Anthropic in Talks to Acquire Decart for $6 Billion

The numbers alone raise eyebrows. Decart was founded in 2023 and has just over 100 employees. When it raised $300 million in May — three months ago — it was valued at $4 billion. The reported $6 billion is 50% above that. Revenue, according to Israeli outlet Calcalist, is in the tens of millions of dollars, and growth has been uneven.

But the stated purpose of the acquisition, as reported by Bloomberg, changes the picture. Decart's technology is reportedly headed for Anthropic's inference team. What Anthropic is buying isn't the video Decart produces — it's the underlying layer that let that video get made so cheaply.

What's Being Bought Isn't Video, It's the Overhead

An AI company's costs break down into two big buckets: training, which builds the model, and inference, which delivers the finished model to users. If training is the one-time construction cost of a building, inference is the monthly maintenance bill that never stops. Construction ends once; the maintenance bill climbs faithfully with every new user.

DOS (Decart Optimization Stack), the product Decart sells, targets exactly that maintenance bill. According to the company, DOS is a software layer that writes custom kernels and tunes compilers to squeeze "the last drop" of performance out of a single chip. Crucially, it isn't tied to any one chip — the company says it handles NVIDIA GPUs as well as Google TPUs, Amazon Trainium, and AMD accelerators in the same way.

Figures disclosed by investor Radical Ventures are fairly aggressive: 1,600 tokens per second in agentic inference — eight times the roughly 200 tokens per second the company cites as an industry average — and MFU above 80% on Amazon Trainium. MFU (Model FLOPs Utilization) measures what percentage of a chip's purchased compute is actually being used. Given that industry-standard figures run in the 40–50% range, this is the equivalent of having paved an eight-lane highway, using only four lanes, and suddenly using seven.

The trick of squeezing double the output from chips you already own, without buying new ones, is not a common one to find for sale in today's AI industry — let alone at this price.

Two 27-Year-Old Brothers and Unit 8200

Decart has three co-founders. CEO Dean Leitersdorf is 27 and finished his computer science PhD at 23; his brother Orian Leitersdorf earned his PhD at 21. The third co-founder, Moshe Shalev, spent 13 years in Unit 8200, Israel's military signals intelligence corps.

The investor roster reflects the company's profile. Decart started with a $21 million seed round led by Sequoia in October 2024, followed by a $32 million Series A that December, and a $300 million round in May of this year led by Radical Ventures with participation from NVIDIA — bringing total funding to around $450 million. Benchmark, Adobe Ventures, and eBay, also a customer, are among the investors.

Anthropic wasn't the only suitor. According to Calcalist, NVIDIA, which had joined the most recent funding round, first pursued acquisition talks before backing off, after which Amazon, Nebius, and SpaceX were floated as candidates. Elon Musk himself dismissed the SpaceX rumor as "fake news."

Revenue Run-Rate: $900 Million to Over $3 Billion in a Year

Anthropic's own circumstances explain why this deal is happening now. The company's stated revenue run-rate rose from $900 million at the end of 2025 to more than $3 billion in 2026. The number of enterprise customers spending over $1 million annually grew from 500 to more than 1,000 in under two months.

The problem is that this growth shows up directly on the bill. In November 2025, Anthropic announced it would spend $50 billion on U.S. infrastructure, and under a next-generation TPU agreement with Google and Broadcom, it will receive additional gigawatt-scale capacity starting in 2027. In the words of CFO Krishna Rao, the company is building "capacity to handle the exponential growth of our customer base."

There are only two ways to add capacity: buy more, or use what you have better. Data centers require power, land, and time, but optimization software starts paying off the week after the contract is signed. Set alongside the fact that Anthropic confidentially filed for an IPO in June and was recently valued at $965 billion, it becomes clear why a card that boosts margins commanded a 50% premium.

Where Lucy and Oasis Point

Looking at DOS alone, this acquisition is pure cost-cutting. But Decart also has two models.

Lucy is a real-time video editing model. It alters specific elements in a live video stream while leaving the rest unchanged, running, according to the company, at 30fps with under 35 milliseconds of latency. A commerce demo letting users try on clothes and accessories in real time is a flagship use case, and streamers on Twitch, TikTok, and YouTube are already using it.

Oasis is different in nature. It's a world model — a playable world generated on the fly based on user input. An early version drew a million users within three days and generated buzz, but the use case the company now emphasizes isn't gaming. It's simulation for training robots and self-driving systems, said to generate "hours of realistic driving."

Dean Leitersdorf has said: "When systems can understand and operate within the physical world, the scope of what's possible expands dramatically — from robotics and autonomous systems to entirely new forms of commerce and live experiences."

For Anthropic, which grew up on text and code, this is an unfamiliar direction. That's why the deal reads on two levels: in the short term, a tool to cut inference costs; in the long term, ingredients for Claude's move beyond the chat window.

Editor's Take

Anthropic's acquisition history shows how different this deal is. The 2025 Humanloop acquisition was effectively an acqui-hire, and bringing on the JavaScript runtime Bun that December was a way to secure a tool to make Claude Code faster — a decision made right after Claude Code hit a $1 billion run-rate just six months after launch.

Both were acquisitions "to build a better product." A $6 billion deal is a different animal. This is an acquisition aimed squarely at the income statement — built on the judgment that, for a company heading toward an IPO, an inference optimization team is the single most reliable thing money can buy to move the margin needle.

As a daily user of Claude Code, looking back over the past year, this tracks. What's been felt more often than the model getting smarter is usage limits and response speed. Throw the same task at it and some days it's smooth, other times of day it's noticeably slow. Users feel in their bones — before they ever see a bill — that the real bottleneck in an exploding service isn't intelligence, but throughput.

There's a lesson here for teams at home, too. When integrating an AI service, most of the meeting time goes to "which model should we use." But the variable that actually determines cost is how many users you can pack onto the same GPU at once. If that's the very variable a frontier lab just paid $6 billion for, it might be worth diverting even half of those model-selection meetings toward the serving stack instead.

To be clear, this is still just a negotiation. Bloomberg left room for it to collapse, and it's a seat NVIDIA once walked away from. But even if the deal falls apart, one fact remains confirmed: in the summer of 2026, the market put a $6 billion price tag not on the ability to build a frontier model, but on the ability to run that model at half the cost of everyone else.