Opinion: The foundry makes all of the logic chips critical for AI data centers, and might do so for years to come.
Large language models (LLMs like ChatGPT) are driving the rapid expansion of data center AI capacity and performance. More capable LLM models drive demand and need more compute.
AI data centers require GPUs/AI Accelerators, switches, CPUs, storage and DRAM. About half of semiconductors are consumed by AI data centers now. This percentage will be much higher by 2030.
TSMC has essentially 100% market share in AI data center logic semiconductors. TSMC makes:
The only essential AI data center chips not made by TSMC are memories: HBM, DDR, flash.
Fig. 1: Created by ChatGPT from text input.
TSMC has the big 4 things AI data centers must have:
Advanced Process Technology: AI data center chips, especially AI accelerators, need the most advanced process technology to get the most transistors on a chip. Foundries like GlobalFoundries were not able to fund the development of advanced finFET nodes. Only Intel, Samsung, and TSMC have 2nm-and-below process technologies and roadmaps.
Advanced Package Technology: LLM models have grown exponentially in size, so a single GPU chip can’t process the model even with the most advanced process nodes and maximum reticle size. Multiple GPUs/AI accelerators are required to run an LLM model. With multiple chips the bottleneck becomes chip-to-chip data rates. The data rates between chips using copper interconnects at 200 Gb/s is very, very slow compared to on-chip data rates.
The solution is advanced packages that integrate multiple GPU chiplets and HBM memories on a multi-reticle substrate with very, very fast chip-to-chip data transfer to create a GPU compute assembly. TSMC constantly develops ahead of customer needs. Co-packaged optic communications will be in use for AI accelerators of the future to replace copper connects. TSMC in 2024 announced their COUPE process for integrating optical engines in advanced packages. Nvidia earlier this year announced their first scale-out switch using this technology.
Fig. 2: HBM, multiple AI chiplets, voltage regulators & CPO in one package. Source: TSMC
To ensure yield and production ramp, ALL of the chiplets in an advanced package need to be made by TSMC. The only chips not made by TSMC are the DRAM memory and the passive.
TSMC started developing multi-chip substrate packaging years ago when HBM started to be used by GPU and has continually increased the size and complexity of what they can deliver. They recently announced a wafer-size substrate:
Fig. 3: Larger substrates increase compute speed with fast chip-to-chip data rates. Source: TSMC
Capacity: TSMC has the most advanced process and package technology capacity in the world with a significant and growing portion in the USA.
In Q4/24 TSMC hit 67% share of the global foundry market, up about 10% from early 2023. Samsung was 11%. Intel was in “other”. (Data courtesy of Counterpoint). TSMC’s share of advanced process technology foundry is much higher.
TSMC is growing the percentage of its capacity in the USA for the most advanced nodes. It plans to produce N2 and A16 nodes in Arizona with a goal that about 1/3 of its total production in these nodes comes out of the USA. It will continue to develop the most advanced nodes in Taiwan first then transfer them to Arizona.
Morgan Stanley estimates that in 2026 TSMC will have 90% of the world’s CoWoS (advanced packaging) capacity with the rest being at Amkor/UMC/ASE – Intel and Samsung are not mentioned. TSMC also has major and growing advanced packaging capacity in the USA.
When Nvidia or Amazon or Google needs an AI Accelerator it needs not only the advanced processes and the advanced packaging but also it needs the capacity and the predictability of TSMC to ensure they can ramp and fill their data centers quickly. TSMC has proven with Apple they can scale multiple complex chips in huge volumes rapidly and reliably.
Financial Strength: TSMC’s market cap is almost $1 Trillion with a very strong balance sheet. Their customers ultimately are the companies developing and providing the frontier LLMs: Microsoft, Amazon, Google, OpenAI who are Trillion $ market cap companies (or on track to be so). They want big, reliable, financially strong vendors.
Management Strength and Depth: Ultimately, TSMC’s #1 competitive edge is the strength and depth of their management and their systems/processes for running the company. My last startup was an IP partner with TSMC for 10 years and we were a customer designing chips for 5 years. We worked with TSMC on a dozen process nodes, with dozens of joint customers and I met regularly with TSMC executives in San Jose and Hsinchu. They are amazingly well run and everyone is well synchronized working in the same direction. Their ability to deal with incredible technical complexity across dozens of fabs in multiple countries is impressive. Their continuity is a key part.
Today the vast bulk of AI compute in data centers is based on Nvidia, and Nvidia is one of TSMC’s largest customers. AMD currently has a small market share, but their GPU offering is catching up to Nvidia; they still have work to do on the scale-up network and software, but it seems likely they will gain share. Customers want choice they are motivated to give AMD a chance, and the June 12 Advancing AI AMD event showcased multiple major players using AMD AI to support their models with good results.
Lisa Su, AMD’s CEO, projects the Data Center AI Accelerator TAM is growing >60%/year and will exceed $500 Billion by 2028! And the mix of inference vs. training is shifting fast. In 2025 it was about 50/50 mix, but by 2028 inference will be 70% of the TAM. The whole semi market is forecast to be $1 trillion in 2030, so AI data centers are becoming the majority of the market.
GPUs are extremely flexible. They can do training and inference. But they’re expensive.
The biggest 4 CSPs (cloud service providers) buy about half of Nvidia’s outputs, or roughly $20B annually each. If the market grows as fast as Lisa and Jensen project, the biggest 4 CSPs will be buying close to $100B of GPUs per year.
This means the biggest Hyperscalers and LLM companies have the scale to design their own chips. They design their own models, so they know exactly what hardware they need, and they can build accelerators that are cheaper, partly because they can leave out hardware they don’t need and partly because they can work with vendors charging lower margins or even work directly with TSMC. (They’ll still need GPUs for their cloud customers who design their models for Nvidia/AMD).
Amazon and Google have been designing their own AI accelerators for almost a decade. The first Google TPU paper was published in 2017. Amazon acquired Annapurna Labs in 2015: this team has designed the generations of Inferentia and Trainium AI accelerators. Amazon is one of TSMC’s top 10 customers.
Now every major Hyperscaler/LLM is building their own AI accelerators, including Microsoft Maia, Meta MTIA, and widely rumored AI accelerator efforts at OpenAI and xAI. (The only exception is Anthropic, which is using Amazon.) Today these companies are working with Broadcom, Marvell, Alchip, and others that have the ability to design 2nm or lower chips with advanced packaging and high-speed PHYs. Like Amazon, some Hyperscalers/LLMs could decide to buy existing players to integrate to tape-out directly with TSMC.
The future is hard to predict, but the most likely scenario in 2030 is that Nvidia still will be the market leader, with AMD’s share becoming much larger than today and a significant share of the market being custom AI accelerators by the big Hyperscalers/LLMs, maybe 10% share for each of Amazon, Microsoft, Google, Meta or the ASIC companies who build the chips for them. This means six of TSMC’s biggest customers in 2030 will be Data center AI providers.
Hyperscalers/LLM companies want to keep down costs but first they have to make sure they get the AI accelerators they need in high volume with high reliability.
Intel and Samsung are working on advanced process nodes and advanced packaging, but all of the big players are currently 100% TSMC. The chip companies will be motivated to evaluate and consider Intel and Samsung, because everyone wants options, but TSMC’s multiple advantages will be hard for Intel and Samsung to overcome anytime soon.
Leave a Reply