AI Video & Creative Generation 2026: From Model Race to Commercial Loop

Our core thesis is unchanged: AI video is moving from a parameter race to a revenue race, and commercial-loop capability is the valuation divide. What changed in mid-to-late July is that three sets of events turned this from inference into observable fact. First, the death of Sora: OpenAI shut the consumer experience on April 26 and retires the API on September 24 with no replacement listed; reports put daily operating costs near $1M against ~$2.1M in lifetime in-app purchases. Second, Kling's closing: on July 2 Kuaishou announced a completed ~$3B round at $18B post-money — roughly 36x March ARR of ~$500M — showing capital still pays up for revenue-generating video models, but with pricing discipline restored versus the rumored $20B. Third, the double squeeze: earnings week pushed Big Tech's combined 2026 capex guidance toward $500B, while Moonshot put the full 2.8T-parameter Kimi K3 weights on Hugging Face for free. Rigid compute costs above, free marginal model capability below — the middle layer is squeezed from both ends. The report uses a neutral value-chain framework and is not investment advice.

Data through August 1, 2026. Period highlights: (1) On July 2 Kuaishou disclosed Kling AI's completed ~$3B round at $18B post-money, with Tencent, Alibaba Cloud and Baidu co-investing and an IPO required by October 30, 2031; Q1 filings show March ARR near $500M. (2) OpenAI's Sora consumer experience closed April 26 and the API retires September 24, reportedly because costs far exceeded revenue. (3) On July 26 Moonshot released the full Kimi K3 weights (2.8T parameters, 1M context) — the largest open-weight model yet. (4) In the July 22-29 earnings week, Alphabet raised 2026 capex guidance to $195-205B, Microsoft's Azure AI hit a $37B run rate, and Meta guided $130-145B while profit fell and the stock slid. (5) Apple sued OpenAI on July 10 over hardware trade secrets; OpenAI denied merit on July 14. (6) The EU's Digital Omnibus defers high-risk duties, but AI-content labeling still applies from August 2. (7) Anthropic reportedly began IPO investor meetings in mid-July at a $965B valuation; Zhipu is up ~13x since its January listing. The remainder follows the original six-chapter framework.

Six hard facts frame this update. One: Kling AI closed ~$3B at $18B post-money — about 36x its ~$500M March ARR, above Anthropic's ~20x public anchor and below the rumored $20B, pricing enthusiasm and discipline in a single deal. Two: OpenAI exited consumer video, with Sora's consumer surface closed April 26 and the API retiring September 24, reportedly deeply loss-making. Three: the full Kimi K3 weights (2.8T parameters, ~1.4TB) went public on July 26, the largest open-weight release to date, cutting production costs across creative tooling. Four: earnings week confirmed rigid compute costs — combined 2026 capex guidance from the big three near the $500B level — while the market punished spending without visible returns; xAI's cleared $1B+ APR Energy deal writes energy into the value chain. Five: the EU deferred high-risk duties but AI-content labeling applies on schedule from August 2, making synthetic-content marking hard compliance for exports. Six: public markets (Zhipu up ~13x, MiniMax listed, Anthropic reportedly in investor meetings) now anchor private AI pricing.

The past three weeks played out this report's central claim: the AI video contest is now about revenue, and for the first time the cost side has hard numbers too.

Confirmed (Kuaishou Q2 results; National Business Daily and Caixin, Aug 19-20): Kling AI booked over RMB 850 million in quarterly revenue, up more than 200% year on year, with over 100 million global users and nearly 50,000 enterprise customers. Annualized, that is roughly $470 million, consistent with the ~$500 million March ARR cited in this report, so the run rate is holding. The bill came due elsewhere. Kuaishou's adjusted net profit fell 30.3% to RMB 3.9 billion, as management effectively traded current earnings for Kling's compute and R&D. The revenue is real. So is the profit toll.

Confirmed (TechCrunch, ContentGrip, Aug 17): Higgsfield raised a $400 million Series B at a $5.4 billion valuation, quadrupling in eight months, with most of its revenue now coming from enterprise customers. Capital is chasing verifiable business revenue rather than benchmark scores, a mirror image of Sora, which is exiting for failing to cover its costs, its API due to shut down on September 24.

Verdict: the thesis holds, with one amendment. The race no longer ends at revenue scale; it ends at unit economics. Whether Kling can turn its revenue curve into a profit curve before listing is the question its next earnings report must answer.

Reported (Bloomberg, September 8, 2026): Runway's annual recurring revenue reached $200 million, with growth coming from enterprise demand for its video generation tools. The Bloomberg article sits behind a paywall; this section records the figure as stated in its headline and in same-day accounts from Investing.com and PYMNTS. Alongside Kling's nearly $500 million ARR as of March, already recorded in this report, the figure gives a comparable revenue sample from each side of the Pacific.

Reported (SMZDM, September 8, 2026, citing a MiniMax open platform announcement dated September 3): MiniMax H3 Max generates a 5-second, 768P clip with audio in under 3 seconds, so generation takes less time than the clip runs. A third-party model page, AI-bot.cn, puts its throughput at about 35 times that of H3. The same review lists MiniMax H3 API pricing per second: from 0.1 yuan at 768P and 0.8 yuan at 2K, so a 5-second 768P clip costs about 0.5 yuan. The 0.1 yuan rate applies to H3, not H3 Max; the third-party page lists H3 Max at $2.40 per minute at 768P, about $0.04 per second.

Reported (SMZDM, September 8, 2026): The same September pricing review compares per-second API prices among Chinese vendors. Alibaba Wan 3.0 is officially priced at 0.3, 0.6 and 1.2 yuan per second for 480P, 720P and 1080P, with 30-second single-shot output; that pricing was published at its August public test. MiniMax H3 is at 0.1 yuan (768P) and 0.8 yuan (2K) per second. On that basis a 5-second clip at roughly 720P-class resolution costs about 0.5 yuan on MiniMax H3 and 3 yuan on Wan 3.0, a gap of about 6 times. Resolution and generation latency are now the main pricing variables.

Impact on this report's conclusions: These items reinforce the central thesis that the sector has moved from a parameter race to a revenue race. Runway's $200 million ARR alongside Kling's nearly $500 million ARR shows that leading video model companies can report verifiable revenue, and OpenAI's exit from consumer video has not slowed monetization across the sector. One correction follows: MiniMax placed "a 5-second video in under 3 seconds" and 0.1 yuan per second in the same week, which signals that latency and unit price now sit next to quality and clip length as competitive variables. The report's product comparison framework, built around duration and consistency, should add real-time performance and cost per second as dimensions.

Verified (OpenAI Help Center, September 24, 2026; corroborated by advance reports from NewsBytes and other outlets on September 21): The Sora API stopped serving requests on September 24 as scheduled, closing the programmatic route for developers to generate video with Sora. OpenAI's help center carries a discontinuation page that directs users to export their material. From the consumer app on April 26 to the API on September 24, OpenAI took five months to withdraw from video generation entirely.

Reported (Bloomberg, September 24, 2026): AI video company Higgsfield says it expects to reach $1 billion in sales over the next 12 months, an annualized run rate of roughly $1 billion. The figure is the company's own forward-looking claim and is not audited; on September 25 the advertising trade publication AdExchanger ran a piece titled "Higgsfield's $1 Billion Run Rate Comes With An Asterisk." Even with that caveat, the claimed scale sits in the same order of magnitude as the near-$500 million ARR Kling disclosed in March.

Reported (Semafor, September 17, 2026): Pika released a tool that automatically decides which underlying model generates each element of a video. Users enter basic details and receive an editable shot-by-shot list, which cuts the prompting work. Pika also released a set of apps, including one for relighting and one that swaps characters or outfits. The approach adds an orchestration layer above the models and matches the emphasis on creative control seen in Kling 3.0 and Seedance 2.5.

Impact on this report's conclusions: These developments reinforce the core thesis that the sector has moved from a parameter race to a revenue race. The three items map to exit, revenue, and product. OpenAI's API shutdown shows a general-purpose model without a revenue loop does not survive. Higgsfield's $1 billion claim, whether or not it is met, shows revenue has become the central metric in how independent startups present themselves. Pika's move to model orchestration shows differentiation at the application layer is shifting from model quality to workflow. One correction: the revenue race is no longer confined to platform-backed players such as Kling. The set of companies this report tracks should include independent startups, and company-stated run-rate figures should be held to independent verification.

Reported (Tech Times, September 30, 2026; Bloomberg also reported the launch on September 28): Kuaishou's Kling AI announced Kling 4.0 on September 28. What is live now is Kling 4.0 Flash, in closed beta for Ultra annual subscribers only, producing 3- to 20-second clips at 720p in 8-bit SDR. The full model is expected in October, with no specific date given, and will support clips up to 30 seconds, up to 4K resolution, up to 10 keyframe reference images and up to 15 multimodal inputs at once.

Reported (Adweek, September 30, 2026): Runway launched Runway Ads inside its web app, a video agent for advertising that can recreate existing ads, generate variations and localize them for different markets, publish human-approved versions to Meta, Google and TikTok, and pull performance data back from those platforms to inform the next round of creative. Lovable, an AI development platform, is the launch partner.

Impact on this report's view: this reinforces the view that the ability to close the commercial loop separates valuations. Kling is giving its new model first to its highest-paying annual subscribers, so new capability goes to heavy paying users first; Runway is pushing the product boundary from "generate a video" to "generate an ad and bring back its performance data," plugging directly into advertising budgets. Thresholds to watch: whether the full Kling 4.0 opens on schedule in October and whether access extends beyond Ultra annual subscribers; and whether Runway Ads produces a verifiable revenue figure driven by customer campaigns.