Has the search for AI computing uncovered the next Cerebra?

The strong demand for computers to run AI models has only increased, but there are two major obstacles that everyone in the business must overcome: getting the right chips, and getting them into data centers where they can start making money.

General Compute, a new inference inference neocloud – a company that rents the power to create AI, based on the field where the models run and respond to users instead of being trained – has answers to questions that illuminate the direction of the AI ​​universe. Those responses helped raise a $15 million seed round at a $60 million funding round, led by FUSE VC and Carya Venture Partners and Village Global Ventures.

First, what is the right chip? Demand for GPUs has gone through the roof, but it’s becoming common wisdom that they aren’t the right chips to run AI models when trained. The part of AI where the model is generating solutions quickly has different computational requirements than training, and a new class of chips is being developed specifically. Nvidia’s $20 billion Groq transaction in December and Cerebras’ $57 billion IPO last week pointed the way.

With power over both companies, General Compute’s founders, CEO Finn Puklowski and CTO Jason Goodison, found another way. They’re turning to special chips built by SambaNova, an Intel-based manufacturer that’s focused on what’s been a bit of a downer in Silicon Valley talk.

This may change when SambaNova releases new chips this year. The architecture is highly flexible and uses large amounts of memory to store data during computation, and SambaNova claims it outperforms not only GPUs but also other specialized chips made by Groq or Cerebras. Puklowski says the new chips will generate 600 to 700 tokens per second, versus 250 tokens per second on GPUs.

General Compute has $300 million worth of SN50 chips for the company and is said to be the first neocloud to ship them.

These chips also help to solve the second big problem – where to put them – for General Compute: They are cooled by air, not water cooled, and they consume less energy, so they can be installed in existing places without new construction costs.

Puklowski is pursuing colocation – an arrangement in which General Compute places its equipment in someone else’s premises – not only with data center providers, but also with crypto miners who want to restore their infrastructure because the cost of producing bitcoin often exceeds its cost.

General Compute launched its cloud platform last week, claiming it is the fastest version of MiniMax 2.7, a powerful open source LLM.

Joe Hasselmann is an investor who went under the radar when he invested in Groq in 2021. This year, he launched a new fund, Evercrest Capital Partners, which focused on the AI ​​space, and made General Compute its first investment. Hassleman sees in SambaNova’s partnership with General Compute parallels to Coreweave’s relationship with Nvidia — and the integration of Groq’s chip manufacturing and its legacy cloud offerings.

“They need a mix of customers who are going to put their chips in high-growth areas,” Hassleman said. “Just as General Compute is betting on SambaNova, SambaNova is betting on General Compute.”

The question is what kind of computer architecture will take the biggest advantage in the future of AI. Inference Clouds is an abstract betting world of multiple brands and agents, in which no single provider controls and the speed and value of ideas become the main types of competition. Think about it $113 million Series B was upgraded to OpenRouter this week, demonstrating the company’s ability to provide customers with multiple access options to optimize their tokens.

Speed ​​is required for that calculation, cost, and capability. Puklowski wants to turn an hour’s work into a five- or ten-minute job, and create audio for customer support, which requires quick adjustments to communicate effectively, cost-effectively.

“If you use ChatGPT and it gives you 50 tokens per second, it’s still faster than we can count,” Puklowski told TechCrunch.