The artificial intelligence gold rush has been about training ever-larger models—until now, perhaps.
DeepInfra, a Palo Alto startup that barely registers on LinkedIn's employee counter, just pulled in $107 million betting on a different proposition: that the real constraint in AI isn't teaching these systems what they know, but running them once they're deployed. The Series B round, announced May 4, drew backing from NVIDIA and a roster of strategic investors who seem to believe inference—the unglamorous work of executing trained models in production—is where the next chapter gets written.
It's a capital-intensive wager. DeepInfra owns hardware across eight U.S. data centers and now processes close to five trillion tokens weekly, all while operating with what LinkedIn shows as a team of 2–10 employees. If that math sounds improbable, well, that's part of the story.
Follow the Chips
The round was co-led by 500 Global and Georges Harik, an early Google engineer who's since made a name in angel investing. But it's NVIDIA's participation that carries strategic weight. DeepInfra is among the first infrastructure partners deploying NVIDIA's Blackwell chips and Vera Rubin systems, alongside the chipmaker's Dynamo inference software and NemoClaw agent framework.
Samsung Next, Supermicro, and Upper90 joined as well, with existing investors Felicis, A.Capital Ventures, Crescent Cove, and Peak6 participating. The roster reads less like a typical venture syndicate and more like a supply chain getting vertically aligned.
DeepInfra claims the NVIDIA co-design delivered cost-efficiency gains of "up to 20x," though the baseline for that comparison isn't specified. A February VentureBeat analysis cited a case study with Latitude, one of DeepInfra's customers, showing 4x lower inference costs moving from Hopper to Blackwell architecture. These aren't abstract metrics—in a market where inference costs determine whether an AI application can survive at scale, marginal improvements quickly become existential.
The Inference Thesis
DeepInfra's pitch rests on a shift in where AI compute gets consumed. Training large models still grabs headlines and burns capital. But Deloitte estimates that inference workloads could represent roughly two-thirds of all AI compute this year—a projection the company is clearly positioning to capitalize on.
The logic goes like this: agentic AI systems, the ones making 50 to 100+ model calls per task, turn inference into the system bottleneck rather than an afterthought. Suddenly, who can run models cheaply and reliably matters more than who can train them fastest.
Revenue has tripled since January, the company says, though specific figures remain undisclosed. Whether that growth reflects real traction or the kind of early adoption bump that fades when competition intensifies isn't yet clear.
The Hardware Gamble
DeepInfra's approach differs from pure software infrastructure plays. The company owns and operates its own hardware—a capital-intensive strategy that carries both risk and potential advantage. The platform currently supports 190+ open-source models through OpenAI-compatible APIs, with private deployment options spanning A100, H100, H200, B200, and B300 GPUs.
Current pricing runs from NVIDIA Nemotron-3-Super-120B at $0.10 per million input tokens and $0.50 per million output, to DeepSeek-V4-Flash at $0.14 and $0.28 per million tokens respectively. On-demand DGX B300 GPUs are listed at $4.20 per instance-hour. Dedicated B300 clusters on five-year terms drop to $1.98 per GPU-hour—advertised as 70% cheaper than cloud alternatives, though those comparisons inevitably depend on which cloud, which workload, and which contract terms.
The company holds SOC 2 and ISO 27001 certifications and claims zero data retention, checking boxes that matter for enterprise deployments where compliance and data sovereignty aren't negotiable.
A Peculiar Efficiency
Then there's the team size.
Third-party directory data can lag reality, and LinkedIn's 2–10 employee count as of late April may not capture contractors or the full picture. Still, it's an unusual profile for a company processing trillions of tokens weekly and operating multi-datacenter infrastructure. The founding trio—CEO Nikola Borisov, Yessenzhar Kanapin, and Georgios Papoutsis—previously built and scaled imo messenger to over 200 million users, so they're no strangers to doing more with less.
Venice AI, a customer quoted in the announcement, praised the performance and reliability. DeepInfra also shows up as a provider on OpenRouter with 81 models in that catalog.
Scaling both the technical platform and go-to-market motion against better-funded competitors with larger teams presents its own set of challenges.
Crowded Territory
The $107 million raise puts DeepInfra in a market where the competitive lines are still being drawn. Together AI raised $305 million in February 2025 for its acceleration cloud focused on open-source and enterprise AI. Baseten secured $150 million at a $2.15 billion valuation last September. Both companies work with NVIDIA's stack and compete for similar customer budgets.
An October CB Insights database entry shows an earlier unattributed venture round for DeepInfra with undisclosed amount—possibly a bridge between the Series A and this latest raise. Total disclosed funding now stands at $133 million: an $8 million seed in November 2023, an $18 million Series A last April, and this new round.
What the Capital Buys
DeepInfra plans to expand global compute capacity, enhance developer tooling, and support next-generation models. The company is pushing beyond its current eight U.S. data centers, though it hasn't announced specific international locations.
The round positions the startup in a market where the economics remain unsettled. If inference truly becomes the constraint as models move from research demos to always-on production systems, owning purpose-built infrastructure could prove prescient. NVIDIA's involvement suggests the chipmaker sees value in having partners who can showcase what its latest silicon can do in real workloads.
Whether a team this lean can execute on that scale—building enterprise relationships, navigating procurement cycles, and maintaining infrastructure reliability—remains the open question. DeepInfra is betting that the shift from training to inference is real, that owning the hardware stack matters, and that a small team with the right technical chops can compete against startups with ten times the headcount and funding.
The market will render its verdict soon enough. In the meantime, five trillion tokens a week keeps flowing.
