Anthropic's $7 Billion Bet on Making AI Cheaper to Run

The next constraint in generative AI is no longer which model is smartest. It is what it costs to run each one.

Anthropic, the company behind the Claude chatbot, is reportedly on the verge of acquiring Decart, a 3-year-old Israeli startup that builds software to make AI training and inference faster and cheaper. Calcalist reports the deal is approaching the signing stage, with parties exchanging advanced drafts of the agreement, though nothing has been finalized and the terms could still change. Reuters and Bloomberg had previously confirmed talks at a roughly $6 billion valuation, before the reported price moved higher.

The price has moved fast. In May 2026, Decart raised $300 million at a $4 billion valuation in a round led by Radical Ventures, with Nvidia, Adobe Ventures, Toyota Ventures and Atreides Management participating. Total funding before the reported Anthropic transaction exceeds $450 million, per Calcalist's coverage. A move to $7 billion in roughly three months is a near-2x step-up, and most of the consideration is expected to be paid in Anthropic shares rather than cash.

That step-up is the signal. Inference is the operation a chatbot company runs every time a user sends a message, and it is the line item that has begun to dominate the AI cost stack. Reuters Breakingviews has framed the math in the simplest possible terms: even modest efficiency gains are worth billions to a frontier lab. A startup that can show repeatable speedups on training and inference earns the buyer the right to grow into the savings, which is why a 3-year-old company with $450 million of prior capital can land at a $7 billion ticket.

Decart's pitch is two-part. The first half is compute-optimization software, the layer that decides how many GPUs and how much memory a given model workload actually consumes. The second half is "world models" for real-time interactive environments, the kind of generative simulation that robotics, autonomous driving, drones and real-time video applications would need if they are ever to run on commodity hardware. Calcalist describes the company as a play on both inference cost and physical-world simulation, the two areas where AI capex is most exposed to unit-economics pressure.

If the deal closes, Decart's team would join Anthropic's inference and performance organization, per Calcalist's reporting. That is where the company's own public claims become load-bearing. Decart's specific training- and inference-speedup figures are company-stated; independent benchmarks in the open record are thin. The Reuters Breakingviews efficiency-gain framing is the cleanest external anchor for why an infrastructure buyer would pay this multiple, and the company's own benchmarks should not be stretched into a settled fact.

Two other pieces of context frame the timing. Calcalist reports Nvidia also bid, possibly at a higher headline price, but Decart's founders and backer Sequoia preferred Anthropic, partly because the acquirer's inference roadmap and equity currency matched Decart's long-term thesis. Anthropic is also reportedly preparing a major initial public offering later this year, with Calcalist putting the company's revenue run-rate at roughly $65 billion. Acquiring a category-leading compute-optimization asset in that window is a concrete way to address investor concerns about inference unit economics before the S-1 lands.

The press characterization of the deal, if it closes, is that it would be Anthropic's largest acquisition to date and one of the largest purchases of an Israeli tech company on record. Both framings are supported by reporting, not by Anthropic or Decart on the record, and the deal has not been signed. The terms could still change or fall apart.

What would survive a collapse of the talks is the category signal. The next phase of the AI race is being fought on the cost of every query, not on which model tops a leaderboard, and the most expensive line item on the P&L is the one a buyer will pay to compress. The next milestone to watch is the per-token inference-cost disclosure in Anthropic's IPO filings, which is where the market will judge whether the buyer priced this right.