[AI influence. on the CLARiiON servernews from Taipei]The ongoing Taipei International Computer Show (COMPUTER 2026) today ushered in the biggest look in the field of hardware architecture. In response to the unprecedented power and cost crisis faced by global data centers due to the outbreak of "Agent AI", Intel CEO Lip-Bu Tan jointly launched custom chip upstart SambaNova and OEM giant Foxconn in his keynote speech to drop a blockbuster intended to reshape data center economics: the three parties announced that they would jointly build a production-level "decoupled reasoning" AI rack infrastructure for the next generation of Intelligence Centers.
The "computing power disaster" in the era of agents
"The entire industry is fully evolving from the 'training era' to the 'reasoning era.'" Alex Katouzian, head of client computing and physics AI at Intel, revealed a surprising set of predictions at the show: By 2030, AI reasoning workloads are expected to consume nearly 40% of the world's data center power supply.
The root cause of this phenomenon isThe comprehensive implementation of "agent AI" in enterprise business flows. Traditional chat robots only need to generate answers once, while agents with independent thinking and planning capabilities need to undergo intensive iterations of "think-plan-action-reflection" when performing tasks. This continuous self-error correction and multiple rounds of dialogue has caused the consumption of reasoning and computing power on servers around the world to increase exponentially. If we continue to rely on Nvidia's traditional GPU clusters, which cost tens of thousands of dollars and are hard to find a card, the company's IT margin will be completely drained.
Heterogeneous decoupling: Putting a "troika" on computing power
In order to break this deadlock, the prescription given by the Intel camp is"Disaggregated Inference System". In the real production rack running on the Computex site, the hardware of the three giants has achieved unprecedented heterogeneous collaboration:
Task scheduling (Orchestration): Using Intel's latest generation of Xeon 6 + processors based on advanced 18A processes. As the "commander" of mankind entering the era of agents, Xeon 6+ specifically optimizes the scheduling of highly concurrent data movement and complex business logic.
Text pre-filled (Prefill): Using the NVIDIA Blackwell GPU, it is responsible for the first stage of high-speed processing when large models receive huge contextual input.
Token decoding (Decode): This is also the most costly and time-consuming core stage, and it is all handed over to SambaNova's newly developed SN40/SN-50 reconfigurable data stream unit (RDU).
By having different chips perform their duties,Pioneering cloud vendors such as Vector Core Compute successfully proved in on-site demonstrations that the decoupling system can increase the comprehensive energy efficiency ratio of AI reasoning several times and significantly reduce the cost of generating a single Token.
Foxconn's participation has pushed for this innovationThe acceleration key of "industrial mass production". As the world's most powerful electronics manufacturing machine, Foxconn will assemble this out-of-the-box AI rack directly on the production line. Industry experts commented that Intel's move combines the power of Purpose-built Silicon beyond GPUs and finds a breakthrough in fighting Nvidia's monopoly on the economic level for the first time. The global computing power market is likely to usher in a drastic price reshuffle.