edgeir.com web signal

DeepInfra raises $107M to scale AI inference

funding inference nvidia ai-infrastructure inference funding

Key insights

  • DeepInfra processes nearly five trillion tokens per week, placing it among the highest-volume inference providers outside of major model labs.
  • Revenue tripled from January to May 2026, suggesting strong demand for cost-efficient inference alternatives to hyperscaler APIs.
  • NVIDIA and Samsung Next participating as strategic backers ties DeepInfra's growth directly to GPU hardware and device ecosystem interests.

Why this matters

Inference infrastructure is separating from model development as a distinct, fundable category, and DeepInfra's scale numbers suggest the market has arrived faster than most roadmaps assumed. For AI practitioners, a purpose-built inference cloud processing five trillion tokens weekly creates real pricing leverage against OpenAI and Anthropic's API tiers, which matters for any team running high-volume production workloads. For founders and technical leaders, NVIDIA's strategic participation signals that the GPU supply chain is actively backing third-party inference to diversify demand beyond the hyperscalers.

Summary

DeepInfra has closed a $107M Series B to expand its cloud inference platform, with 500 Global and angel investor Georges Harik leading the round. NVIDIA, Samsung Next, and Felicis Ventures joined as strategic backers, signaling that inference infrastructure between hyperscaler APIs and edge hardware is attracting serious capital. The company now processes nearly five trillion tokens per week and has tripled revenue since January 2026. That growth positions DeepInfra as a credible alternative for developers and enterprises priced out of or throttled by OpenAI, Anthropic, and Google's own inference endpoints. Essentially: (DeepInfra, 500 Global, Georges Harik) are betting that high-volume, cost-sensitive AI workloads will migrate away from hyperscaler APIs toward purpose-built inference clouds. - Five trillion tokens per week puts DeepInfra in the same throughput conversation as major model providers, not just resellers. - NVIDIA's participation is strategic, not incidental: more inference demand at scale means more GPU sales. - Revenue tripling in under five months suggests the unit economics of third-party inference are compressing faster than critics expected. The round reflects a broader market sorting where inference is becoming its own infrastructure layer, distinct from model development and distinct from raw cloud compute.

Potential risks and opportunities

Risks

  • If OpenAI, Anthropic, or Google aggressively cut inference API pricing in the next 90 days, DeepInfra's cost-advantage positioning erodes before the Series B capital is deployed.
  • NVIDIA's dual role as investor and primary hardware supplier creates a dependency risk: any GPU allocation tightening could constrain DeepInfra's capacity expansion disproportionately relative to hyperscalers.
  • Customers running sensitive workloads may face compliance friction routing data through a third-party inference cloud that lacks the SOC 2 Type II or FedRAMP certifications that AWS, Azure, and GCP already hold.

Opportunities

  • Model labs without strong inference infrastructure (Mistral, Cohere, AI21) could partner with or acquire DeepInfra to close their API reliability gap against OpenAI.
  • Enterprise software vendors building AI features (Salesforce, ServiceNow, HubSpot) gain a credible second-source inference vendor to use as pricing leverage in hyperscaler negotiations.
  • Observability and cost-management vendors targeting AI infrastructure (Weights and Biases, Helicone, Langfuse) see an expanded addressable customer base as DeepInfra onboards cost-sensitive, high-volume enterprise accounts.

What we don't know yet

  • Which specific model families account for the five trillion weekly tokens, and whether DeepInfra is licensed to serve closed models or running open-weight alternatives exclusively.
  • What gross margin profile underlies the tripled revenue, given that inference at scale is GPU-cost-intensive and pricing competition with hyperscalers is aggressive.
  • Whether Georges Harik's involvement brings distribution ties to his prior networks (Google, early AdSense) that could accelerate enterprise customer acquisition.

Originally reported by edgeir.com

Read the original article →

Original headline: DeepInfra Closes $107M Series B Led by 500 Global and Georges Harik, Processes Nearly Five Trillion Tokens Per Week