Most people who use these models never think about the economics of how they got here, but the economics shape almost everything about which models exist, who builds them, and what the end-user experience feels like. We will spend a chapter on this because it is essential context for understanding why the field looks the way it does.
The cost of training a frontier model
Let us start with concrete numbers, with the caveat that most of these are estimates because labs guard their training cost data carefully.
Stable Diffusion 1.5 in 2022 cost an estimated $600,000 to train. This used a cluster of 256 NVIDIA A100 GPUs running for about 150,000 GPU-hours. At the time, A100 cloud rentals were running about $4 per GPU-hour, though with reserved capacity and bulk discounts the effective cost was lower.
By 2023, SDXL pushed costs higher, probably $2 to $3 million for the full training run, because the model was bigger and the training data was more carefully curated. Stable Diffusion 3, released in 2024, used a multimodal diffusion transformer architecture and is believed to have cost $5 to $10 million to train, with another large fraction of that for the smaller distilled variants.
FLUX, released in August 2024, is a 12-billion-parameter model trained by a team at Black Forest Labs that included Robin Rombach and the original Stable Diffusion architects. The training cost has not been disclosed, but industry estimates put it in the $10 to $30 million range based on the model size and the compute that would be required. The fact that this team raised $31 million in seed funding, then $300 million in a Series B in late 2025 at a $3.25 billion valuation, gives you some sense of the scale of capital being deployed.
Video models are an order of magnitude more expensive. Sora is widely estimated to have cost $100 million or more to train, based on the model size and the duration of the training runs. Veo 3, with its audio-video joint generation and 4K output, is presumably in the same range or higher. Kling, Hailuo, Wan, and Seedance are all proprietary, but the major Chinese labs (Kuaishou, MiniMax, Alibaba, ByteDance) have substantial GPU resources and have been spending aggressively. The open-source Open-Sora 2.0 project reported training a 'commercial-level video generation model for only $200,000,' which they presented (correctly) as a remarkable feat of efficiency, and that gives you a sense of how much more closed-source models must be spending to maintain their quality lead.
The cost of inference, and why it matters
Training cost is a one-time expenditure. Inference cost, the cost of actually generating an image or video for a user, is recurring, and it determines the unit economics of every product built on these models.
For images, inference is cheap. A FLUX-quality 1024x1024 image takes a few seconds on a single high-end GPU and costs a few cents in compute. This is why image generation has become a commodity in consumer products, you can give people effectively unlimited images for a $20-per-month subscription and still make money.
For video, inference is brutal. A five-second 1080p clip from a frontier video model can take several minutes of compute on multiple high-end GPUs and cost anywhere from twenty cents to several dollars in raw compute. When you see a tool like Runway charging twelve credits per second of video, or Veo 3 priced at the equivalent of a few dollars per clip, that is not pure markup, most of it is actual compute cost. This is why video tools all use credit systems instead of unlimited subscriptions, and why pricing varies so much across providers.
The inference cost difference is also why many products give you a 'fast' or 'turbo' option that runs a distilled version of the model. The distilled variant might use a quarter of the compute per generation, which means the provider can charge less or give you more generations per credit. The trade-off is some quality loss, but for many uses (especially iteration and exploration) the trade-off is worth it.
Who has raised what
To give you a sense of the capital flowing through this space, here is a rough snapshot as of early 2026. Numbers are total disclosed funding.
OpenAI has raised tens of billions of dollars in cumulative funding, mostly from Microsoft and other strategic and financial investors. Sora is one product among many. Google DeepMind is funded by Alphabet and does not raise externally. Meta's generative imagery work (Movie Gen, Emu) is funded internally.
Among independent labs: Runway has raised more than $300 million across several rounds, valued at around $1.5 billion in early 2024 and likely higher now. Stability AI raised about $100 million but went through significant turmoil in 2024 (founder Emad Mostaque resigned, the original Stable Diffusion team left, the company nearly collapsed, then was rescued by new investors). Black Forest Labs, the spinoff from the Stable Diffusion team, raised a $31 million seed in August 2024 from Andreessen Horowitz, then a $300 million Series B in December 2025 at a $3.25 billion valuation, co-led by Salesforce Ventures and AMP. Pika Labs has raised about $135 million. Luma Labs has raised about $70 million. Midjourney has famously never raised external venture capital and is bootstrapped from its own subscription revenue.
On the Chinese side, Kuaishou is a publicly traded company on the Hong Kong Stock Exchange and Kling is a division within it. Alibaba (Wan, Qwen-Image), ByteDance (Seedream, Seedance), MiniMax (Hailuo), and Tencent (Hunyuan) are all funded internally by their large parent companies or backed by major investors. The Chinese AI image and video generation lineage has access to enormous compute and engineering talent, and it shows in how rapidly their models have caught up to and in some cases surpassed Western labs.
Why open weights matter
There is one more economic distinction that matters enormously: which models are released as 'open weights' and which are kept proprietary. An open-weights model is one where the trained parameters are published publicly, so anyone can download the model and run it on their own hardware. This is what Stable Diffusion did in 2022 and what FLUX [dev] and [schnell] did in 2024. It is what Wan and Hunyuan and Qwen-Image have done. It is what Black Forest Labs continues to do for some of their models even as they sell paid API access to others.
Closed-weights models, by contrast, are only accessible through the lab's API. You cannot download Sora. You cannot download Veo. You cannot download Kling, even though it has open-weights cousins. You can only generate through the official endpoint, which means the lab controls who can use it, what content is allowed, and how much it costs.
The economic implications of this split are profound. Open-weights models become substrates that an entire ecosystem can build on. ComfyUI, custom training tools, fine-tunes, LoRAs, derivatives, integrations, none of this is possible without the underlying weights being available. Stable Diffusion's open release in 2022 is what allowed the explosion of creative tooling that followed; it is why ComfyUI exists, why Civitai (the model-sharing community) exists, why there are hundreds of thousands of fine-tuned variants for every imaginable use case. FLUX has done the same for the post-2024 generation.
Closed-weights models, on the other hand, can capture more revenue from each user because there is no alternative. Runway can charge per credit because you cannot rent the model, and Sora charged $200/month for its Pro tier on the same logic before OpenAI exited consumer video in 2026. The trade-off, from a lab's perspective, is between ecosystem leverage (which open weights provide) and direct monetization (which closed weights enable). Black Forest Labs has tried to do both, open-weights [dev] and [schnell] for the community, paid API [pro] and [max] for revenue. So has Stability AI in its Stable Diffusion 3.5 release strategy. Whether this hybrid model is sustainable is one of the open questions in the field's economics.
The platform layer: Krea, Flora, fal, Replicate
Sitting on top of all these models is a layer of platforms that aggregate access to multiple providers in a single interface. Krea, Flora, fal.ai, Replicate, and others all play this role, they negotiate API access with the model labs, build user-facing interfaces on top, and resell generation as a service. From the user's perspective, this means you can use Veo and Kling and FLUX and Nano Banana from a single dashboard without managing multiple accounts. From the platform's perspective, it means they take a margin on every generation, in exchange for handling the integration work and providing a better User Experience (UX).
This aggregation layer is now the dominant way that working creative professionals interact with the field. Almost nobody who uses these tools day-to-day goes directly to Google's Veo, or directly to Kling's web app, or any single lab's own interface. They use Krea or Flora or fal because those tools let them flow between models without friction, compare outputs, build workflows that chain multiple models together, and collaborate with teams. The economic value being captured at this layer is significant, Krea has raised meaningful venture capital and Flora raised $6.5 million in 2025 to expand its team, and it is a layer that did not really exist eighteen months ago.
There is a useful analogy here to the early days of cloud computing. In the mid-2000s, there were a small number of compute providers (AWS, Google, Microsoft) and a much larger number of platforms built on top of them that abstracted away the underlying complexity and made the compute usable. The same pattern is now playing out with generative imagery. The model labs are AWS. The Kreas and Floras of the world are Heroku and Vercel. Both layers will continue to exist, but the user experience will increasingly be shaped by the platform layer, and the model labs will increasingly compete on capability, price, and openness rather than on user interface.
The hardware underneath, and what generation costs
Every image and every second of video is a bill for GPU time, and in 2026 that bill is denominated in NVIDIA Blackwell.
A B200, Blackwell's workhorse, rents for roughly three to six dollars an hour at cost-focused clouds and up into the mid teens at hyperscalers; a GB200 superchip runs higher, roughly ten to eighteen dollars an hour. The generational point is that Blackwell cut the cost per generated frame or token materially against the previous Hopper generation, with a single B200 delivering several times an H100's inference throughput. Treat the exact multiplier as a range, not a fixed number, because it depends on the model and the benchmark.
At the API layer, prices are falling and vary by tier, but a few reference points help. Black Forest Labs lists FLUX.2 pro from about three cents per image and its small klein tier near a cent and a half, both from official pricing. On the video side, per-second pricing is the norm and it moves fast; as of mid-2026 the frontier video APIs land very roughly between ten and seventy cents per second depending on model, tier, and resolution. Take the video figures as approximate and check the vendor's current page before you build a budget on them.
The lesson is the one from the economics chapter made concrete. Your unit cost is a GPU-hour divided by how many images or seconds you can squeeze out of it, and both halves are moving. Distilled few-step models and newer silicon push the cost down; reasoning models and native audio push it back up. Reprice quarterly, not annually.