The proposed deal centres less on expanding Anthropic’s consumer product portfolio than on improving the economics of computing. Decart develops software designed to extract greater performance from existing AI chips, potentially allowing Anthropic to serve more users without increasing hardware expenditure at the same pace. Its engineers would be expected to join Anthropic’s inference and performance operations if an agreement is completed.
Decart’s optimisation technology has become the main attraction as AI developers face mounting pressure to make inference — the process of running trained models to answer user requests — faster and cheaper. Its Decart Optimization Stack, or DOS, operates across Nvidia GPUs, Google TPUs and Amazon’s Trainium hardware, giving customers more flexibility over which processors they use.
The company says DOS 2.0 can process more than 1,600 tokens a second for agent-based AI workloads, roughly eight times the industry average cited by Decart. The performance claim has not been independently audited, making it an important qualification when assessing the potential savings for Anthropic. Decart also says its systems can process high-definition world-model workloads at up to 100 frames per second.
The acquisition talks come as Anthropic attempts to demonstrate that the rapid growth of Claude can eventually translate into software-style profitability despite enormous expenditure on computing capacity. Its gross margins have remained far below the levels typical of mature software companies, while projections circulating among investors envisage margins rising towards 77 per cent by 2028. Achieving that target depends heavily on falling unit costs for inference and improved utilisation of expensive processors.
That makes Decart strategically significant. Even modest improvements in the amount of useful computing output generated by each processor could produce substantial savings at Anthropic’s scale. The Claude maker is spending heavily to train new models and handle accelerating demand from businesses, developers and users of its coding products. Efficiency gains could therefore affect both operating costs and the amount of infrastructure required for future growth.
Decart was founded by brothers Dean and Orian Leitersdorf and has developed both infrastructure software and so-called world models, which learn representations of physical environments and can generate interactive simulations. Its Lucy models transform live video, while Oasis is designed for interactive environments that can be used in areas including robotics, autonomous systems and simulation.
The company raised $300 million in May in a funding round led by Radical Ventures, taking its valuation to about $4 billion. Nvidia joined the financing alongside other technology and venture investors. Decart has raised more than $450 million altogether, meaning a $6 billion sale would represent a substantial premium to its valuation only months earlier.
Anthropic has traditionally relied heavily on external computing suppliers. Bringing Decart’s optimisation expertise in-house could give it greater control over the software layer connecting Claude models with processors supplied by Nvidia and cloud partners. It could also reduce dependence on simply purchasing additional chips whenever demand rises.
The timing adds another dimension. Anthropic has been preparing for a possible stock-market listing, with investors closely examining whether explosive revenue growth can ultimately overcome the high costs inherent in training and operating frontier AI systems. A large infrastructure acquisition before a flotation would signal that management views computing efficiency as central to improving the company’s long-term financial profile.
Competition is also intensifying as AI companies search for ways to deliver comparable model performance at lower prices. Cheaper models and improvements in inference technology have increased pressure on developers to focus on cost per query rather than model capability alone. Anthropic, OpenAI and other frontier laboratories consequently face a growing requirement to improve utilisation of GPUs and other accelerators while maintaining model quality.
Decart’s technology offers one possible route. Its optimisation software works closer to the hardware layer, addressing kernels, memory management and model execution rather than depending entirely on more powerful processors. Decart also markets its architecture as hardware-independent, potentially allowing workloads to shift more easily between Nvidia, Amazon and Google systems.
Follow Arabian Post
Select Arabian Post as your preferred source on Google and MSN News for trusted business news and Arab politics and updates.
