Anthropic Bets Big on AI Compute: Multi-Billion Deals, Custom Chips, and the Future of Claude

Fueled by robust enterprise adoption, Anthropic's flagship model, Claude, has propelled the company's annual revenue run-rate to over $30 billion—up from just $9 billion at the end of 2025. This explosive growth is closely linked to a doubling of large business customers; over 1,000 enterprise clients now spend at least $1 million per year, up from 500 only two months ago, underscoring both the exponential uptake of generative AI and the dependence of customers on advanced model infrastructure ^[2],^[9].

Anthropic’s internal projections highlight the critical nature of reliable compute. Company leaders, including CFO Krishna Rao, frame the current moment as one of unprecedented scaling demands: “We are building the capacity necessary to serve the exponential growth we have seen in our customer base while also enabling Claude to define the frontier of AI development” ^[4],^[7]. As business demand intensifies, so too does the need to secure dedicated, next-generation AI hardware.

In April 2026, Anthropic announced a sweeping, multi-year agreement with Google and Broadcom, cementing access to 3.5 gigawatts of Google’s latest TPU (Tensor Processing Unit) compute from 2027 onward. Most of this hardware will be sited in the US, aiming to deliver hyperscale resources for Claude and other advanced models ^[4],^[7],^[8]. Broadcom will build these custom chips in collaboration with Google, extending their long-standing relationship through 2031 and offering significant competitive differentiation for all parties involved ^[2],^[5],^[6].

The scale is extraordinary—Anthropic’s secured compute would rival supercomputing facilities, reflecting the sky-high resource needs of frontier AI development. Analysts at Mizuho Securities project that Broadcom could generate as much as $42 billion annually from AI-related business by 2027, largely on the back of these megadeals ^[9]. Furthermore, this chip deal positions Anthropic to diversify away from Nvidia GPU dependency, while Broadcom’s CEO Hock Tan contends that their engineering capabilities will allow them to challenge Nvidia’s longstanding dominance in AI accelerator markets ^[2],^[6],^[9].

While Anthropic’s major deals reflect an appetite for external chip supply, the company is also actively exploring the possibility of designing its own AI silicon. Sources indicate that Anthropic’s discussions are in early stages—no specific design or engineering team has yet been assembled, and the project direction remains under review ^[1],^[3]. This mirrors a broader industry trend; both OpenAI and Meta have moved to develop custom chips to hedge against global shortages and optimize for specialized AI workloads ^[1],^[3].

The stakes are high, as the cost of bringing a cutting-edge AI chip to market can reach $500 million, factoring in specialized talent and rigorous fabrication standards. For Anthropic, which currently leverages Google TPUs, Amazon’s Inferentia and Trainium, and Nvidia GPUs, developing proprietary chips could offer a competitive edge in efficiency and performance, as well as insulation from volatile global supply chains ^[3],^[5]. Yet, committing to internal chip design also involves significant long-term risk, and may shift internal technical priorities in substantial ways.

Nvidia has long dominated the AI accelerator market, but an accelerating industry pivot toward custom, application-specific integrated circuits (ASICs) may erode this lead. The Anthropic-Broadcom-Google collaboration and rumored OpenAI-Broadcom talks reflect a strategic move by major AI developers to control their fate and costs at the silicon layer ^[5],^[9]. MediaTek is also expanding its presence, deepening collaborations with Google and emerging as a challenger in cloud-centric AI chip markets ^[5].

For hyperscalers and AI labs alike, vertical integration is no longer a luxury but essential for scaling next-generation models efficiently. This trend is likely to reshape supplier relationships, profit structures, and cloud AI economics. Broadcom’s commitment to building not just chips, but full-stack AI infrastructure, signals a fundamental restructuring of the traditional supplier-customer dynamic in the semiconductor space ^[9]. For practitioners, this means anticipating chip heterogeneity in cloud offerings and investing in cross-platform model optimization strategies.

Amid Anthropic’s infrastructure push, regulatory and security concerns have come to the fore. The launch of its new Mythos model—which the company claims can identify and exploit software vulnerabilities across major operating systems and browsers—prompted urgent meetings with US Treasury officials and banking sector CEOs ^[3]. Access to Mythos is now tightly restricted to about 40 leading tech firms, including Microsoft and Google. This decision underscores new operational risks and responsibilities for AI labs operating at the cyber-capable frontier.

Simultaneously, Anthropic has navigated political tension in Washington, particularly with the Pentagon, over the permissible uses of its technology. The company has refused to support applications in mass surveillance or fully autonomous weapons, defending a principled stance even as government pressure mounts ^[4],^[8]. For AI practitioners, the lesson is clear: as model capabilities escalate, so too do expectations for proactive risk management, controlled deployments, and coordinated engagement with state actors.