Ayomide Daniels, founder of Alchymia AI, has unveiled the Alchymia Bar, a prototype pocket computer designed to run the company’s AI models locally. The device is intended to combine on-device computing with a projected visual interface. The company says the screenless device uses projected visuals to provide an interactive interface. If demonstrated, the approach could offer an alternative to conventional displays for local AI applications.
Alchymia says the project aims to make locally run AI more accessible in markets where computing infrastructure, electricity supply or internet connectivity can constrain deployment.
Alchymia AI holographic computer’s footprint
The device reflects an alternative approach to AI deployment: running a model on local hardware rather than sending every request to a remote service. Alchymia says its model architecture is designed to reduce computing requirements, with the aim of lowering energy use during training and inference. This strategy opens new avenues for deploying sophisticated AI directly at the point of need.
The energy demands of artificial intelligence have become a growing concern for global power grids and sustainability.Google estimated that a median text prompt in the Gemini app used 0.24 watt-hours of electricity in May 2025. Although the estimate is small for a single prompt, aggregate electricity use depends on the number of requests, their complexity, the models used and the infrastructure serving them.
More complex AI workloads can require substantially more computation than routine text generation, although the difference depends on the model, task and hardware. Any specific claim of a hundredfold or thousandfold increase should be supported by a directly relevant IEA source or technical study. Rising AI workloads could add pressure to electricity infrastructure, particularly in areas where supply is already constrained. Daniels argues that the way AI models are designed can influence their computing and energy requirements.
Alchymia says its work on AI architecture dates back to 2016, when the organisation was incorporated as Immersive Research. Daniels told TechCabal that model architecture is a major factor in determining computing requirements and energy use. This contrasts sharply with the industry trend of applying surface-level optimisations to established transformer designs.
Regent and Darkhorse: Architectures for Efficiency
Alchymia developed Regent as an alternative to conventional transformer-based language models. According to TechCabal’s report, Regent uses a Mamba-2 state-space architecture with attention mechanisms at selected layers. Alchymia says Regent is designed to maintain a fixed recurrent-state memory as context grows, with attention used selectively. This differs from standard full-attention transformers, whose attention computation can grow quadratically with sequence length during training.
Alchymia says its smaller open-source Regent models can run on local hardware, with the hardware requirements depending on the model size and configuration. Alchymia says Regent produces confidence scores during generation and can retrieve information from its structured knowledge store when confidence falls below a configured threshold. This may support verification, but the confidence scores should not be treated as proof that an answer is correct.
Alchymia is developing Darkhorse, which it describes as a model designed to allocate computation according to task complexity. According to Alchymia, Darkhorse is intended to allocate less computation to routine queries and more to complex tasks. The effect on inference costs would need to be demonstrated through benchmarks that compare energy use, latency and output quality across equivalent workloads.
Alchymia reports that its internal benchmarks show Darkhorse using approximately one-third as much training energy as comparable models and around half as much energy per token during inference. The article should identify the models used for comparison, the hardware, the benchmark tasks and the measurement methodology. Alchymia says it has secured more than US$2.3 million in non-dilutive funding over five years.
Engineering AI for the Physical World
Alchymia’s plans extend beyond language models to what Daniels describes as ‘physical AI’—systems that use AI to interact with devices and the physical environment. Physical AI generally refers to AI systems that perceive or act in the physical world, including through robots, drones and industrial machinery. Daniels aims to transcend the typical screen-based interaction with AI chatbots, bringing intelligence into the physical environment.
Alchymia aims to make its models deployable across different hardware platforms, although the extent of portability will depend on each device’s computing resources, software compatibility and control interfaces. This adaptability is key to widespread adoption, particularly in industrial settings where specialised hardware often dominates. The firm’s ambition aligns with the increasing integration of AI in robotics and automation.
The Alchymia Bar is a prototype computer intended to run Alchymia’s models locally and provide a projected visual interface. Daniels describes the prototype as approximately the size of a Samsung Galaxy smartphone. The company says it projects an interactive visual interface, although the display technology and the extent of its interaction capabilities should be explained in more technical detail.
The device runs the Darkhorse model locally, capable of handling tasks such as coding, writing, document processing, 3D workflows, and media consumption. Alchymia is targeting a commercial launch in 2027, although pricing, final specifications and availability have not been established in the article. Daniels says the device is intended to provide another way to access locally run AI, rather than replace traditional laptops.
The Holographic Interface: A New User Paradigm
The Alchymia Bar is designed to replace a conventional screen with a projected visual interface. The proposed interface could allow users to interact with visual content in three dimensions. The article should explain how the projection works and what input methods are available before describing it as a spatial-computing system.
The broader holographic display market is experiencing substantial growth, reflecting increasing interest in such interfaces. One market-research estimate values the global holographic-display market at US$4.3 billion in 2025 and forecasts growth to US$40.8 billion by 2035, representing a compound annual growth rate of 25.5%. This expansion highlights the appetite for immersive, screenless visual experiences.
If the 71.3% forecast is retained, cite the specific report and clarify how it defines ‘electro-holographic’ displays. Avoid using that market-share figure to imply that the Alchymia Bar uses the same technology. The Looking Glass Go, announced in December 2023, is an example of a compact light-field display designed to show glasses-free 3D content. It provides a useful comparison for portable visual interfaces, although its technology and functions differ from the Alchymia Bar as described here. These devices convert 2D and 3D visuals into holographic projections.
A portable 3D interface could have potential applications in design visualisation and field diagnostics, provided it can display the required data clearly and operate reliably in the intended environment. For engineers and operations professionals, interactive 3D models directly projected into their workspace could streamline complex tasks. The Alchymia Bar pushes this frontier, bringing advanced AI into a tangible, accessible form.
Addressing AI’s Energy Demands and Accessibility
The energy consumption of artificial intelligence is a critical global challenge. The International Energy Agency’s 2025 outlook projected that global data-centre electricity consumption would reach about 945 terawatt-hours (TWh) by 2030, more than double its estimated 2024 level. The IEA’s subsequent outlook put consumption at around 485 TWh in 2025 and approximately 950 TWh in 2030. If the 17% annual increase and 50% growth figure are retained, identify the exact series and whether the latter refers to AI-focused data-centre electricity consumption, capacity or the number of facilities.
While much discussion centres on training advanced AI models, day-to-day usage, or inference, accounts for a substantial 80% to 90% of total energy demand. The IEA projects that electricity consumption from accelerated servers, largely associated with AI workloads, will grow by around 30% annually in its base case. These servers account for almost half of the projected net increase in data-centre electricity consumption between 2024 and 2030.
Alchymia says its model architectures are designed to reduce computing requirements. Whether they deliver lower energy use in practice will depend on benchmark results and the hardware used. Ayomide Daniels emphasised the company’s mission: to build foundational models suitable for regions across Africa and Asia. These markets often lack access to the vast data centres and substantial financial resources typically required for high-end AI deployment.
By enabling advanced AI to run on edge devices, Alchymia AI makes sophisticated computational power available in environments where traditional cloud-based AI is impractical. This focus on decentralised, energy-lean intelligence aligns with broader goals of technological inclusivity and sustainable industrial development. Local deployment could reduce dependence on internet connectivity and recurring cloud-inference fees, although hardware costs, electricity supply, licensing and technical support will still affect accessibility.
Broader Implications for Industrial Computing
The Alchymia Bar could have applications in industrial computing if its local AI capabilities, display and hardware performance meet the needs of field and factory environments. Local AI can support applications such as equipment troubleshooting, fault detection and predictive maintenance where internet connectivity is limited. Whether the Alchymia Bar can perform these tasks will depend on the model’s capabilities, available sensors, integration with industrial systems and validated performance.
For field engineers, a portable computer with a three-dimensional display could provide access to schematics and diagnostic information at a worksite. Support for simulations and interactive models would need to be confirmed against the device’s final specifications. Possible applications include viewing 3D designs on-site and consulting AI-assisted troubleshooting tools. The effect on productivity and downtime would need to be assessed through practical trials.
The broader development of physical AI could support new applications in automation and robotics, although these uses require reliable integration with the relevant hardware and control systems. If AI models can adapt across different hardware platforms, as Alchymia intends, it simplifies the integration of advanced intelligence into diverse industrial machinery. This could accelerate the adoption of smarter automation across various sectors.
More efficient models and local deployment may improve the feasibility of AI in areas with limited connectivity or computing infrastructure. However, a local device still needs a reliable power source, adequate processing capacity and technical support. This addresses a critical barrier to industrial advancement in many emerging markets. Alchymia’s approach aims to broaden access to AI by reducing dependence on cloud infrastructure. The extent to which it achieves this will depend on the performance, cost and availability of its products.
