<div id="readability-page-1"><div><h2>Oracle Just Raised Its FY2026 Capex By $15 Billion. The TikTok Cloud Deal Explains Why.</h2>A TikTok account called Chloe vs History posts short videos of a tattooed influencer who travels back in time. She visits Ancient Rome. She films from the deck of the sinking Titanic. She alerts historical figures to incoming disasters. Within a few weeks of launch, the account crossed 80,000 followers.Chloe is not real. The account is built by Jonathan Laramy, a British creator at Majestic Studios, using ByteDance's Seedance video model and Anthropic's Claude. The character does not exist outside a generation pipeline. The history-loving girl from L.A. is a sequence of prompts.Most coverage of synthetic creators treats them as a culture story. Authenticity, identity, what the next generation of creators will look like. Those are real questions. They are not the question that matters for investors.The compute bill does not disappear when the creator is unknown. Every Chloe video is rendered somewhere. Every viewer who scrolls past her is hitting a data center. The economics of synthetic video are not decided at the prompt. They are decided at the cloud.That decision, inside the United States, increasingly belongs to a database company in Austin.<h2>How A British Creator Got Past Hollywood</h2>Two specific releases changed the production math in early 2026. ByteDance launched Seedance 2.0 on February 12, with a unified audio-video architecture that generates 15-second clips with synchronized sound in a single pass. A week earlier, Kuaishou shipped Kling 3.0 with similar capabilities, including extended duration and native audio across multiple languages and dialects.The leap is not a marginal benchmark improvement. The previous generation of models produced silent video and left creators to add audio in post. The new generation handles motion, camera planning, environmental physics, and audio together. A single creator on a laptop can now produce content that looks closer to a polished commercial than to a phone-shot vlog.The Chinese model lead is structural, not accidental. ByteDance and Kuaishou have absolute control over short-form vertical video. That format is action-dense by construction. It is exactly the training data a motion-generation model needs. Google and Meta have larger media libraries, but those libraries are tangled in copyright, creator licensing, and regulatory review. Chinese platforms own the pipeline end to end.Laramy is exactly the kind of user the new models are built for. He is one person. He uses the tools, ships the videos, watches the followers come in. The cost of production has collapsed.The cost of hosting has not.<h2>What Happened To Sora</h2>OpenAI's Sora app was the most prominent US entry in consumer video generation. It launched, peaked at around a million active users, and shut down on March 24, 2026. The cited reason was simple. The product was burning roughly $15 million per day in inference costs against $2.1 million in lifetime in-app revenue. OpenAI could not afford to host its own application.This is the part that should reframe the discussion. The Sora shutdown is not evidence that AI video is unprofitable. It is evidence that the model layer is commoditizing faster than the compute layer is getting cheaper. The most expensive part of the stack is the compute. The compute lives in a data center. The data center belongs to someone.OpenAI's compute capacity sits on Microsoft Azure and, increasingly, on Oracle Cloud Infrastructure through the Stargate buildout. When Sora shut down, the question it raised did not go away. If OpenAI, with the most aggressive consumer rollout and the deepest compute partnerships in the industry, could not make the unit economics work, the next question is who can.The current answer is the platforms whose business model already monetizes attention. TikTok is the largest of them in the United States. And TikTok runs on Oracle Cloud.<h2>Project Texas Just Became An Equity Position</h2>The TikTok USDS Joint Venture closed on January 22, 2026, concluding a six-year geopolitical dispute that had nearly produced a US ban. The new ownership structure is 80.1% US and international investors, 19.9% ByteDance. Oracle holds 15% of the equity. Silver Lake holds 15%. Abu Dhabi's MGX holds 15%. ByteDance keeps a non-controlling stake to comply with federal law.Oracle did not arrive in this deal by accident. TikTok's US data has been hosted on Oracle Cloud Infrastructure since the Project Texas arrangement was put in place years earlier. The equity stake formalizes a relationship that was already operationally embedded on the cloud side. According to Oracle's filings, TikTok contributed roughly $800 million, around 5%, of OCI's fiscal 2025 revenue. That share is the floor, not the ceiling.Here is the structural piece. As Chinese video models like Seedance and Kling get integrated into TikTok's creator pipeline, the compute load on Oracle Cloud Infrastructure scales with the format, not the user count. Every audio-video generation pass is a multi-second rendering job on a dense GPU cluster. The denominator that matters is hours of synthetic video watched, not users on the app.The AI Pyramid frame applies cleanly here. Applications commoditize. Models commoditize. The layer that captures the bill is the data infrastructure underneath. Oracle owns the part of that layer that maps directly onto the largest US short-video distribution platform, and now owns 15% of the equity sitting on top of it.<h2>The Capex Number Wall Street Missed</h2>In September, Oracle guided FY2026 capital expenditures to approximately $35 billion. By the end of Q2 FY2026, the company had revised that figure to roughly $50 billion. A $15 billion upward revision in under six months is not the spending pattern of a mature database vendor.Oracle's remaining performance obligations crossed $455 billion at the end of Q1 FY2026 and pushed past half a trillion dollars by Q2. The Solstice supercomputer at Argonne National Laboratory, anchored on 100,000 Blackwell GPUs, is the most visible single project. The broader buildout is what the capex line funds. Multicloud database revenue is up over 1,500% year over year. Stargate participation alone implies sustained multi-year capacity demand.The market has priced some of this in. Oracle's stock has compressed since the late 2025 peak on broader concerns about hyperscaler capex sustainability, on tariff-cycle fears, and on the routine skepticism Larry Ellison has met for two decades. The compression has not changed what the order book is telling you.A company writing $50 billion in capex against $500 billion of contracted forward revenue is committing capacity to demand that already exists on paper. That is different from building speculative capacity and hoping demand arrives. The Chloe accounts of the world are the marginal demand on top.<h2>What This Tells Investors</h2>The synthetic creator story is being read through the wrong lens. The interesting question is not whether AI influencers will replace human ones, or whether Chinese video models will out-compete US labs. Those are application-layer questions, and the application layer is the part of the stack that gets eaten over time.The interesting question is where the workload lives.Companies like Oracle, sitting at the bottleneck between commoditizing application layers and consumer-scale viewing demand, are positioned to capture the part of the bill that does not commoditize. Microsoft Azure benefits from the same structural pattern through its OpenAI exposure and its enterprise base. Broadcom captures it one level down through the custom silicon that goes into the racks. The thread connecting them is that the application layer can collapse in production cost, and the layer underneath still has to absorb the rendered output.What is worth watching from here is straightforward. The TikTok-attributable revenue line on OCI, which Oracle has disclosed once and will likely keep updating as the USDS entity matures. The capex cadence into FY2027. The growth rate of the multicloud database business, which sell-side analysts have consistently underestimated for two years.Wall Street has spent the last two years asking which model wins the AI video race. The Sora shutdown made clear that the race itself is the wrong question. The compute lives somewhere. The companies that own where it lives are the ones that get paid.</div></div>

甲骨文

基准数据与基础设施房地产 SCTR ETF - Pacer

2 倍做多 ORCL ETF - Defiance

云计算 ETF - GlobalX

DTCR

甲骨文将其 2026 财年的资本支出提高了 150 亿美元，这主要是由于与 TikTok 的合作，后者一直在利用甲骨文云进行数据托管。合成视频创作者的兴起，例如 TikTok 账号 Chloe vs History，正在推动对云计算资源的需求。尽管面临 OpenAI 的 Sora 应用等竞争对手的挑战，TikTok 对先进视频生成模型的整合预计将显著增加甲骨文的收入，TikTok 预计将为甲骨文 2025 财年的收入贡献约 8 亿美元

TikTok 的 AI 视频是如何推动甲骨文最大规模的云扩展的