Will AI growth grind to a halt due to bottlenecks?
AI is rocketing ahead. It is the biggest industrial revolution of our age. AI adoption is growing, but still most are at early stages of learning.
Anthropic, the leading frontier model provider with an annualized revenue run rate (ARR) of ~$47 billion with OpenAI close behind at ~$30 billion (Forbes). Google Gemini revenues aren’t broken out but Google Gemini processes over 3.2 quadrillion tokens per month (Google I/O, May 2026), which is up 7x year over year.
Anthropic says it could grow faster with more compute. Everyone says they cannot get enough compute. Hok Tan, CEO of Broadcom, said on CNBC June 24 that customer demand is insatiable through 2028. OpenAI has developed its own AI accelerator, Jalapeño, to deliver more throughput with less power and silicon and as a way to add to their compute resources.
The major players are buying or renting AI compute wherever they can find it. Older Nvidia GPUs are still fully utilized because they provide much-needed incremental compute in a constrained environment. Hyperscalers are building their own compute to optimize for lower cost and power, but also to provide additive capacity to what they can buy from Nvidia and AMD. Anthropic and OpenAI have learned to run their models on any kind of AI compute that they can buy or rent in scale.
Still, there isn’t enough AI compute due to multiple supply chain bottlenecks.
Figure 1: AI Growth faces high hurdles to growth (Source: ChatGPT prompt)
Almost all AI Compute is manufactured by TSMC (Taiwan Semiconductor Manufacturing Co.) because TSMC has the advanced processes needed, in the volume needed, and with the advanced packaging required to integrate multiple chiplets on interposers and substrates. It’s an almost-monopoly for data center AI. TSMC’s revenues are expected to grow 4X from 2023 to 2028. Unlike the memory companies, it has not radically raised prices, though its margins are trending upward into the mid 60% range.
Recently TSMC raised its outlook for the global semiconductor market to $1.5 trillion by 2030, up from its earlier estimate of $1 trillion (Reuters). AI is expected to be 55% of the total market, followed by 20% for smartphones and 10% for automotive.
TSMC’s CEO CC Wei says, “it will be a long time before we can meet customer demand” (Bloomberg). This is despite TSMC investing heavily for multiple new fabs and packaging facilities in USA and in Taiwan. Its CoWoS (Chip-on-Wafer-on-Substrate) packaging capacity, essential for AI compute, is growing at 80% CAGR from 2022 to 2027. TSMC is outsourcing some CoWoS packaging to ASE and Amkor to help with demand. AI accelerator wafer demand is growing >11X in the same period.
TSMC is developing CoPoS (Chip-on-Panel-on-Substrate) packaging to replace CoWoS using glass core substrates to cut costs and boost wafer utilizations (wccftech.com).
TSMC has done an amazing job of developing advanced processes and packages while ramping numerous manufacturing plants in multiple locations worldwide. But even TSMC can only grow so fast – people are the ultimate limiter.
TSMC is expected to prioritize Nvidia, AMD, Broadcom, and its other major strategic customers. TSMC spends a lot of effort to carefully evaluate their customer’s wafer demands against all the market data they can find to validate that orders are realistic. The foundry is also careful to take care of smaller companies it sees as having high potential.
There are few good alternatives to TSMC:
Given the huge demand, it seems likely that major players with deep pockets will try building chips with Intel and Samsung, perhaps their lower volume and/or lower complexity products, especially if that frees up their TSMC allocations for products that can’t be outsourced. Rumors are Google has given Intel (or Samsung?) orders for millions of TPUs (Tensor Processing Units). President Trump recently said Apple will use Intel, though Apple has not confirmed this. It seems likely that one of, or perhaps both, Intel and Samsung may emerge as a viable #2 to TSMC, but it will be a long road.
All foundries, except SMIC, are bottlenecked by the Netherlands’ ASML (Advanced Semiconductor Materials Lithography), which is the only supplier of advanced lithography manufacturing equipment. It doesn’t seem to be the bottleneck in capacity expansion, though.
All foundries are also bottlenecked by ABF (Ajinomoto Build-up Film), which is used to connect GPU dies to HBM stacks. Japan’s Ajinomoto makes more than 95% of the global supply of ABF. It raised prices 30% in 2026, and projects a supply gap of >20% in 2027 (Wall Street Journal).
In the 1990s there were dozens of manufacturers of DRAM. Today, three companies dominate — SK hynix, Samsung, and Micron.
There is a rising DRAM supplier in China, CXMT, with revenues of $8 billion in 2025. Its sales are primarily in China and its products lag those of the industry leaders in specs. Yangtze Memory is close behind, building three new factories in China to double its current capacity. (Wall Street Journal)
The critical shortage is in HBM (high-bandwidth memory) DRAM, which is critical for all GPUs/XPUs made by Nvidia, AMD, Google and Amazon. About 80% is made in South Korea.
Manufacturing more HBM requires DRAM wafer fab capacity, but also complex packaging capacity. HBM is comprised 8-, 12-, or 16-high DRAM stacks, with through-vias to ever-more memory in a small space. This is very challenging to manufacture.
DRAM has traditionally been a brutal boom and bust business — high prices and margins during shortages, then years of oversupply, low prices and losses. This makes DRAM companies very conservative to add capacity. They worry they will swing back into oversupply. DRAMs are very, very complex, so a new DRAM process takes >5 years to develop, and a fab several years.
Unlike TSMC, DRAM companies are raising prices to meet demand. They are working on strategic supply agreements, such as recently announced between Micron and Anthropic, to lock in customer demand for the long term, typically with upfront cash. As a result, all three companies are now >$1 trillion market cap.
JEDEC recently approved a new version of HBM that uses more standard packages and glass substrates that can avoid the ABF shortage. This will take years to phase in.
The biggest problem for data centers is Power. Amazon’s CEO named this as the #1 constraint (Wall Street Journal).
Data centers are using whatever power they can get, wherever they can get it — grid power, natural gas, Bloom Energy fuel cells, Babock & Wilcox steam turbines, solar+batteries, …
The grid can’t add capacity fast enough for the demand, and there is growing community resistance in many locations to data centers increasing their local electricity prices.
Three companies — Tesla, Sunrun & Renew Home, a Google spin-off — recently announced they were working together to free up enough electrical capacity to meet the needs of 17 large data centers during periods of high demand. They will do this by getting consumers to opt in and let them tap their home batteries during periods of high demand.
Hyperscalers are contracting natural gas at the production point where supply exceeds what pipelines can take to markets. Chevron recently struck a 20-year agreement to sell electricity to Microsoft working with Joulent, which is building a 2.7GW power generation plant on a campus in the heart of the Permian Basin oil-and-gas field in West Texas. (Wall Street Journal)
It’s not just the power sources in short supply. All the other items needed to power a data center are in short supply — the transformers, high-voltage breakers, etc. GE Vernova, the leading gas turbine supplier, is sold out through 2029.
Solar and batteries are components of an off-grid power solution, but they are not economic as a solution by themselves because the worst-case scenario is a winter cloudy day without enough solar to charge the batteries for the long winter night. The capital required is multiples of what is required for solar on a sunny summer day with a short summer night. So solar and batteries will be used in combination with natural gas, especially if natural gas or the turbines are in short supply.
There is also growing resistance in many communities and states to data centers. Even in Texas, the Texas Tribune recently reported most Texans oppose data center construction. The biggest concerns are that data centers will drive up electricity prices and use up scarce water. Cerebras’ CEO, Andrew Feldman, says that the California almond industry uses more water than all the data centers in the USA (CNBC).
This is perhaps the least daunting bottleneck because the bottleneck is not here yet.
Lasers are used in scale-out pluggable transceiver optical links between all the top-of-rack switches in the data center.
Lasers will also be used in scale-up CPO (co-packaged optics) links that will become prevalent over the next 2 to 5 years. There are 10 to 100X more scale-up links than scale-out.
The largest 3 laser suppliers are Coherent, Lumentum, and Sumitomo, which together have 68% share (source: fact.MR). They have their own manufacturing facilities in multiple locations. Lumentum is the largest supplier. Coherent is the first to move to 6-inch wafers for InP (Indium Phosphide). There are many other suppliers, including Broadcom, Mitsubishi, MACOM, Applied Opto, and Landmark.
AI laser leaders Lumentum and Coherent are now over $60 billion market cap each, 10X what they were just a year ago. Both are sold-out. Both require up-front cash to get capacity. In March, Nvidia announced a $2 billion investment in each of them to secure supply chain capacity. This happened shortly before Nvidia’s GTC, where Jensen Huang showed CPO starting on the Nvidia roadmap from 2028.
At its briefing at OFC (Optical Fiber Conference) in March, Lumentum showed it is rapidly growing InP capacity, but demand is growing even faster.
Figure 2: Lumentum InP demand is growing fast, but demand grows faster (Source: Lumentum)
Coherent also did an investor briefing at OFC showing its InP output capacity doubled in 2026, and will more than double in 2027 – and will continue to grow.
The more suppliers, the smaller CapEx required for an InP fab, and the shorter lead time to build an InP fab. That means this bottleneck is the most manageable of the big four bottlenecks.
TSMC’s capacity is primarily in Taiwan, which is very close to a much larger China, with the world’s second-largest military and an intention to re-unite Taiwan with China, by force if necessary.
South Korea makes 80% of the world’s HBM, but all of the fabs and packaging factories are within range of the missiles and nuclear weapons of the world’s least predictable country, North Korea.
Rapid AI growth is stressing numerous parts of the supply chain. The supply chain is responding, but these bottlenecks are likely to be with us for years.
Next month we’ll explore strategies for coping with these bottlenecks to keep AI growth strong.
Leave a Reply