With the price of performant GPUs sky high, and their availability almost rock bottom, hyperscalers and large enterprises are getting ratty about the rising costs of dealing with AI workloads. Enter Sweden-headquartered ZeroPoint Technologies, which is making a small buzz around the industry with promises of greatly improved server capacity utilization – by trying to solve a memory problem.
The company is a spin-out from the renowned Chalmers University of Technology in Gothenburg. In April, ZeroPoint joined Intel Ignite, the global accelerator program for early-stage deep tech startups. It is not messing about on the patent front either, with 38 already awarded for its technology.
In May, with no fully commercial products yet on the market, ZeroPoint closed a €5 million Series A funding round, led by Europe deep tech fund Matterwave Ventures. The capital raised will be used to scale sales of existing products, bring additional “hardware-accelerated memory” products to market, and grow the company’s existing teams in Sweden and the US.
“[O]ur products are in demand by datacenter operators looking to overcome the mounting challenge of memory bottlenecks,” said CEO Klas Moreau. “In fact, leading hyperscalers, such as Meta and Google, are now specifically calling for hardware-based compression technologies related to CXL-connected memory.”
Such hyperscalers are potentially already in the sales pipeline being built by ZeroPoint, which gave an update on its approach at an IT Press Tour in Istanbul, attended by Blocks & Files and other press and analysts.
Typically, up to 70 percent of stored data is redundant. ZeroPoint’s products reduce that waste, we’re told, which frees up memory capacity and increases bandwidth. This is accomplished by the company’s “hardware-accelerated” compression and compaction solution. It combines lossless ultra-fast data compression with real-time data compaction and transparent memory management.
ZeroPoint claims its technology can increase memory capacity by 2-4x while also delivering up to 50 percent more performance per watt. “In combination, these two effects can reduce the total cost of ownership of datacenter servers by up to 25 percent,” said Moreau.
He further claims the technology is 1,000x faster than other compression technologies on the market, which allows the firm to compress data across the entire memory hierarchy – “all the way from cache to storage”. The technology is agnostic to data load, processor type, architectures, memory technologies, and processing node, and the company’s IP has “already been proven” on a TSMC 5nm node.
Moreau told us memory is a “bottleneck” that has developed much slower than processors, and that 30 percent of memory transactions can be avoided. He added that memory was “expensive” too as a large and increasing part of both capex and opex in datacenters.
The business model for ZeroPoint is based on royalty fees by licensing IP, with 24-36 month sales processes. It licenses the hardware IP core and software driver to SoC customers worldwide.
The main income is from royalty fees it charges 24-36 months after tape out and manufacture, from first contact with the customer. “This business model is industry standard, with Arm as the most successful example,” said Moreau.
From almost nothing now, the company is predicting sales of $110 million by 2029. By then it believes it will have achieved a degree of critical mass in the industry, and be ready to tackle a market worth billions of dollars, according to Moreau.
Analyst Matt Kimball, of Moor Insights & Strategy, said of the potential: “For enterprises grappling with the twin imperatives of driving performance and reducing costs, ZeroPoint provides a compelling solution.”
It’s a relatively long run-up from a customer contact to customer sales, but Moreau is adamant the company will be successful. He claimed: “Hyperscalers are paying an absolute fortune for their GPUs, and can only use half of them for their AI workloads, we meet them and they tell us that.”
