A Top Video Model's Half-Life Is 30 Days. fal Runs 600 Anyway

Sonya Huang, hosting Sequoia's *Training Data*, sits down with fal, the developer platform that lets people hit more than 600 generative media models through one API: OpenAI's Sora, Google's Veo, open-weight Kling, and a long tail behind them. The conversation is stacked like the company, from kernels up through models and workflows to who is paying. The number that hangs over the hour comes from fal's own Q2 and Q3 2025 traffic. The half-life of a top-five model on the platform was 30 days. Huang's reaction is a wince: that is a tough depreciation schedule for the people who train these things.

The guests are co-founders Gorkem Yurtseven and Burkay Gur, and Batuhan Taskaya, who leads the inference engine. Taskaya is 22. He started on compilers at 14, moved from Python tooling into the core compiler, parser, and interpreter, and became, he says, the youngest core maintainer of the language at the time. Huang asks his age as a way of explaining why the company is so hardcore about the metal. The rest of the hour is the argument that generative video is not a side quest for LLM shops. It is a different optimization problem, a different marketplace, and, they keep saying, no longer a toy.

Huang's first question is why so few companies are focused on generative video when video already eats more than 80 percent of internet bandwidth. Yurtseven's answer is that image and then video were overlooked in this phase of AI for two reasons. There was no crisp industry use case the way vibe coding, search, or customer support organized the LLM market. Research dollars three years ago were thinner here and ramped more slowly. The models felt like toys, a fun side market that would stay small. They are now capable enough, he says, that real industry use cases are visible, and the customers look nothing like the LLM crowd.

Gur remembers the actual sequence. They were working on Python compute infrastructure when DALL-E 2 landed, then ChatGPT, then Llama. The original itch was not a grand thesis. They were curious whether they could run image models much faster. A few customers showed up and grew so fast the team said, what the heck is going on, and decided to double down. Around them, the AGI story was soaking up dollars and talent. They had a niche that was growing quickly and, Gur says, they thought: don't tell anyone. They changed the website copy to "generative media platform." Two or three months later, Sora was announced.

Karpathy, Gur notes, recently tweeted that people are visual animals and produce far more video than walls of text. Gur does a lot of learning in ChatGPT, through text. If a concept could be compressed into a 15-second video instead of 10,000 characters, the form factor would win. Quality is the gate. Yurtseven says the education market is almost untouched by video generation, waiting on quality and predictability. They sent Huang a generative-video Bible app; she found it a much better way to learn some of the lessons than reading the original. Capture attention where people already are, she says. They are at the beginning.

When fal started serving image models, Taskaya's team refused to put their eggs in one architecture. There were three image models then; upscalers and video were obviously coming. They built a tracing compiler that watches a model run, finds patterns that fit their templated kernels, and swaps in specialized kernels at runtime. A ten-person performance team writes kernels that are about 95 percent of the way there but stay general enough to reuse. The abstractions are meant to be mathematically sound, so output quality does not drop a frame, which is a hard bar in media. They claim the number-one spot on the performance benchmarks that matter to them.

The constraint is not the same as in language. An autoregressive LLM's bottleneck is how fast you can move giant weights from memory into SRAM while predicting the next token. A 600-billion-parameter model does attention over a handful of previous tokens. A diffusion video model denoises thousands or tens of thousands of tokens at once, saturating compute bandwidth rather than memory bandwidth. Scale that to a 600-billion dense video model and a clip of 100,000 tokens, denoised 50 times, with attention over all of them at every step, and the bill becomes insane. At the batch sizes people actually run, LLMs stay memory-bound, which is why the industry pours energy into speculative decoding. Video is trying to use every watt of the GPU.

Yurtseven puts numbers on it. Call a 200-token SOTA LLM request one unit of teraflops. One image is about 100x. A five-second clip at 24 frames per second is about 120 frames, another 100x, so 10,000x for a low-standard-definition video. 4K is another 10x. Parameter counts, for now, go the other way. Image models often fit on a single GPU. Open-source video models sit around 30 billion parameters, against rumors of GPT-4 in the trillions. That will not last: they hear rumors about Veo and Sora growing, which will force more distributed serving. Huang asks what it takes to stay on top of the leaderboard when other smart teams exist. Taskaya's answer is focus. From Nvidia to other inference shops, everyone is obsessed with one more token per second on DeepSeek. fal is on a different lane. Benchmark last year's fal engine against the latest PyTorch and they underperform, because Torch caught up. A lead of three to six months is the most you can keep. Hardware moves too. They adapted to B200s earlier than anyone else and ran video models faster all year because of that obsession.

Yurtseven says people under-appreciate the other half. A foundation lab has to run its own model, maybe a couple of versions. fal has to run 600, and run each one better than the lab that trained it. GPU kernels are not enough. The cluster has to load and unload models at the right times, route traffic to GPUs that already have a warm cache, pick which chips run which models, and chase customer traffic that never sits still. Taskaya calls the result distributed supercomputing, a name he says people do not like. A month before the taping they were in 28 data centers; by the conversation, about 35. The machines are heterogeneous, with different specs and networking, and they have to schedule as if it were a homogeneous hyperscaler cluster. It is not. They spent three years building their own orchestrator and even their own CDN, putting racks in carrier facilities to tap scarce GPUs wherever they sit.

Image generation used to take 15 seconds; video takes 20 to 30. Counterintuitively, longer jobs are easier on the infrastructure because you are not shaving 50 milliseconds. The new pain is real-time video: 24 frames per second streamed off the GPUs as someone types. They spent a year cutting latency on speech-to-speech with a globally distributed fleet, routing to the closest GPU, and they are applying the same tricks to interactive video. Huang's instinct is that hyperscalers should own giant clusters. Taskaya splits the answer. Inference optimization is about three years old; the clouds have no backlog of expertise. On infrastructure, GPU demand no longer fits the growth curves hyperscalers budgeted for. Even Microsoft, he says, has started buying from the new class of clouds. The tables have turned a bit.

Gur adds the price knife. The same compute through a hyperscaler can be two or three times more expensive. Public clouds carry operational cost, better uptime, better SLAs, and a cloud margin the market expects them to keep. A three-year-old neocloud is private and can compete on infinite demand against limited capacity, which pushes prices down. Hyperscalers can keep list prices high, fill the boxes, and take better economics. Perfect price competition, Huang says, sits on the other side of that split.

Three years ago, investors told them omni-models would generate video, audio, image, and code, and a unimodal edge would not last. Yurtseven says the jury is still out, but it is clearer that you are better off optimizing for a type of output. The best upscaling model only upscales. The best image-editing model is not the best text-to-image model. Special tasks need their own weights, even inside a family, which is why the ecosystem is diverse and why fal has a business. Taskaya notes that even language models have personas — Claude's code is not GPT-5's — and media multiplies that by category: upscaling, editing, text-to-video. You get close to 50 models active at any moment, then a long tail people keep because they like that persona.

What customers actually buy has been stable for almost two years even as names rotate. There is one expensive, best-quality video model — Veo, Kling, or Sora taking turns — and a cheaper workhorse used at higher volume. Which workhorse and which flagship depends on the app. Capabilities split further. OpenAI's image-editing model had superior text rendering, so anything with a lot of type went there. People like Kling for visual-effects workflows because that kind of data was in the set. ByteDance's Seedance is strong on detailed textures and artistic diversity. Taskaya drops the 30-day half-life number here. Yurtseven is calm about it: hopefully the labs are building on prior work, so the end result is additive.

Personal favorites diverge. Yurtseven likes Kling for video but it has no sound; for audio it is currently Veo 3 and Sora, and he wants more variety. Taskaya still uses Reve and Flux, a year on, a little nostalgic. Gur picks Midjourney, which is not on fal and has no API, because of how it navigated the space: photorealism when nobody else had it, then an artsy niche once photorealism was cracked and nobody cared. On the closed-model side, labs partner because fal collected developers who refuse to be tied to one model. Labs use the platform as distribution; the next lab sees the developers and wants in. They call it marketplace-plus-plus: infrastructure for labs and for developers, plus a marketing machine that turns every launch into co-marketing and, sometimes, exclusive day-zero access. Kling and MiniMax have given them that. Model providers want the biggest platform. Increasingly, they say, that is theirs.

Open source is livelier in video and image editing than in text, where open models sit a generation behind. Yurtseven traces it to Stability. Stable Diffusion got insane adoption; almost the same people founded Black Forest Labs knowing the ecosystem was the point. LoRAs, adapters, free marketing, stickiness: people still use Stable Diffusion because they like that world. Flux kept some models open and some closed. Later video labs copied the playbook. Gur adds the domain difference. Llama 2 spawned fine-tunes you could not tell apart. Language is general; a small weight change in vision can explode into a new style. Even a model trained on every aesthetic may miss the exact one a brand wants, so you still fine-tune. Sora 2, he says, is heavily post-trained toward social-looking content. A customer-support chatbot wants vanilla. Filmmakers and marketing teams want style and brand control. Those are different worlds.

fal's top 100 customers use 14 different models at the same time. They are often chained: text-to-image, then an upscaler, then image-to-video, or a more complicated graph. fal built a no-code workflow builder with Shopify for PMs and marketing teams who need to try and compare models; the same graph can ship behind an API. Huang's straw-man, type a prompt and get a five-minute commercial, is not how it works. Studios and serious marketing teams prefer open-source models because they can swap pieces and steer. The graphs look like giant ComfyUI boards. There is no single Photoshop workflow, Gur says, and there will not be a single generative one. A creator they like, PJ Ace, posts a new workflow almost every month because a new model arrived. Best practices will settle when professionals standardize. Nobody knows the final form yet.

A common path looks like Pixar pre-production. You lock aesthetics on an image model, generate a series of storyboard frames, then let video models interpolate between them. Storyboards used to exist because they were cheaper than shooting. They still make sense as speed. AI makes the loop stranger: a new text-to-image model drops, someone presses a button, every combination regenerates, then every video regenerates. Yurtseven has seen a rerun cost a thousand dollars. Individuals spend a ton on creator platforms. He has seen bills of half a million dollars from a single person, maybe more inside a small production studio.

Customers are not one industry. Adaptive Security, Brian Long's company, generates phishing-style training on the fly so each employee sees different content. Faith, the Bible app, ranks high on the App Store with produced stories for each passage. Canva and Adobe are wiring new models into older design tools. Ads are the biggest category: UGC-style product demos, higher-production spots like the recent Coca-Cola ad (Yurtseven notes the controversy), and programmatic ads personalized down to putting you in the spot. Huang is most excited about education, a market that matters and has rarely had a compelling business. The bottleneck was always high-quality content at learner scale. Gur thinks ChatGPT is already solving learning in the wrong form factor. To realize the models, you have to go visual: more compact, more approachable. Crack visual learning, he says, and it actually hits people.

Six months before the taping, the fal team thought Hollywood was too slow and AI-native studios would simply go around it. This summer the weather changed. They have been talking to the usual suspects. At fal's first generative media conference, Jeffrey Katzenberg, former CEO of DreamWorks, compared the moment to computer animation. Hand-drawn people revolted; the technology happened anyway. You are part of it or you are not. In the medium term, Yurtseven thinks IP holders are well positioned: technical people who actually care, the IP itself, and storytelling craft. Filmmaking will still be expensive. AI may make it a little cheaper. The combination of deeply technical filmmakers plus IP will matter for the next few years. The opposite is also true. They had little toys of Italian Brainrot characters: no owner, no IP, generated by an internet community, then toys and a Roblox game and real money. Once generation is cheap, permutations that catch the zeitgeist become new IP from nothing.

Huang asks how you avoid the infinite slop machine. Gur's bet is that taste prevails. People are drawn to what feels personal and human. Meta's Vibes landed as a slot machine of characters you have no relation to. Sora 2 made it about friends and Cameos — now you can Cameo your pets — and the reception was different. Taskaya, asked about a Hollywood-grade feature with no human filming, wants the question tightened. Editing is allowed. In less than a year he expects feature-grade shorts under 20 minutes; even today the models are close enough, and the missing ingredient is time invested. Photorealism is what everyone targets. Yurtseven thinks the first breakthrough will be animation, anime, cartoon, as far from photoreal as possible, even fantasy. Shooting photoreal is already cheap. The expensive part of movies is the non-photoreal work. *Toy Story*, *How to Train Your Dragon*, *Shrek*, *Ratatouille* work because of story, not because they look like cameras. AI, he says, will be to film what animation was: a new medium. VFX such as explosions and collapsing buildings is already nearly perfect. Facial expressions are still hard. Gymnastics, Huang jokes, they can do. Good thing there is a lot of Olympic footage, Yurtseven says.

Gur is bullish on discardable, one-time-playable hyper-casual games. Text-to-game is the natural next step after text-to-video: make the video interactive. AAA generated by a model is at least three or four years out, though LLMs already change dialogue, pre-production, and mini-games from existing IP. Getting to the promised land, Taskaya says, needs architecture, not only scale-up. Image models left pixel space for latents that fold 64 pixels into one. Video's time compression is about 4x; why not 24x? More compression would help training and inference. Google catching up fast showed that money and effort still move the needle, and smaller labs can still win on data work. Yurtseven says the research community has done the simplest thing: caption images and train. Video editing needs real data engineering. Luckily, they will run out of compute before they run out of video. Gur's earlier math still holds: 4K real-time needs another 100x or more. The best models still make you wait 40 seconds, sometimes two or three minutes, which is not acceptable in an on-demand world. Hardware getting faster is not enough. Architecture has to get better, or the timelines get long.