Anthropic Eyes $6 Billion Decart Acquisition in Bid to Control Inference Costs

Anthropic is in advanced negotiations to acquire Israeli AI startup Decart for approximately $6 billion, according to a Bloomberg report published August 13, 2026. If the deal closes, it would mark the largest acquisition in Anthropic’s history and signal a dramatic strategic shift: frontier AI labs are no longer just competing on model quality — they are racing to own the infrastructure layer that determines how cheaply and quickly those models can run.

The Target: Decart’s Optimization Stack

Founded in 2023 by Dean Leitersdorf and Moshe Shalev, Decart has built what it calls the Decart Optimization Stack (DOS) — a software engine designed to extract peak performance from GPUs, TPUs, and custom AI accelerators. The platform tackles the thorniest problem in AI infrastructure: the abstraction layer between models and silicon. By optimizing how AI workloads are scheduled, compiled, and executed across different hardware vendors, DOS can significantly reduce both latency and total cost of ownership for inference workloads.

What makes Decart particularly valuable is its hardware-agnostic approach. The company’s software helps developers switch between different processors — including Nvidia GPUs, AMD accelerators, and AWS Trainium chips — with minimal re-engineering. In an industry where GPU vendor lock-in has become a multi-billion-dollar pain point, this portability layer is strategically priceless. The Wall Street Journal noted in May 2026 that Nvidia itself participated in Decart’s $300 million funding round despite the startup’s mission to make it easier to move away from Nvidia hardware — a paradox that underscores just how critical optimization technology has become to the entire AI ecosystem.

Decart also operates a research division focused on real-time world models. Its flagship product, Lucy 2.5, generates live video transformations at 1080p and 30 frames per second, enabling applications from virtual try-on to live AI-generated visual effects. While the Lucy product line has generated significant consumer attention, it is the underlying optimization technology that makes Decart worth $6 billion to a company like Anthropic.

Why Anthropic Needs This

The acquisition talks come against a backdrop of escalating compute costs across the AI industry. By mid-2026, inference — not training — accounts for an estimated 80% of all AI GPU spending. As models grow larger and agentic workloads multiply the number of API calls per user interaction, the economics of serving AI at scale have become existential. A single agentic task might trigger dozens of model invocations, each consuming tokens and compute cycles that must be served in real time.

Anthropic has already made aggressive moves on the compute front. In April 2026, the company announced an expanded partnership with Amazon, committing $100 billion to AWS over the coming years and securing up to 5 gigawatts of Amazon Trainium compute capacity. Anthropic committed to running its large language models on Trainium chips for the next decade, betting that custom silicon from AWS could deliver inference at roughly half the cost of competing GPU-based solutions. The Trainium2 platform runs at an effective committed cost of approximately $0.50 per chip-hour, according to industry analysts.

But buying raw compute capacity is only half the equation. The other half is software optimization — squeezing maximum throughput out of every chip, routing workloads to the most cost-effective hardware available at any given moment, and minimizing the overhead of cross-platform model deployment. That is precisely what Decart’s DOS engine delivers. By bringing this technology in-house, Anthropic could potentially reduce its inference costs by a meaningful margin across its entire fleet of Trainium, GPU, and future accelerator deployments.

According to a source familiar with the discussions, if the deal closes, Decart’s team would join Anthropic’s inference and performance organization — a signal that Anthropic views this not as a talent acquisition or a product play, but as a foundational infrastructure investment.

The Competitive Bidding War

Anthropic was not the only suitor. Decart had reportedly been in advanced talks with multiple parties, with SpaceX — which absorbed Elon Musk’s xAI earlier in 2026 — Amazon, and cloud provider Nebius all expressing interest at valuations ranging from $6 billion to $7 billion. The involvement of SpaceX is particularly notable: Musk’s AI operations require massive inference infrastructure, and Decart’s optimization layer could provide a critical edge in the increasingly bitter rivalry between xAI and Anthropic.

The fact that Anthropic appears to be winning the bidding war speaks to its financial firepower. Backed by Amazon’s multi-billion-dollar investments and reportedly preparing for an IPO, Anthropic has the capital to outbid strategic competitors. A $6 billion price tag would represent a roughly 50% premium over Decart’s last private valuation of nearly $4 billion, set during its May 2026 Series C round led by Radical Ventures with participation from Nvidia, Sequoia Capital, Benchmark, and Zeev Ventures.

Broader Industry Implications

The Anthropic–Decart talks reflect a fundamental restructuring of the AI industry’s value chain. For the past two years, the dominant narrative has been about scaling laws — bigger models, more data, more training compute. But as the industry matures, the battlefield is shifting to the inference layer, where the economics of serving models at billion-user scale will determine which companies can sustain their growth trajectories.

Google’s Gemini app recently crossed 1 billion monthly active users, while ChatGPT has surpassed 1 billion weekly users. At that scale, even marginal improvements in inference efficiency translate into hundreds of millions of dollars in savings annually. Companies that control their optimization stack — from custom silicon to kernel-level software — gain a structural cost advantage that compounds over time.

This acquisition also signals that the era of AI labs relying entirely on third-party infrastructure is ending. Anthropic’s moves — the AWS Trainium commitment, the Theseus Infrastructure data center partnership with Macquarie and GIC, and now the potential Decart acquisition — collectively represent a sweeping vertical integration strategy. The company is building ownership across every layer of the AI compute stack: chips, data centers, optimization software, and the models themselves.

For Decart’s investors, a $6 billion exit would represent a remarkable return. The company raised just over $450 million in total funding across its three-year history, meaning the acquisition would deliver roughly a 13x return on invested capital at the headline valuation. Sequoia Capital, which led Decart’s early rounds, and Radical Ventures, which led the $300 million Series C, stand to see the largest absolute gains.

What Comes Next

The deal remains in its early stages, and Bloomberg cautioned that talks could still fall apart. Regulatory scrutiny is likely to be intense — a $6 billion acquisition by one of the world’s most prominent AI companies, of a startup backed by Nvidia and serving multiple competitors, will inevitably draw antitrust attention from both U.S. and European regulators.

If the deal does close, it will accelerate an industry-wide consolidation of the AI infrastructure layer. Expect other frontier labs — OpenAI, Google DeepMind, Meta — to respond with their own optimization acquisitions or aggressive internal buildouts. The inference wars have only just begun, and the companies that control the software stack between models and silicon will hold the keys to profitable AI at planetary scale.