OpenAI’s Sora Shutdown and Seedance 2.0 Blocked in the US: Is the United States Losing the Global AI Video Race to China?

0:00 Sora Shutdown & Seedance Blocked: A Seismic Shift in AI Video Picture the tension in that boardroom. I mean, it is early this morning, March 24th, 2026. Executives from Disney are sitting right across from the leadership at Open AI, and they are just pulling the plug on a $1 billion investment in licensing deal. 0:16 Speaker 2 Right, just a billion dollars vanishing in an instant. 0:18 Speaker 1 Exactly vanishing. They aren't just, you know, tweaking a contract. Disney is walking away because the product they were banking on, Open AI, is highly anticipated. Video generation app Aura is officially dead. They killed it today. And at the exact same moment, like half a world away, Bytedance's groundbreaking video Model C Dance 2 Point O is facing this intense global backlash and a completely suspended rollout. 0:45 Speaker 2 It really is a massive earthquake. I mean the landscape of synthetic media just be fractured today. Two of the most powerful engines driving the future of video. It essentially hit brick walls on the exact same day. 0:55 Speaker 1 Which is exactly why we are tearing into this topic today. For this deep dive, our mission is to map out this tectonic shift happening right now in the US China AI race. 1:03 Speaker 2 Yeah, there is a lot to cover. 1:05 Speaker 1 Oh, there really is. We are going to explore why the US is suddenly hitting the brakes on consumer AI video, how Hollywood is launching an all out war against Chinese models. And you know what this actually means for the future of filmmaking and your own creative projects. 1:20 So OK, let's unpack this. 1:23 Speaker 2 What's fascinating here is that to understand the magnitude of today, we really have to recognize that SORA shutting down and C dents being delayed, They aren't isolated incidents. We're watching the collision of three massive forces. You've got astronomical compute costs that are just breaking corporate budgets, fierce intellectual property battles that are literally rewriting the law, and two completely divergent national strategies for artificial intelligence. 1:49 Why OpenAI Shut Down Sora: The Astronomical Cost of Inference Let's start with that immediate fallout, because Open AI shutting down the Sora app today is just staggering reversal. It really is if you look back at those viral videos from last year. Like, remember Sam Altman supposedly stealing graphics cards from Target? 2:02 Speaker 2 Oh yeah, that was everywhere. 2:03 Speaker 1 Right. It felt like they were about to just hand everyone a Hollywood studio on their phone, but instead they're completely abandoning consumer video generation. They're pulling all those resources and pivoting hard toward Productivity Tools. 2:17 Speaker 2 Yes, specifically building out their Atlas web browser and their codecs coding platforms. 2:22 Speaker 1 Exactly. So why walk away from a billion dollar Disney partnership just to build another web browser? Well. 2:28 Speaker 2 It basically comes down to a brutal reality check regarding inference costs. 2:34 Speaker 1 Wait, let's pause right there. What exactly do we mean by inference costs in this context? Because I hear that thrown around a lot. 2:40 Speaker 2 Right. So when we talk about AI, there are two main phases. First, there's the training phase, where you feed the model mass amounts of data so it learns how to generate video. 2:49 Speaker 1 And that's the expensive part upfront. 2:51 Speaker 2 Right. Yes, exactly. That costs 10s of millions of dollars upfront. Yeah, but then there's the inference phase, that is the actual cost every single time a user types in a prompt and hits generate. And video generation requires an astonishing amount of computational power for every single frame produced. 3:09 O Open AI basically looked at the math and realized that running SORA at scale for millions of consumers would essentially burn cash faster than they could ever recoup through a standard subscription fee. 3:21 Speaker 1 So it feels like Open AI spent years building this cutting edge experimental hypercar only to realize that, well, the jet fuel required to run it is way too expensive and the insurance is just a nightmare, so they're pivoting to selling tractors instead. 3:36 You know, web browsers and coding assistance. That's the reliable, safe money. 3:40 Speaker 2 That is a perfect way to look at it actually. Enterprise Productivity Tools have very predictable compute costs. There's a clear return on investment for businesses. Consumer video generation, though, it's just a bottomless pit of server costs. 3:53 Ethical Boundaries Shattered: Seedance 2.0 and Hollywood's Backlash But then we look at Byte Dance. I mean, they launched Seed Dance 2.0 in February and the cost didn't seem to bother them at all. 4:00 Speaker 2 Not even slightly. 4:01 Speaker 1 No, and the launch was chaotic in like the best and worst ways. We saw cinema quality photorealistic clips just flooding the Internet. Remember that viral video of Tom Cruise and Brad Pitt getting into a fight? 4:16 Speaker 2 Oh, or the new version of Will Smith eating spaghetti. 4:19 Speaker 1 Yes, which finally lacked those usual tells, right? Like the text in the background wasn't garbled. There weren't extra fingers everywhere. 4:26 Speaker 2 The leap in capability was undeniable. I mean the physics, the way the light wrapped around objects, the temporal consistency. 4:33 Speaker 1 Leaning objects didn't morph into weird shapes when the camera. 4:36 Speaker 2 Pan exactly. It was a true generational leap. 4:39 Speaker 1 Yet an absolute emergency brake just got pulled today. Sea Dance is abruptly suspended globally and the catalyst wasn't server costs like with open AI. 4:48 Speaker 2 No, it was something much darker. 4:50 Speaker 1 Yeah, it was a terrifying feature discovered by a media founder named Pantianhong. He found out you could generate a highly accurate personal voice clone using only a facial photograph. 4:59 Speaker 2 Just a photo. 5:00 Speaker 1 Right, Let that sink in for you listening. Not a voice sample, just a picture of your face. And the model extrapolates your voice in under 60 seconds without any authorization. How is that even technologically possible? How does a picture make a voice? 5:15 Speaker 2 It relies on something called cross modal correlation. 5:18 Speaker 1 OK, what is that? 5:19 Speaker 2 Well, because Sedence was trained on such a massive, completely unfiltered data set of videos, meaning literally millions of hours of people talking, the AI learned to map the physical structure of a person's face directly to their voice. 5:34 Speaker 1 Like their jawline? 5:35 Speaker 2 Yeah, their jawline, the shape of their vocal tractor nasal cavity, it maps those visual markers to the acoustic qualities of the voice that usually accompanies that physical structure. 5:45 Speaker 1 So it's literally guessing what you sound like based on the bone structure of your face, and it's doing it accurately. 5:51 Speaker 2 Frighteningly accurately. And it does it seamlessly in the background. So Byte Dance didn't stop because the tech was too expensive to run. They paused because the technology she completely shattered our current understanding of privacy and identity protection. They collided head first with ethical boundaries, and the global panic surrounding deep fakes and identity theft forced their hand. 6:11 Speaker 1 Which brings us to exactly who is panicking the most right now. The fallout from C Dance 2 Point O has triggered a fury from the global entertainment industry unlike anything we've seen. 6:20 Speaker 2 Total panic. 6:21 Speaker 1 Netflix has actively labeled the model a high speed piracy engine. We are seeing perfect, completely unauthorized replications of characters and settings from shows like Stranger Things and Squid Game. 6:33 Speaker 2 Yeah, because the model clearly ingested those specific visual languages, right? When a user prompts C dance for a, you know, a dystopian Korean game show, the model isn't inventing an original concept, it's retrieving and reassembling the exact copyrighted aesthetic of Squid Game because that data was scraped and heavily weighted during the training phase. 6:52 Speaker 1 And the legacy studios are not taking this lying down. The Screen Actors Guild, the SAG, FTRA and the Motion Picture Association are firing off cease and desist letters to Bite Dance as we speak. 7:01 Speaker 2 It's a full legal assault. 7:03 Speaker 1 They are citing existential threats to actors, careers and just a blatant disregard for consent and this is escalated all the way to Capitol Hill. US Senators Marshall Blackburn and Peter Welch wrote directly to Byte Dance CEO Liang Rubo demanding an immediate shutdown due to copyright and likeness violations. 7:22 Speaker 2 Yeah, it's intense. 7:23 Speaker 1 I mean, a screenwriter for Deadpool literally went on the record and said, I hate to say this, but we're done. But I have to push back a little here, though. Sure. Hasn't technology always disrupted Hollywood? I mean people thought CGI was going to ruin movies and instead it just became a new tool for the studios. 7:39 Is Hollywood overreacting or is this truly an existential threat? 7:43 How US Copyright Laws Stall Domestic AI Innovation I'd. 7:43 Speaker 2 Argue it is a foundational threat to their entire business model because this isn't just a new tool like CGI, OK, it's a replacement for the studio machinery itself. Generative AI models require vast amounts of premium, high quality data to achieve realism, and that data is owned by fiercely protective legacy studios. 8:03 Speaker 1 Right, the Disney Vault. 8:04 Speaker 2 Exactly. The tech industry operates on an ethos of frictionless scaling. You know, scrape everything, ask for forgiveness later. But the entertainment industry relies on deeply entrenched intellectual property laws to monetize their catalogs. 8:19 Speaker 1 So they are fundamentally opposed. 8:22 Speaker 2 Right. If an algorithm can perfectly synthesize a Marvel quality aesthetic on demand for literally pennies, the immense financial value of the original studio infrastructure completely evaporates. 8:35 Speaker 1 Which perfectly explains why US lawmakers are jumping into demand of foreign companies shut its model down. But the way the US government is actually trying to manage this landscape domestically is incredibly complex. 8:46 Speaker 2 Very messy. 8:47 Speaker 1 Let's look at the national policy Framework for artificial intelligence, which the Trump administration released just a few days ago on March 20, 2026. And before we dive into the details, I want to be very clear to you listening. We are looking at this framework strictly objectively to understand its impact on the tech landscape. 9:03 We are simply unpacking the policy documents, not endorsing or taking sides on the political strategies behind them. 9:09 Speaker 2 And objectively speaking, this framework is a critical piece of the puzzle for understanding why the US market is hesitating right now. 9:18 Speaker 1 So the biggest take away from the framework is its stance on copyright. It actually adopts the position that training AI on copyrighted material is not a violation. Wow. Yeah. The administration is essentially saying they're not going to intervene with new federal legislation. 9:34 They are completely leaving the messy, complex fair use battles to the judicial system to figure out why is that such a big deal for developers. 9:43 Speaker 2 Because the judicial system moves at a glacial pace compared to software development. Very true. By explicitly leaving the copyright issue to the courts, the framework creates profound legal ambiguity. You mentioned open AI pivoting earlier. Imagine being an executive there. 9:59 You cannot afford to scale a consumer product like Sora if a judge might rule, say, three years from now that your entire foundational training data set was illegal. 10:07 Speaker 1 That would expose you to billions and damages. 10:10 Speaker 2 Exactly that ambiguity forces established American companies to play it incredibly safe. 10:16 Speaker 1 The framework also pushes heavily for federal preemption. This is an aggressive move to override state level AI laws. For example, the Colorado AI Act that went into effect earlier this year. It's essentially wiped out under this plan, replaced by a lighter single national standard. 10:34 And if you are listening right now and wondering why you should care about a state law being overridden, it's because this dictates the actual software you are legally allowed to download and use tomorrow. The stated goal here is to ensure American AI dominance by just removing local red tape. 10:49 Speaker 2 And they are pairing that deregulation with something called the Ratepayer Protection Pledge. 10:53 Speaker 1 Right, the pledge where tech giants like Amazon, Google, Open AI, and Microsoft have agreed to cover their own massive data center electricity costs and infrastructure upgrades, rather than passing those costs onto residential utility citizens. What's the strategy behind that specific pledge? 11:10 Speaker 2 It's an attempt to mitigate a growing populist backlash. I mean, these AI models require gigawatts of power. 11:17 Speaker 1 It's insane how much power they draw. 11:18 Speaker 2 It really is, and communities were starting to see their local energy grid strained and their utility bills spiking simply because a tech company built a massive server farm down the road. So the pledge is designed to keep the public from turning against AI infrastructure expansion. 11:34 Speaker 1 But the roll out of this framework is facing intense friction. State lawmakers and attorneys general are pushing back hard against having their local regulations overridden. 11:44 Speaker 2 Yeah, it's a huge fight. 11:45 Speaker 1 And their internal administration clashes, too, specifically between tech advisor David Sachs and vice presidential aide Mike Davis over the sheer pace and scope of this deregulation. So what does this all mean? We have a federal framework that is explicitly hands off when it comes to copyright training data, yet we have US senators simultaneously writing letters demanding a foreign company be shut down for copyright violations. 12:10 Speaker 2 If we connect this to the bigger picture, the paradox is glaring. The US government is actively trying to foster rapid domestic innovation by removing red tape. But the resulting legal vacuum at the federal level makes domestic companies incredibly cautious. 12:25 And simultaneously, lawmakers try to erect walls against foreign models that just don't share that caution. 12:31 Walled Gardens vs. Open Source: The US-China AI Divide And that vacuum sets up the most crucial dynamic we are seeing today, the divergent playbooks of the US and China. While American companies are pivoting to enterprise tools or getting bogged down in copyright courts, Chinese companies are absolutely flooding the zone with highly accessible, incredibly capable video tools. 12:51 Speaker 2 The contrast in how the two nations are deploying this technology couldn't be starker. 12:56 Speaker 1 Let's break down the US strategy first. It is heavily focused on high quality, safe and controllable frontier models. We were talking about tools like Runway Gen. 4.5 and Google VO 3.1. They are highly polished, they integrate perfectly with professional editing workflows. But they are walled gardens, they are expensive, they can feel a bit sterile, and they are heavily restricted by IP safeguards to prevent exactly the kind of copyright backlash Cedence is facing. 13:21 Speaker 2 They are building enterprise tools designed to be sold directly into the legacy studio system. They want to be the software that Disney uses, not the software that replaces Disney. 13:30 Speaker 1 Exactly. Now look at the China strategy. It is hyper competitive, focused on rapid deployment, and incredibly cheap for the end user. Models like cling 3 point O from Kuaishu and Hilu 2.3 from minimax are consistently dominating in blind tests right now. 13:46 They really are. They excel at natural human motion, realistic physics, and generating longer coherent clips. But the absolute showstopper today is Tencent. They just released an open source model called Honey One Video 1.5. 13:59 Speaker 2 This is arguably the most disruptive piece of technology released this year. It's. 14:02 Speaker 1 Monumental Honey One video 1.5 is an 8.3 billion parameter model. But here is the crazy part. It can run locally on a single consumer grade RTX 4090 GPU which? 14:14 Speaker 2 Is wild. 14:15 Speaker 1 It is and they achieved this by using an architecture trick called Selective and Sliding Tile attention or SSTA. OK, I have. 14:21 Speaker 2 To stop you before we go for it, what exactly is a spatiotemporal token and what does pruning them actually mean for the person rendering the video? OK, let's simplify. 14:28 Speaker 1 It think of a video frame as a grid of tiles, like a mosaic. 14:31 Speaker 2 Right, exactly like a mosaic in an older model. The AI analyzes every single tile in every single frame to figure out what happens next. If you have a video of a person talking against a brick wall, the AI is constantly calculating the math for the brick wall even though it isn't moving. 14:47 That sounds really. 14:48 Speaker 1 Inefficient it is. 14:50 Speaker 2 So SSTA allows the model to selectively ignore the parts of the video that are static, the redundant tiles. It prunes them from its active memory, focusing all its computational power only on the tiles where movement is happening, like the person's mouth or hands. 15:07 So it drastically. 15:08 Speaker 1 Speeds up rendering times because it just stops overthinking the background. Exactly. 15:12 Speaker 2 And they combine that with a highly optimized architecture. You'll often hear the term mixture of experts or Moe alternatives in these models. Let's translate that. 15:20 Speaker 1 Two, what is a mixture of experts Imagine a. 15:23 Speaker 2 Massive corporation. When a specific problem arises, you don't call an all hands meeting with all 10,000 employees right? No, that would be. 15:30 Speaker 1 Chaos. Exactly. 15:32 Speaker 2 You route the problem to the specific department, the experts who handled that exact issue. A mixture of experts architecture does the same thing for the neural network. 15:43 Speaker 1 Instead of waking. 15:44 Speaker 2 Up the entire massive 8.3 billion parameter model for every single pixel. It routes the prompt to small specialized sub networks. This radically reduces the amount of memory and compute power needed to run the model. So think about what this. 15:59 Speaker 1 Means for you, the listener, if you are a creator, an indie filmmaker, or just someone trying to make a 15 minute short film right now on a tight budget, you aren't waiting around for Sora anymore. It's gone, it's done. You might subscribe to an American tool like Runway for very specific polished camera control on a few complex shots, but you are almost certainly stitching the bulk of your film together using Cling or Hilu because the generation credits are cheaper. 16:22 The Human feel more alive and the rendering times are way faster and this. 16:27 Speaker 2 Highlights a fascinating unintended consequence of global tech policy. Over the last few years. the US placed heavy export controls on advanced AI chips, specifically trying to slow Chinese advancement by cutting off their supply of raw compute power. But it didn't stop. 16:42 Speaker 1 Them at all? 16:44 Speaker 2 It actually forced them to adapt because Chinese developers couldn't rely on brute forcing their models with massive clusters of the absolute newest, most powerful chips. They had to innovate on software efficiency. They had to invent things like SSTA and optimize their routing architectures. 17:00 The result of those export bans is highly optimized models that are democratizing video generation for the average consumer, while the US remains focused on high end walled garden Hollywood pipelines. The US. 17:12 Speaker 1 Basically built the polished expensive studio tools and China effectively built the indie film makers camera. 17:18 Beyond Visuals: Cryptographic Authenticity and the Future of Entertainment It has been an absolute whiplash of a day. So let's recap the madness. Open a eyes. Sora is officially dead, pivoting to enterprise productivity because the inference costs are a nightmare bite. Dances, sedens 2 point O is suspended globally after an incredibly creepy voice cloning feature pushed Hollywood and regulators to the brink. 17:37 The US regulatory environment is leaving copyright issues to the courts, creating this legal ambiguity that is stalling domestic innovation. And ultimately, the US and China are driving down to totally different paths toward the future of synthetic media and looking. 17:52 Speaker 2 At this entire landscape, there is a deeper implication that goes beyond just software and copyright law. Yeah, we've spent this deep dive talking about tech giants, US Senators and legacy movie studios fighting over control of this technology. They're fighting over who owns the training data and who gets to deploy the models. 18:10 But with open source. 18:11 Speaker 1 Models like Hanyuan running locally on standard gaming rigs, the cat is already out of the bag completely and. 18:17 Speaker 2 That brings up a massive question about the future currency of the entertainment industry. Yeah, if we are entering an era where visual spectacle is totally democratized, where anyone with a laptop can generate a Marvel quality explosion or a photorealistic alien world for $0.00, then visual spectacle loses its premium value. 18:35 So what's? 18:35 Speaker 1 Left well, the future. 18:37 Speaker 2 Of film studio might not be about owning massive rendering farms or controlling distribution pipelines. Instead, the real power might shift entirely to cryptographic authenticity. Wait. 18:49 Speaker 1 Cryptographic authenticity? You mean proving a human actually made it? Precisely. 18:53 Speaker 2 When AI can generate any face, any voice, and any scene perfectly, the only scarce resource left is human truth. The future business model of Hollywood might pivot from selling visual effects to licensing verified cryptographic signatures. 19:09 That is wild. 19:10 Speaker 1 An actor won't. 19:11 Speaker 2 Just sell their performance. They will sell the blockchain verified proof that they actually participated in a project. Audiences might eventually pay a premium, not because a movie looks amazing, because everything will look amazing, but because they have mathematical proof that a real human being experienced the emotions portrayed on screen. 19:26 It turns. 19:26 Speaker 1 Actors into localized data holding companies and it completely redefines what we consider a premium movie experience. The visual tracks of the industry might be gone, but that train is hurtling and completely uncharted territory. Thank you so much for joining us on this deep dive into the US China AI race. 19:43 We'll be keeping a very close eye on this, but until next time, keep exploring and keep creating.