Anthropic is in talks to acquire AI startup Decart for about $6 billion, Bloomberg reported Wednesday, a deal that would be the Claude maker's largest known acquisition and a sign that major AI labs are racing to cut computing costs as they scale infrastructure ahead of potential IPOs.
Decart, founded in 2023, builds software to make AI chips run more efficiently. Its Decart Optimization Stack (DOS) optimizes training and inference across Nvidia GPUs, Amazon's Trainium, and Google's TPUs. That capability is increasingly valuable as inference costs rise alongside demand for tools like Claude, making the startup strategically important for Anthropic.
The talks remain ongoing and could fall through, Bloomberg reported. If completed, the deal would come as Anthropic ramps up spending on computing infrastructure ahead of a Wall Street debut that reports say could happen next month or in early October.
AI labs face ballooning infrastructure costs
The acquisition talks suggest Anthropic is looking for ways to extract more performance from each chip rather than just adding more hardware. The company has committed billions toward expanding compute capacity in recent months.
In November, Anthropic purchased $30 billion of Microsoft Azure compute capacity and announced a $50 billion investment in data centers with Fluidstack. In April, the company expanded its Amazon partnership to secure up to 5 gigawatts of additional compute, committing more than $100 billion to AWS technologies over 10 years.
OpenAI faces a similar burden. Its Stargate project with Oracle includes 4.5 gigawatts of additional US data-center capacity, with total capacity under development exceeding 5 gigawatts and involving more than 2 million chips.
Anthropic and OpenAI's recent deal activity
Anthropic has focused its acquisition strategy on small technical teams. It acquired Stainless in May 2026, an SDK and MCP tooling company, and Vercept in February 2026, a startup building AI systems that interact with computers. Before that, it bought Bun in December 2025, a high-performance JavaScript runtime and developer toolkit.
OpenAI has pursued larger and more frequent deals. Its largest was the $6.5 billion acquisition of Jony Ive's io in May 2025, bringing the hardware startup's team into OpenAI to develop consumer devices. More recently, OpenAI agreed to buy Promptfoo in March 2026, an AI security and testing platform, and Astral later that month, a maker of Python developer tools.
IPO watch and competition
Spending at both companies remains high, though as private firms they don't disclose cash burn. Anthropic raised $65 billion at a $965 billion valuation in May, with revenue reaching a $47 billion annualized run rate. OpenAI's annualized revenue run rate stands at about $25 billion.
The race for cheaper compute has become as critical as the race for more powerful models. For development teams, that means the tools they use to build and deploy AI applications could shift as labs invest heavily in software that optimizes hardware performance rather than just add more chips.
Why this matters for IT and development teams
As companies like Anthropic push more AI work onto existing hardware, developers can expect more training for AI alongside software developers on multi-architecture optimization - particularly around Nvidia, Amazon, and Google chips. Tools like Decart's DOS treat compute efficiency as a software problem, and the trend means development teams will increasingly need to understand how their code performs across different hardware stacks that are still being managed. The race to train bigger models is also a race to run them cheaper, and that pressure will eventually come to your production budgets.
Your membership also unlocks:
