improvement Training

According to WSJ reports, Anthropic called for a global pause on training experiments that could significantly enhance AI self-improvement capabilities in a policy briefing, stating that current alignment technologies are insufficient to address the uncontrollable risks posed by "recursive self-improvement."

This is the first time a mainstream large model provider has formally proposed a training pause initiative, which is of significant importance:
  1. Coming from the least "conservative" player (Anthropic is the most aggressive in commercialization)
  2. The timing is intriguing—coinciding with Anthropic's own IPO filing
  3. Forming a "open-source defense + call for pause" combination with Anthropic's open-sourcing of defending-code-reference-harness last week

During the same period, Anthropic also open-sourced an AI vulnerability discovery framework on GitHub, which can be fine-tuned to create a "security audit Agent" targeting their own code repositories.

Six, Products and Ecosystem: AI Agent Fully Embedded Implementation

Apple Approves Poke as the First AI Agent on Messages for Business

Apple has allowed a third-party AI Agent to access the Messages for Business platform for the first time. This is a landmark event marking Apple's official opening of its channel to the AI Agent ecosystem.

Meta Pushes WhatsApp Business AI Agent Globally

WhatsApp Business AI customer service Agent exits pilot and opens to global merchants, supporting multilingual automatic responses, product recommendations, and ordering processes.

ByteDance Bernini Open Source: Equipping DiT with an "LLM Strategist"

ByteDance open-sources unified framework Bernini, focusing on "understand first, then act" AI video editing capability—using large language model as the decision-making hub, collaborating with diffusion models to enhance controllability in video generation.

Baidu PaddleOCR-VL-1.6 Achieves 96.33% Accuracy, Setting New SOTA for Same Size

Baidu ERNIE team releases PaddleOCR-VL-1.6 document parsing model with 96.33% accuracy, consolidating advantages in multimodal document understanding.

Doubao Officially Launches Paid Services in Late June, Integrating with Douyin E-commerce

ByteDance Doubao large model will officially launch paid subscriptions in late June, accelerating integration with Douyin e-commerce ecosystem. Simultaneously launching the "Doubao Auto" solution, targeting the mainstream family car market segment priced at 100,000-200,000 yuan.

VII. Numbers Worth Attention

MetricValueSignificance
Volcano Engine MaaS Target150 billion yuanUpward adjusted by 50% from 100 billion yuan at end of 2025
Seedance 2.0 Monthly Revenue10 billion yuanFastest AIGC video monetization record
Claude Mythos Deployed Countries15+Critical infrastructure-level AI deployment
Tesla Optimus Annual Production Target10 million unitsMass production race officially begins
Meta Tent Data CenterLaunch time compressed to a few monthsComputing power construction model undergoing dramatic change
BYD Xuanji A33x 2100 TOPSNew benchmark for domestic automotive-grade chips
Anthropic Valuation (last week)$965 billionIndustry enters the trillion-dollar club

Overall Analysis

This week's most significant signal in the AI industry: Commercial scaling and industrial embedding are accelerating simultaneously, with regulation and governance moving from "discussion" to "action".

From the commercial perspective, ByteDance Volcano Engine's MaaS target has been raised to 15 billion, Seedance reached 1 billion monthly, Anthropic officially filed for IPO, and Claude Mythos is deployed in critical infrastructure across 15 countries—the AI industry's first phase of "technology validation" has ended, entering the era of dual validation for "cash flow and use cases." The particularly important aspect of Volcano Engine's monthly Seedance reaching 1 billion: It is the first time proving that AIGC video is not a "money-burning" business, but one that can generate real cash flow. From the hardware perspective, the industry's focus has shifted from "building better chips" to "getting chips deployed faster." Meta setting up tents, Anthropic deploying across 15 countries, Nvidia acquiring Kumo, OpenAI hardware ahead of mass production—AI has entered the "engineering delivery" phase, where supply chain speed itself is the moat. From embodied intelligence perspective, Tesla Optimus's annual production capacity target of 10 million units is taking shape, CVPR 2026 physical AI topics are being led by Chinese teams, LeCun betting $1 billion on world models—robots moving from demos to production lines is undeniable. Companies like Li Auto, Unitree, Daimeng, Xingchen, and Zhiwei Chip are raising funds intensively, with post-2000s domestic entrepreneurs taking center stage, reflecting a new wave of "high-stakes, high-consumption, high-valuation" gameplay taking shape. From a governance perspective, Anthropic's rare call to pause AI self-improvement training was unthinkable in the past. This is both Anthropic's own compliance positioning (risk management before IPO) and the industry's turning point from "hard alignment" to "proactive limits." Anthropic's open-source vulnerability framework combined with the pause call shows that leading companies are beginning to proactively shape the regulatory narrative. From a product perspective, AI becoming comprehensively "embedded" has become the main theme—Apple approved Poke for Messages for Business, Meta WhatsApp Business Agent opened globally, Alibaba Qwen3.7-Plus replicating desktop software workflows, Microsoft Project Solara building a system for AI Agent hardware—users no longer need to "go find AI"; AI directly appears in existing workflows. In one sentence: In June 2026, the AI industry completed its shift from "hundred model wars" to "scenario wars," Chinese AI gained global standard-setting power for the first time in embodied intelligence and video generation.

Looking Ahead to Next Week