Anthropic PBC is in advanced talks to acquire Decart AI for roughly $6 billion, Bloomberg reported on August 13, citing people familiar with the matter. Decart makes software that reduces the cost of training and operating AI by helping chips work more efficiently, and the potential acquisition is intended to help Anthropic's existing computing infrastructure absorb surging demand for its software. The deal would be Anthropic's largest known acquisition and, according to The Next Web, would mark its fifth of the year. The deal has not been finalized and could fall through.
Founded in 2023 by Israeli brothers Dean and Orian Leitersdorf and Moshe Shalev, Decart's central asset is DOS, the Decart Optimization Stack, a vertically integrated platform that operates at the lowest level of the software hierarchy between AI models and the silicon that runs them. Most AI software frameworks generate GPU execution code using general-purpose compilers that apply the same heuristics to every chip without exploiting specific memory hierarchies or parallel-compute architectures, with the result that a substantial fraction of a chip's raw computing capacity sits idle during AI workloads. By rewriting how AI models communicate with GPU and accelerator chips, DOS delivers over 1,600 tokens per second for agentic inference, compared with an industry average of roughly 200, and achieves over 80 percent of a chip's raw computing capacity doing productive work versus typical production rates of 40 to 50 percent. Crucially, DOS is hardware-agnostic, working on Nvidia GPUs as well as competing chips from Google (TPUs) and Amazon (Trainium). If the deal closes, Decart's team would join Anthropic's inference and performance organization.
Anthropic submitted a confidential S-1 filing to the SEC in early June in preparation for a widely anticipated IPO. The filing targets an October 2026 Nasdaq listing, with Goldman Sachs, JPMorgan, and Morgan Stanley leading an offering expected to raise more than $60 billion. The margin arithmetic is the context that makes the acquisition's timing legible. Anthropic's gross margin currently sits at roughly 40 percent; the company is targeting 77 percent by 2028, but it plans to spend approximately $19 billion on training and inference compute in 2026, roughly matching its full-year revenue, and margins compressed after inference costs ran 23 percent over projections. That improvement depends on GPU costs declining and inference efficiency improving. Bloomberg Deals reporter Ryan Gould noted that while $6 billion might seem modest in the current deal landscape, it is significant for Anthropic because such a deal could help the company better communicate its cost-efficiency story to potential investors. Anthropic already pays xAI $1.25 billion per month to train and run AI models in its data centers, underscoring how acute the compute-cost problem is that Decart's technology is meant to address.
Skeptics see a different arithmetic. Decart's revenue is described only as "several tens of millions of dollars," with uneven growth driven largely by custom engagements with chip manufacturers; at the reported acquisition price, any buyer is primarily paying for the technology and the team, not for current cash flows. The most recent funding round, closed in May, valued Decart at nearly $4 billion, meaning a $6 billion deal would represent a premium of roughly 50 percent over that mark in under three months. Neither company has confirmed the negotiations, and early acquisition talks have a habit of changing direction, as an apparent prior pivot away from SpaceX already demonstrates. Separately, analysts have flagged that Anthropic's gross-versus-net revenue accounting practice inflates headline figures relative to net-reporting peers, a question the full S-1 will need to resolve.
For anyone trying to price Anthropic's IPO, the blunt question the Decart deal raises is whether a 50-percent acquisition premium on a company with tens of millions in revenue is a genuine engineering solution to a margin problem or an expensive way to paper over one in the prospectus.