Not every task an AI agent performs needs to be handed off to the largest and most expensive model available. Earlier in August, Nvidia unveiled NeMo Switchyard, an open-source system designed to address the need to switch between frontier and lower-level AI models.
As the Wall Street Journal put it at the time, Switchyard is effectively a traffic controller for AI models, automatically choosing which model should handle different parts of an agent's workload. Nvidia says the technology can balance quality, latency and cost, sending routine jobs to cheaper models while reserving frontier systems for the difficult stuff.
But Nvidia isn't the only company pursuing that idea.
This week, a considerably smaller startup, aixplain, is making a credible case that it has been building around the same underlying principle for several years.
The company was founded in San Jose, California, in 2020 and is headquartered there.
Its founder, Hassan Sawaf, is steeped in the US tech industry, having previously worked at Meta, Amazon, eBay and Leidos.
In 2023, it raised $8 million led by California-based Calibrate Ventures and San Francisco-based Transform VC. Wa’ed Ventures, the venture arm backed by Saudi Aramco, participated in a subsequent round in 2024, and it now has a regional headquarters in Riyadh.
Now, in an unusual David-and-Goliath twist, while Nvidia's Switchyard remains officially experimental, aixplain is now taking its product to a highly unusual market: Syria.
This week, Aixplain announced a partnership with Syrian telecom operator Syriatel at the LEAP technology conference in Riyadh.
The company says its infrastructure is now powering AI services across Syriatel's network, which serves more than 12 million subscribers. It describes the move as the first deployment of US/Saudi AI-based infrastructure inside Syria.
The timing is significant. The US formally rescinded Syria's designation as a State Sponsor of Terrorism on August 24 2026, following broader sanctions relief that has progressively reopened the country to international business.
Aixplain pitches itself as an "Agentic OS”, or infrastructure for building and operating AI agents across different models, tools and deployment environments.
Its Syria deal is all about the hot new topic ‘AI sovereignty’ which arose when Trump abruptly put export licenses on some AI models.
Aixplain says Syriatel will be able to deploy AI while retaining control of its infrastructure and data, with particular emphasis on Arabic-language applications.
Meanwhile, Nvidia's Switchyard is designed to sit between an AI application and the available models, deciding where individual requests should go. Nvidia says developers can optimise routing around cost, latency, or quality, and use combinations of its own models, open-source models, and proprietary systems from other providers.
The results Nvidia is publishing are striking. LangChain reported a 74% reduction in model costs in one test by allowing only 7% of calls to reach a frontier model, albeit at a six-percentage-point reduction in accuracy. Ramp reported cutting costs by 58% while reducing runtime by 33%, while Cognition achieved near-frontier coding performance at 28% lower mean cost.
Yet the underlying idea of model routing predates Nvidia's announcement by some distance.
In November 2022, aixplain's release notes introduced a product called AutoMode, describing it as an ensemble model that would route individual inputs to the system considered optimal according to a chosen quality measure.
Initially, the technology was limited to areas such as speech recognition and machine translation, rather than to today's general-purpose LLMs. But the basic idea was there: don’t choose one AI model for everything but, instead, dynamically decide which model should process each job.
By 2025, aixplain was explicitly describing "Dynamic Model Auto-Routing" for LLMs. The company said its infrastructure could automatically switch between models when performance deteriorated or availability changed, without developers having to rewrite their applications.
Its current Agentic OS documentation goes further.
aixplain says its Asset Serving layer executes and routes calls across models and tools, while a dedicated router directs requests to an "optimal endpoint" at runtime, incorporating supplier fallbacks, retries, timeouts and rules such as automatically moving work to the lowest-cost model. Its platform currently exposes more than 170 LLMs behind a common interface.
In other words, it bears some resemblance to the Switchyard proposition.
However, there are important differences. Switchyard is open-source, while aixplain's platform is a commercial product.
Nvidia has also paired Switchyard with another piece of technology that aixplain isn't claiming to have replicated: Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model designed specifically to handle high-volume agent tasks cheaply and quickly.
And, of course, Nvidia has an enormous footprint and ecosystem.
That said, while Nvidia's own GitHub repository currently describes Switchyard as "pre-alpha" aixplain, by contrast, is presenting its technology as production-ready infrastructure.
And is now attempting to prove it not in Silicon Valley, but in Syria.