Anthropic's Historic IPO Approaching? Plans $6 Billion Acquisition of Decart, Valuation Challenges $2 Trillion

On Thursday, June 13, a highly anticipated potential acquisition emerged in the artificial intelligence (AI) industry. According to international media outlets such as Bloomberg and Reuters, AI startup Anthropic is in discussions to acquire Israeli AI startup Decart, with the deal potentially reaching $6 billion. If finalized, this would mark the most significant valuation leap for Decart since its founding nearly three years ago.

Meanwhile, the Financial Times reported that some of Anthropic’s investors expect the company could go public as early as October this year, with an IPO valuation potentially reaching $2 trillion or higher.

If this expectation materializes, Anthropic would surpass SpaceX, which went public in June at a valuation of approximately $1.77 trillion, becoming the company with the highest IPO valuation in history.

Analysts suggest that Anthropic’s most pressing challenge is no longer simply expanding revenue, but rather managing the rapidly growing computational demands behind its high-speed growth.

What Decart possesses is precisely the capability to make AI models run more efficiently across different chip architectures.

Why is Anthropic interested in Decart?

Anthropic’s growth pace this year has been remarkable.

In February, the company completed a $30 billion Series G funding round, achieving a post-money valuation of $380 billion, with an annualized revenue of approximately $14 billion at the time. By April, annualized revenue reportedly reached $30 billion.

By the end of May, Anthropic completed another $65 billion Series H funding round, pushing its post-money valuation to $965 billion, with annualized revenue surpassing $47 billion.

In just a few months, Anthropic’s valuation more than doubled, and its annualized revenue increased over threefold.

On June 1, Anthropic secretly filed an IPO draft application (Form S-1) with the U.S. Securities and Exchange Commission (SEC), shifting market attention to whether the company can sustain its rapid growth in the public markets.

The Financial Times noted that some Anthropic investors anticipate the company’s annualized revenue could reach $100 billion to $120 billion by the end of 2026. One investor even estimated, based on revenue growth and valuation multiples, that Anthropic’s theoretical valuation could reach $3 trillion.

This rapid growth has rapidly inflated the wealth of Anthropic’s early shareholders. According to Forbes’ late May estimates, seven co-founders, including CEO Dario Amodei and President Daniela Amodei, each hold slightly over 1.6% of the company. At a $965 billion valuation, each founder’s stake is worth approximately $15.5 billion.

However, for Anthropic, the flip side of rapidly rising valuation and revenue is equally fast-growing costs.

Computational demands for large models can be divided into two main parts: training and inference. Model training requires massive data and chip resources, while every user query, code generation request, or document processing task handled by Claude after deployment consumes computational power again.

Unlike training, which is concentrated in a specific phase, inference demand grows continuously with the number of users and usage frequency. The more widely Claude is used, the greater the inference compute Anthropic must invest.

Therefore, Anthropic has been expanding its compute infrastructure over the past few months.

In April, the company signed a new agreement with Amazon (AMZN-US), committing to invest over $100 billion in AWS technology over the next decade, securing up to 5GW of compute capacity. The same month, it expanded cooperation with Google (GOOGL-US) and Broadcom (AVGO-US), securing several GW of next-generation TPU capacity. In May, Anthropic reached a compute collaboration with SpaceX (SPCX-US).

Currently, Claude runs simultaneously on Amazon Trainium, Google TPU, and NVIDIA (NVDA-US) GPU platforms.

The problem is that different chips use different hardware architectures. Migrating a model from one chip platform to another isn’t as simple as swapping hardware—the underlying code, computation methods, and data transmission and communication mechanisms may all need reconfiguration and optimization.

This is precisely where Decart’s core expertise lies.

One of Decart’s key technologies is the DOS (Decart Optimization Stack), which functions as a performance optimization layer between AI models and chips. Its core function is to reconfigure model computation, chip invocation, and data transmission methods for different hardware architectures, enabling the same AI model to run more efficiently across different platforms.

Currently, DOS already supports NVIDIA GPUs, Google TPUs, and Amazon Trainium.

In other words, regardless of which AI chip Anthropic adopts, it will face the challenge of fully leveraging hardware performance, and Decart possesses exactly the technical and engineering expertise for this.

Just on August 5, Anthropic announced it had begun assembling its own chip design team, recruiting engineers with both hardware and software capabilities, aiming to improve Claude’s computational efficiency through custom chips while maintaining a multi-chip strategy.

A few days later, news of the potential Decart acquisition surfaced.

According to Reuters, if the deal is completed, the Decart team will be integrated into Anthropic’s inference and performance division.

Analysts believe that what Anthropic is after may not just be Decart’s AI products, but rather an engineering team that deeply understands models, chips, and low-level system optimization.

For a company preparing to enter the public market with a potential valuation challenging $2 trillion, model capability is certainly important, but the cost per inference is equally critical.

The more people use Claude, the more Anthropic must consider how to increase compute investment while maximizing output per dollar of compute spending.

Decart: From GPU Optimization to Market Presence

Decart was founded in September 2023.

Co-founder and CEO Dean Leitersdorf previously conducted postdoctoral research at the National University of Singapore, where he and Moshe Shalev prepared their entrepreneurial team. Earlier, Leitersdorf studied at the Technion – Israel Institute of Technology, earning a Ph.D. in computer science at age 23. During his studies, he also served in Unit 8200 of the Israel Defense Forces.

It was in Unit 8200 that Leitersdorf met Moshe Shalev.

However, their backgrounds differ significantly. Leitersdorf followed an academic and technical research path; Shalev grew up in an ultra-Orthodox Jewish family in Israel, worked various jobs in his youth, and studied accounting at night while working. He enlisted at age 23, joined Unit 8200, served for 13 years, and gradually transitioned into AI system construction and management.

After founding Decart, the company did not immediately focus entirely on building its own large AI models but instead first helped other companies improve AI model inference efficiency on GPUs.

According to a previous report by The Information, about three months after its founding, Decart secured a multi-million-dollar contract from a GPU cloud service provider to help improve AI model GPU inference efficiency. By October 2024, related business annualized revenue had exceeded $10 million.

More notably, even as Decart began investing in training its own video foundation model, the company maintained positive free cash flow.

This allowed Decart to follow a different path from many AI startups: first generating revenue by leveraging its deep understanding of GPU and model efficiency, then reinvesting that revenue into its own model development.

Shaun Maguire, a partner at Sequoia Capital, described Decart’s understanding of GPUs as reaching extremely low-level hardware operation layers.

Sequoia even compared this capability to Google’s early development model. Google’s competitive advantage stemmed not only from PageRank but also from its distributed systems capability, enabling it to build systems from large numbers of relatively inexpensive hardware while maximizing the efficiency of each unit of compute.

In Sequoia’s view, Decart may be following a similar path in the AI era: starting from low-level efficiency and transforming that capability into products.

The Real-Time Generation Power Behind Oasis and Lucy

Decart’s first product to gain market attention was Oasis, launched in October 2024.