OpenAI, MiniMax and Nvidia Set the Tone for March in AI: Sora, GPT-5.4 and the Bet on Mira Murati

March has shown that the AI market is once again shifting from isolated impressive demos to competition for real-world use cases: who will give models access to the computer, who will accelerate and cheapen inference, who will turn an assistant into a full-fledged tool, not just another chat. One of the main themes of the month was OpenAI's strategic pivot. If previously Sora looked like a symbol of generative wow-factor, now attention is increasingly moving toward practical products.

Signals about closing or restructuring Sora in its current format show that even a flashy launch doesn't guarantee longevity if the solution doesn't integrate into users' daily workflows. Against this backdrop, the release of GPT-5.4 with native computer access looks very telling.

The market increasingly values not a model that answers requests beautifully, but a system capable of opening the right window, executing a sequence of actions, checking the result, and saving a user time on a real task. Google is responding to this race from their side. Gemini 3.

1 Flash-Lite is positioned as a lighter, faster model, which sends an important signal to the entire market: victory will go not only to the most powerful, but also to the most cost-effective solutions. In parallel, Google is making embeddings multimodal, and this is no longer a marketing story, but one about a new architecture for search and recommendations. If previously text, images, interfaces, and documents often existed in separate domains, now they increasingly fall into a unified space of representations.

This means that enterprise search, RAG systems, assistants in products, and analytics tools will gain more precise understanding of mixed content. Anthropic also strengthened their pragmatic vector in March: Claude gained more freedom on the desktop. And a particularly resonant moment was Donald Knuth's reaction, whom Claude managed to truly surprise with the quality of its work.

Another strong thread of the month is connected to MiniMax and the idea of self-learning systems. The story around MiniMax M2.7 heated up interest in the question of how far one can go in using a model for its own development, evaluation, or preparation of synthetic data.

If such practices take hold, the industry will get recursive acceleration: a model helps create the next version of the model, shortening the cycle between experiment and release. But here lies the main risk. Self-learning without strict control can not only accelerate progress but also scale errors, artifacts, and false confidence.

Therefore, what matters is not the fact of the model's participation in development, but the quality of checks, filters, and external evaluation criteria. Equally telling is the Nvidia story, which is essentially betting not just on chips, but on future product power centers. Interest in Mira Murati's project reads as an attempt to stake a claim in the next cycle of AI companies ahead of time.

For Nvidia, this makes sense: infrastructure alone is no longer enough if the market is moving toward a tighter coupling between hardware, models, tools, and application interfaces. For the entire industry, this is also an important marker. Major players are no longer waiting for a new leader to emerge on their own; they're trying to participate in its formation as early as possible.

Taken together, March's events paint a very clear picture. The AI market is maturing and moving away from competition in isolated wow-demos toward a battle for speed, cost, integration, and depth of implementation into everyday work. Models are becoming more multimodal, gaining more rights within the computer, learning to help in their own development, and becoming increasingly tightly integrated with the infrastructure of large technology companies.

For users and teams, this means one thing: choosing tools now will have to be based not on announcement hype, but on how well they integrate into real processes. It is here, in the coming months, that it will be decided who is truly setting the agenda in AI.