Canva and Uber pivot to in-house models to slash enterprise AI infrastructure costs

The shift toward financial sovereignty indicates that streaming and media enterprises are moving past the experimentation phase to focus on sustainable unit economics. By architecting internal routing gateways, companies can treat frontier models as specialized tools rather than the entire control plane, preventing margin erosion from unpredictable token usage. This trend forces a realignment in the vendor ecosystem, where providers must compete on specific outcomes rather than just raw capability. As infrastructure becomes a hybrid of rented and owned assets, the industry will prioritize platforms that offer portability and exit paths. Watch for whether major cloud providers introduce more aggressive reserved-capacity pricing to prevent further enterprise repatriation of AI workloads.

The trend toward enterprise AI infrastructure cost control is accelerating across the technology sector, with companies increasingly building internal routing layers to manage model selection and pricing. In mid-2026, OpenRouter emerged as a key intermediary for enterprises seeking to route queries across multiple AI models based on cost and performance, reflecting the broader industry shift away from single-vendor dependence. This routing approach allows companies like Canva and Uber to treat frontier models as interchangeable components rather than locked-in dependencies, directly addressing the token economics problem highlighted in the SiliconANGLE analysis.

The business case for AI sovereignty is being reinforced by shifting licensing dynamics in adjacent markets. The New York Times signed its first AI licensing deal with Amazon in 2026, a multi-year agreement covering real-time summaries and model training rights, while simultaneously maintaining its copyright lawsuit against OpenAI and Microsoft. This dual-track approach, where publishers monetize content through paid partnerships while litigating against unauthorized use, mirrors the broader enterprise calculation around AI economics. Companies are discovering that negotiating favorable terms with specific vendors, rather than accepting default pricing, produces better unit economics. The NYT has reportedly spent $4.4 million on its copyright lawsuit in a single quarter, underscoring the real costs of the alternative path.

Technical implementation of cost-optimized AI stacks increasingly relies on retrieval-augmented generation and vector database architectures to reduce token consumption. Product managers are being advised that efficient chunking, good embeddings, and smart filtering directly control per-query costs in LLM-based systems, since each token processed requires compute and longer prompts with RAG context multiply expenses. This technical reality explains why enterprises like Canva and Uber are investing in internal model routing rather than simply sending all queries to the most expensive frontier models. The vector database and RAG stack is becoming an increasingly important part of the enterprise AI product architecture, providing both cost control and the foundation for more agentic AI workloads that can operate across multiple information sources while maintaining budget constraints. Bilibili and miHoYo are also leading this charge by developing proprietary models to reduce reliance on external vendors.