Inference Chips Drive New Capital Models

The feverish hunt for AI compute is shifting, moving beyond the standard reliance on high-end GPUs to address the mounting costs of running models. As companies seek to optimize the price of tokens, the infrastructure landscape is evolving, marked by a significant $400 million loan from Upper90 to the AI inference startup General Compute.

Capitalizing on Specialized Silicon

General Compute, led by CEO Finn Puklowski, is leveraging this capital to construct an inference-focused neocloud using SambaNova silicon. Unlike the expansive, general-purpose infrastructure managed by giants like AWS, neoclouds are engineered specifically for AI workloads. By utilizing chips that avoid the need for water-cooling systems, the company aims to expedite deployment across data centers while focusing on power-efficient operations.

Reframing the Value of Collateral

The deal represents a milestone in financial engineering: it is believed to be the first instance where inference-specific hardware has been used as collateral for a loan. Upper90 CEO Billy Libby, a veteran of quantitative trading at Goldman Sachs, views this as a strategic expansion of a playbook his firm pioneered in 2021 when it financed GPU purchases for Crusoe.

“When we financed Nvidia GPUs as the first group to do that, the market was inefficient. We could really put together something as an early participant, and kind of get compensated for the risk.”

— Billy Libby, co-founder and CEO at Upper90

Shifting Away From Nvidia Dominance

The market is showing a clear appetite for alternatives that challenge Nvidia's current hold on the sector. With models such as Kimi’s K3 demonstrating competitive performance against established frontier labs, providers like TensorWave are exploring partnerships with hardware makers like AMD. This fragmentation of the compute market suggests that providers capable of bypassing traditional ecosystem lock-ins may secure a critical advantage in total cost of ownership.

• $400 million: The total value of the loan secured by General Compute from Upper90.
• $15 million: The amount raised by General Compute in a seed round during May.
• 16 times: The improvement in inference speed over GPU-based clouds that General Compute claims its new chips will provide.

Implications for the AI Infrastructure

For organizations, this shift suggests that the era of relying solely on generic GPU clusters is being challenged by more granular, cost-conscious hardware strategies. As specialized silicon becomes a viable asset class for financiers, businesses should prepare for a more fragmented cloud environment. This trend underscores a broader move toward inference-optimized infrastructure, forcing enterprises to balance the performance of frontier models against the operational efficiency offered by these emerging, cost-optimized alternatives.