Anthropic은 Decart 인수 계획을 철회하며 Claude 운영을 위한 컴퓨팅 비용 격차 해법을 미결 상태로 남겼다.

Anthropic has abandoned a roughly $6 billion acquisition of Israeli AI startup Decart following the completion of due diligence, Bloomberg reported Monday. The move ends what would have been the Claude developer’s largest purchase and leaves a critical compute-efficiency gap unresolved—a question Wall Street investors are expected to press ahead of the company’s planned Nasdaq debut next month.

The decision blindsided Decart’s founders, who had previously rejected a reportedly larger, primarily cash offer from Nvidia to pursue Anthropic’s stock-heavy bid. Their wager was that equity in a company poised for one of the most anticipated U.S. IPOs would ultimately yield greater value. That calculation has not yet materialized, leaving Decart without a buyer. Neither company has issued a public statement. Representatives for both Anthropic and Decart declined to comment when contacted by Bloomberg.

**What Decart Built — and Why Anthropic Wanted It**

Founded on September 7, 2023, Decart was established by Dean Leitersdorf and Moshe Shalev, both veterans of Israel’s elite signals intelligence unit Unit 8200, alongside Chief Technology Officer Orian Leitersdorf. CEO Dean Leitersdorf holds a doctorate in computer science from the Technion—Israel’s premier technical university—which he earned at age 23, making him one of the institution’s youngest doctoral graduates. He later completed a postdoctoral fellowship at the National University of Singapore. CPO Moshe Shalev spent approximately 13 years in Unit 8200, where he oversaw AI operations, having grown up in a Haredi community and studied accounting part-time.

While Decart’s consumer-facing applications—Lucy, a near-real-time live video transformation platform, and Oasis, an AI world model for physical simulation that garnered one million users within 72 hours of its Minecraft-inspired October 2024 demo—generated significant attention, they were not the primary draw for Anthropic. Instead, the company targeted DOS: the Decart Optimization Stack.

DOS is a hardware-agnostic platform designed to optimize both inference and training. Unlike most AI software engineered for specific chip architectures, DOS operates seamlessly across Nvidia GPUs, Google TPUs, and Amazon Trainium processors. This flexibility allows AI laboratories to shift workloads between different silicon suppliers without requiring ground-up re-optimization.

According to Radical Ventures, which led Decart’s latest funding round, DOS 2.0 can process over 1,600 tokens per second for AI agents, against an industry average Decart estimates at roughly 200. If validated across Anthropic’s specific model architecture and workloads, this would equate to an eightfold increase in inference throughput. However, Decart’s Cogito API—which exposes DOS to external developers—reports real-world performance exceeding 1,000 tokens per second, indicating that the 1,600-token peak may apply only to highly optimized scenarios. These figures originate from investor pitch materials and corporate disclosures; no independent audit has verified them at scale.

**The Margin That Explains Everything**

The strategic importance of an eightfold inference gain boils down to a single metric: compute cost per revenue dollar. In the first quarter of 2026, Anthropic allocated 71 cents of every revenue dollar to computing expenses. For the second quarter, the company projects that ratio will drop to 56 cents. This 15-cent quarterly reduction is central to Anthropic’s forecast of generating its first-ever Q2 operating profit of approximately $559 million against roughly $10.9 billion in revenue.

This margin expansion is being driven by major hardware partnerships: a $25 billion infrastructure commitment from Amazon, access to millions of Google TPUs, and a July agreement with AMD. Under the AMD deal, the chipmaker will invest up to $5 billion in Anthropic, while Anthropic commits to purchasing up to two gigawatts of AMD’s latest Instinct MI450 processors, with deployment slated for no earlier than the first half of 2027.

DOS offered an alternative, potentially accelerated pathway: a software-layer solution capable of extracting significantly higher throughput from Anthropic’s existing hardware inventory, bypassing the wait for next-generation silicon. However, McKinsey’s 2026 inference cost analysis noted that no single technological breakthrough will likely provide the step-change required for frontier model providers to achieve scalable positive margins. Sustained progress will instead depend on coordinated innovation across the entire supply chain, spanning software optimization to advanced silicon design. Across the industry, inference currently represents an estimated 80 to 90 percent of a production AI system’s lifetime compute expenditure.

**Why Investors Are Watching — and Asking Hard Questions**

Investors positioning for Anthropic’s upcoming Nasdaq listing have voiced clear concerns. During August investor meetings focused on margins and competitive dynamics, prospective backers questioned executives on the durability of gross margins amid intensifying competition from lower-cost alternatives and escalating data center infrastructure costs.

Anthropic confidentially submitted its S-1 registration statement to the SEC on June 1, 2026, triggering a regulatory quiet period that limits public commentary on financial performance. This restriction also complicates M&A activity: a $6 billion acquisition would necessitate amending the S-1 filing and could invite additional SEC scrutiny precisely when the company needs to maintain a steady narrative ahead of its public offering.

According to investors speaking with the Financial Times, market participants anticipate an IPO valuation of $2 trillion or higher. Such a figure would surpass SpaceX’s $1.77 trillion debut in June and establish the largest U.S. public offering in history. Achieving that benchmark requires Anthropic to convincingly demonstrate a continued downward trajectory in compute costs—proving that each incremental dollar of revenue becomes progressively less expensive to generate.

By walking away from a target that claimed the market’s most dramatic inference cost improvements, Anthropic has left a critical question unanswered—one it cannot address publicly during the quiet period: What is its explicit software-layer strategy for compute efficiency, and when will it bridge the gap DOS was intended to close?

**Decart's Strategic Miscalculation**

For Decart, the collapsed transaction represents a particularly sharp setback. The startup had secured approximately $450 million in total capital, highlighted by a $300 million Series B round in May 2026 led by Radical Ventures, which placed a roughly $4 billion valuation on the company. Its investor syndicate features prominent firms including Sequoia Capital, Benchmark, Nvidia, Adobe Ventures, Toyota Ventures, Zeev Ventures, eBay Ventures, Atreides Management, and Valor Equity Partners.

Founders Dean Leitersdorf, Moshe Shalev, and Orian Leitersdorf retained approximately 64 percent of the company during negotiations—an exceptionally high founder ownership percentage for a venture that had raised nearly half a billion dollars. This concentrated equity provided substantial leverage in selecting among competing suitors.

The founding team ultimately selected Anthropic over Nvidia, reportedly turning down a $7 billion to $8 billion predominantly cash offer in exchange for Anthropic’s $6 billion stock-denominated proposal. The rationale was straightforward: pre-IPO equity in a company targeting a $2 trillion valuation promised significantly greater upside than immediate liquidity.

Following the rejection, Nvidia pursued its own strategic moves. On September 2, 2026, the chipmaker finalized a $12.9 billion acquisition of Hugging Face—a transaction that may dampen its appetite for pursuing another large-scale deal at the valuation thresholds Decart’s stakeholders would likely demand.

**What Comes Next**

Bloomberg’s reporting, citing individuals familiar with the situation, indicates that neither side has closed the door on future cooperation. Rather than a full acquisition, the companies may explore alternative collaborations. A licensing agreement, commercial partnership, or minority equity stake would grant Anthropic access to DOS while sidestepping the regulatory complications of a merger during the IPO quiet period.

Anthropic’s internal compute strategy is advancing across multiple vectors. An August 2026 initiative to recruit engineers for custom chip co-design aims to solve the same inference efficiency challenges addressed by DOS, albeit over a longer development horizon. Meanwhile, the rollout of AMD’s Instinct MI450 processors will not commence until the first half of 2027, extending well beyond the IPO window.

The broader AI inference landscape continues to evolve rapidly. McKinsey projects that scaling down inference costs will demand a synchronized wave of innovation spanning software model optimization, advanced silicon architecture, and next-generation memory systems—advancements unlikely to be captured through a single acquisition. Nevertheless, an optimization stack capable of delivering an eightfold speedup on existing hardware, if validated against Anthropic’s specific workload profiles, would remain one of the most direct routes to near-term efficiency gains.

Decart, for its part, now faces the immediate challenge of recalibrating its commercial strategy without either strategic partner, while managing the expectations of a deeply involved investor syndicate.