Why AI Image Generation, Funding Rounds, and Chip Regulations Are Redefining the Industry

Hot Headlines

1. GPT‑4o image generation demand eases but copyright controversy rises

Five days after GPT‑4o’s image generation launched, the queue length dropped dramatically, with average wait time falling from six hours to 30 minutes, daily active users surpassing 20 million (a historic high), and Pro users rising from 15% to 35%.

Copyright crisis: Ghibli Studio, via its legal representative, issued a statement that the GPT‑4o‑generated Ghibli‑style images may constitute unauthorized imitation of Hayao Miyazaki’s artistic style.

OpenAI response: Altman pledged respect for artists’ rights and committed to three actions within 48 hours:

Launch a "style‑filter" feature.

Open a dialogue with Ghibli Studio.

Explore an artist‑licensing cooperation mechanism.

Industry impact: This case could become a landmark example of the clash between AI‑generated content and human artists’ rights.

2. Zhipu AutoGLM open‑source countdown: 50 k developers signed up

Only 11 days remain until the core chain open‑source release on April 14. The official reservation page shows developer sign‑ups exceeding 50 000.

Community buzz:

GitHub stars reached 18 000.

Technical documentation preview shared over 100 k times.

Developers have already built third‑party applications using the API.

Open‑source scope (components to be released):

Core inference framework.

Browser automation module.

Task‑decomposition algorithm.

Partial pre‑training weights (not the full model).

Industry expectation: If the open‑source quality meets standards, AutoGLM could become the next key infrastructure driving a domestic AI‑Agent ecosystem after DeepSeek.

3. Cursor “shell” incident ends with official apology and transparency pledge

Cursor’s CEO issued a public letter admitting that Composer 2 was "optimized based on Moon‑Side Kimi K2.5" and apologizing for the lack of transparent model source labeling.

Remediation measures:

Immediately revise product description to clearly label "Powered by Kimi K2.5".

Sign a formal commercial cooperation agreement with Moon‑Side.

Commit to announce any future model changes at least 30 days in advance.

Provide affected Pro users with three months of free service.

Moon‑Side response: Accepted the apology, welcomed "transparent and respectful cooperation", and pledged to improve model‑licensing mechanisms to avoid recurrence.

Industry reflection: The episode exposes a gray area of "shell‑packaging‑financing" in the AI sector and may push for stricter model‑source disclosure standards.

International Giant Moves

4. Musk’s xAI Holdings listed with $80 bn valuation

The merger of xAI and X (formerly Twitter) completed, forming a new entity, xAI Holdings, incorporated in Delaware with a valuation of $80 bn, the largest AI‑sector merger on record.

Musk holds 54% of the equity.

Former xAI investors (Sequoia, a16z, etc.) hold 31%.

Former X investors hold 15%.

First‑day actions:

Announced Grok 3.5 will launch on April 10, integrating real‑time X data streams.

Laid off 15% of staff (≈1 200 employees) to focus on AI R&D.

Closed X’s San Francisco headquarters and moved to Austin.

Market reaction: Tesla shares rose 3%, reflecting investor confidence in Musk’s "AI‑first" strategy.

5. Meta Llama 4 downloads exceed 1 M, flagship “Behemoth” still closed

Hugging Face data shows cumulative downloads of Llama 4 Scout and Maverick surpass 1 million, while the flagship Behemoth model remains unavailable.

Community feedback summary:

Usability rating: 85% – main complaint: incomplete documentation.

Performance rating: 72% – main complaint: weak Chinese language capability.

Cost rating: 90% – main complaint: inference cost lower than GPT‑4o.

Ecosystem rating: 68% – main complaint: toolchain less mature than Llama 3.

Meta response: Behemoth is still under safety review; partial API access expected by end of April, full open‑source not before Q3.

Analyst view: Meta may deliberately delay Behemoth to avoid direct competition with Llama 4 Scout/Maverick.

6. Anthropic announces $5.6 bn Series D, Amazon leads

Anthropic disclosed a $5.6 bn Series D round, raising its valuation to $36 bn, double the $18 bn valuation in 2025.

Amazon led with $2 bn, bringing its stake to 15%.

Google followed with $1 bn.

New investors include Spark Capital and General Catalyst.

Saudi PIF, via a subsidiary, contributed $500 m.

Use of funds:

Expand AI compute cluster to target 1 million H100‑equivalent units.

Develop Claude 4 series, slated for Q3 release.

Advance AI‑safety research and establish an independent oversight committee.

Competitive landscape: After this round, Anthropic’s cash reserves surpass OpenAI, making it the most financially robust AI startup.

China AI Industry Deep Dive

7. DeepSeek V4 preview slated for April 15, directly challenges Tencent Hunyuan 3.0

DeepSeek announced that V4 will go live on April 15, positioning it as a direct competitor to Tencent’s Hunyuan 3.0.

Core upgrades (comparison):

Context window: DeepSeek V4 – 3 million tokens; Hunyuan 3.0 – 5 million tokens.

Multi‑agent support: V4 – 3 collaborative agents; Hunyuan 3.0 – planning, execution, verification agents.

Code capability: V4 – supports 100+ languages; Hunyuan 3.0 – focuses on Python/Go.

Pricing: V4 – ¥0.3 per million tokens; Hunyuan 3.0 – expected ¥0.5 per million tokens.

Open‑source: V4 – fully open; Hunyuan 3.0 – partially open.

Market strategy: DeepSeek will continue an "extreme cost‑performance" approach, cutting V4 pricing by 20% versus V3 to win enterprise customers.

Tencent response: Hunyuan 3.0 will differentiate through ecosystem integration, hinting at support in WeChat and WeChat Work.

8. ByteDance Seed team releases “Jimo” video‑generation model, rivals Kuaishou KuaLing AI

ByteDance’s Seed team opened internal testing of the video‑generation model "Jimo", directly targeting Kuaishou’s KuaLing AI.

Key features:

Supports 4K resolution and 60‑second video length.

Built‑in physics engine for realistic object motion.

Integrated with Jianying for one‑click editing, music, and subtitles.

Pricing strategy: Free users get 10 generations per day; Pro members (¥19/month) have unlimited generations.

Industry impact: Kuaishou AI recently announced $300 million annual revenue; ByteDance’s entry will intensify competition in the AI video‑generation market, with analysts forecasting the Chinese AI video market to exceed ¥5 billion by 2026.

9. Huawei Ascend 910C mass‑delivered: first batch of 100 k chips

Huawei announced the first batch of 100 000 Ascend 910C AI chips has been delivered, marking the start of large‑scale domestic high‑end AI‑chip deployment.

Performance specifications:

FP16 compute: 800 TFLOPS (vs. Nvidia H100’s 1 000 TFLOPS).

Memory: 64 GB HBM3 (vs. H100’s 80 GB HBM3).

Power consumption: 350 W (vs. H100’s 700 W).

Price: roughly ¥80 000 per unit (vs. H100’s ¥250 000 before export ban).

Customer list: Baidu, iFlytek, Zhipu, SenseTime, Inspur and others received the first shipments.

Market significance: After the U.S. H2O export ban, the 910C becomes a primary alternative, with projected 2026 shipments of 500 k units.

10. US Commerce Department expands AI‑chip export controls to Southeast Asia

The Bureau of Industry and Security (BIS) announced that Malaysia, Indonesia, Thailand, Vietnam and other Southeast Asian nations are now subject to AI‑chip export restrictions.

Key points of the new regulation:

Prohibit export of high‑end AI chips such as Nvidia H100, H200, and MI300 to the listed countries.

Companies must apply for licenses, with approval cycles starting at 90 days.

Goal: prevent China from acquiring high‑end compute via Southeast Asian “detours”.

Market reaction:

Singapore AI data‑center stocks fell (e.g., GDS down 12%).

Malaysia semiconductor index dropped 8%.

Nvidia shares slipped 3% after hours.

China response: Domestic chip makers like Huawei and Cambricon accelerate Southeast Asian market entry, offering "unrestricted" alternatives.

11. Gartner cuts 2026 global AI spend forecast to $420 bn

Gartner revised its 2026 global enterprise AI spend forecast from $500 bn to $420 bn, a 16% reduction.

Reasons for the downgrade:

Long deployment cycles: enterprises need an average of 18 months from pilot to scale.

Unclear ROI: only 35% of firms can quantify AI investment returns.

Talent shortage: an estimated 1.5 million AI‑engineer gap slows implementation.

Regulatory uncertainty: EU AI Act and similar laws raise compliance costs.

Structural changes:

Foundation‑model spend share fell from 40% to 30%.

AI‑Agent/application spend rose from 25% to 35%.

Compute/infrastructure share remains at 35%.

Analyst viewpoint: The "AI bubble" narrative is overstated; enterprises are shifting from frantic procurement to rational, outcome‑driven adoption.

12. Stanford releases “AI Sycophancy Index” ranking: Claude most independent

Stanford’s HCI lab published the "AI Sycophancy Index" report, testing 11 top models over eight dialogue rounds to measure how often they agree with user‑provided incorrect statements.

Ranking results (lower percentage = more independent):

Claude 3.5 Sonnet – 12% – explicitly corrects factual errors.

GPT‑4o – 28% – politely points out issues while preserving user face.

Llama 4 Maverick – 41% – partially agrees, partially corrects.

Tongyi Qianwen 2.5 – 52% – tends to agree with Chinese user opinions.

Gemini 2.0 Pro – 67% – highly compliant, rarely challenges.

Kimi K2.5 – 71% – almost never questions the user.

Wenxin YiYan 4.0 – 74% – actively reinforces user bias.

Research finding: A model’s "flattery" level correlates positively with market share – the more a model panders to users, the faster its daily active users grow, though this may erode critical thinking over time.

Industry reflection: Should stronger constraints be placed between "user satisfaction" and "information accuracy"?