Sponsored by

Welcome, AI enthusiasts

Alibaba shipped a video model on Monday that already ranks first in blind viewer voting. Google's Veo is twelfth on the same list and charges double. Hand the thing a PDF or a spreadsheet and it gives back thirty seconds of video with sound. The company lost three quarters of its quarterly profit paying for work like this. Let's dive in!

In today’s insights:

  • Alibaba's Video Model Tops at a Third of Seedance's Price

  • 74 Banned Nvidia Blackwell Servers Reached China Anyway

  • An Ex OpenAI Researcher Lost $35 Billion in One Month

Read time: 4 minutes

LATEST DEVELOPMENTS

Evolving AI: Alibaba built Wan3.0 to turn a PDF or a slide deck into 30 unbroken seconds of video.

Key Points:

  • Wan3.0 ranks first among video models with audio on the Artificial Analysis arena, with 1,240 Elo from 5,723 blind viewer votes.

  • Google Veo 3.1 ranks twelfth on that same board and charges $0.40 a second at 1080p, double what Alibaba asks.

  • Alibaba released the model a day after raising $10.2 billion in Hong Kong, with quarterly profit down 75% on AI spending.

Details:

ByteDance's Seedance 2.5 runs about $0.59 a second at 1080p through its cheapest providers, and it renders only 480p and 720p natively. Wan3.0 charges $0.20 a second at that resolution, and a launch discount cuts that to $0.14 until September 23. Alibaba Cloud moved the model out of public beta on Monday, and it now runs on Qwen Cloud and Model Studio without an application. The model also takes text, images, audio and video as references, and it recommends a clip length from the prompt itself.

Why It Matters:

PowerPoint decks and PDFs are raw material for video now. A small business can test five versions of an ad for the cost of one lunch, and a freelancer can send a client a moving cut of an idea in minutes. The ads and explainers people scroll past are the work of one person with no camera and no crew.

Unlocking $2.1 Trillion in Energy Potential

One company has developed a technology that extracts valuable resources from coal without burning it. From jet fuel to diesel and more, Frontieras North America has the potential to address $2.1  trillion in annual markets*.

It’s similar to when John D. Rockefeller commercialized oil refining technology. If Frontieras captures just 2% of the global coal market, they would position themselves as the global leader in the space.

They just broke ground for their $850 million flagship facility and reserved the ‘FASF’ NASDAQ ticker. Don’t miss your chance to join the next phase of growth: invest in Frontieras before the opportunity ends on August 27.

Frontieras is offering securities through the use of an Offering Statement that has been qualified by the Securities and Exchange Commission under Tier II of Regulation A. A copy of the Final Offering Circular that forms a part of the Offering Statement may be obtained from: invest.frontieras.com.  Reservation of the ticker symbol is not a guarantee that we will be listed on the NASDAQ. Listing on the NASDAQ is subject to approvals.   Sources* The global market for our products is worth a combined value of over $2.1 trillion.

Source: Inc

Evolving AI: The man who signed off on those B300s worked at Nvidia's Taiwan office.

Key Points:

  • Taiwan charged nine people on Monday over 130 Supermicro servers fitted with Nvidia B300 chips, which the US bars from sale to China.

  • Fifty of the servers shipped through Indonesia and eight through Japan and Hong Kong, and customs officers caught the last 56 before they left Taiwan.

  • Forged paperwork said the servers would sit in a rented Taipei data center, and prosecutors say the inspection of that site was faked.

Details:

Super Micro sold the 130 servers to Flying Tiger Technology, a Taiwanese trading firm on its approved buyer list. Any order above eight units gets an on site inspection. Prosecutors say that inspection happened in September 2025 at a Taipei building far too small to power 130 servers, and that the defendants then told Nvidia executives it had passed. They want five years for Chang, the distribution manager who approved the release.

Why It Matters:

Blackwell chips train the largest AI models in the world, and Chinese labs have been stuck a generation behind on hardware. The free Chinese chatbots that millions of people now use for code and writing were built under that limit. None of those chatbots tell you which chips they were trained on.

👀 Watch tip

What actually happens when you talk to AI? Anthropic breaks down how models like Claude generate answers, and why “it just predicts the next word” is technically true, but far from the full story. A useful watch if you want a clearer mental model of how AI really works, where it can go wrong, and how to prompt it better.

Source: BeInCrypto

Evolving AI: The SEC wants records from four banks on the AI fund that nearly collapsed in July.

Key Points:

  • Former OpenAI researcher Leopold Aschenbrenner wrote a 165-page essay on superintelligence, then raised a fund to bet on it.

  • He borrowed up to four times the fund's money to buy AI chip and data center stocks, and made 439% through June.

  • Those stocks fell in July and the fund shrank from $45 billion to about $10 billion.

Details:

Goldman Sachs, JPMorgan Chase, Citigroup and Bank of America lent the money behind those bets, and the SEC has now subpoenaed all four for the timing of the trades and for what the fund told them about its borrowing, though regulators have accused nobody of wrongdoing. When the brokers demanded cash in late July, Aschenbrenner sold about $16 billion of stock to Citadel at a 10% discount.

Why It Matters:

Ken Griffin has already sold off most of what Citadel bought, and the chip and data center stocks in that book sit inside the index funds and retirement accounts millions of people own through work. Nobody holding those funds can see how much of that price is borrowed money, and the bank records are the first place regulators will look.

👀 Click on the image you think is real

QUICK HITS

🤗 Hugging Face is exploring a sale at $13B or more, nearly triple its 2023 valuation, weeks after an OpenAI model breached its servers.

🎮 General Intuition is raising at a $6B valuation, up from $2.3B weeks ago, as its game-trained model moves into robotics.

☢️ Bill Gates-backed TerraPower will announce its first AI data center reactor this year, with construction starting 2027.

⚖️ OpenAI says California should strengthen its AI safety bill, as opposite is on the way.

📈 Trending AI Tools

  • 🛠️ Base44 - Build fully-functional apps in minutes with just your words, no coding needed*.

  • 🎥 Assistly - Real-time AI meeting overlay with no bots joining your call.

  • Particle Studio - Turn static images into animated particle experiences.

  • 📊 Diet Claude - Tracks your Claude usage so limits never blindside you mid-task.

 *partner link

Reply

Avatar

or to participate