Daily: TSMC-Intel-Nvidia-AMD-Broadcom-Qualcomm JV?; Meta's in-house chip test; Cerebras' new data centres
6 min read.
Highlights
TSMC-Intel-Nvidia-AMD-Broadcom-Qualcomm JV? Reuters reports on more information re: the TSMC-Intel JV question. This was first reported about a month ago and there’s been mixed reactions from the US and decidedly negative reaction from China.
Reuters’ reporting shows the TSMC would not own more than 50% of the JV that would have TSMC operating Intel’s fabs. They reportedly were in talks with Nvidia, AMD, Broadcom, and also seemingly Qualcomm, on how they might do this Intel JV. It would be interesting to see how it all shakes out but is definitely a creative solution to the question of US-made chips for self-sufficiency / national security reasons. But with such big players, it would have to clear anti-trust hurdles first, and likely other legal and logistical challenges. And even then, there’s no guarantee that you can copy+paste TSMC’s culture onto Intel’s to get up to TSMC’s profitability.
Meta’s in-house chip test. Meta is testing their in-house AI chips, fabricated by TSMC 5nm process. Historically, the semiconductor industry has consolidated over time (lots of fabs to just a few, lots of memory companies to just a few). Right now, there is a flourishing of customised/in-house AI chips. When it’s all said and done, will the industry consolidate, winner takes all? Or would is there space for all the in-house chips to fill their own niches.
Cerebras’ new data centres. Cerebras adds six new AI data centres in North America and Europe. The article says it “could be bad news for Nvidia.” Though this may be true, Cerebras’ differentiator is that they make wafer-scale chips, so it’s a bit of a misleading comparison. There’s questions about Cerebras’ chips’ yield rate but otherwise, it’s true that their chips run very fast.
I suppose it’s also good that they’re diversifying out of the Middle East. Almost 90% of their revenue comes from G42, an Abu Dhabi AI company. This is holding up their IPO, which they filed for last year, as they come under scrutiny by the US government.
Thanks for reading.
1. Policy and Geopolitics
1.1
Chosun Biz (03/12): Government extends special work hours for semiconductor research to 6 months
The government will establish a provision that allows for a special extension of working hours for semiconductor research positions to be increased from the current 3 months to 6 months per instance. Employers are required to provide mandatory health check-ups for workers who worked under this arrangement every 6 months.
Initially, the government and the ruling party had been pushing for the enactment of the 'Special Law for Semiconductors,' which would contain provisions exempting the semiconductor industry from the 52-hour work week. However, faced with opposition from the opposition party and the labor sector that hampered the legislative process, they proposed a plan to extend the special working period per instance.
The special extension work system allows for work beyond the legal extended hours up to 64 hours a week if necessary. Applications can be made for reasons such as research and development. The maximum authorization period per instance is 3 months, and it can be extended up to 3 times for a total of 12 months.
1.2
TrendForce (03/12): Samsung-Baidu 2nm Chip Deal Reportedly Stalls Amid U.S.-China Tensions
Baidu previously aimed to produce AI chips with Samsung’s 2nm process, but talks stalled after Trump took office. It appears that balancing U.S. concerns while securing major clients has become a tough dilemma, even for tech giants.
Samsung had planned to mass-produce AI accelerators for Baidu in 2025 using its 2nm node. The plan is also said to include integrating Samsung’s HBM products for better performance.
2. Economy, Finance, and Business
2.1
Reuters (03/12): Exclusive: TSMC pitched Intel foundry JV to Nvidia, AMD and Broadcom, sources say
TSMC has pitched U.S. chip designers Nvidia , AMD and Broadcom about taking stakes in a joint venture that would operate Intel's, according to four sources familiar with the matter.
Under the proposal, the Taiwanese chipmaking giant would run the operations of Intel's foundry division, which makes chips adapted for the needs of customers, but it would not own more than 50%, the sources said. Qualcomm has also been pitched by TSMC, according to one of the sources and a separate source.
2.2
Bloomberg (03/11): Fidelity Backs Chip Startup Celestial AI at $2.5 Billion Value
Celestial AI, a startup that seeks to use light to speed up data transfer inside servers, has raised $250 million to better position itself in a race to lower AI computing power requirements.
Fidelity Management & Research Co. led the latest funding round, with participation from new investors including BlackRock Inc., Maverick Capital, Tiger Global Management and Lip-Bu Tan, according to the startup. Tan, whose name has been floated as the next Intel Corp. chief executive officer, also joined the board of directors.
Celestial AI’s valuation rose to more than $2.5 billion following the latest financing, according to its CEO David Lazovsky. The startup has raised more than $515 million since its founding in 2020.
2.3
Bloomberg (03/12): Chip-Design Firm SkyeChip Said to Consider Malaysia IPO This Year
Semiconductor designer SkyeChip Sdn. is considering an initial public offering in Malaysia as early as the second half of this year, people familiar with the matter said, riding on the country’s push to make its own chips.
The company, founded in 2019, is targeting a price that would value it at more than 1 billion ringgit ($226 million), the people said, asking not to be named because they weren’t authorized to speak publicly.
The size and timing of the offering may change as deliberations are ongoing, the people said. SkyeChip didn’t respond to a request for comment.
3. Technology
3.1
Reuters (03/12): Exclusive: Meta begins testing its first in-house AI training chip
Facebook owner Meta is testing its first in-house chip for training artificial intelligence systems, a key milestone as it moves to design more of its own custom silicon and reduce reliance on external suppliers like Nvidia, two sources told Reuters.
The push to develop in-house chips is part of a long-term plan at Meta to bring down its mammoth infrastructure costs as the company places expensive bets on AI tools to drive growth.
One of the sources said Meta's new training chip is a dedicated accelerator, meaning it is designed to handle only AI-specific tasks. This can make it more power-efficient than the integrated graphics processing units (GPUs) generally used for AI workloads.
3.2
VentureBeat (03/11): Cerebras just announced 6 new AI datacenters that process 40M tokens per second — and it could be bad news for Nvidia
Cerebras Systems, an AI hardware startup that has been steadily challenging Nvidia’s dominance in the artificial intelligence market, announced Tuesday a significant expansion of its data center footprint and two major enterprise partnerships that position the company to become the leading provider of high-speed AI inference services.
The company will add six new AI data centers across North America and Europe, increasing its inference capacity twentyfold to over 40 million tokens per second. The expansion includes facilities in Dallas, Minneapolis, Oklahoma City, Montreal, New York and France, with 85% of the total capacity located in the United States.
3.3
TrendForce (03/12): Samsung Reportedly Brings High-NA EUV Machine to Hwaseong Campus in Early March amid 2nm Race
Samsung Electronics brought in its first High-NA EUV equipment, the EXE:5000, to its Hwaseong campus earlier this month.
This equipment, priced at approximately 500 billion KRW (around 500 million USD), is exclusively supplied by ASML worldwide, as the report notes. The report suggests that the deployment of this equipment could be a game-changer for Samsung Electronics, as High-NA EUV is regarded as an essential tool for next-generation processes below 2nm.
3.4
Yonhap News (03/12): Samsung, SK hynix to showcase latest updates of AI memory, solutions at GTC 2025
Chip giants Samsung Electronics Co. and SK hynix Inc. will showcase their latest advancements in artificial intelligence (AI) memory and solutions at GTC 2025, an annual tech conference hosted by Nvidia Corp., industry sources said Wednesday.
Samsung Electronics will also host a presentation on its collaboration with Nvidia, highlighting the integration of its GDDR7 graphics memory into Nvidia's graphics processing units (GPUs).
Meanwhile, SK hynix will discuss the expanding role of HBM chips in AI computing.