In the AI era, time is becoming the most expensive asset.
On August 13, a potential $6 billion acquisition thrust a nearly three-year-old AI company into the spotlight. According to Bloomberg, Reuters, and other foreign media, Anthropic is in talks to acquire Israeli AI startup Decart, with a potential deal size of about $6 billion (approximately 40.5 billion RMB). At the same time, the Financial Times reported that six Anthropic investors expect the company could go public as soon as October this year, and believe its valuation could reach $2 trillion or even higher at that time. If that happens, Anthropic would surpass SpaceX, which went public in June this year at a valuation of approximately $1.77 trillion, to become the company with the highest IPO valuation in history.
Decart, which Anthropic wants to acquire, was founded in 2023. In May this year, Decart had just completed a $300 million financing round at a valuation of nearly $4 billion; less than three months later, Anthropic is willing to offer a potential acquisition price of $6 billion.
Decart's CEO Dean Leitersdorf earned his computer science Ph.D. from the Technion at 23; his younger brother Orian Leitersdorf later completed his doctorate at 21, breaking his brother's record, and now serves as Decart's CTO. Co-founder Moshe Shalev took a different path—he grew up in an ultra-Orthodox Jewish family in Israel, enlisted in the military at 23, joined the IDF's Unit 8200, and served there for 13 years. In less than three years since founding the startup, they are now sitting at a $6 billion M&A negotiation table. What exactly does Anthropic see in Decart?
1. What Exactly Does Anthropic Want to Buy Before Its IPO?
Anthropic is not short of growth right now. In February this year, Anthropic completed a $30 billion Series G round at a post-money valuation of $380 billion, with annualized revenue of $14 billion disclosed at the time; in April, its annualized revenue was reported to have reached $30 billion; by the end of May, the company completed a $65 billion Series H round at a post-money valuation of $965 billion, with annualized revenue exceeding $47 billion. In just over three months, its valuation more than doubled, and annualized revenue more than tripled.
On June 1, Anthropic confidentially submitted a draft registration statement (Form S-1) for an initial public offering to the U.S. Securities and Exchange Commission. Now, investors are betting that this growth curve can continue. The Financial Times reported that some Anthropic investors expect the company's annualized revenue could further reach $100 billion–$120 billion by the end of 2026. One investor even extrapolated based on revenue growth rates and valuation multiples, arriving at a theoretical valuation of $3 trillion.
The valuation surge has also quickly created a number of super-billionaires. Forbes estimated at the end of May that the seven co-founders of Anthropic, including CEO Dario Amodei and President Daniela Amodei, each hold slightly more than 1.6% of the company's shares, which at a $965 billion valuation would make each worth about $15.5 billion.
The closer Anthropic gets to the public market, the more realistic the questions it faces. How much money will it take to sustain such rapid growth? The computing power consumption of large models can be broadly divided into two parts. The first is training—using massive data and chips to "train" the model; the second is inference—the actual answering of user questions after the model is deployed. Every time a user asks Claude to write code, process a document, or answer a question, the chip must perform a new computation behind the scenes. Training often occurs intensively during a certain phase, but inference occurs continuously as users interact. The more Claude users and the more frequent the calls, the more inference computing power is required.
Anthropic has already been expanding capacity frantically. In April this year, the company signed a new agreement with Amazon, committing to invest over $100 billion in AWS technology over the next decade in exchange for up to 5 GW of compute capacity; that same month, it expanded its partnership with Google and Broadcom to obtain several GW of next-generation TPU capacity. In May, Anthropic also reached a computing partnership with SpaceX. Currently, Claude runs simultaneously on multiple chip architectures, including AWS Trainium, Google TPU, and Nvidia GPU.
Problems also arise. Nvidia GPUs, Google TPUs, and Amazon Trainium use different hardware architectures. Migrating a model from one chip to another is not simply a matter of "changing computers"; the underlying programs, computational approaches, and communication systems often need to be re-optimized. This is precisely where Decart excels. Its core product, DOS (Decart Optimization Stack), can be understood as a "performance optimization layer" between AI models and chips. It rearranges how models compute, how they invoke chips, and how data is transmitted according to different hardware, so that the same AI model can run as efficiently as possible on different platforms. Decart has already adapted DOS to Nvidia GPUs, Google TPUs, and Amazon Trainium. In other words, no matter which vendor's chips Anthropic buys, Decart happens to be researching how to squeeze more out of those chips.
On August 5, Anthropic announced it had begun building its own chip design team, recruiting engineers who understand both hardware and software, hoping to make Claude run faster and more efficiently through custom chips while maintaining a multi-chip strategy. A few days later, the acquisition news about Decart emerged. More importantly, if the deal goes through, Reuters reported that the Decart team would be integrated into Anthropic's inference and performance division. What Anthropic really wants to buy is not an "AI video company", but a team that understands models, chips, and underlying system optimization. For an AI company preparing to hit a $2 trillion valuation, no matter how powerful its model, if every call is too expensive, that will eventually show up in cost structure, gross margins, and cash requirements. The public market will not only ask how smart Claude is, but also: for every $1 spent on computing, how much value can Anthropic get back?
2. What Does Decart Have?
Decart was founded in September 2023. Before starting the company, Dean Leitersdorf conducted postdoctoral research at the National University of Singapore; during that time, he had already begun preparing the entrepreneurial team with Moshe Shalev for Decart's establishment. Earlier, he studied computer science at the Technion and earned his Ph.D. at 23. While studying, he also served in the IDF's Unit 8200. Unit 8200 is one of Israel's most important signals intelligence and cyber technology units, long responsible for communications intelligence, cybersecurity, and technology R&D, and has produced many Israeli tech entrepreneurs. Dean met Moshe Shalev there.
Dean's life has almost been on "fast-forward": he completed his studies early and got his Ph.D. at 23. Moshe took a different path. He grew up in an ultra-Orthodox Jewish family in Israel and did various jobs when young, even attending night school in accounting while working. He did not enlist until age 23, then joined Unit 8200 and served for 13 years. Over the years, he moved from the technical front line to AI system construction and management, becoming a key technical aide to then-unit commander Yossi Sariel.
The Information previously reported that about three months after Decart was founded, it won a multi-million-dollar contract from a GPU cloud service provider, mainly to improve the inference efficiency of AI models on GPUs. By October 2024, this business had generated over $10 million in annualized revenue for the company. More remarkably, Dean said that even as the company had begun investing funds to train its own video foundation model, it maintained positive free cash flow. This is quite rare in the AI startup world. The first thing many AI startups do after raising money is buy GPUs, rent data centers, and train larger models; Decart, by contrast, first made money by helping others improve GPU efficiency, then used that income to train its own models.
Sequoia Capital partner Shaun Maguire summarized Decart's technical capabilities: "They understand how GPUs work down to the level of electrons"—meaning their understanding of GPUs goes down to the most fundamental hardware operating mechanisms. Sequoia even likened this capability to that of early Google. Google's truly difficult-to-replicate advantage was not only the PageRank algorithm, but also its underlying distributed systems capability: when competitors bought expensive servers, Google found ways to combine large amounts of cheap hardware and squeeze out every unit of computing power. Sequoia believes Decart is following a similar path in the AI era—first perfecting underlying efficiency to the extreme, then turning that efficiency into a product.
Their first breakout product was Oasis. In October 2024, Decart released Oasis. Its experience is somewhat like Minecraft, but the operating logic is completely different: the visuals are not pre-built by a traditional game engine, but generated in real time by an AI model based on the player's keyboard and mouse actions. Every time a player makes a move, the model generates the next frame, and the entire world changes accordingly. If the player walks left, the world generates to the left; if a block is broken, the model continues to predict what should happen in the next frame. Decart officially said Oasis surpassed 1 million users within 72 hours of launch.
Later, the company developed a new product, Lucy. Lucy applies this real-time generation capability to video. Unlike traditional AI video that must be generated before playback, with Lucy, users can change characters' appearances, clothing, backgrounds, and surroundings in real time during live streaming or video calls. According to Decart, Lucy 2.0 can continuously generate images at 1080p resolution and 30 frames per second.
In June this year, Decart launched Oasis 3, pushing the world model further into autonomous driving and robotics. Simply put, a "world model" enables AI not only to know "what is in front of it", but also to predict "what will happen next". If a car turns, how will the surrounding roads change; if a robot reaches out and touches a cup, how might the cup move—the model needs to generate a world that changes continuously with actions. This allows autonomous vehicles and robots to first experience a large number of extreme situations in the virtual environments created by AI before entering real roads, factories, and homes.
In summary, Decart has a very clear technology chain: DOS is Decart's underlying technical foundation, responsible for improving model training and inference efficiency; Lucy uses real-time generation capability for video on top of that; Oasis further targets robotics and autonomous driving, generating interactive virtual environments that can change in real time.
Image
Capital quickly followed. In 2024, Sequoia Capital led Decart's $21 million seed round; the company subsequently raised additional funding, bringing total funding to $53 million and valuation to about $500 million. In August 2025, Decart completed a $100 million financing round, with valuation jumping to $3.1 billion. In May this year, the company completed a $300 million round at a valuation of nearly $4 billion, led by Radical Ventures, with new investors including Nvidia, Adobe Ventures, Toyota Ventures, and eBay Ventures, and existing investors Sequoia Capital, Benchmark, and Zeev Ventures continuing to participate. According to Decart, total funding has exceeded $450 million, with valuation after the latest round approaching $4 billion. Three months later, the potential acquisition price became $6 billion.
Dean's ambition actually goes further. In a Fortune interview in 2025, he said: "We started the company to build a consumer application with 1 billion users." He also publicly stated he wanted to turn Decart into a trillion-dollar company. Now, on one side is a trillion-dollar dream, and on the other is a potential $6 billion acquisition. This may be the choice Dean has to make.
3. In the AI Era, Giants Are Increasingly Willing to Pay for Time
The negotiations between Anthropic and Decart also reflect an increasingly obvious trend in the AI industry: the growth and value-realization cycle of startups is being compressed rapidly. In June this year, SpaceX and Anysphere, the parent company of AI coding tool Cursor, reached an M&A agreement at an equity value of approximately $60 billion for Cursor. Cursor was founded in 2022; a few 00s MIT alumni went from founding to sitting at a $60 billion M&A negotiation table in just four years.
More notably, the pricing pace of giants and capital for AI startups is also accelerating. In the past, a company often needed years to prove revenue, market, and business model; now, as long as a team has first solved a problem that giants will inevitably have to face in the next few years, large companies may no longer be willing to recruit from scratch, build teams, and iterate repeatedly; they may acquire directly. What they buy is not just technology, but also time. Decart spent nearly three years adapting its models to different hardware such as Nvidia GPUs, Google TPUs, and Amazon Trainium, and developed its own optimization systems and engineering team. If the $6 billion deal ultimately closes, Anthropic will acquire not only Decart's technical capabilities, but also the R&D and trial-and-error that Decart has already completed.
Therefore, AI startups are entering high-valuation or even M&A stages earlier and earlier, and this is not just a result of capital enthusiasm. The faster technology iterates, the earlier scarce teams and critical capabilities are priced in. Wealth is the most intuitive reflection of this change. According to Forbes data released in March this year, among the 2026 global billionaires list, 86 people derived their primary wealth from artificial intelligence, and 45 of them became billionaires in the past 12 months, from a group of relatively young AI companies such as Perplexity, Mercor, Mistral, Cursor, Lovable, and Cognition.
What AI is truly accelerating is not just the pace of wealth accumulation, but the entire cycle from discovery, pricing, to realization of technology. In the past, a company might need ten years to prove its value; today, as long as they possess capabilities that giants urgently need but cannot easily replicate in a short time, three years are enough to bring them to a negotiation table worth billions of dollars.
The name Decart comes from the philosopher Descartes. Descartes tried to answer "how do people understand the world", and what Decart does is make AI compute and generate images more efficiently, further simulating a world that can be continuously interacted with and changed. In nearly three years since its founding, its latest valuation was around $4 billion, and now it has entered negotiations for a potential $6 billion deal. What is truly being re-priced behind this is technology and time: if a team has already solved the problems that giants must address in the coming years, then its value depends on how scarce this capability is and how long it would take competitors to replicate it. In the AI era, time is becoming the most expensive asset.
