Summary

  • Patent 12,775,572 describes paired DPUs forecasting accelerator availability and moving data between accelerator memories; it contains 20 claims.
  • SOAR AI pitches NeuralSync as a product and licensing path, but its release gives no measured performance result and does not say the first Ohio GPU service will use the design.

The first thing a GPU wants is data. The second is more data, preferably before the expensive processor has finished waiting for the first batch. SOAR AI’s new patent takes aim at that queue.

U.S. Patent 12,775,572 B2, issued on 27 October, is titled Remote Memory Access Systems and Methods. Its cover names Aligned.co as assignee. SOAR AI’s release says the patent covers technology underlying NeuralSync, the system it hopes to develop alongside its data-centre and accelerated-cloud operations.

The patent’s design is more specific than the phrase “smarter AI infrastructure.” Each computing system pairs an accelerator with a data processing unit, or DPU. The DPU forecasts when its accelerator will be available within a forward window and makes that forecast visible to another system. A requesting DPU can use the forecasts to select a target accelerator, request data held in its memory and receive that data for the local accelerator. The grant’s 20 claims include a system claim and a method claim describing this exchange.

That is a dispatch desk for memory traffic: estimate which accelerator can take the next job, then arrange the handoff. If the prediction and transfer work well, less paid compute time could be spent waiting for data. That is the business hook. GPU count sells capacity; useful output per paid GPU-hour is what makes that capacity earn its keep.

SOAR AI says the technology is designed to reduce idle accelerator cycles and that NeuralSync could become a product or licensing opportunity. Those are now testable commercial aims, rather than a brand name floating free of a technical document. But the release reports no workload, baseline, throughput gain, utilization change, latency result, external licence or NeuralSync revenue. The patent cover lists Aligned.co as assignee; SOAR’s release connects the grant to NeuralSync.

There is a near-term clock elsewhere in the company’s plan. On 20 October, SOAR said its first 4,096 MI355X GPUs were scheduled to enter service at its Ohio campus on 31 December, with revenue under its first order expected from January 2027, subject to delivery and acceptance. The company described two four-year take-or-pay orders at US$1.17 billion in aggregate value and forecast US$175 million to US$200 million of 2027 revenue. It did not say that this first service will run NeuralSync.

That separation matters. The campus milestone can show whether the company brings contracted GPU capacity online; a NeuralSync deployment would test the patented dispatch mechanism. A useful result would put the workload, cluster size, baseline, data-wait time and useful throughput on the same page. It would also name the customer deployment or licence that turns the technology into a sale. Without those particulars, a patent tells readers what the design claims to do. It cannot tell them how many extra jobs got through the door.

Sources: USPTO patent 12,775,572 B2; SOAR AI’s 27 October NeuralSync announcement; SOAR AI’s 20 October Ohio campus update.