While the tech world continues its obsession with hoarding GPUs to train massive artificial intelligence models, a quiet bottleneck has emerged on the other side of the AI equation: inference. Generating the actual answers for end-users is becoming increasingly power-hungry, expensive, and slow.
Enter General Compute.
The new startup which pitches itself as the world’s fastest inference neocloud, today announced a massive committed debt facility of up to $400 million from hybrid investment firm Upper90 Capital Management.
The financing, which unlocks an initial $100 million and scales alongside customer demand, gives General Compute the dry powder to rapidly scale its infrastructure. But unlike the first wave of AI cloud providers, General Compute isn’t buying up Nvidia GPUs. Instead, the company is building an empire on Application-Specific Integrated Circuits (ASICs), specifically, specialized inference silicon from SambaNova.
The GPU Dilemma and the Pivot to Inference
To understand General Compute’s massive debt raise, you have to look at where the AI market is heading. Goldman Sachs recently predicted that AI token consumption will grow 24x over the next three and a half years.
While traditional GPU neoclouds were vital for alleviating the initial AI training bottleneck, they are notoriously inefficient for inference workloads. Pushing queries to end-users on top-tier GPUs is akin to using a semitruck to deliver groceries. The newest GPU racks are absolute power hogs, demanding upwards of 120 kilowatts of power. Worse, they require specialized liquid-cooling retrofits, trapping much-needed new compute capacity behind multi-year data center construction timelines.
General Compute bypasses this physical bottleneck entirely. By utilizing SambaNova’s SN40 and SN50 inference chips, the startup operates on just 20kW per rack. Because the hardware is entirely air-cooled, it requires no specialized water-cooling infrastructure.
“We are the only neocloud that can serve premium tokens: frontier-level intelligence on the largest models, served fast,” said Finn Puklowski, co-founder and CEO of General Compute. “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.”
According to the company, this allows them to deploy in existing colocation facilities in a matter of weeks, rather than years.
By the Numbers: 16x Faster and 6x More Efficient
For developers and AI companies, the pitch boils down to raw performance and cost economics. General Compute claims its infrastructure delivers results up to 16x faster than standard GPU clouds, alongside a 7x faster time-to-first-token.
Throughput is also significantly higher, boasting 8.5x the output of standard clouds and reaching speeds of 1,000 tokens per second. Because the ASICs require 6x less energy than current GPUs, General Compute can offer inference at a radically more affordable price point.
The platform is designed to be plug-and-play. Developers can switch their workloads to General Compute in under 30 seconds, running frontier models from OpenAI, DeepSeek, and MiniMax. In a nod to the growing trend of autonomous AI, the platform is also Agentic-First, meaning it is API and MCP compatible to allow AI agents to provision compute on a user’s behalf.
Financing the Infrastructure Layer
Building a cloud from scratch is a highly capital-intensive endeavor, which is why debt facilities are becoming the financing tool of choice for compute-heavy startups looking to avoid massive equity dilution.
Upper90, which manages over $1.2 billion in assets and is also an equity investor in General Compute, is effectively aligning its capital stack behind the startup’s growth. General Compute says it already has more than $300 million of secured, price-protected supply, positioning it to be the first neocloud to deploy ASIC silicon at this scale.
“Upper90 has been early to the AI infrastructure opportunity while remaining selective about where we deploy capital,” said Billy Libby, Co-Founder and CEO of Upper90. “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.”
The $400 million debt facility follows a previously closed $15 million seed and pre-seed round backed by FUSE VC, Village Global, Carya Venture Partners, NZVC, Matterscale Ventures, and Mana Ventures.
As the AI industry matures from the training phase into the deployment phase, the race to serve tokens efficiently is heating up. By heavily backing purpose-built ASICs over general-purpose GPUs, General Compute is making a $400 million bet that speed and power efficiency will be the ultimate winning metrics in the inference wars.
Companies can switch to General Compute today at https://www.generalcompute.com/.
