TSMC is dominating the manufacturing of processors to the Datacenter. Based on Intel’s datacenter business operating margin, it as safe to assume the company’s business is mainly CPU business rather than AI that would command a higher margin. If this is true, TSMC has total domination in the semiconductor manufacturing of AI GPU’s to the datacenter.
If you get value from our articles, please consider giving us a like or a share
The revenue TSMC derives from Datacenter processing is reported in the High Performance Computing (HPC) segment of the company’s P&L and is currently at 43% or 8.5B$/qtr of TSMC’s revenue.
As the the only company manufacturing AI processors, TSMC should be an obvious benefactor of the AI rally together with its key GPU customers, AMD and Nvidia. However, TSMC does not seem to have seen the same success as its clients.
While TSMC does not reveal any customer data, it is well known that AMD and Nvidia are the company largest customers in the HPC segment.
It looks like TSMC is not gaining from the AI boom as you would suspect.
A simple revenue comparison, shows that the combined revenue of AMD and Nvidia has skyrocketed, while TSMC High Performance Computing have had more modest revenue growth. It looks like TSMC is not gaining from the AI boom as you would suspect.
It will come as no surprise that the bulk of AI value in the Semiconductor supply chain is not derived from the manufacturing of GPU’s rather in the business of the fabless customers of TSMC.
This should be a warning to Intel that the IFS strategy might be less valuable than intel’s efforts in design of their own AI processors.
TSMC also have two other significant HPC customers. Apple is (was) the only 3 nm customer so the total Apple revenue is significant. The 3nm revenue is for two of Apple’s platforms, the A17 processor for iPhone 15 Max and Max Pro and the M3 processors for the MAC Pro.
Based on the die sizes of the 3 types of M3 processors, we estimate that TSMC HPC revenue from the Apple MAC is around 500M$/Qtr.
It is harder to quantify the Google Tensor business as there is no public metric that can be used so we assume that the size of the business is lower than Apples purchases from TSMC as the Tensor only supplies Googles own Data centres where also AMD and Nvidia is present.
While Broadcom business is not large yet, it is possible to quantify so we have included it into the analysis.
There are likely more TSMC HPC customers, but the AMD, Nvidia & Broadcom business is sufficient to take a deeper dive into the numbers.
The supply chain view
The supply chain view of the relationship between TSMC HPC financials and AMD/Nvidia financials can be seen graphically below.
TSMC revenue equates to Cost of Goods Sold (COGS) in the financial reporting from AMD and Nvidia. Both companies have other COGS than what they pay to TSMC. The companies still need staff associated with buying and planning the manufacturing but the expenses are not increasing significantly with revenue increases so we can assume a significant part of the COGS is material cost that equal TSMC HPC revenue.
A comparison between TSMC HPC revenue and the combined COGS of Nvidia, AMD and Broadcom does indeed show a strong connection between the two metric. It is pretty much verified that AMD and Nvidia account for the bulk of TSMC’s HPC revenue.
While this is not perfect science, this analysis can help uncover some interesting changes in the supply chains of AMD and Nvidia.
Subtracting the two metrics shows an interesting inflexion point - it looks like something is changing in the business between TSMC and the company's key HPC customers.
Is TSMC’s HPC business changing?
It looks like TSMC has derived a larger and larger revenue stream outside the 3 companies over time. Apart from Apple and Google, this could come from notebook processors manufactured for Mediatek and Qualcomm although their revenue is not significant and not changing at the moment.
So what causes the drop after the inflexion point? It could be some of the same customers that have created the increase before the inflexion point have had a significant drop in business or has been lost to other foundries.
The largest of TSMC HPC customers outside Nvidia and AMD is Apple’s Mac production. This is pretty stable (with some seasonality) around 6Mpcs sold every quarter while the rest of the HPC customers are not sufficiently large to drive the change.
As COGS is calculated based on goods sold - and not on goods purchased, it is necessary to look at the inventories of the companies involved.
While there has been some jitter, it is likely not sufficient to be a good explanation of the inflexion point - there must be other things going on at the same time. If it is indeed inventory movements, it will soon be evident.
The Changing Supply Chain
The inflexion point also seem to coincide with the introduction of the H100 based on the Hopper GPU silicon and the acceleration in Nvidia revenue. In their Q3 conference call, Nvidia reveal that almost all of the datacenter revenue came from H100 with a small contribution from A100. Both the A100 and H100 has a memory configuration of up to 80GB HBM2E where the H200 will have 140GB.
The move to HBM is changing the supply chain significantly. This could change TSMC’s revenue if they lost some of the packaging revenue to Foxconn although right now it seems that TSMC is the only company capable of doing mass production CoWoS. This could become a future scenario but is not likely to impact the current TSMC business.
Nvidia’s changing bill of material
Most of Nvidias GPU revenue comes from the H100 GPU, with two main flavours:
The DGX is a finished system where the total GPU cost is a fraction of the total cost of the system. The system consists of 35.000 parts of which, many are not semiconductors.
HGX is a modular AI board with GPU’s, Memory, Networking & Power components where the semiconductor costs are a significant proportion of the total BOM.
The cost/price of the DGX system is a magnitude higher than the HGX board.
While Nvidia does not reveal the split between the two systems, it is clear that if the DGX is capturing a significant part of the revenue it would directly impact TSMC revenue as the total share of semiconductors would decline.
The increasing importance of Networking in AI applications
Nvidia reports in their customer segments rather than their product segments and although the company does not report the numbers, they leave sufficient breadcrumbs to give a complete picture.
Nvidia’s networking products originates from the Mellanox acquisition and is a vital part of the AI GPU products. While the networking revenue shown, is what Nvidia sells to the data center customers and not what it consumes internally, It is obvious that the AI revolution also impacts networking and could shift TSMC’s HPC revenue.
The networking revenue would also benefit TSMC (both from Nvidia and other networking component suppliers) but it would be reported in another revenue segment than TSMC HPC.
While it is still too early to conclude what is changing the business between TSMC HPC and its two largest clients, we believe something interesting is going on. We will keep monitoring for insights that can enlighten the increasingly complex AI Semiconductor supply chain.