Infinigence Token Calls Surge 20x as Inference Overtakes Training

The shift from training to inference changes the economics of AI compute. Training requires massive, concentrated bursts of GPU time. Inference is more distributed, latency-sensitive, and runs continuously. Platforms that can manage that routing efficiently capture a growing share of spending. Infinigence said its flywheel works like this: more token calls attract more model providers, which improves routing options, which draws more developers, which drives more calls.

Competition is not standing still. Cloud hyperscalers offer their own inference services, often bundled with storage and networking. Some model providers, like OpenAI and Anthropic, sell direct access to their models. Infinigence's pitch is that developers want a layer that works across all of them without favoring one. The company said its growth rate suggests that pitch is resonating, at least among AI agent builders who need to switch models frequently.

The company did not disclose revenue or profit margins. It said it remains focused on expanding the number of supported models and hardware backends. Infinigence plans to add support for more specialized inference chips, including those from startups that target lower power consumption per token.

For now, the token volume numbers are the clearest signal. A 20x increase in six months, driven by inference overtaking training, points to a market where middleware can capture value without owning the underlying compute or the model. Whether that holds as hyperscalers tighten their bundles is the open question. Infinigence says its independence is the reason customers choose it, not a weakness.