Post
Chip Ambitions
Taiwan remains the epicenter of AI hardware infrastructure, which is why this month’s presidential election was in part a bellwether for the security of AI chip supply chains. Lai Ching-te of the Democratic Progressive Party (DPP) was elected President of Taiwan in a three-way contest with 40% of the vote, marking a third term for the pro-independence ruling party. However, the DPP has lost its majority in the Legislative Yuan, Taiwan’s parliament, which will hinder consensus on divisive issues like defense spending policies pertaining to Taiwan’s independence. One of the issues discussed on the campaign trail was TSMC’s overseas investments, with the VP for candidate for Taiwan’s opposition party (Kuomintang) commenting:
If Taiwan does not have a peaceful environment, nobody will dare invest,” Jaw said, blaming the ruling Democratic Progressive Party (DPP) for tensions with China… “Our TSMC wants to run off overseas. Taiwan plus one – one factory in Taiwan, one overseas, hollowing out our Taiwan.”
TSMC is building factories in Japan, US (Arizona), and planning another in Germany. The company generated $19B of revenue in Q4’23, which gives it ~60% market share amongst other foundries, with Samsung in a distant second at 12% market share (GlobalFoundries, UMC, and SMIC account for the other 3 in the top 5). Most will know this story, but in its early years, TSMC was founded by Morris Chang with a clear mission: provide semiconductor manufacturing services to integrated circuit designers who lacked their own fabrication facilities. Initially, TSMC primarily focused on producing chips using existing process tech. Then in the late 1990s, the company introduced the 0.25-micron process. TSMC continued shrinking feature sizes, leading to the production of smaller, faster, and more power-efficient chips. By the mid-2000s, the company had firmly established itself at the forefront of semiconductor technology with its 90nm and 65nm processes. Industry giants like Apple, NVIDIA, Qualcomm, and many, many others have come to rely on TSMC’s advanced manufacturing capabilities to bring their chip designs to life. Notably TSMC began production of its 3nm semis in in December ‘22 (entirely for Apple), and is now expecting 80% production capacity later this year with greater demand from the broader market for its second generation 3nm process (N3E). Samsung has been working to close the yield gap with TSMC on 3nm and 4nm nodes, but for now, TSCM reigns supreme:
They are the chips that power everything from mobile phones to electric cars—and they make up 15% of Taiwan’s gdp. Taiwan produces over 60% of the world’s semiconductors and over 90% of the most advanced ones. Most are manufactured by a single company, Taiwan Semiconductor Manufacturing Corporation (tsmc)…The semiconductor industry is called Taiwan’s “silicon shield”, giving the world a big reason to defend the island.
What makes TSMC an even more critical suppliers for tech companies around the globe is that replicating its infrastructure is enormously costly – the company spent $11B on capex in Q4’23 alone, with 70% allocated towards high-end technology development. But to be more specific about what this requires and what exactly is at stake should any serious conflict ensue between Taiwan and China: constructing a state-of-the-art semiconductor manufacturing facility demands an enormous investment in physical infrastructure and talent. This includes clean rooms, specialized equipment, chemical handling facilities, and utilities such as ultra-pure water and gas supplies. These facilities rely on highly specialized equipment, including photolithography machines, etching tools, and wafer processing systems. These machines are costly and require continuous upgrades to keep up with advancements in semiconductor technology – TSMC’s established relationships with equipment suppliers and their ability to purchase in bulk provide them with cost advantages. TSMC has spent decades building and upgrading this infrastructure, giving them a substantial advantage over newcomers. Additionally, developing cutting-edge semiconductor manufacturing processes requires substantial investment in R&D (20% of CapEx goes towards mature and specialty process development). Catching up with TSMC would require not only replicating their current capabilities but also investing heavily in R&D to push the envelope and develop next-generation chip technologies. Even with this physical infrastructure, semi competitors are coming up against a deep pool of experienced engineers, technicians, and researchers who have honed their skills over many years. And ultimately, it can take several years, if not a decade or more, to design, construct, and validate a foundry capable of producing advanced chips. This is evident just in the herculean effortrequired for TSMC to build out its own plant in Arizona:
…The chip factory in Arizona that TSMC is building to etch 3 nanometer chips in 2026 costs on the order of $28 billion, more than twice the $12 billion the company earmarked for the 4 nanometer foundry in the same location. This is a crazy capital intensive space, and it takes a hoard of engineers with all the intellectual property in the world to make it all work properly… Small wonder Intel wants to be in the foundry business and not just make chips for itself.
Sam Altman, OpenAI’s founder, is allegedly exploring a new venture to develop and fabricate chips, in part to reduce the company’s reliance on NVIDIA, and is speaking with TSMC and Middle Eastern investors about a potential partnership:
Altman has spoken to some of the wealthiest investors in the region about funding the ambitious new project to develop chips required to train and build AI models, and the plants required to fabricate them, according to people with direct knowledge of the talks… He is also talking to Taiwanese chipmaker Taiwan Semiconductor Manufacturing Co about a partnership to fabricate the chips, they said.
It wouldn’t surprise us if Sam has a broader strategy in mind to insulate OpenAI from the geopolitical challenges surrounding TSMC’s Taiwan-based facilities. One potential approach could involve supporting the establishment of TSMC fabrication plants in the United States. This would allow OpenAI to reduce its dependence on TSMC’s Taiwan operations and enhance its supply chain resilience. It would also align with ongoing efforts by the US government and the tech industry to strengthen domestic semiconductor manufacturing capabilities. While it presents challenges such as significant investments, technology transfer, and regulatory considerations, it would be a crucial step in ensuring a stable and secure supply of custom-built chips for its AI models. And more broadly, the TSMC conundrum poses a unique issue for AI companies like OpenAI. TSMC’s dominance in chip manufacturing is undeniable, as is the fact that its geographic location in a geopolitically sensitive area is a severe supply chain risk. Diversifying this risk is paramount, but doing so is an incredibly high-stakes, costly, and uncertain endeavor. The Arizona plant does paint a picture of what diversification would require (and all the problems that come with it). It’s unclear exactly what Sam is planning, but a long-term vision for chip fabrication outside of Taiwan seems like necessity to secure “chip sovereignty” and the stable supply of tailor-made chips for AI models that would come with that.