DataVerge, a company that operates and owns a co-location facility for AI inferencing, is expanding its Brooklyn, New York, facility with Mathpix, a company that converts documents into machine-readable text.
Mathpix will be deploying Nvidia B300s GPUs for AI training and inference at the DataVerge Brooklyn facility. Mathpix will use its own hardware at the facility instead of using distant cloud computing to cut latency to New York City-area customers.
DataVerge will supply the power density and cooling to support the Mathpix infrastructure and plans for more capacity in 2027.
The advantages of co-locating the data center in New York City allow Mathpix to service New York City-based users faster instead of uploading data to a cloud in perhaps another state. This will improve latency. And by filling the DataVerge data center with their own hardware, Mathpix can control how much power it needs, how fast it moves and it has complete control over the access.
Mathpix converts documents like PDFs, handwritten notes, equations and more into a readable text that can then be fed into AI applications and automated workflows in enterprise, research and financial markets.
"For Mathpix, AI training and inference performance are product features,” said Nico Jimenez, CEO of Mathpix. “Our customers expect near-instantaneous document conversion, which means our infrastructure needs to be close to them and built for the demands of modern AI workloads. For both fine-tuning models and real-time inference, milliseconds matter when processing user uploads, enterprise batch jobs, and API-driven workflows. DataVerge gives us the ability to deploy the B300s with the power density, connectivity, and hands-on support we need at a cost structure that makes sense."
Why it matters
AI inference is fast becoming latency sensitive. GPU scarcity is making owned hardware more attractive.
These two features are making co-locating of data centers more feasible to smaller firms to gain an edge. While Mathpix will not own the data center, they own the hardware using the local warehouse or facility from DataVerge. This gives them more flexibility while also accelerating inference in the local NYC area.
For these companies, traditional cloud AI data centers no longer can service users fast enough. Combined with energy and cost pressures, many companies are scrutinizing cloud bills more closely.
While Mathpix bears the initial cost of buying the hardware, in this case the Nvidia B300 GPUs, they will gain faster response time for users and lower ongoing costs than going to cloud computing.
For example, if Mathpix went with Amazon Web Service (AWS), they would use its hardware, its network paths and its GPU availability (which could be constrained). With DataVerge co-locating, Mathpix picked the B300s specifically, configured them how it wants and aren't competing with other tenants for capacity.
While co-locating is not a new concept, the idea of pushing compute closer to users generally to not just one city but distributed across many locations has been building for years. It looks like this is now reaching data centers as well.
