Even if the reported deal with Disney brings new cash and access to marquee characters for Sora, it won’t address OpenAI’s central challenge: the brutal economics of operating frontier AI at internet scale.
A splashy media partnership may juice brand perception and put a headline or two in the press, but not much to shift the cost curve of compute, power, and infrastructure that will define OpenAI’s destiny.
Editorial Note: The author is a volunteer at OpenAI.
Industry scuttlebutt has the deal at about $1 billion, and it comprises well-managed licensing so users can create brand-safe clips. That’s real money, and the stalwart approval of one of the world’s most valuable entertainment companies means something. But at OpenAI’s current level of spending, $1 billion is weeks of runway — not a strategic reset. The difference is calculated in orders of magnitude, not press releases.
The math behind the hype and the hard costs of AI scale
Generative AI is an infrastructure business in consumer app clothing. Training models may cost tens to hundreds of millions, but the real bill comes in at inference — every prompt, every video frame rendered, every call against a safety system. Analysts at SemiAnalysis and other research shops have repeatedly cautioned that inference expenses, which are in large part GPU time and electricity costs, can exceed subscription revenue unless usage is rationed or prices increase.
Think about the hardware: NVIDIA’s H100 chips have been quoted at $25,000–$40,000 each depending on volume and build, and modern multimodal clusters require thousands of them plus high-speed networking. Then there’s the power. The International Energy Agency has projected that global data center electricity consumption could nearly double by the middle of the decade, with AI one key contributor. That energy consumption, of course, translates directly into operating costs that no media licensing agreement can offset.
OpenAI’s revenues have grown rapidly — The Information said it had a multibillion-dollar annual run rate by the middle of the year — but so have costs. Margins are still hostage to GPU supply, unit prices, and how price-onerous OpenAI’s deal with its cloud partner is. Cash from a content deal comes in handy. It’s not transformative.
What Disney Gets and What OpenAI Will Get
The appeal for Disney is fairly obvious. It maintains a tight grip over its intellectual property, imposes guardrails for how characters are portrayed in AI-generated media, and gets to sit front and center as new kinds of fan interaction abound. Where the deal involves an equity component, Disney also gains on any future uplift without having to carry the heavy compute cost. It’s a clever way to dip a toe into the waters of AI while safeguarding a 100-year-old brand.
OpenAI wins itself some prestige, a library that lures creators to its platform, and a proof point that major rights holders like the NFL will license content to its platform. That’s important for talent pipelines and wooing other studios to sign on. But it’s strategic polish, not a P&L answer. We’ve seen comparable dynamics in previous licensing “deals” with publishers like the Associated Press and Axel Springer: important for legitimization and a game-changer for data access, not much good as a panacea for infrastructure economics.
Platform risk versus compute reality for OpenAI’s margins
OpenAI still runs almost entirely on Microsoft’s cloud for training and serving, thus its gross margins depend on multiyear pricing from a single hyperscale partner. The focus has been a double-edged sword: It helps speed model development, but makes it harder to cut costs without custom silicon or owning data center buildouts.
Sam Altman has openly threatened radical fixes, including new chip manufacturing capacity and multitrillion-dollar financing schemes reported by the Financial Times and the Wall Street Journal. Those are the right-size ideas for this problem. By contrast, there aren’t even that many nine-figure media deals, and the few we do have are tactical. They don’t secure wafers, they don’t reduce GPU unit costs, and they don’t end up locking in long-term power at steady pricing.
Competitive pressure compounds the challenge. Google, Meta, and Anthropic are all pursuing multimodal systems as they build or buy access to custom accelerators and renewable-heavy data centers. If competitors can reduce their cost per token or cost per generated minute faster than OpenAI, the company will face price wars that will hurt whoever has the weakest infrastructure position.
What would actually move the needle for OpenAI’s costs
OpenAI requires structural levers:
- Custom or near-custom silicon to break reliance on scarce cloud GPUs.
- Multiyear power arrangements tied to new builds.
- Sovereign and enterprise contracts with non-paltry minimums.
- Tighter product bundles that shoo high-cost workloads toward pricier tiers.
- Ongoing diversification beyond consumer chat into workflows where customers evidence ROI and tolerate premium pricing.
Landmark IP can be licensed to speed adoption and enhance content diversity. It can’t rewrite the cost function of generative AI. The winners will be companies that rein in their compute costs, source cheap power, and continue to ship useful models under predictable unit economics.
The Disney pact, in that light, is best read as a brand statement and content accelerant. It could also be a boon to the top line and momentum of OpenAI. It won’t rescue the company from the math of AI.
