All news

Million-token models move the moat from plumbing to data

A long ribbon of blank white paper unrolled across a dark table into the distance, held by a small cobalt glass weight

Gemini 1.5 reads a million tokens and Sora makes minute-long video. Tooling around model limits loses value; proprietary data and workflow gain it.

A single prompt can now hold about an hour of video, 11 hours of audio or more than 700,000 words. Google's Gemini 1.5 Pro, introduced on 15 February, reaches 1 million tokens of context in a private preview for selected developers and enterprise customers, up from 32,000 in the first Gemini release.

Each jump in context length removes a layer of engineering that startups have been selling, and this jump is a large one. Splitting documents into chunks, ranking them and stitching the answers together became a product category last year. When a model can read an entire contract set, codebase or filing archive at once, much of that layer becomes a feature of the model provider. Value moves toward what the provider cannot supply: proprietary data, the workflow around the answer and the customer relationship.

Google and OpenAI each showed a step change that day in what a single model can take in and put out. Gemini 1.5 Pro is a mixture-of-experts model that Google says matches its largest model, 1.0 Ultra, while using less compute, and it ships with a 128,000-token context window as standard. OpenAI previewed Sora, which generates video up to a minute long from a short text description and is open for now only to safety testers and a group of visual artists, designers and filmmakers. Video shows the same pattern from the other side. If a minute of credible footage costs a prompt, production costs in advertising, previsualisation and games fall fast, and the margin moves to whoever owns distribution and rights.

The safe assumption for any product is that context, quality and modality keep expanding, and the useful question is what the product is worth when they do. A company whose main asset is clever handling of model limits is exposed with every release. One that holds data no lab can buy, or sits inside a regulated workflow with audit trails and integrations, gains from each upgrade, because the upgrade improves its product at no cost to it. In video the near-term opening is the tooling around generation: editing, consistency of characters and scenes across shots, rights management and provenance labels for clients who need them.

A million tokens per call is only useful at a cost that fits a real workflow, and that cost is unknown until Google publishes its price tiers. Latency and accuracy at full length still need independent tests rather than launch demonstrations. In video the open question is when Sora leaves the preview and on what terms, including how OpenAI handles labelling of generated video and the licensing of training material.

Investing over a long horizon means underwriting the layer that compounds with model progress, not the one progress absorbs. Fanaura looks for AI-native companies whose moat is data, distribution and trust, and for them each frontier release is a free upgrade rather than a threat. Every release absorbs more plumbing, and none can absorb a customer's data.

For founders

Building an AI-native product on proprietary data that gets better with every model release? Send us the project.

Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.

Apply in five minutes
Subscribe

Get the news by email

One short note each weekday. No marketing, and you can unsubscribe at any time.