Due to a global shortage of RAM spurred by the demands of artificial intelligence, the cost of various gadgets, from Apple's MacBook line to Xbox and PlayStation consoles, has significantly increased in South Africa, making 2026 a challenging period for buyers.
The main culprit behind this phenomenon is the global scarcity of RAM, unified memory, and storage space, which is caused by the insatiable demand of AI workflows. Modern users require more memory to run resource-intensive applications such as ChatGPT, Google Gemini, Microsoft Copilot, Claude, and Perplexity AI.
Although Apple was the last major tech corporation to raise prices, it managed to avoid an 'RAM apocalypse' until recently thanks to long-term supply contracts signed in advance. However, over the past year, Windows-based laptops have shown multiple price increases, which briefly made Apple a more attractive brand, but now the old order has been restored, and Apple is once again setting high prices for its hardware.
For instance, the 13-inch MacBook Air M5, which cost R19,999 before the recent price hike, is now selling for R26,499, representing a sharp jump of over 32%. Furthermore, Apple's most affordable 2026 model, the MacBook Neo, has risen from an initial price of R11,999 to R13,999, an increase of R2,000 or 16.67% in a short time. This means that the entry-level MacBook Air is almost twice as expensive as the Neo, and budget buyers are likely to opt for the Neo despite performance compromises.
The problem is not limited to computers. The Xbox Series S and Series X consoles, released about six years ago, have also seen price increases. The Series S model launched at $300 (equivalent to approximately R5,100 at the standard exchange rate), and the Series X cost $500 (about R8,600). Although local import costs initially made them slightly more expensive in South Africa, the rise was not as dramatic. The author purchased an Xbox Series S in March 2023 for R5,999, but today the same console costs R9,998, an increase of R4,000 for a six-year-old device.
Instead of the expected price drop, the price continues to climb. To get the top-tier Series X, it will cost around R14,999. When the author bought the Series S, the Series X price of R11,000 seemed too high given the R5,000 difference between the models, which would allow buying almost two Series S units for the price of one. Microsoft is not alone in this situation: Sony's base PlayStation 5 Digital Edition has risen from $400 at launch in 2020 to $500 by 2025. PlayStation maintains its lead in the console race, with about 75 million units sold compared to 31 million for Xbox, meaning millions of gamers are feeling the same pressure.
Thus, purchasing almost any new electronics during 2026 will be expensive. However, there is uncertainty about when prices might stabilize, making the current moment both the best and worst time to make a purchase.
NueroDance has introduced its cross-modal data platform, NeuroAI, along with the ND1000 and ND8 series devices, at an event in Beijing. The company positions itself not merely as a manufacturer of brain-computer interface equipment, but as a provider of data infrastructure for NeuroAI.
NueroDance is betting that neural data will become the next critical modality after text, images, speech, video, and code. The goal is to improve AI's understanding of human states such as intention, fatigue, engagement, collaboration, and consensus. The company notes that existing data used to train AI only captures the results of human expression, but does not reflect the neurophysiological state that precedes and accompanies that expression.
The company's product architecture is built on a three-level scheme. The first level is represented by the ND1000 system, which supports 14 types of physiological signals, including EEG, EOG, EMG, ECG, EDA, and respiration. This system provides wireless synchronous data collection with millisecond-level alignment and has received national registration certification as a medical device for use in research, medical, and serious industrial scenarios.
The second level is designed to ensure accessibility in real-world conditions and includes the portable ND8 series devices: the ND8 model with 8 channels and the 21-gram ND8 Ultra. The latter model is also equipped with voice recording, real-time transcription, pose detection, and local storage functions, allowing EEG collection in classrooms, rehabilitation centers, sports venues, corporate training, and daily life.
The third level is the NeuroAI platform, which aggregates multi-user, multi-site, multi-device, and multimodal data into unified timelines and real-world scenario frameworks. This platform serves both for individual experiments and for continuous accumulation of data necessary for training scenario models. A group of a thousand people demonstrated the capabilities of quick device donning, real signal acquisition, synchronization across different locations, real-time visualization, and collaborative creation of anonymized data.
Commercialization extends beyond research to cover education, rehabilitation, sports, corporate training, and neuroaesthetics. The common point is that behavioral outcomes are visible, while state processes remain hidden. NueroDance's strategy is to use hardware as the entry point, scenarios as the channel, data as the asset, and the platform as the ultimate goal.
It is reported that the company is preparing for a new round of equity financing to support iterations of core technology, modernization of product systems, expansion of market channels, and team building. Key risks identified include signal stability in real-world conditions, consistency between different devices, comfort during prolonged wear, willingness to pay from industrial clients, data standards, compliance mechanisms, and building the platform ecosystem. For integrating neural and EEG data into educational, medical, and corporate scenarios, informed participation, data anonymization, and adherence to usage norms are required as the foundation for industrialization. The company believes that its 'hardware first, platform later' strategy aligns with the model of other infrastructure companies that started as niche tools before becoming indispensable platforms.
The WAIC 2026 exhibition shows a rise in token factories as computational processes shift from renting hardware to delivering tokens. Currently, models of both soft and hard infrastructure are competing to standardize the measurement of artificial intelligence performance.
The main trend at WAIC 2026 is not any specific model or chip, but the sale of tokens. Token factories have become the most common topic on the exhibition floor, reflecting a fundamental transition of the AI industry from renting computing power in the form of hardware to providing computation in the form of tokens. The token, initially the smallest unit of consumption and output for AI models, is turning into the primary indicator of AI performance. This shift indicates the industry's move from centralized training to inference scaling, opening up possibilities for new business models that decouple computation from specific hardware configurations.
Token factories exist in two main forms. The hard infrastructure model involves creating specialized data centers or utilizing reclaimed capacity from outdated GPUs. iSoftStone presented the Beijing Token Factory No. 1, which aims to produce 1.4 trillion tokens daily, in collaboration with the Greater Bay Area public computing platform and Shaoguan factory. Jiuzhang Cloud plans to create more than ten global AI computing centers with a total capacity of 30 EFLOPS, capable of generating 50 billion tokens daily. These centers operate in a closed loop with a dual engine: one factory handles training using reinforcement learning, while the other packages models into standardized services. Tiangang Zhisuan refurbishes degraded GPUs by using computation optimization to increase performance from 20% to 40-60% of the original, offering low-cost computing to price-sensitive clients.
The soft service model, represented by SiliconFlow, involves renting higher-level computing resources to generate tokens for end consumers. SiliconFlow supports over 170 major AI models on Nvidia, Huawei Ascend, MUSA, and Moore Threads hardware, achieving a throughput of over 1 trillion tokens per day. The flexibility of this model is achieved by directing batch processing to unused supercomputers at night. The actual efficiency of token output is determined by system optimization, including scheduling and inference mechanisms.
The benefit for the client is obvious: it is cheaper than cloud providers and more universal than individual model APIs. Token factories aggregate open-source models like GLM and DeepSeek alongside interfaces compatible with OpenAI and Anthropic, providing a single API entry point for multiple models with pay-as-you-go billing. Intelligent routing automatically parses multimodal requests requiring text, image, and video generation, directing each subtask to the most efficient model, ensuring task alignment with the model rather than human selection of the model. Jiuzhang Cloud further segments tokens into three levels: consumer for general scenarios, professional with embedded industry knowledge for finance and manufacturing, and frontier for complex reasoning and multi-step planning tasks.
Standardization issues remain. Jiuzhang proposed using the Data Compute Unit (DCU), which transforms various chips, including Nvidia, Ascend, and Hygon, into a unified measurement of 312 TFLOPS × 1 hour. In turn, iSoftStone released the world's first token factory performance benchmark based on chaotic workload characteristics for long-term agent testing. The consensus reached at WAIC was clear: the competitive logic has shifted from the number of GPUs you own to how many tokens those GPUs produce and whether those tokens translate into real performance. Token factories are transforming computation from a capital-intensive hardware asset into a standardized, tradable unit of performance, changing how AI computing is bought and sold.
IT giant HCLTech has announced plans to build its first artificial intelligence (AI) data center at the upcoming Sovereign AI Park in Bhubaneswar, Odisha. According to a filing with BSE, this center will be implemented through a tripartite partnership between the AI startup Sarvam, the Government of Odisha, and HCLTech itself.
The company stated that the planned capital expenditure for this project will amount to approximately 14,257 crore rupees, including financial support from the Government of Odisha. This facility is intended to support India's sovereign AI ecosystem by combining HCLTech's full-fledged AI capabilities with Sarvam's foundational models to develop specialized AI applications for government bodies and enterprises.
Furthermore, the project will enable the provision of multilingual AI services and meet government requirements regarding data sovereignty. A Memorandum of Understanding (MoU) for this project was signed in the presence of the Chief Minister of Odisha, Mohan Charan Majhi, Chairman of HCLTech, Roshni Nadar Malhotra, and CEO and Managing Director C Vijayakumar.
Roshni Nadar Malhotra, Chairman of HCLTech, stated: 'We aim to be a catalyst for India's sovereign AI ecosystem by offering differentiated solutions across the entire value chain through targeted investments and partnerships. We are delighted to embark on this venture in partnership with the Government of Odisha and Sarvam.'
This announcement followed a report by Business Standard on Wednesday that HCLTech planned to establish an AI-optimized data center and a Global Development Center (GDC) in Bhubaneswar with investments of around 730 crore rupees. According to that report, both facilities will allow HCLTech to provide advanced services in AI, cloud computing, cybersecurity, digital design, and enterprise technologies from Odisha. The proposed AI data center is also expected to be HCLTech's first proprietary center in India. Previously, the company indicated that investments in its data center business would be made through several newly established subsidiaries, and the network was expected to expand to a capacity of 50 MW.
