Musk warned of a chip glut from China in 2028, but experts aren’t so sure

[NEWS ANALYSIS]

Elon Musk and Jensen Huang are at odds over the future of the memory chip market. The Tesla CEO sees a potential supply glut looming after 2028, while the Nvidia CEO believes demand from a persistent AI boom will outpace supply for years to come.

The disagreement has rattled investors, reviving concerns over whether the massive sums being poured into AI infrastructure can sustain the semiconductor industry's extraordinary performance.

Some experts readily agree with the Nvidia chief and extend the timeline by two more years, saying supply will stay tight until at least 2030. Even then, they say, China's ability to flood the global market will be constrained by limited access to advanced chipmaking equipment, while demand from AI data centers continues to absorb much of the available supply.

“Hyperscalers are earning operating margins of 35 percent to 39 percent in their existing businesses,” said Roh Geun-chang, the CEO of Semicon Research Lab and a former head of the semiconductor research center at Hyundai Motor Securities. “Still, operating cash flow alone is not enough for them, so they are borrowing to build AI data centers, which is why free cash flow is negative. That’s a cash flow problem, not a profitability problem.”

Despite growing pushback against the AI infrastructure buildout, demand forecasts remain high, and continued AI advances mean more memory will be needed.

Musk recently took a heavy swipe at the pricing power of Samsung Electronics and SK hynix in a reply to a post warning that Korea's chip boom could face oversupply from 2029. The Tesla CEO predicted that Chinese chipmakers would begin producing large volumes of random access memory (RAM) in 2028, hinting at a threat to the dominance of the two Korean companies.

The remarks helped fuel a sell-off in Korean chip stocks on Oct. 15, with Samsung shares plunging 10.56 percent and SK hynix tumbling 12.76 percent.

Huang, however, has pushed back against fears of an impending glut, arguing that memory supplies are unlikely to exceed demand even by 2029. He also expects high bandwidth memory (HBM), a specialized type of memory essential to AI processors, to remain in heavy use for years.

Market sentiment has since rebounded following Anthropic's public filing, with Samsung shares gaining 4.23 percent and SK hynix rising 5.65 percent on Tuesday from the previous trading session.

The greater risks may lie with smaller, debt-funded cloud infrastructure providers such as CoreWeave, which have less financial flexibility than established technology giants. Some of these companies could struggle if financing becomes more expensive or AI demand falls short of expectations, potentially setting off a shakeout in the industry.

For now, however, demand for computing capacity remains robust. Executives at Amazon and Google have said their own data center capacity is fully booked, forcing them to turn to outside providers such as SpaceX to meet demand.

“Next-generation memory is increasingly complex and difficult to manufacture at economic yields,” said Song Zhe, a senior investment specialist for Asia ex-Japan and Emerging Market Equities & Greater China Equities at BNP Paribas Asset Management. “China is emerging as a serious supply-side competitor, but scaling production and climbing the technology curve will take time, particularly in HBM.”

Local media reported Tuesday that Chinese cloud service providers including Alibaba, Tencent and ByteDance are seeking long-term server memory supply deals with Samsung and SK hynix. Some have asked to extend the terms to 2035, well beyond the three-to-five years of typical long-term contracts. The products they want are server double data rate (DDR) and low-power DDR memory.

The biggest bottleneck in chip manufacturing is extreme ultraviolet (EUV) lithography equipment, which Chinese memory makers such as CXMT and YMTC are barred from obtaining under U.S. regulations. Even though ASML’s EUV production capacity is expected to rise from 44 units in 2025 to more than 80 in 2027, the acquisition ban would impose a substantial barrier to significant technological competitiveness for Chinese firms, according to Roh.

“The problem with the China-led supply outlook is that significantly expanding capacity requires securing chipmaking equipment, and TSMC, Samsung, SK hynix and Micron are buying up all of it,” Roh said. “Even if Chinese companies announce expansion plans, the actual increase will inevitably be very limited.”

Because of the U.S.-led ban on ASML equipment, CXMT is focusing on technologies that work around the process requiring such tools. CEO Cao Kanyu said earlier this month that the company has developed a transistor called a vertical surround-gate, which addresses longstanding problems such as electrical leakage. He said this lays the groundwork for a “4F-squared” design, and that CXMT will launch a DDR5 server memory module using it by the end of the year.

The 4F-squared design shrinks the area of a memory cell, the smallest unit that stores data in DRAM. It cuts cell area by about 33 percent from the current mainstream 6F-squared design, allowing more data on a chip of the same size. SK hynix unveiled a 4F-squared road map last year, and Samsung aims to apply it in mass production to DRAM below 10 nanometers by 2028.

Peter Lee, the managing director and head of Korea research at Citigroup Global Markets Korea Securities, said the technology “is not new.” Chinese companies have not been able to adopt EUV, so this is their only option, he said. Samsung and SK hynix can keep using EUV and do not need to move to 4F-squared right now.

“The problem with 4F-squared is that after one or two generations, it becomes hard to make the next transition,” Lee said. “To shrink further, you would need to use EUV again.”

BY LEE JAE-LIM [[email protected]]