AI labs go chip-to-chip: OpenAI, anthropic, and china’s zhipu escalate silicon arms race amid Nvidia squeeze

OpenAI's collaboration with Broadcom to develop the Jalapeño custom AI inference chip marks a strategic pivot aimed at reducing reliance on premium Nvidia GPUs by the end of 2026. This chip is engineered to accelerate large language model (LLM) inference speed while significantly cutting costs, embodying CEO Sam Altman's vision for scalable and affordable AI infrastructure. By fast-tracking Jalapeño, OpenAI is not only enhancing performance but also positioning itself as a formidable competitor in the AI hardware landscape.

The Jalapeño chip exemplifies the critical trend of AI labs moving into custom silicon to manage soaring inference costs and sustain large-scale operations. As compute demands escalate, OpenAI’s move underscores that even top-tier software-centric labs must now control their hardware stack to maintain efficiency and scalability. This shift reflects a broader industry imperative to tightly integrate hardware and software for optimized AI workloads.

Designed with a deep understanding of AI workloads, Jalapeño maximizes tokens per watt by co-designing hardware tailored specifically for inference tasks, a major and growing portion of compute usage. This focus on token efficiency directly addresses the global power constraints threatening AI infrastructure growth, enabling OpenAI to deliver more compute output with less energy. By doing so, Jalapeño redefines the AI hardware arms race, intensifying pressure on Nvidia and the sprawling ecosystem amid ongoing power and supply chain challenges.

While Jalapeño cuts AI inference costs by an impressive 50%, challenging Nvidia’s GPU dominance, the latter’s vast scale and entrenched CUDA software ecosystem remain formidable competitive advantages. OpenAI’s chip thus sharpens the hardware competition but does not yet displace Nvidia’s entrenched position, highlighting that efficiency gains must be balanced against ecosystem maturity and scale in this rapidly evolving market.

Anthropic’s strategic partnerships with Samsung Electronics, Broadcom, and Google mark a decisive push to develop custom AI chips leveraging Samsung’s cutting-edge 2-nanometer fabrication process. By tapping into Samsung’s advanced manufacturing capabilities without the prohibitive costs of building their own fabs, Anthropic aims to significantly reduce inference costs and lessen its heavy reliance on Nvidia’s hardware, reflecting a broader industry trend toward hardware-software co-design to optimize AI performance and efficiency.

This hardware initiative signals a pivotal shift in AI competition, where control over the silicon infrastructure becomes as critical as software innovation. Anthropic’s move to integrate custom processors tailored for models like Claude not only promises enhanced energy efficiency and deployment flexibility across enterprise AI services but also positions the company to challenge Nvidia’s dominance amid intensifying rivalry with OpenAI and others for the next wave of AI leadership.

Chinese AI labs such as DeepSeek and Zhipu AI are aggressively pursuing the development of proprietary AI inference chips as a strategic response to US export restrictions that limit access to Nvidia hardware. This push is driven not only by geopolitical constraints but also by the fragmented nature of China's domestic AI market, where reducing reliance on both foreign and even domestic suppliers like Huawei is critical for achieving compute sovereignty. DeepSeek, for instance, has quietly expanded its chip design team and engaged with foundries and memory suppliers while preparing a major $7 billion funding round to support its ambitions, signaling a serious and well-resourced commitment to vertical integration within China’s AI ecosystem.

The development of custom AI inference chips by Chinese firms reflects a broader industry trend toward tightly integrated hardware-software stacks, where silicon is viewed as a strategic extension of AI model development rather than mere infrastructure. This approach aims to optimize inference efficiency, lower per-token operational costs, and enhance energy efficiency—key competitive advantages as AI workloads increasingly shift from training to inference. As Arisa Liu from Taiwan Industry Economics Services notes, the core motivation lies in achieving greater hardware-software synergy and reducing long-term operating expenses, a necessity given the recurring costs and deployment constraints inherent in AI inference.

Despite the ambitious drive toward compute sovereignty, Chinese AI chip projects face significant challenges including long development timelines, limited access to cutting-edge semiconductor manufacturing, and the need to adapt existing software ecosystems to new hardware architectures. Zhipu AI’s ASIC development, for example, could take over two years to complete, underscoring the complexity of building competitive AI chips from the ground up amid ongoing geopolitical tensions. Nevertheless, these efforts form a critical part of China’s strategic push to overcome supply chain vulnerabilities and establish a self-reliant AI hardware stack capable of competing globally.

Zhipu AI’s active pursuit of custom chip innovation is emblematic of China’s broader ambition to dominate its rapidly expanding domestic AI market through compute sovereignty. By developing proprietary silicon tailored to its GLM model family and pushing the boundaries of inference efficiency, Zhipu AI aims to secure a competitive edge amid fierce global AI hardware competition. This reflects a wider pattern where Chinese AI firms are not only responding to export controls but also strategically positioning themselves to lead in the next phase of AI deployment and commercialization.

The global AI chip supply chain remains severely constrained by a narrow semiconductor manufacturing pipeline dominated by a handful of key players, notably TSMC and Dutch lithography equipment maker ASML. Despite TSMC's multibillion-dollar investments to expand capacity, demand for cutting-edge AI chips continues to outstrip supply, creating a critical bottleneck that affects all major AI players. This scarcity extends beyond fabrication to advanced packaging and high-bandwidth memory, underscoring the fragility of the AI hardware ecosystem amid surging global demand.

US-China geopolitical tensions and export controls have intensified supply chain challenges, compelling Chinese AI firms like Alibaba, ByteDance, and DeepSeek to pivot away from Nvidia GPUs toward domestic alternatives and custom chip development. While Beijing has partially eased restrictions to allow limited imports of Nvidia's high-end H200 chips to alleviate acute compute shortages, this pragmatic move is balanced by a strategic emphasis on semiconductor self-reliance, with Huawei's upcoming AI chips poised to rival Nvidia's capabilities. This dual approach highlights China's urgent need to bridge gaps in manufacturing capacity, memory supply, and trustworthy software ecosystems amid ongoing export constraints.

The intensifying AI hardware arms race is driving fragmentation and strategic realignments as leading AI companies seek compute sovereignty to mitigate geopolitical risks and supply bottlenecks. OpenAI's collaboration with Broadcom on the Jalapeño chip, Meta's mass production plans for its Iris chip co-designed with Broadcom and manufactured by TSMC, and DeepSeek's $7.5 billion investment in custom inference chips exemplify this shift. As Forrester analyst Mike Gualtieri observes, 'If you rely on someone else's chips, you can never become a true AI giant,' underscoring the imperative for vertical integration amid a constrained and politically fraught semiconductor landscape.