Anthropic Nears $7B Acquisition of AI Infrastructure Startup Decart Ahead of IPO

Anthropic is finalizing negotiations to acquire Israeli artificial intelligence infrastructure startup Decart in a transaction valued at approximately $7 billion, according to reporting from Calcalist and Reuters. The acquisition, expected to be settled primarily in Anthropic equity, would mark the Claude developer's largest purchase to date as it prepares for a planned initial public offering. Founded in September 2023 by Dr. Dean Leitersdorf and Moshe Shalev, Decart specializes in hardware-ag

2 min
Anthropic Nears $7B Acquisition of AI Infrastructure Startup Decart Ahead of IPO

Anthropic is finalizing negotiations to acquire Israeli artificial intelligence infrastructure startup Decart in a transaction valued at approximately $7 billion, according to reporting from Calcalist and Reuters. The acquisition, expected to be settled primarily in Anthropic equity, would mark the Claude developer's largest purchase to date as it prepares for a planned initial public offering.

Founded in September 2023 by Dr. Dean Leitersdorf and Moshe Shalev, Decart specializes in hardware-agnostic inference optimization software alongside real-time generative world models. The 100-person startup has raised $450 million to date, backed by investors including Sequoia Capital and Nvidia.

Hardware-Agnostic Inference and Real-Time World Models

Decart's core technology focuses on extracting higher throughput from AI accelerators without requiring architecture-specific rewrites. The company's optimization stack operates across Nvidia graphics processing units, Google Tensor Processing Units, and Amazon Inferentia and Trainium silicon. Internal benchmarks indicate the software can increase inference execution speeds by up to eightfold compared to baseline serving configurations.

In addition to compute optimization, Decart developed generative world models designed for physical AI and interactive simulations. Its Lucy model performs real-time video editing and high-resolution synthetic rendering, while its Oasis architecture generates dynamic interactive environments for training autonomous robotics systems.

Anthropic Decart Multi-Chip Architecture

Bidding Dynamics and Israeli R&D Footprint

The agreement follows competitive interest from major industry players. Decart previously held advanced acquisition talks with Nvidia, which offered a higher headline valuation than Anthropic. However, Decart's founders and lead investor Sequoia Capital selected Anthropic's proposal due to strategic alignment and the upside potential of Anthropic equity. Google and SpaceX also evaluated potential bids.

Under the proposed transaction structure, Decart's team of approximately 89 engineers in Tel Aviv and 17 in the United States will join Anthropic's inference and performance organization. The Tel Aviv facility will become Anthropic's first engineering and research center in Israel, serving as its second primary international development hub after London, where the company operates a 200-person facility.

Decart hired boutique investment bank Catalyst to advise on the sale. Anthropic is working with J.P. Morgan, Morgan Stanley, and Goldman Sachs on transaction structuring and its broader capital markets roadmap.

Compute Economics Ahead of Public Listing

The acquisition arrives at a critical juncture for Anthropic's capital strategy. The company submitted a confidential S-1 registration statement to the U.S. Securities and Exchange Commission in June, targeting a public listing as early as September or October 2026 with an anticipated valuation between $1 trillion and $2 trillion.

Investor presentations indicate Anthropic generated $11.5 billion in second-quarter revenue, representing a fourteenfold increase over the $787 million reported in the second quarter of 2025, while achieving positive EBITDA. Industry data places Anthropic's annualized revenue run rate between $47 billion and $65 billion.

As enterprise AI demand transitions from model pre-training toward high-volume production inference, optimizing token serving costs across diverse accelerator fleets has become central to operating margins. Incorporating Decart's compiler and runtime optimizations directly into Anthropic's infrastructure stack is intended to reduce per-query compute overhead across Claude's hosted developer and enterprise tiers.

Sources

Written by

More to read

  • Sparse Attention and BigBird: How Window, Global, and Random Graphs Preserve Turing Completeness in Linear Time

    Standard self-attention in transformer architectures scales quadratically with sequence length. Computing full pairwise interactions between n tokens requires evaluating an n x n attention matrix, yielding O(n^2) computational complexity and memory consumption. While hardware accelerators and IO-aware tiling algorithms like FlashAttention optimize memory traffic, the quadratic compute and KV footprint remains a barrier for processing long contexts, document-level summarization, and genomic seque

    1 min
  • Oxford Study Details Chinese Gray-Market Proxies Reselling Claude Tokens at 90% Discounts

    An investigation by the Oxford China Policy Lab reveals that Chinese developers routinely access Anthropic's frontier Claude models at discounts between 70% and 90% below list price, bypassing geographical blocks, payment filters, and biometric identity verification through a decentralized network of API proxies known locally as "transfer stations" (中转站). The analysis, authored by Oxford researcher Zilan Qian and published via ChinaTalk, outlines the modular supply chain and economic mechanics

    1 min
  • Low-Precision Quantization Kernels in Production: Comparing Marlin, ExLlamaV2, FlashInfer, and BitBLAS

    Low-Precision Quantization Kernels in Production: Comparing Marlin, ExLlamaV2, FlashInfer, and BitBLAS Architecture, Memory Bandwidth, and Decoding Throughput Autoregressive large language model (LLM) serving operates under two distinct compute regimes: a compute-bound prefill phase and a memory-bandwidth-bound decode phase. While processing the initial prompt involves matrix-matrix multiplications (GEMM) with high arithmetic intensity, generating tokens one by one requires matrix-vector multip

    1 min