Anthropic’s Decart talks point to a new AI moat: buying inference-efficiency, not just more compute

Reuters reports that Anthropic is in talks to buy Nvidia-backed Decart AI; the report describes the deal as early stage and notes that Anthropic declined to comment while Decart did not respond to Reuters outside regular hours. Reuters also describes the acquisition as part of Anthropic’s broader efforts to scale computing power and improve Claude’s performance efficiency.

On the technology side, Decart publicly markets an “Optimization Stack” designed to squeeze more performance out of GPUs and other accelerators for both inference and training, emphasizing faster, cheaper operation and higher utilization.

Decart’s positioning helps explain why an AI lab might pay acquisition-sized attention to it.

At a high level, compute supply for model inference is not just about chip availability. It includes (1) scheduling and runtime efficiency, (2) kernel/compiler-level performance, (3) workload-to-hardware mapping, and (4) utilization during real traffic (where latency budgets and throughput targets collide).

Decart’s publicly described Optimization Stack emphasizes extracting peak performance from multiple accelerator types and focusing on latency, throughput, and cost requirements for continuous, real-time AI. That implies the “unit economics lever” is not only token pricing—it’s tokens per unit time per accelerator, i.e., what fraction of expensive silicon time actually turns into useful output.

For frontier labs, scaling revenue is often constrained by how quickly inference can be produced under real-world traffic.

If optimization improves effective throughput or reduces tail latency, two things can happen in practice: 1) the same cluster can serve more concurrent users (or larger context windows) without proportionally increasing silicon demand; and 2) the lab can better keep latency within product targets, reducing churn and increasing usage per account.

The moat angle comes from iteration speed. When an optimization layer is developed externally (as a vendor or open software), a lab’s ability to customize it to its own model architectures, runtime patterns, and workload mix can be slower. Acquiring the team and stack can shorten that feedback loop.

So the investor bet is less “Decart is good at optimization” and more Anthropic can internalize the utilization lever and lock in better cost-to-output conversion.

Decart’s “Optimization Stack (DOS)” page describes squeezing every ounce of performance from chips across inference and training, operating across GPUs and other accelerator families (and emphasizing “latency, throughput, and cost requirements of continuous, real-time AI”).

On top of that, Decart’s main site also emphasizes very high efficiency and includes an example collaboration claim with Comcast, describing sub-35ms generative AI experiences at the edge powered by Nvidia GPUs.

These public claims don’t prove the economic outcome inside Anthropic—but they do establish that Decart’s core competence is the exact kind of layer that can translate into better effective inference capacity.

In the short term, investors should watch for signals that Anthropic can accelerate engineering integration: team retention, migration timelines, and whether Anthropic begins productizing internal efficiency improvements faster than peers.

In the long run, the thesis depends on whether the efficiency layer becomes a compounding advantage across model updates—not a one-time optimization bump. The strategic question is whether Anthropic builds a repeatable performance toolkit that scales with successive model generations and new workload types (coding agents, multimodal assistants, edge/real-time experiences).

If Anthropic’s acquisition succeeds, it could change bargaining power across the AI supply chain. Labs that internalize inference efficiency can become less dependent on “brute force” capacity expansion and can potentially negotiate compute arrangements differently.

For infrastructure providers, the consequence is twofold: they may face increased competition on effective cost per token, but they also gain customers who can better scale workloads (including enterprise offerings) when latency and throughput improve.

This is why the right comparison set is not only GPU manufacturers, but also large cloud platforms that sell managed inference and AI platforms.