During a fireside chat at an industry summit on September 14, Intel CEO Lip-Bu Tan disclosed that the explosive growth of AI agents has driven CPU demand far beyond the company's current supply capacity, with Intel currently able to fulfill only about half of its leading-edge customers' requirements.

He also revealed major progress in process technology, noting that the 18A node has entered volume production, while the 14A node is scheduled to begin manufacturing in the first quarter of next year.

Tan said multiple CEOs of major tech firms have directly called to request more CPU supply, adding, "I've had to apologize because our capacity simply cannot keep up." He attributed this shortage to a surge in CPU demand within AI inference scenarios, where processors hold distinct advantages over GPUs in agent orchestration, control plane operations, and multi-threaded task processing.

Additionally, Tan shared that Intel is nurturing new specialized chips based on dataflow architecture and wafer-scale expansion solutions, which can deliver equivalent performance in specific inference scenarios while consuming only 1/10 to 1/15 of the power used by GPUs.

On the market capitalization front, Tan revealed that when he assumed the CEO role, Intel's valuation stood at approximately $90 billion, and has since climbed to around $500 billion.

Regarding process yield improvements, Tan disclosed that by opening factory data to equipment vendors and data analytics partners, Intel has achieved annual yield gains of about 7 percent. He admitted that factory yields were "terrible" during his early tenure but stressed that breaking down internal barriers and seeking external assistance was crucial to turning the situation around.

The agent era triggers a computing power reversal, with high-end CPUs in severe short supply

There is a widespread misconception that AI development relies solely on GPUs. Tan pointed out that while GPUs excel at model training, computing power demands are reversing as the industry shifts into agent-based and reinforcement learning inference phases. He stated, "When it comes to orchestration, control flow, and complex single-threaded or multi-threaded scheduling, CPUs are the best choice."

Tan noted that leading frontier model developers are urgently seeking high-end CPUs, which has directly led to extremely tight capacity at Intel. He acknowledged the immense delivery pressure, saying, "I can only provide about 50 percent of what customers need. Many tech giant CEOs have called me to queue up for orders, and I have to apologize because I can't manufacture enough."

He firmly predicted that as millions or even trillions of agents come online in the future, computing infrastructure, including CPUs, memory, and network bandwidth, will remain in a "supply-constrained" environment.

Cultural rebuilding forms the foundational logic of Intel's transformation

Reflecting on the challenges he faced upon becoming CEO, Tan said the depth and breadth of the company's problems far exceeded what he had perceived during his two years as a board member, describing them as "at least ten times worse than I expected."

He prioritized cultural change, focusing on two key areas: establishing a culture of accountability and genuinely listening to customers. He recounted that shortly after his appointment, a major client listed 15 areas where Intel had failed, telling him, "Your team doesn't listen to us. They just lecture us, so we designed you out of our products."

Tan also emphasized that having engineering teams report directly to him allows him to access genuine frontline information rather than a filtered "sanitized version." He views maintaining ongoing connections with startups and the venture capital community as a critical mechanism to prevent Intel from missing the next wave of technology.

18A enters full production, 14A slated for Q1 next year

Foundry operations sit at the core of Intel's revival, but Tan stressed that the essence of foundry work is "customer service," which demands a complete cultural shift. He explained, "Foundry is not about being aloof. You need to serve customers with a pleasing and humble attitude, seamlessly accommodating their preferred EDA tools or specific third-party IP libraries."

To tackle yield issues in advanced processes, Tan took decisive action, firmly dismantling the ingrained habit of engineers hiding data. He recalled the impactful process: "I had to change the culture, fired some engineers, and ensured we were completely transparent. When we published the real data, two equipment vendor CEOs who are my friends called me and said, 'Lip-Bu, I have bad news for you. The initial yield is terrible.' I replied, 'That's exactly why I'm asking for your help.'"

Through these collaborative efforts, Intel's advanced process technology has achieved substantial progress. Tan officially confirmed the latest timeline and key data, stating, "The good news is that I'm now seeing yield improvements of about 7 percent per year. Our Intel 18A process has fully entered volume production, and our Intel 14A process will officially commence manufacturing in Q1 next year."

Advanced packaging is the 'ultimate holy grail,' with massive investments landing in the U.S. to resolve supply chain risks

With Moore's Law slowing down, Tan regards advanced packaging as the "ultimate holy grail" for the industry's future and has designated it as a top priority for the next five years.

He believes that system-level packaging, which integrates CPUs, large memory, and I/O, is key to boosting computing density. However, he issued a severe supply chain warning. He noted that the global supply of critical packaging substrate materials is monopolized by a handful of suppliers (two Japanese and two Taiwan-based companies) and requires huge advance payments to secure capacity. Tan stated, "95 percent of the world's advanced packaging capacity is concentrated in a single region, which is extremely dangerous."

Intel is now investing massive capital expenditures to accelerate bringing advanced packaging and manufacturing capabilities to U.S. soil.

A 'ten-year disruption plan' and the quantum computing timeline

Regarding current AI infrastructure development, Tan raised serious concerns about power consumption. He noted, "Jensen Huang has done an excellent job in the training domain, but the entire industry is facing an insurmountable issue: the limits of electricity consumption."

To break through this bottleneck, Tan revealed a "ten-year forward-looking project" he has been quietly developing. He pointed out that the human brain processes extremely complex cognitive tasks using only dozens of watts of power, suggesting there is "potential technology headroom to reduce energy consumption by 10,000 times" in future computing.

Currently, Intel is supporting startups based on dataflow architecture and wafer-scale expansion, such as SambaNova and Cerebras, which "in specific inference scenarios can deliver equal performance while consuming 1/10 to 1/15 of a GPU's power." The company is also making major investments in neuromorphic computing and brain-inspired MPUs.

As for the more advanced field of quantum computing, Tan provided a clear commercialization timeline. He said the company is currently highly focused on solving the problem of quantum error correction: "In the next 3 to 5 years, quantum computing will generate substantial industry impact. It will form a heterogeneous hybrid computing network with CPUs and GPUs, collectively solving high-dimensional computing challenges."