The AI Chip Rebellion: Why General Compute is Ditching Nvidia to Win the Inference War

Everyone knows the tech world is fighting a brutal war for computing power. Building artificial intelligence requires massive server farms, but the industry faces two major roadblocks. First, companies struggle to buy the right computer chips. Second, even if they get the chips, they struggle to find data centers equipped to run them. A new startup called General Compute thinks it has the answer to both problems. The company acts as an inference neocloud. This means they rent out specialized server access specifically for running AI models that are already built, rather than training new ones from scratch.

Investors are paying close attention. General Compute recently closed a seed funding round for fifteen million dollars, giving the company a valuation of sixty million dollars. FUSE VC led the round, with Caryn Venture Partners and Village Global Ventures jumping in as well. This funding helps the team chase a massive opportunity in the global inference market.

When you talk about AI hardware, most people instantly think of Nvidia GPUs. However, GPUs are not actually the best tool for every job. Training an AI model requires different math than asking that model a question. The industry calls the question-and-answering phase inference. General Compute CEO Finn Puklowski and CTO Jason Goodison recognized that relying on standard GPUs for inference wastes money and time. Instead of following the crowd, they signed a massive deal with SambaNova, an Intel-backed chip manufacturer.

SambaNova is releasing a brand new chip this year called the SN50. General Compute placed a massive three hundred million dollar order for this hardware. The founders plan to be the first cloud provider to deploy these chips to the public. The SN50 features a flexible architecture that stores context highly efficiently during inference calculations. Puklowski expects these new chips to spit out six hundred to seven hundred tokens per second. A standard GPU usually struggles to break two hundred and fifty tokens per second.

Buying fast chips only solves half the problem. You still need a physical building to plug them into. Modern Nvidia GPUs run incredibly hot and require expensive, complex water-cooling systems. You cannot just put them in a standard server rack. The SambaNova chips fix this physical bottleneck. They run on standard air cooling and require significantly less electricity.

Because the SN50 chips run cool, General Compute can install its hardware inside almost any existing data center. Puklowski is actively signing colocation deals with standard facility operators. He is also pitching cryptocurrency miners. Many crypto operations are losing money right now because mining Bitcoin costs more than the coin is worth. These miners own massive warehouses with heavy electrical hookups. General Compute can easily move its air-cooled AI servers into those empty crypto facilities.

Veteran investors see the potential. Joe Hasselmann, who invested early in the AI chipmaker Groq, recently started Evercrest Capital Partners. He chose General Compute as his fund’s first official investment. Hasselmann views the partnership between General Compute and SambaNova as a brilliant strategic move, comparing it to the highly successful alliance between CoreWeave and Nvidia.

General Compute launched its cloud offering just last week. The team already claims they run the open-source MiniMax model faster than any competitor. This speed is vital for the next phase of the internet. Right now, a standard chatbot generates text faster than a human can read. But the next wave of software relies on AI agents. These agents will talk to other agents, write software code, and analyze massive databases in the background. AI agents do not read at human speeds. They require instant data delivery. By pushing hardware limits, General Compute is building the exact engine the next generation of software needs to operate.