On October 26, Unconventional AI, a startup valued at $4.5 billion, announced an undisclosed cash-and-stock deal to acquire NeuReality, an Israeli chip manufacturing company, as Reuters reported. The acquisition hands the company something its experimental chips alone cannot provide: engineers capable of turning silicon into products that can be sold to customers.

From a five-year plan to a buying spree

CEO Naveen Rao initially anticipated a timeline of three years of research and five years before launching a product. AI-assisted engineering has instead allowed Unconventional to complete its first test chip and come close to finishing a second. Rao said that “things have gone a lot faster than anticipated,” adding that the company had now “built up the research bones” and must now “build up the production bones.”

Rao has a track record in the sector: he sold Nervana to Intel for more than $400 million and MosaicML to Databricks for $1.3 billion. Unconventional’s $475 million seed round, backed by Lightspeed and Andreessen Horowitz, supports the computing vision the company laid out in its launch announcement.

What NeuReality was built to fix

NeuReality was founded in 2019 by Moshe Tanach, Zvika Shmueli and Yossi Kasus. The company has spent years working to lower the cost of transmitting data and managing jobs during AI inference — the stage at which trained models generate answers for users. Its NR1 architecture employs a network engine alongside a hardware hypervisor, a component that manages and isolates computing jobs, to perform work that would otherwise be handled by host CPUs.

That experience is what Unconventional hopes to gain from the acquisition. About 40 Israeli employees will join the company, although some sales, marketing and finance staff will lose their jobs, Globes reported. Both companies’ boards have approved the deal, which is expected to close in the fourth quarter of 2026, according to NeuReality’s statement. The expected close lands just as Unconventional’s second test chip nears completion — the handoff from the “research bones” Rao described to the “production bones” the deal is meant to supply.

The purchase price remains undisclosed, but Globes estimates it at $150 million–$200 million. That compares with NeuReality’s $220 million valuation in July 2023. The Israeli startup had raised roughly $105 million since its founding.

The 1,000x promise against measured results

Unconventional believes that using analogue circuits that replicate biological systems may improve the efficiency of AI computing by a factor of 1,000 over standard designs. It is an ambitious goal — and the company has yet to deliver proof of these advantages in a commercially viable product.

NeuReality has more tangible results, although they come with limitations. In its 2024 trials, the company reported average cost-per-query savings of 90% using NR1-S systems, built on the NR1 architecture and paired with Qualcomm Cloud AI 100 accelerators, compared with selected CPU-hosted Nvidia H100 and L40S systems. However, NeuReality conducted the tests itself, using different hardware configurations. The reported cost savings have not been independently verified and should not be confused with energy savings.

For a broader industry comparison, September’s MLPerf v6.1 results, published by MLCommons, the industry group that maintains the benchmark, showed a 5.7-fold improvement over the previous year for the best per-accelerator DeepSeek-R1 result. That finding shows how quickly inference performance is advancing, but it does not validate Unconventional’s 1,000x target or NeuReality’s reported cost savings. Benchmarks like MLPerf are where vendors’ own numbers meet standardized, comparable measurement — the kind of outside validation Unconventional’s 1,000x figure has yet to face.

Why the race matters for Nvidia, clouds and power bills

TrendForce expects ASIC-based systems — chips designed for specific workloads rather than general-purpose GPUs — to account for 27.8% of AI server shipments in 2026 and nearly 40% by 2030. Agentic AI is also increasing CPU demand: a separate TrendForce forecast projects shipments could rise more than 40% in 2027.

The growing demand for AI computing is also putting pressure on electricity supplies. The IEA projects data-center electricity consumption to rise from 485 TWh in 2025 to 950 TWh by 2030 — nearly doubling within five years. That pressure is helping attract investment into Nvidia alternatives promising more efficient inference. One example is rival chipmaker Etched, which reached a $20.3 billion valuation, as Cryptopolitan reported.

But investor enthusiasm does not guarantee better performance or lower energy costs. For Unconventional, the real challenge is turning its prototype chips into commercial products — and independently demonstrating that its promised efficiency gains hold up in practice.