Anthropic in Talks to Buy Israeli Startup Decart for About $6 Billion

# Anthropic in Talks to Buy Israeli Startup Decart for About $6 Billion

Anthropic is in advanced talks to acquire Decart, a three-year-old Israeli startup that specializes in wringing more speed out of AI chips, in a deal valued at about $6 billion, according to a Bloomberg report on Thursday that was quickly matched by Fortune and other outlets. If completed, it would be by far the largest acquisition in the history of the Claude maker, and it would land at an extraordinary moment: just weeks before Anthropic is expected to attempt one of the biggest initial public offerings on record.

The talks are still at an early stage and could collapse, people familiar with the matter cautioned. Neither company responded to requests for comment. But the price alone signals how seriously Anthropic is treating the acquisition. A $6 billion deal would represent a premium of roughly 50% over the nearly $4 billion valuation Decart commanded in May, when it raised $300 million in a Series B round led by Radical Ventures, with Nvidia, Atreides Management, Valor Equity Partners and Adobe Ventures participating. Decart has raised roughly $450 million in total since emerging from stealth in October 2024 with a $21 million seed round.

What Anthropic Would Be Buying

Founded in September 2023 by Dean Leitersdorf and Moshe Shalev, both veterans of Israel's elite Unit 8200 signals-intelligence corps, Decart describes itself as "an efficiency-focused AI research lab." It employs about 100 people out of relatively modest offices on Yitzhak Sadeh Street in Tel Aviv. Leitersdorf, who earned a doctorate in computer science at 23, has said his ambition was nothing less than "to become the Google or Apple of AI."

The company's core asset is software that squeezes dramatically more performance out of existing hardware. Its optimization stack, marketed as DOS, co-designs model architecture to match specific chip constraints and pairs it with hand-tuned GPU execution kernels. By industry estimates, the technology can run AI models up to eight times faster than standard frameworks on the same silicon — and, crucially, it is hardware-agnostic, working not only with Nvidia GPUs but also with competing chips from Google and Amazon. Decart also builds so-called world models: Oasis, a real-time, playable open-world system first released in October 2024, and Lucy, a video model that edits live streams frame-by-frame at 1080p and 30 frames per second, already used by eBay, an investor and customer, for virtual clothing try-ons.

If a deal closes, Decart's team would fold into Anthropic's inference and performance organization, and the Tel Aviv operation could become the anchor of Anthropic's first substantial development center in Israel — a foothold the country's tech industry has long sought. Decart drew a crowded field of suitors in recent weeks; Calcalist, which first reported the sale talks, said interested parties ranged from Nvidia to Elon Musk's SpaceX.

Why It Matters

The logic of the deal is inference economics. As the AI industry pivots from the one-time expense of training frontier models toward the recurring, at-scale cost of running them for hundreds of millions of users, the ability to extract more output per chip becomes a strategic weapon rather than a nice-to-have. Anthropic's inference costs reportedly ran 23% over budget in 2025, even as demand for Claude surged. Buying efficiency technology outright — rather than simply pouring more capital into data centers — reflects a bet that cost discipline, not just capacity, will define the next phase of the AI race.

"Once systems can understand and operate within the physical world, the range of what becomes possible expands dramatically — from robotics and autonomous systems to entirely new forms of commerce and live experience," Leitersdorf said in May when Decart announced its Series B. "The most immediate test is what happens when you put it in front of a game designer, an advertiser or an engineer building a robot — and watch what they reach for."

The timing sharpens the calculus. The Wall Street Journal reported this week that Anthropic could go public as soon as September or October, with private-market transactions valuing the company around $1 trillion and public-market chatter already floating a $2 trillion debut that would be the largest in history. Once Anthropic begins reporting to public shareholders, its ability to convert explosive demand into sustainable gross margins will face intense scrutiny. Owning technology that meaningfully lowers the cost of every Claude query is exactly the kind of story a pre-IPO company wants to tell — and exactly the kind of talent it does not want a rival to grab first.

The scale of the surrounding infrastructure arms race underscores the point. This week alone, Nvidia and Wall Street heavyweights including Blackstone and Apollo unveiled a roughly $500 billion financing arrangement to expand computing capacity. Against that backdrop, as Calcalist's Sophie Shulman wrote, "a company capable of making existing hardware significantly more efficient becomes strategically important." For Israel, the deal carries symbolic weight too, echoing Intel's Mobileye purchase and Nvidia's Mellanox deal.

There is also a note of unfinished business. A $6 billion sale after just three years is a substantial return, but not an extraordinary one by the standards of today's AI market, and it would come before Decart built the durable business model its founders originally envisioned. The company still generates revenue largely through projects rather than a clearly defined commercial engine — one reason a well-resourced acquirer may make sense for both sides.

What to Watch

The immediate question is whether the talks survive the crowd. With Nvidia, SpaceX and others reportedly circling, Anthropic may face a bidding contest that pushes the price higher. Watch for confirmation of deal terms, any regulatory or export-control complications given Decart's Unit 8200 lineage, and how the acquisition is framed in Anthropic's eventual IPO prospectus. Most telling of all will be the margin math: if Anthropic can show public investors that Decart's optimization stack measurably bends its inference-cost curve, the $6 billion may look less like a premium and more like a down payment on the economics of running AI at planetary scale.