QBT moves its AI Oracle from laboratory work to manufacturer-backed ASIC testing
Source report: 2026-09-28 · Editorial analysis published: 2026-09-29
Quantum Blockchain Technologies says its first AI Oracles were trained on data from an ASIC manufacturer and showed an advantage over traditional mining in tests through August. The company also says more validation, compression and a larger dataset are required before commercial deployment.
Analysis and practical implications
This section is our analysis and illustrative calculations, separate from the source report.
The interim report marks a change in test environment
Quantum Blockchain Technologies published interim results on September 28, 2026 covering the six months ended June 30. The company says its Method C AI Oracle progressed from laboratory validation to live testing with hardware supplied by an unnamed ASIC manufacturer. A mining rig and Mining Development Kit arrived in March, after which the Milan University team built an operational test environment connected to QBT servers. This is a more relevant setting than a stand-alone simulation, although the manufacturer, machine model, test protocol and complete dataset remain undisclosed.
First models were trained on manufacturer data
According to the report, the first AI Oracles trained on the manufacturer’s own data were produced in July. Testing continued through August and QBT says results showed a consistent advantage over traditional mining across the evaluated time intervals. The release does not publish a peer-reviewed methodology, hashrate uplift, energy-normalised result or confidence interval. Operators should therefore treat the statement as a company progress update rather than proof that a commercial miner will earn a specific percentage more Bitcoin.
Method C attempts to guide nonce search
QBT describes Method C as an AI Oracle intended to improve the probability of finding useful Bitcoin hashes compared with an undirected search process. Any claim of improved search efficiency has to be assessed against the enormous output space of SHA-256 and the exact work performed by the ASIC. A valid comparison needs equal difficulty, equal energy, enough hashes and clearly defined start and stop conditions. Pool-side accepted work and wall power are more useful than a short demonstration based only on internal counters.
Hardware deployment requires model compression
The company says its team is working to improve predictive performance, compress the Oracle for hardware deployment and expand the training dataset. Compression matters because an ASIC control path has strict latency, memory and power limits. A model that performs on a server may not retain the same result when reduced and placed beside a miner. Engineers also need to measure whether communication, preprocessing and control overhead consume more time or energy than the proposed search advantage creates.
The ASIC partner remains confidential
QBT refers to an ASIC manufacturer operating under a non-disclosure agreement and does not identify the company or the development kit. That limits independent replication and prevents buyers from knowing which models might eventually be compatible. The arrangement nevertheless matters because access to manufacturer data and a development interface can expose signals that ordinary firmware cannot use. Commercial relevance will depend on whether the final method can be licensed, integrated and supported on shipping hardware rather than one protected test platform.
The report also mentions ASIC UltraBoost
QBT says it secured formal US patent pre-approval for its ASIC UltraBoost technology and created a new mining subsidiary named BlocKeeper. Patent status and corporate structure do not establish production readiness. A miner operator should look for a granted claim set, a reproducible prototype, firmware or board compatibility, thermal behaviour and a support model. The company’s own wording says further validation is required before the AI technology is ready for commercial deployment, which is the key qualification in the announcement.
A proper field trial needs a control fleet
For a convincing operating test, identical ASICs should be divided into Oracle and control groups under the same pool, firmware, temperature and power conditions. The trial should record wall energy, accepted and rejected shares, effective hashrate, downtime, restarts and every configuration change. Results need several difficulty periods or a statistically justified sample, because pool luck and short intervals can make one group appear better. Raw logs and the selection rule for evaluated intervals should be preserved for independent review.
Commercial claims must include total cost
Even if Method C improves successful work per unit of time, farms need to know the licence fee, extra compute, bandwidth, integration effort and failure behaviour. A higher gross hashrate can still reduce margin when control infrastructure or power overhead is expensive. The relevant outcome is net revenue after electricity and all software costs. Operators would also need a safe fallback if the Oracle server, network connection or model update becomes unavailable so that hashing continues within approved thermal and electrical limits.
The update is promising but still pre-commercial
The September 28 report provides a concrete sequence: manufacturer hardware in March, an operating environment, first models in July and tests through August. It also clearly identifies the remaining work of better prediction, compression and broader training data. That makes the announcement useful for tracking development without assuming a finished product. The next evidence to watch is a documented hardware benchmark, named compatibility, deployment terms and results that can be reproduced outside QBT’s controlled environment.
Source: Quantum Blockchain Technologies ↗
Mining calculator ↗
