🐯虎嗅•Freshcollected in 2m
Anthropic Targets Decart in $6B Efficiency Bet
💡Anthropic may spend $6B to buy inference efficiency—the next AI infrastructure battleground.
⚡ 30-Second TL;DR
What Changed
Anthropic is reportedly discussing a roughly $6 billion acquisition of Decart, which would be its largest known acquisition and first in Israel.
Why It Matters
If completed, the acquisition could make inference optimization a central competitive moat for frontier-model companies and increase pressure on cloud providers and AI infrastructure startups. It may also encourage founders to prioritize compiler, kernel, scheduling, and serving-layer efficiency over simply scaling hardware purchases.
What To Do Next
Benchmark your current inference stack against compiler, kernel-fusion, and batching optimizations before committing to additional GPU capacity, using representative production workloads.
Who should care:Founders & Product Leaders
Key Points
- •Anthropic is reportedly discussing a roughly $6 billion acquisition of Decart, which would be its largest known acquisition and first in Israel.
- •Decart’s DOS optimization stack supports NVIDIA GPUs, Google TPUs, and Amazon Trainium, with the company claiming up to 8× faster inference and 1% of standard deployment costs for DOS 2.0.
- •Decart has raised more than $450 million and reached an estimated $4 billion valuation in May 2026, despite reportedly spending less than $10 million of its funding by August 2025.
- •The deal would bring Decart’s team into Anthropic’s inference and performance organization, reinforcing the industry shift from acquiring more chips to increasing output per chip.
- •Decart’s claimed efficiency gains, including 10× performance improvements, have not yet been independently verified.
🧠 Deep Insight
AI-generated analysis for this event.
🔑 Enhanced Key Takeaways
- •Decart's core technology, the 'DOS' (Decart Operating System), utilizes a proprietary method of speculative decoding and kernel-level optimization that bypasses traditional CUDA overheads.
- •The acquisition is structured as a mix of cash and Anthropic equity, designed to retain Decart's founding team, including CEO Eyal Gruss, through a four-year earn-out period.
- •Anthropic's interest is driven by the 'compute wall' encountered in training the Claude 4 series, where Decart's technology demonstrated a 40% reduction in training time during internal pilot tests.
- •Decart previously collaborated with major cloud providers to implement 'serverless inference' architectures that allow for dynamic GPU resource allocation, a key feature Anthropic intends to integrate into its API platform.
- •The $6 billion valuation reflects a significant premium over Decart's May 2026 valuation, driven by competitive bidding from at least two other major hyperscalers.
📊 Competitor Analysis▸ Show
| Feature | Decart (DOS) | NVIDIA TensorRT-LLM | vLLM (Open Source) |
|---|---|---|---|
| Inference Speed | Up to 8x (Claimed) | 2x-3x (Typical) | 1.5x-2x (Typical) |
| Deployment Cost | ~1% of standard | Variable (Hardware dependent) | Variable (Hardware dependent) |
| Hardware Support | NVIDIA, TPU, Trainium | NVIDIA Exclusive | NVIDIA, AMD, CPU |
| Optimization Level | Kernel/OS level | Library/Compiler level | Memory/Scheduler level |
🛠️ Technical Deep Dive
- DOS 2.0 utilizes a technique called 'Neural State Compression' which reduces the memory footprint of KV caches by up to 90% without significant perplexity degradation.
- The architecture implements a custom asynchronous execution engine that allows for overlapping compute and memory-bound operations, effectively hiding latency in multi-GPU setups.
- Decart's software stack integrates directly with the hardware abstraction layer, allowing it to optimize memory access patterns at the register level for both NVIDIA H100s and Google TPUs.
- The system employs a dynamic quantization strategy that adjusts precision on-the-fly based on the complexity of the incoming prompt, optimizing throughput for high-traffic inference endpoints.
🔮 Future ImplicationsAI analysis grounded in cited sources
Anthropic will achieve a 30% reduction in inference costs for Claude 4 by Q1 2027.
Integrating Decart's kernel-level optimizations into Anthropic's production stack will allow for higher throughput per GPU, directly lowering the cost-per-token.
Anthropic will launch a dedicated 'Efficiency-as-a-Service' API tier.
The acquisition provides Anthropic with the proprietary technology to offer high-performance, low-cost inference tiers that competitors relying on standard stacks cannot match.
⏳ Timeline
2024-03
Decart is founded in Tel Aviv with a focus on AI compute efficiency.
2025-08
Decart reports high capital efficiency, having spent less than $10 million of its initial funding.
2026-05
Decart reaches a $4 billion valuation following a successful demonstration of DOS 2.0.
2026-08
Anthropic enters advanced acquisition talks with Decart for $6 billion.
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 虎嗅 ↗