Intel CEO: CPUs only meet 50% of demand; 14A to enter production in Q1 next year; new architecture may reduce inference power consumption to 1/15 of GPU.

On September 14, Intel CEO Chen Liwu said in a fireside chat at an industry summit that with the rapid expansion of AI agents, CPU demand has far exceeded Intel’s current supply capacity, and it can only meet about 50% of the demand from cutting-edge customers.

At the same time, he revealed that the company has made significant progress in process technology, with the 18A node already in mass production and the 14A node scheduled to go into production in the first quarter of next year.

Chen Liwu stated that several CEOs of major technology companies have directly called, requesting more CPUs. "I have to apologize, because our production capacity cannot keep up."

He attributed this shortage to the surge in CPU demand in AI inference scenarios. CPUs possess advantages that GPUs cannot easily replace in agent orchestration, control plane operations, and multi-threaded task processing.

In addition, Chen Liwu revealed that Intel is supporting new dedicated chips based on dataflow architecture and wafer-level expansion solutions, which can provide equivalent performance with 1/10 to 1/15 the power consumption of a GPU in specific inference scenarios.

In terms of market capitalization , Chen Liwu revealed that when he took over as CEO, Intel's market capitalization was approximately $90 billion, and it has now risen to approximately $500 billion.

Regarding improvements in process yield , Chen Liwu disclosed that Intel has achieved an annual yield improvement of approximately 7% by opening up factory data to equipment manufacturers and data analytics partners. He admitted that the factory yield was "extremely bad" in the early days of his tenure, but emphasized that breaking down barriers and seeking external assistance were key to turning the situation around.

The era of intelligent agents has triggered a reversal in computing power, leading to a severe shortage of high-end CPUs.

There is a common misconception in the market that the development of AI depends solely on GPUs.

Chen Liwu pointed out that although GPUs perform excellently in model training, the computational power requirements are reversing once the agent and reinforcement learning inference stages begin. Chen Liwu stated:

When orchestration, control flow, and complex single-threaded or multi-threaded scheduling are required, the CPU is the best choice.

Chen Liwu pointed out that cutting-edge model manufacturers are urgently seeking to purchase high-end CPUs, which has directly led to an extreme shortage of Intel's production capacity.

He admitted that the current delivery pressure is extremely high:

I can only supply 50% of the CPU orders my customers need. Many CEOs of tech giants called me to queue up and buy them, and I had to apologize to them because I hadn't manufactured enough.

He clearly predicted that as millions or even trillions of intelligent agents go online in the future, computing infrastructure, including CPU, memory, and network bandwidth, will continue to be in a "supply-constrained" environment.

Cultural reconstruction is the underlying logic of Intel's transformation.

When reviewing the challenges of taking over as CEO, Chen Liwu said that the depth and breadth of the company's problems far exceeded his judgment during his two years as a board member, "at least ten times more than I expected."

He prioritized cultural change, focusing on two key areas: establishing a culture of accountability and genuinely listening to customers.

He revealed that shortly after he took office, a major client listed 15 areas where Intel had made mistakes.

They told me that your team doesn't listen to us and only lectures us, so we designed and sent you out.

Chen Liwu also emphasized that the engineering team reported directly to him, enabling him to grasp the real, first-hand information, rather than a "beautified version" that had been filtered through layers of bureaucracy.

He sees maintaining ongoing connections with startups and the venture capital community as a key mechanism to prevent Intel from missing out on the next wave of technology again.

The 18A is already in full-scale mass production, and the 14A will go into production in the first quarter of next year.

The foundry business is central to Intel's revitalization, but Chen Liwu emphasized that the essence of foundry is "customer service," which requires Intel to completely change its past culture. Chen Liwu said:

Outsourcing isn't about being aloof; you need to serve your clients with a pleasing and humble attitude, seamlessly integrating their preferred EDA tools or specific third-party IP libraries.

To address the yield issues in advanced manufacturing processes, Chen Liwu took decisive action, resolutely breaking the long-standing practice of internal engineers concealing data. He recalled this highly impactful process:

I had to change the culture, lay off some engineers, and make sure we were completely open. After we released the real data, two equipment manufacturers' CEOs, who were friends with me, called and said, 'Liwu, I have bad news for you; the initial yield is terrible.' I replied, 'That's why I asked you for help.'

Through concerted efforts, Intel has made significant strides in its advanced manufacturing processes. Chen Liwu officially confirmed the latest timeline and key data highlights:

The good news is that we are now seeing a 7% annual improvement in yield. Intel's 18A process is now in full-scale mass production, and Intel's 14A process will officially enter production in the first quarter of next year.

Advanced packaging is the "holy grail," with massive investments in the US aimed at resolving supply chain crises.

Against the backdrop of Moore's Law slowing down, Chen Liwu regards advanced packaging as the "ultimate holy grail" for the industry's future and has set it as the top priority for the next five years.

Chen Liwu believes that system-level packaging that integrates CPU, large memory and I/O is the key to improving computing power density.

However, he issued an extremely stern supply chain warning. He pointed out that the global market for key packaging substrate materials is monopolized by a very few manufacturers (two in Japan and two in Taiwan), and huge upfront payments are required to secure production capacity. Chen Liwu stated:

It is extremely dangerous that 95% of the world's advanced packaging capacity is concentrated in a single region.

Intel is investing heavily in capital expenditures to accelerate the localization of its advanced packaging and manufacturing capabilities in the United States.

"Ten-Year Dimensional Reduction Strategy" and Quantum Computing Nodes

Regarding the current development of AI infrastructure, Chen Liwu expressed deep concern about electricity consumption.

Jensen Huang has done an excellent job in training, but the entire industry is facing an insurmountable problem—the limit of power consumption.

To break through this bottleneck, Chen Liwu revealed his "Ten-Year Forward-Looking Project," which he is secretly planning.

He pointed out that the human brain only requires tens of watts of energy to process extremely complex cognition, which means that there is "potential technological space for reducing energy consumption by 10,000 times" in the future computing power.

Currently, Intel is supporting startups (such as SambaNova and Cerebras) that are based on dataflow architecture and wafer-level scaling solutions, " which can provide equivalent performance at 1/10 to 1/15 the power consumption of a GPU in certain inference scenarios."

At the same time, the company is heavily investing in neuromorphic computing and brain-like MPUs.

Regarding the more cutting-edge field of quantum computing, Chen Liwu provided a clear timeline for commercialization. He stated that they are currently highly focused on solving the quantum error correction problem.

It is expected that quantum computing will have a substantial industrial impact in the next 3 to 5 years. It will form a heterogeneous hybrid computing network with CPUs and GPUs to jointly solve the problem of high-dimensional computing.

Risk warning and disclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.