Anthropic Looks to Buy Decart for $6 Billion

Our top story is Anthropic’s reported talks to acquire Decart AI for roughly $6 billion, potentially its largest deal. Beyond world models, Decart could help squeeze more work from Nvidia hardware, TPUs, and Amazon chips as vendors scramble for compute. With an IPO possibly near, is Anthropic buying missing stack pieces?

Anthropic has been disciplined. Dario Amodei has never sounded eager to yolo into huge compute spending, and its acquisition record is quieter than rivals’. A deal this large would be a notable, targeted shift into efficiency.

They’ve been looking, including at hardware-layer opportunities. Decart is a scarcity play: few scaled companies improve inference and chip efficiency. Before a public listing, this may be easier—and give investors a clearer cost-efficiency story.

You don’t have to own mega-caps for AI. The buildout reaches power, grid equipment, water, and treatment. Mid-caps offer more dispersion and trade near 17 times earnings, versus 21 for large caps and 25 for small caps. Hubbell has grid modernization plus data-center demand. Robinhood handles about 75 percent of customer-service calls with AI. Western Digital matters because more models and data mean a long storage runway.

Cisco’s $7.5 billion AI revenue outlook was okay; it had already pointed to $6 billion. Investors are hung up on margins: fourth-quarter gross margin was 66 percent, with guidance implying a 100- to 150-basis-point decline. Its $9.3 billion in AI orders converts over time, though supply constraints could push revenue into fiscal 2028.

At IVP, we like companies that can go from Series B to the S and P, not just companies we can sell quickly. Decart is interesting because inference optimization—software making chips work better—is hot. Build something hard others avoid, and you can compound for a long time. Perplexity was an easy call: Arvind is exceptional and the team ships fast. Browser, deep research, and agents expand usage. AI starts with individual productivity; the bigger payoff comes through company workflows. Venture is a picking business: one or two winners can drive most returns.

Our first-party cloud, deploying our equipment and renting compute, grew 281 percent year over year. Some hardware customers lacked data-center capacity; demand isn’t the issue. We have 600 megawatts signed or live, a multi-gigawatt pipeline, and expect revenue to more than triple next year. We’re expanding manufacturing more than tenfold, with TSMC support. GPUs bring throughput, Cerebras brings speed; together they can improve economics and customer experience.

That does it for Bloomberg Tech: big AI spending, widening infrastructure bets, and markets still pricing the winners. Catch the recap on the podcast, the terminal, or wherever you listen.