Daily: UAE wants AI chips; U.S.-China chip war; Micron's new structure
7 min read.
Highlights
UAE wants AI chips. UAE is optimistic about getting more AI chips from the United States, after pledging US$1.4 trillion of investment. That’s a lot more than the $500b Nvidia had pledged. Per the AI Diffusion Rule, the UAE is currently in the second tier, which carries export restrictions and hinders the country’s high tech goals to transform the Middle East. In the ideal case, the UAE would be upgraded to the first tier of no restrictions. Cerebras, the wafer-scale AI chip startup, would be happy since the vast majority of their revenue comes from selling to the UAE. The startup’s IPO was paused due to national security concerns.
U.S.-China chip war. A good, long piece in the NYT on the U.S.-China saga around AI chips. With the latest Trump restrictions on H20 and equivalent chips, US companies are quickly ceding domestic Chinese market share to Huawei. I think Huawei is still some way away re: AI chips, and critically they don’t have EUV chipmaking tools, so it is very difficult for them to manufacture <7nm nodes, making them a whole category behind ex-China counterparts. TSMC had 7nm manufacturing ~2018, so China could be said to be maybe 7 or so years behind, though it is contingent on them making the technological breakthrough of EUV, which is more than just a matter of time.
Chinese firms are making do by disregarding the commercial need to have a high yield rate (i.e., subsidised by China) and being creative with their software (e.g., DeepSeek’s efficient AI models). Huawei’s Ascend 920 reportedly equivalent to Nvidia’s H20 chips and may be ready for mass production in the second half of this year.
Micron’s new structure. Some more details on the Micron company restructure, that I included last week. They will establish a new “Cloud Memory Business Unit,” which aims to deliver customised memory to hyperscalers.
Thanks for reading.
1. Policy and Geopolitics
1.1
Bloomberg (04/19): UAE Sees Progress on AI Chips Access After $1.4 Trillion Pledge
The United Arab Emirates is optimistic about securing access to cutting-edge chips currently under US export restrictions, after pledging $1.4 trillion to strengthen ties with Washington.
The country is making “very good and tangible progress” toward securing advanced semiconductors from the US, according to Peng Xiao, chief executive officer of G42, the UAE’s leading artificial intelligence firm.
The Emiratis have been stepping up investments in US technology and energy sectors to gain access to chips critical to its aspirations of becoming an AI powerhouse. The UAE currently falls within the second tier of the AI diffusion rule, which restricts the number of processors that can be exported to most nations.
People familiar with Sheikh Tahnoon’s plans ahead of that meeting said he would seek easier access to chips and highlight the UAE’s plans for tech infrastructure built on American soil. That includes funding from MGX, which the Abu Dhabi royal chairs, for a $100 billion AI infrastructure venture that Trump unveiled during his first week in office.
1.2
Nikkei (04/18): Rapidus' chip plans are 'good for world,' IBM CEO says
Rapidus' effort to mass-produce advanced chips will benefit all sides, said Arvind Krishna, the CEO of IBM, which a partner of the Japanese company.
"Can we depend on only one supplier in one country, especially one that carries some geopolitical risk?" Krishna told Nikkei in an interview. "The answer is no."
2. Economy, Finance, and Business
2.1
NYT (04/18): U.S. Chipmakers Fear They Are Ceding China’s A.I. Market to Huawei
The semiconductor industry has lobbied two presidential administrations to go easy with restrictions on selling cutting-edge computer chips to China. Jensen Huang, the chief executive of Nvidia, the world’s leading artificial intelligence chip maker, even traveled to Mar-a-Lago this month to discuss policy with President Trump.
But with the Trump administration putting new curbs on A.I. chip sales this week, it is clear that the industry’s pushback has failed. The fallout has set off a scramble among chipmakers to reset expectations for a future with less sales to China and prompted fears that their retreat could turn the Chinese tech giant Huawei into a global chip-making powerhouse.
In the two days after the limits became public, shares of Nvidia, the world’s leading A.I. chipmaker, fell 8.4 percent. AMD’s shares dropped 7.4 percent, and Intel’s were down 6.8 percent.
“For the U.S. semiconductor industry, China is gone,” said Handel Jones, a semiconductor consultant at International Business Strategies, which advises electronics companies. He projects that Chinese companies will have a majority share of chips in every major category in China by 2030.
2.2
TrendForce (04/21): TSMC’s U.S. Fab Posts NT$14.3 Billion Loss, While China Operations Deliver Steady Profit
According to its 2024 annual report, TSMC’s Arizona facility posted a loss of nearly NT$14.3 billion last year, making it the company’s most costly overseas site. In contrast, its Nanjing operations recorded a strong profit of nearly NT$26 billion during the same period.
TSMC’s Arizona subsidiary reported consecutive losses of NT$4.81 billion, NT$9.43 billion, and NT$10.924 billion in its 2021, 2022, and 2023 annual reports, with cumulative losses over the past four years exceeding NT$39.4 billion. The report notes that 2025 will mark the Arizona fab’s first full year of mass production, which could help narrow losses. However, it also warns that ongoing investment might prolong losses.
2.3
TrendForce (04/18): Micron Creates Dedicated HBM Business Unit, Reportedly Eyes Custom Memory for Big CSPs
As HBM has emerged as a key battle ground for memory giants amid the surging demand for AI, Micron has created a dedicated business unit for HBM. The company says in its press release that the new structure will be in place by early fiscal Q4, starting May 30, 2025, and financial reporting will follow the new format from that quarter onward.
The newly established Cloud Memory Business Unit (CMBU), in particular, will focus on memory solutions for hyperscale cloud providers and oversee Micron’s HBM business. This is seen as a strategic move to deliver customized memory to key clients like Microsoft and Amazon Web Services (AWS), the report indicates.
3. Technology
3.1
TrendForce (04/21): Huawei’s Ascend 920 Reportedly Set for 2H25 : Concerns Grow Over SMIC’s Capability and HBM Access
As the U.S.-China chip war heats up, China is going all-in on AI ASICs, with Huawei’s flagship chips stealing the spotlight. But with rumors of the Ascend 920 entering mass production in 2H25, Commercial Times raises doubts about SMIC’s N+3 process and the chip’s access to top-tier HBM.
Huawei’s upcoming AI chip, the Ascend 920, could serve as a potential replacement for NVIDIA’s H20, which now faces indefinite export restrictions to China. Market chatter suggests the chip will be built using SMIC’s N+3 process and paired with HBM3 memory.
But according to Commercial Times, SMIC’s N+3 process—despite claiming similar transistor density to TSMC’s 6nm—still lacks the backup of EUV tools and relies on DUV multi-patterning, which likely drags down yield and efficiency, making it hard to match U.S. AI chips on cost.
3.2
TrendForce (04/21): Samsung and ASML Reportedly Scrap EUV Research Center in Hwaseong, Eyeing New Alternatives
As EUV machines are now essential for pushing sub-2nm chips into mass production in the angstrom era, partnerships with ASML remain a key industry focus. However, according to etnews, ASML and Samsung appear to have dropped plans for their joint EUV research center in Hwaseong, as ASML has begun selling off the designated site.
While the move has sparked questions about the strength of the ASML-Samsung partnership, etnews reports that the two are actually eyeing a new location for the joint R&D center—outside of Hwaseong. To speed things up, they’re also considering more efficient alternatives, including setting up the facility within a Samsung campus, the report indicates.
3.3
TrendForce (04/21): JEDEC Finalizes HBM4 Standard, Potentially Easing Pressure on Hybrid Bonding Adoption
As leading memory giants ramp up preparations for the next-generation high bandwidth memory (HBM) battle, JEDEC—the standards-setting body for the microelectronics industry—has officially released the JESD270-4 HBM4 standard, a new memory standard designed to meet the surging demands of AI workloads.
Notably, one of the major improvements in HBM4 is the doubling of independent channels per stack—from 16 in HBM3 to 32. Each channel is further split into two pseudo-channels, allowing for faster and more flexible memory access, as stated by JEDEC.
Meanwhile, HBM4 delivers transfer speeds of up to 8 Gb/s over a 2048-bit interface, reaching a total bandwidth of 2 TB/s. It also supports lower power consumption and improved energy efficiency, according to JEDEC’s press release.
3.4
Tom’s Hardware (04/20): New chip cooling technique is 7X more more effective than standard approaches
A research team from the University of Tokyo has devised a new cooling solution that uses water’s changing phases to make it more efficient at removing heat. SciTech Daily says that water absorbs seven times the energy when it changes phase from liquid to gas (i.e., boiling the water), allowing it to capture and dissipate more heat versus the traditional method of using flowing water. However, since the coolant flows through tiny capillaries built directly into the chip, the steam would often have difficulty flowing through these narrow channels. This often makes it less efficient than traditional methods.
The researchers solved this issue using 3D microfluidic channels with a capillary structure and a manifold distribution layer. They discovered that the shape of the microchannels and how the coolant is distributed throughout the system significantly impact its thermal and hydraulic performance. By ensuring the continuous flow of water and steam, the team achieved a coefficient of performance (COP) of 100,000 — about ten times greater than what single-phase water cooling can achieve.
-