Anthropic's $6 Billion Decart Gambit: Buying Speed on the Road to a $2 Trillion IPO

Three weeks before what investors expect to be the largest initial public offering in corporate history, Anthropic is reportedly negotiating the largest acquisition in its own history. According to Bloomberg, Reuters, and Calcalist, the Claude maker is in talks to buy Israeli AI startup Decart for approximately $6 billion — a deal that would dwarf every prior purchase by the San Francisco company and signal a sharp strategic pivot in how frontier labs plan to compete.

The deal is not final. Sources describe the discussions as early-stage, with one telling Reuters that if completed, Decart’s team would be folded into Anthropic’s inference and performance organization. But the sheer size of the number — $6 billion for a three-year-old startup that raised its $21 million seed round only in late 2024 — tells its own story about where the AI industry believes the next competitive battlefield lies: not in training bigger models, but in making the models we already have run faster and cheaper.

What Decart Actually Does

To understand why Anthropic would pay what amounts to a 50% premium over Decart’s last private valuation, you have to look past the consumer demos that made the company famous and into the unglamorous machinery underneath.

Founded in 2023 by Dr. Dean Leitersdorf and Moshe Shalev — both alumni of the Israel Defense Forces’ Unit 8200 signals-intelligence unit — Decart built its reputation on two viral products. Oasis, released in late 2024, was a fully playable AI-generated Minecraft clone that rendered every frame in a neural network rather than a game engine. Lucy (originally Mirage), released in 2025, lets users edit live video in real time with text prompts. Both were impressive proof points for real-time generative AI, and both were, in a sense, marketing for the company’s real product.

That product is the Decart Optimization Stack, or DOS — a software engine that optimizes AI training and inference at the kernel, memory, and hardware level. In practice, DOS does two things that matter enormously to a company like Anthropic. First, it squeezes dramatically more throughput out of existing GPU fleets, attacking the single largest line item in any frontier lab’s budget. Second, it makes AI workloads portable across different chips: Nvidia GPUs, Google TPUs, and Amazon Trainium, with Nvidia itself participating in Decart’s May 2026 funding round as both investor and partner.

Decart’s business momentum tracks its technical claims. The company raised a $21 million seed from Sequoia Capital in late 2024, a $32 million round in early 2025, and $153 million from Sequoia in August 2025 at a $3.1 billion valuation, before landing a $300 million Series B led by Radical Ventures in May 2026 at roughly $4 billion — bringing total funding to around $450 million and counting Amazon among its strategic backers. The reported $6 billion price represents a 50% markup over that last mark, in under three months.

The Math That Justifies the Price

Why would Anthropic pay it? The answer lies in the brutal economics of inference at scale.

Anthropic’s quarterly revenue reportedly exceeded $11.5 billion in Q2 2026 — a 14-fold increase year over year — with annualized run-rate figures cited between $47 billion and higher as the company prepares for an IPO that backers hope will value it at $2 trillion or more. Revenue at that scale means inference at that scale: every percentage point of efficiency gained across Anthropic’s compute fleet translates into hundreds of millions of dollars in deferred capital expenditure.

Bloomberg’s reporting points at the core rationale directly: the acquisition is intended to help Anthropic’s existing computing infrastructure absorb more demand. In an industry where GPU supply remains the binding constraint on growth — and where Anthropic, unlike Google or Amazon, owns no silicon of its own — a team that can make the same hardware serve more tokens is worth roughly whatever it costs.

The strategic symmetry extends further. OpenAI just launched its Ultrafast tier, running GPT-5.6 Sol on Cerebras silicon at up to 750 tokens per second — a 14x throughput claim that put raw inference speed at the center of the competitive conversation. Acquiring Decart gives Anthropic a credible answer on both fronts: more speed from existing Nvidia fleets, and a software layer that reduces dependence on any single chip vendor. The chip-portability angle may ultimately matter more than the speed angle, as frontier labs look for leverage in negotiations with Nvidia and hedge against supply concentration.

Timing: Efficiency as an IPO Story

The timing is hard to miss. Anthropic’s CFO Krishna Rao has been holding early IPO meetings with investors, and reports point to a potential October listing. A public-market audience will scrutinize gross margins and capital intensity far more ruthlessly than private investors did. An in-house optimization stack that measurably improves both is exactly the kind of asset that makes a $2 trillion valuation narrative easier to defend in a roadshow.

There is also a talent-and-capability argument. The Unit 8200 alumni network has produced a disproportionate share of the world’s top low-level systems engineers, and Decart’s roughly 50-person team is dense with exactly the profile that is nearly impossible to hire: people who can write GPU kernels and distributed-systems code at the frontier. For Anthropic — which has historically grown through research hiring rather than acquisition — buying a pre-assembled infrastructure team is a new playbook, and one that mirrors how Google absorbed DeepMind-era systems talent.

What Could Still Go Wrong

Skepticism is warranted on three fronts. First, the deal is early and could fall apart — Bloomberg’s own sourcing notes the talks may not conclude, and Israeli press reports suggest the price under discussion spans $6–7 billion, indicating negotiation is still live. Second, $6 billion is a lot to pay for a company whose consumer products (Oasis, Lucy) sit far from Anthropic’s enterprise focus; the deal only makes sense if DOS’s datacenter-side value is real and durable, rather than a set of benchmark-specific tricks. Third, Decart’s investor list — Sequoia, Benchmark, Radical, Nvidia, Amazon — means several of Anthropic’s own ecosystem partners are on the cap table, which complicates the geopolitics of chip-portability software that helps customers move off Nvidia.

The counterargument: in a market where Anthropic’s valuation has compounded from $61.5 billion to nearly a trillion dollars in eighteen months, a $6 billion purchase that protects the margin structure of a $47 billion run-rate business is not a splashy acquisition. It is insurance. And if the October IPO lands anywhere near the numbers investors are whispering about, it may look in retrospect like the cheapest infrastructure Anthropic ever bought.

For the broader industry, the message is unambiguous. The training-scaling era rewarded whoever could spend the most on pretraining runs. The inference era rewards whoever can serve the most tokens per dollar — and the labs are now willing to pay billions, in cash and stock, for the software that tilts that equation in their favor. Expect more deals like this: as models commoditize, the optimization layer underneath them is becoming the moat.