The AI computing power battle enters the second half: from competing on single-chip XPU performance to system integration, electricity, cooling, and networking become the new battlegrounds

The logic of competition in AI infrastructure is undergoing a fundamental transformation.

According to Wind Trading Desk, a research team from BofA Securities released a report after tracking the OCP Global Summit, stating that the basic computing unit of AI processing power has been upgraded from a single chip to a full rack system. Power supply, cooling capacity, and network interconnection are now replacing chip performance itself as the core variables determining the competitiveness of AI infrastructure.

AMD CEO Lisa Su revealed at the summit that over the past two years, monthly consumption of AI tokens has increased by a staggering 95 times. Nvidia, in turn, has shifted its system design framework from chip computing power metrics to focusing on token throughput, tokens per watt, and cost per token. This shift in terminology reflects a profound migration in the competitive dimensions of the entire industry. BofA Securities highlights that AI capital expenditure is evolving from single-product cycles to complete infrastructure construction cycles.

Discussions at the summit showed no signs of slowing AI infrastructure demand. OCP is preparing technology reserves for MW-level racks, GW-level data centers, and even multi-GW campuses. Broadcom CEO Hock Tan previously stated that capital is currently the greatest constraint on AI expansion, but the technical roadmap discussions at the summit were undeterred by this limitation.

The Rack Becomes the New Computing Unit, System-Level Optimization Replaces Chip-Level Competition

The BofA Securities research team attended the OCP Global Summit 2026 in San Jose and summarized four key themes across the main sessions and technical sub-forums in their report: rack-level systems are becoming the new computing units for AI infrastructure; AI token demand continues to grow exponentially, driving synchronous expansion of computing power and storage needs; optical interconnect and Ethernet architectures are extending from traditional scale-out scenarios to vertical scale-up scenarios; and 800V DC power supply, memory disaggregation based on CXL and HBF (for KV cache), along with open ecosystems and control frameworks, are emerging as key enabling technologies for the next generation of AI factories.

The report emphasizes that the focus of industry dialogue is shifting from individual chip performance to holistic system optimization encompassing compute, memory, networking, power, and software orchestration. The ability to efficiently scale AI infrastructure is becoming the core competitive differentiator between enterprises.

Optical Interconnects Fully Penetrate Scale-Up, CRDO’s Short-Reach Optical Competitiveness Emerges

Networking and interconnects are among the most persistent structural topics at this OCP Summit. Broadcom continues to advocate for Ethernet as the unified foundational network spanning scale-up, scale-out, and scale-across AI architectures, with multiple vendors also highlighting the increasing integration of optical solutions both within and between racks.

One focus of summit discussions is the technological evolution of scale-up interconnects from copper cables to Near Package Optics (NPO), and eventually to Co-Packaged Optics (CPO). CRDO stated that its short-reach optical products (ZF ALC) are continually gaining competitiveness over copper cables in terms of reliability, power consumption, and total system cost.

BofA Securities points out that the market underestimates Credo Technology’s (CRDO) opportunity in the scale-up optics sector—the firm expects CRDO’s optical business to exceed $600 million in total scale by FY27, with 6.4T scale-up NPO products entering the ramp-up phase in FY28.

Rise of 800V DC Power Architecture, Power and Cooling Become New Infrastructure Bottlenecks

Power supply is another core topic at the summit. Delta and Nvidia jointly presented the technical pathway for transitioning to 800V DC (800VDC) architecture—As rack power density advances from the current approximately 145kW towards 300 to over 500kW, the new generation of power architecture aims to enhance power conversion efficiency, reduce conversion losses, and accelerate the deployment pace for GW-scale AI factories.

Discussions across OCP’s sub-forums confirmed one conclusion: the future bottleneck in expanding AI infrastructure will increasingly be the availability of power, cooling capacity, and DC capability, rather than the mere supply of silicon chips. These structural constraints underpin a multi-year infrastructure buildout cycle across power, cooling, networking, and compute, rather than a short-term procurement cycle focused solely on individual chip products.

Persistent Visibility of Demand, AI Capital Expenditure Enters Long Cycle

Despite ongoing debates among investors regarding AI ROI, infrastructure financing models, and the long-term sustainability of AI capital expenditure, discussions at this OCP Summit showed no signs of slowing demand.

Lisa Su from AMD highlighted the explosive growth in AI inference demand by citing a 95-fold monthly increase in token consumption over the past two years. Nvidia has anchored the core metrics of system design to token throughput and cost efficiency, rather than just raw compute specifications.

From a capacity planning perspective, OCP is preparing technology reserves for 1MW-level racks, GW-level facilities, and multi-GW campuses, indicating that AI capital spending cycles have evolved from single product iterations into a full wave of infrastructure construction.