Intel CEO Lip-Bu Tan on the one thing troubling his customers the most: I think in terms of the AI, the biggest challenge is...
Intel CEO Lip-Bu Tan says memory is now the biggest problem for companies racing to build AI infrastructure, and it is only getting worse. Speaking at Splunk's .conf26 conference in Denver last week, Tan said memory prices have already climbed five to seven times over earlier levels, and that limited production is forcing businesses to delay projects.
Tan has been making this point all year. At an Intel event earlier, he put it plainly: "I think in terms of the AI, the biggest challenge, I think, for a lot of my customer is memory." AI, he added, is "sucking" up much of the memory the computer industry has. He has also said the memory makers he speaks with regularly have given him a clear timeline. "Lip-Bu, there's no relief until 2028," he recalled them telling him.
AI chips are taking memory away from the rest of the industryThe root of the problem is high-bandwidth memory, or HBM, the stacked memory that sits alongside AI accelerators from Nvidia and AMD. Samsung, SK hynix and Micron make almost all of the world's advanced memory, and demand for HBM has outrun what all three can produce. To keep up, they are moving existing production lines to HBM, which leaves less capacity for regular DRAM across the rest of the industry.
That shift is what Tan means when he says projects are being held up. Companies building data centres can secure GPUs and still stall because the memory to pair with them is not available in time. The wait is also a long one. New memory capacity takes years to build and qualify, and much of what comes online is already committed to AI accelerator orders. That is why the 2028 date keeps coming up from Tan and others across the industry.
Intel says CPUs, not just GPUs, are now running short because of AIMemory is not the only thing in short supply. In the same Denver conversation with Cisco President Jeetu Patel , Tan said Intel's processors are running short too. "CPU demand is so high that we can only supply 50% of customers," he said. He added that several CEOs have called him directly asking for more chips, and he has had to apologise because Intel's capacity cannot keep up.
Tan linked this demand to AI agents. GPUs are well suited to training models, he explained, but tasks like orchestration and scheduling run best on CPUs. With millions, possibly trillions, of AI agents expected to come online, he expects computing, memory and network capacity to stay tight. Intel is reportedly set to raise PC processor prices by around 10% from October 5, its third hike since late 2025.
Intel CEO says depending on Taiwan for chip packaging is dangerousTan also flagged packaging substrates as a pressure point. Only four major suppliers, two in Japan and two in Taiwan, control the market, and Intel has to prepay to secure capacity. "Relying 95% on one company, especially one based in Taiwan, is very dangerous," he said.
On manufacturing, Tan confirmed that Intel's 18A process is in mass production and 14A will begin production in the first quarter of 2027. The company's second-quarter revenue rose 25% to $16.1 billion, its fastest growth in 15 years. Intel shares rose as much as 10% after his remarks.
Tan has been making this point all year. At an Intel event earlier, he put it plainly: "I think in terms of the AI, the biggest challenge, I think, for a lot of my customer is memory." AI, he added, is "sucking" up much of the memory the computer industry has. He has also said the memory makers he speaks with regularly have given him a clear timeline. "Lip-Bu, there's no relief until 2028," he recalled them telling him.
AI chips are taking memory away from the rest of the industryThe root of the problem is high-bandwidth memory, or HBM, the stacked memory that sits alongside AI accelerators from Nvidia and AMD. Samsung, SK hynix and Micron make almost all of the world's advanced memory, and demand for HBM has outrun what all three can produce. To keep up, they are moving existing production lines to HBM, which leaves less capacity for regular DRAM across the rest of the industry.
That shift is what Tan means when he says projects are being held up. Companies building data centres can secure GPUs and still stall because the memory to pair with them is not available in time. The wait is also a long one. New memory capacity takes years to build and qualify, and much of what comes online is already committed to AI accelerator orders. That is why the 2028 date keeps coming up from Tan and others across the industry.
Intel says CPUs, not just GPUs, are now running short because of AIMemory is not the only thing in short supply. In the same Denver conversation with Cisco President Jeetu Patel , Tan said Intel's processors are running short too. "CPU demand is so high that we can only supply 50% of customers," he said. He added that several CEOs have called him directly asking for more chips, and he has had to apologise because Intel's capacity cannot keep up.
Tan linked this demand to AI agents. GPUs are well suited to training models, he explained, but tasks like orchestration and scheduling run best on CPUs. With millions, possibly trillions, of AI agents expected to come online, he expects computing, memory and network capacity to stay tight. Intel is reportedly set to raise PC processor prices by around 10% from October 5, its third hike since late 2025.
Intel CEO says depending on Taiwan for chip packaging is dangerousTan also flagged packaging substrates as a pressure point. Only four major suppliers, two in Japan and two in Taiwan, control the market, and Intel has to prepay to secure capacity. "Relying 95% on one company, especially one based in Taiwan, is very dangerous," he said.
On manufacturing, Tan confirmed that Intel's 18A process is in mass production and 14A will begin production in the first quarter of 2027. The company's second-quarter revenue rose 25% to $16.1 billion, its fastest growth in 15 years. Intel shares rose as much as 10% after his remarks.
Next Story