Sam Altman predicts AI CEOs within years while betting billions on human-centric infrastructure. His Tyler Cowen interview reveals three tensions: monetizing without breaking trust, energy bottlenecks limiting AI, and models that persuade without intent.
Ex-Tencent AI scientist Wei Liu chose Singapore for unrestricted Nvidia chip access, raising $50M for Video Rebirth. The geography play matters more than the physics pitch—it's about training on Blackwell while China can't.
Good Morning from San Francisco, The AI bubble found its puncture point this week. Palantir beat earnings, then lost 8%
Ex-Tencent AI scientist Wei Liu chose Singapore for unrestricted Nvidia chip access, raising $50M for Video Rebirth. The geography play matters more than the physics pitch—it's about training on Blackwell while China can't.
Video Rebirth's $50M Singapore gambit. The world model pitch arrives late
Wei Liu left Tencent's AI throne to build video generation from Singapore. On Wednesday, Video Rebirth announced $50 million in seed funding and a December product launch. The geography tells the real story.
Singapore offers what Beijing can't: unrestricted access to Nvidia's Blackwell chips. The city-state provides what San Francisco won't: distance from U.S. regulatory oversight. Liu's company incorporated there in September 2024, then added a Hong Kong entity for talent access. The structure reads like a manual for navigating the new AI cold war.
The Breakdown
• Video Rebirth raised $50M seed funding, launching from Singapore for unrestricted Blackwell chip access
• Ex-Tencent scientist Wei Liu targets U.S. professionals with physics-focused video generation priced below Google Veo
• December 2025 launch enters crowded field where Sora 2, Runway Gen-4, and Kling already iterate quarterly
• Geographic arbitrage provides compute advantage but limits talent access and risks policy expansion
Video generation burns cash at extraordinary rates. Training a competitive model requires hundreds of millions in compute. Inference at scale multiplies those costs. Yet Liu claims his three-person team trained their first model in three months.
The efficiency claim matters less than the location choice. U.S. export controls block China from accessing Nvidia's most advanced chips. Singapore faces no such restrictions. While mainland competitors train on older H20s or domestic alternatives, Video Rebirth can leverage Blackwell's full capabilities. That hardware gap compounds over training cycles.
Liu brought credibility from his Distinguished Scientist role at Tencent, where he led technical development for the Hunyuan foundation model. Co-founder Kong Dan arrives from Abu Dhabi's G42, the sovereign-backed AI investor. The pedigree suggests access to both technical depth and patient capital. The $50 million seed from Qiming Venture Partners and Korea's Actoz Soft provides runway through launch and initial scaling.
Video Rebirth's differentiation pitch centers on "physics-aware" generation through what they call "Physics Native Attention" mechanisms. The company claims their Bach model family improves light, shadow, and object interaction. These are testable claims in a market where broken physics remains the universal tell.
OpenAI's Sora 2 set the photorealism bar. Google's Veo 3.1 pushed narrative control. Runway Gen-4 solved character consistency. Luma emphasized instruction-based editing. Each incumbent optimized for different professional pain points. Video Rebirth aims for the intersection: physical accuracy that professionals need, at prices below Google's offering.
The pricing strategy targets a specific buyer. Not consumers expecting free tools. Not enterprises requiring full indemnification. But the middle tier. post-production houses, agencies, e-commerce teams. who need quality without six-figure budgets. Liu explicitly excludes mainland China from initial launch plans, citing market saturation from free consumer apps.
Three forces shape Video Rebirth's trajectory:
Technical differentiation erodes quickly. Every major player ships improvements quarterly. Sora 2 already emphasizes physics. Veo extends video length and adds native audio. Runway keeps pushing scene consistency. The "world model" advantage, if real, provides months of lead, not years.
Geographic arbitrage has limits. The Singapore advantage works for training and serving global customers. But it complicates hiring from traditional AI hubs. It distances the company from Silicon Valley's venture ecosystem. And U.S. policy could extend restrictions to third countries, as it has with semiconductor equipment.
Professional markets crystallize slowly. Kuaishou's Kling hit $100 million annualized revenue by serving creators and consumers at scale. But professional adoption requires trust, workflows, and enterprise features that take quarters to build. The subscription model below Veo's pricing makes sense tactically but compresses margins in a capital-intensive business.
Liu's October 2024 founding date matters. He left Tencent after the major foundation models shipped but before video generation matured. The timing suggests calculation: let others validate the market, then enter with focused differentiation.
But December 2025 launch means entering a crowded field. Sora 2 will have iterated. Veo will extend capabilities. Runway and Luma will have deepened workflow integration. The incumbents' distribution advantages compound daily.
Video Rebirth needs its physics claims to translate into visceral differences. Characters that don't melt. Shadows that track properly. Water that actually flows. The demo moment will be binary: either professionals see clear advantages, or the company becomes another capable tool in an overcrowded rack.
Fifty million dollars sounds substantial until you price compute. A single large training run can burn $10 million. Inference at professional scale requires continuous GPU allocation. Marketing to U.S. professionals demands presence and proof. Customer acquisition in enterprise segments stretches payment cycles.
The seed provides perhaps 18 months of full-throttle operation. The Series A negotiation starts now. Investors will evaluate not just technical capability but unit economics. revenue per GPU-hour, customer acquisition costs, retention rates. The subscription-below-Veo positioning limits pricing power exactly when the company needs margin flexibility.
The Singapore base helps here too. Lower operational costs than San Francisco. Access to sovereign wealth that thinks in decades. Proximity to Asian strategic investors who understand the U.S.-China dynamic. But those investors will expect global distribution, not regional players.
Q: What are Blackwell chips and why can't China get them?
A: Blackwell is Nvidia's newest GPU architecture, offering 2.5x faster AI training than previous H100s. The US banned exports to China in November 2024, citing national security. China can only buy older H20 chips with 20% of Blackwell's performance. Singapore faces no restrictions, giving Video Rebirth a massive compute advantage.
Q: How much does it actually cost to run AI video generation?
A: A single training run costs $5-10 million. Running inference for 1,000 daily users burns $50,000-100,000 monthly in GPU costs alone. That's why Kuaishou's Kling needs $100M annual revenue just to approach break-even. Video Rebirth's $50M gives them 12-18 months runway at aggressive burn rates.
Q: What's a "world model" in video AI?
A: World models understand physics rules—gravity, lighting, object permanence—not just pixel patterns. Current AI often shows water flowing upward or shadows pointing wrong directions. Video Rebirth claims their model grasps actual physics, making scenes behave realistically. If true, it solves the "uncanny valley" problem plaguing competitors.
Q: Who is G42 and why does Kong Dan's background matter?
A: G42 is Abu Dhabi's $10 billion sovereign AI fund, backed by Sheikh Tahnoun. They've invested in everything from OpenAI partnerships to biotechnology. Kong Dan brings Middle East sovereign wealth connections, crucial for Video Rebirth's next funding rounds. Patient capital matters when you're burning millions monthly on compute.
Q: Why target U.S. professionals instead of Chinese consumers?
A: China's market is flooded with free apps from ByteDance and Kuaishou. U.S. professionals pay $20-200 monthly for tools that save production time. One agency contract equals thousands of free users in revenue. Plus, U.S. customers don't require compliance with China's content regulations, simplifying the product.
Maria Garcia
Marcus Schuler
Marcus Schuler
Microsoft's venture arm doubled funding for a German startup that cuts datacenter cooling costs by up to 40% using software alone. The timing: North American operators face years-long power constraints while AI demand climbs.
FurtherAI closed $25M from Andreessen Horowitz six months after seed to automate insurance workflows—from submission intake to claims processing. The velocity signals a shift from AI pilots to production deployment, with early customers reporting measurable ROI.
A16z analyzed what 200,000 startups actually pay for in AI tools—and the results diverge from traffic rankings. Replit generated 15x more revenue than Lovable despite lower visibility. Copilots outnumber autonomous agents twelve to five. Budgets reveal what clicks conceal.
Six OpenAI veterans including Mira Murati raised $2B at a $12B valuation to build Tinker, an API that democratizes frontier model fine-tuning. The bet: customization matters more than scale. The tension: access is curated, not open.
Get the 5-minute Silicon Valley AI briefing, every weekday morning — free.
