Anthropic Wants to Buy Decart for $6 Billion. Compute Has Become the Product.

Anthropic has apparently reached the stage of startup shopping where the shopping list includes “a company that makes the computers less expensive.” This is what happens when your chatbot becomes popular enough to develop the power appetite of a small weather system.

Axios reported that Anthropic is in talks to acquire Decart for roughly $6 billion, citing Bloomberg. The deal is not closed, the terms could change, and private negotiations remain the natural habitat of the phrase “roughly.” But if it happens, it would be Anthropic’s largest known acquisition and one of the clearest signals yet that frontier AI companies are trying to own more of the machinery underneath the model.

Decart is not just another AI application startup. It builds DOS, an optimization stack designed to make inference and training run faster and more efficiently across Nvidia GPUs, Google TPUs, Amazon Trainium, AMD accelerators, and other hardware. It also builds live world models, including Lucy, which edits video in real time, and Oasis, aimed at interactive physical environments. Anthropic would therefore be buying a curious combination of compiler engineers, systems optimizers, and people who can make your video look like you are trying on a handbag in a parallel universe.

Anthropic Is Buying the Part of AI That Sends the Invoice

Most AI stories begin with the model. This one begins with the bill.

Training is the spectacular expense: vast clusters, long runs, and enough electricity to make a utility executive experience spiritual growth. Inference is the quieter expense that never stops. Every Claude answer, coding task, long-context analysis, and agentic workflow consumes compute after the model already exists. If Anthropic wants to serve more customers, add more capable features, and keep response times tolerable, it needs more capacity—or it needs every chip to do more work for the same money.

That is Decart’s pitch. Its Decart Optimization Stack says it squeezes more performance from each chip across inference, training, and hardware, with a hardware-agnostic approach covering the major accelerator families. In plain English, it is software that tries to turn an expensive pile of silicon into a less wasteful expensive pile of silicon.

That may sound less glamorous than a new reasoning model. It is also closer to where the strategic pressure is. The model is the product customers talk about. The inference stack is the part finance teams eventually ask about when the token meter starts spinning like a slot machine.

DOS Is a Speed Layer for a Very Expensive Personality

Optimization is not one magic switch. It is a long series of decisions about how models use memory, how operations are scheduled, how kernels are compiled, how workloads are batched, and how hardware is kept busy without turning latency into a hostage situation.

Decart says DOS works across different chips rather than locking customers to one vendor. That matters because the AI industry is currently discovering that dependence on a single supplier is strategically uncomfortable, especially when the supplier is also an investor, a platform, and the owner of the most intimidating roadmap in the building.

Decart’s Lucy product page reports sub-40-millisecond latency and 100 frames per second for its live video system, while describing the model as a way to make video programmable in real time. Those are company claims, not independent validation, but they show the technical heritage Anthropic would be acquiring. Decart has optimized for continuous, low-latency workloads where a delay is not merely annoying; it breaks the experience.

Claude is a different kind of workload. A coding agent can spend minutes reasoning, call tools, read files, and generate a long response. It does not need to transform a livestream before the viewer blinks. But the underlying problems rhyme: keep the hardware utilized, minimize wasted work, and make the model feel faster than the infrastructure deserves.

The World Models Are the Shiny Object. The Kernels Are the Business.

Decart’s world-model work is the part most likely to get the glossy acquisition video. Lucy can alter live video, insert products, change environments, and restyle scenes while the stream is happening. Oasis is aimed at interactive worlds and physical AI. It is easy to imagine the keynote: a person walks through a virtual shop, the machine changes the furniture, and an executive says “spatial intelligence” with the confidence of someone who has never assembled a bookshelf.

Anthropic may care about those models. The more immediate prize is likely the optimization layer and the team behind it. Axios described the deal as a way for Anthropic to control compute costs or even help design its own chips as it heads toward a reported autumn IPO. The acquisition would place Decart’s staff inside Anthropic’s inference and performance organization if completed.

That is an important distinction. Anthropic would not be buying a cute video app to diversify Claude’s personality. It would be buying expertise that could make Claude cheaper to run, easier to deploy across available hardware, and less exposed to the bottlenecks of a single accelerator ecosystem.

As SiliconSnark’s AI infrastructure guide keeps insisting, the plumbing is the point. The model gets the stage. The plumbing decides whether the stage can afford another performance.

This Is What Happens When Your Vendor Becomes Your Competitor

There is a strategic irony here. Nvidia is one of Decart’s reported investors, and Anthropic relies heavily on the broader accelerator ecosystem. A company that began life as a model lab is now considering buying the layer that helps models run across Nvidia, TPU, Trainium, and AMD hardware. The customer is looking at the supplier’s menu and quietly asking whether it should learn to cook.

That does not mean Anthropic is about to become Nvidia. It means the frontier labs increasingly have to care about systems engineering, procurement, scheduling, memory, networking, and utilization. The software abstraction is leaking. It turns out “just call the model” is not a complete business plan when the model costs a great deal of money to answer.

We have already watched Etched make specialized inference hardware the entire product, while Broadcom, Apollo, and Blackstone turned compute into a finance product. Anthropic buying Decart would add another layer: the model company buying the optimization team so it can negotiate with physics from a slightly stronger position.

Six Billion Dollars Is a Lot of Optimization

The price is where the story stops being tidy.

Decart reportedly raised more than $450 million, including a $300 million round in May at a valuation near $4 billion. A roughly $6 billion acquisition price would be an enormous premium for a company whose most valuable asset may be a specialized engineering team and a set of optimization systems that must keep evolving as models and hardware change.

That can still be rational. If better optimization saves Anthropic billions in capacity, improves response times, and reduces the need to overbuy scarce accelerators, the purchase could pay for itself in the least cinematic way possible: through fewer invoices. If the team helps Anthropic run across more hardware, it also gives the lab bargaining power and resilience when supply, prices, or export rules shift.

But acquisitions do not magically preserve speed. The startup that moved quickly as an independent team becomes part of a frontier lab with security reviews, research priorities, product roadmaps, and a growing collection of people who need to be invited to the meeting about the meeting. The optimization advantage has to survive integration, model changes, and the gravitational pull of a company preparing for a very large public-market narrative.

Verdict: A Real Shift Wearing a Very Expensive Hoodie

Anthropic’s reported Decart talks feel like a real strategic shift, even though they are still only talks. The frontier AI race is moving from “who has the smartest model?” toward “who can run useful intelligence at acceptable speed, cost, and scale?” That is a less romantic question and a much more important one.

Decart gives Anthropic an intriguing answer: optimize the stack, diversify the hardware, and keep the live-model experiments around in case the future turns out to involve video that edits itself before you can complain about it.

My verdict is cautiously impressed. The world-model layer is shiny, the acquisition price is bonkers, and the deal may never happen. But the core logic is sound. Anthropic is not merely shopping for another model. It is trying to buy control over the cost of being a model company.

That is the moment AI becomes infrastructure: when the chatbot stops asking what it can do and the CFO starts asking which chips it can run on. The future may still be intelligent. It is also apparently going to be optimized, benchmarked, and invoiced by the millisecond.