General Compute has secured a committed debt facility of up to $400 million from Upper90 Capital Management to expand its cloud infrastructure for artificial intelligence inference workloads.
The financing begins with an initial commitment of $100 million and can increase in line with customer demand. General Compute plans to use the capital to purchase and deploy specialized AI chips, expand its computing capacity, and build one of the largest cloud platforms focused specifically on inference.
Upper90 is also an equity investor in General Compute, aligning its ownership interest with the debt financing supporting the company’s growth.
General Compute operates what is commonly described as a neocloud, a specialized cloud infrastructure provider built around high-performance computing workloads rather than general-purpose enterprise applications.
The company focuses on inference, the process through which a trained AI model receives a request and generates an output for an end user. Inference powers activities such as chatbot responses, AI agents, image generation, software development assistants and real-time enterprise applications.
General Compute said its infrastructure can deliver AI inference up to 16 times faster than standard graphics processing unit cloud platforms. The company also claims its systems can produce the first token of a response up to 7 times faster and achieve an output throughput of up to 1,000 tokens per second.
The platform is built using specialized inference chips developed by SambaNova Systems, including its SN40 and SN50 processors.
Unlike many AI cloud providers that rely primarily on GPUs, General Compute is developing its infrastructure around application-specific integrated circuits (ASICs).
These processors are designed for more narrowly defined computing workloads and can provide performance or efficiency advantages when used for the applications they were built to handle.
General Compute believes inference represents an area where specialized silicon can outperform general-purpose GPUs in speed, power consumption, and deployment flexibility.
The company does not maintain a large existing GPU allocation, allowing it to select processors based on inference performance without needing to protect investments already made in a particular supplier’s hardware.
General Compute has secured more than $300 million of price-protected chip supply. The company expects to become the first neocloud provider to deploy specialized ASIC infrastructure at large scale for commercial AI inference.
The debt facility from Upper90 will allow General Compute to finance the equipment as demand grows, rather than funding the entire expansion with equity.
Debt financing can enable an infrastructure company to purchase revenue-generating assets while reducing dilution for founders and existing shareholders.
The facility’s structure also allows General Compute to increase borrowing as additional customers and workloads support the need for more capacity.
General Compute previously raised $15 million in seed and pre-seed funding from investors including FUSE VC, Village Global, Carya Venture Partners, NZVC, Matterscale Ventures, Mana Ventures, and Upper90.
The company is expanding as AI developers increasingly shift their attention from model training to the cost and performance of operating models in production.
Training an advanced AI system requires substantial computing resources, but the model may be trained only periodically. Inference occurs every time a user, application or AI agent sends the model a request.
As more organizations deploy AI products and the number of user interactions increases, inference can become a much larger and more recurring source of computing demand.
General Compute cited a Goldman Sachs forecast estimating that global token consumption could increase by 24 times over the next three and a half years.
Tokens represent the units of text processed or generated by AI models. Higher token consumption generally means greater demand for computing infrastructure capable of delivering model responses quickly and economically.
Many of the first AI neocloud providers were built to address shortages of GPUs used for model training. General Compute believes those platforms may be less efficient for inference, particularly when serving large models that must respond to users in real time.
The company is positioning its infrastructure around what it calls premium tokens, which refer to outputs generated by large, frontier-level AI models at high speeds.
General Compute said its platform supports models from providers including OpenAI, DeepSeek and MiniMax.
Customers can connect through application programming interfaces designed to make the transition from another inference provider relatively simple.
The company claims developers can switch selected workloads to its platform in less than 30 seconds, although the practical migration time will depend on each customer’s systems and integration requirements.
General Compute’s service is also compatible with the Model Context Protocol, enabling AI agents and applications to connect with external services and tools through a standardized interface.
The company describes the platform as agentic-first because it is designed for workloads generated not only by human users but also by autonomous software agents.
AI agents may generate substantially more inference activity than conventional consumer applications because they can perform multi-step tasks, call models repeatedly and interact with several external systems during a single workflow.
As organizations deploy more of these agents, infrastructure providers will need to support larger volumes of low-latency model requests while keeping the cost per generated token manageable.
General Compute said SambaNova’s chips can provide up to six times greater power efficiency than traditional GPUs for the inference workloads targeted by its platform.
The company’s systems operate at approximately 20 kilowatts per rack, compared with more than 120 kilowatts for some newer GPU configurations.
Power availability has become one of the largest constraints facing AI infrastructure developers.
High-density GPU racks may require major electrical upgrades, specialized cooling and data centers designed around significantly greater power consumption than traditional computing environments.
Some new GPU systems use liquid cooling because conventional air cooling may not remove enough heat from the equipment.
Building or retrofitting data centers for these systems can take several years, delaying the deployment of additional AI capacity.
General Compute’s infrastructure uses air-cooled hardware and does not require water-based cooling systems.
The company said this allows its equipment to be installed inside existing colocation data centers within weeks rather than waiting for newly constructed facilities.
Colocation facilities provide power, cooling, security and network connectivity for computing equipment owned or operated by outside customers.
Using established colocation capacity could help General Compute expand into additional markets more quickly, provided suitable electrical capacity and space are available.
Air cooling can also simplify maintenance and reduce the infrastructure required to support each deployment.
General Compute believes this combination of lower power consumption and faster installation will allow it to offer inference at a lower cost than GPU-based cloud providers.
Its economic advantage will depend on several factors, including chip costs, equipment utilization, customer demand, software efficiency and the performance of supported models.
The company will also need to demonstrate that specialized hardware can provide consistent compatibility as AI developers release new architectures and increasingly complex models.
General Compute is addressing this challenge through a software layer designed to make its infrastructure accessible through familiar APIs without requiring customers to manage the underlying chips.
The platform must translate model workloads into instructions optimized for SambaNova’s processors while maintaining the performance, reliability and developer experience expected from a cloud service.
The company’s financing reflects broader investor interest in alternatives to GPUs for AI computing.
GPUs remain the dominant processors for training and operating large AI models, but rising demand, high equipment costs and power constraints have created opportunities for specialized chip developers and cloud providers.
ASIC-based systems may be particularly competitive for mature, high-volume inference workloads where performance requirements are well understood and the hardware can be optimized around a narrower set of operations.
Upper90 focuses on providing a combination of credit and equity financing to technology companies between the seed and Series B stages.
The firm manages more than $1.2 billion in assets and has provided capital to companies including Crusoe, Stax Engineering, DeepInfra, Octane, PayJoy, Apptronik and Clutch.
Upper90’s model is designed to help companies finance assets and growth without relying entirely on additional equity rounds.
For General Compute, the debt facility provides access to capital that can scale with the company’s customer contracts and infrastructure requirements.
The company plans to deploy its specialized chips into existing data centers, increase available inference capacity and support a growing range of production AI applications.
General Compute is betting that the next major AI infrastructure bottleneck will not be the ability to train models, but the ability to operate them quickly, efficiently and affordably for millions of users and software agents.
KEY QUOTES:
“Upper90 has been early to the AI infrastructure opportunity while remaining selective about where we deploy capital. We focus on differentiated approaches that make compute cheaper, faster and more accessible for customers. General Compute’s partnership with SambaNova gives it access to inference silicon that is up to six times more power efficient than traditional GPUs. Our role is to ensure the company has the capital it needs to scale alongside rapidly growing demand.”
Billy Libby, Co-Founder and CEO of Upper90
“We are the only neocloud that can serve premium tokens: frontier-level intelligence on the largest models, served fast. That is where the workloads are moving. Because we are new, we are not handcuffed to a single chip supplier. This facility lets us deploy the fastest inference silicon into existing data centers and meet that demand now.”
Finn Puklowski, Co-Founder and CEO of General Compute
