Chip manufacturers: I'm sorry, but it will be sold out next year

Text: The semiconductor industry is vertical

Recently, there have been frequent rumors of sell-outs from major chip manufacturers.

It's not common for chips to sell out, but it can happen under some specific circumstances. For example, in 2021, the global chip shortage wave, in this year, the global semiconductor industry is facing a serious chip shortage, and the products of many chip companies are in tight supply or even sold out.

Unlike most chips that were out of stock during the epidemic period, this time it was AI chips in subdivided fields that were missing.

Sold out chip manufacturers

Recently, Nvidia's Blackwell GPUs were all sold out for the next 12 months.

According to Morgan Stanley analysts who met with Nvidia's management, Nvidia's traditional customers (such as Amazon, Dell, Google, Meta, Microsoft, etc.) have purchased all of the Blackwell GPUs that Nvidia and its partner TSMC will be able to produce in the coming quarters. This means that new buyers who place orders now will have to wait until the end of next year to receive their goods.

The Blackwell GPU series was released by Nvidia in March of this year. Blackwell GPUs are manufactured using a custom dual-reticle extreme 4NP TSMC process, and the GPU chip is connected to a single unified GPU via a 10TBps chip-to-chip link with 208 billion transistors. This is an increase from the 80 billion in the Hopper series and includes a second-generation transformer engine and a new 4-bit floating-point AI inference feature.

In August, Nvidia CFO Colette Kress told analysts on the company's third-quarter 2024 earnings call that the company expects to deliver "billions of dollars in Blackwell revenue" in the final quarter of 2024. The sell-out of Blackwell GPUs is partly a reflection of the continued popularity of AI.

Not long ago, Foxconn's parent company Hon Hai Precision also announced that it is building the world's largest NVIDIA GB200 chip manufacturing plant in Mexico to cope with the hot demand for Blackwell chips in the AI market. Liu Yangwei, chairman of Hon Hai, said that the market demand for Blackwell chips has reached a "crazy level". The company plans to produce 20,000 NVIDIA NVL72 cabinets by 2025.

This isn't the first time Nvidia has sold out.

The last time Nvidia sold out was last year's H100 GPU. At that time, since ChatGPT sparked a generative AI boom, the demand for NVIDIA's H100 GPU, which is good at training large models, has increased significantly. Starting in April 2023, the GPU market supply became tight, and by around August, there was a massive shortage of Nvidia H100 GPUs, and even if they were ordered immediately at that time, they would not be available until the first or even second quarter of 2024.

AI companies such as OpenAI and Meta, cloud computing companies such as Amazon and Microsoft, and many technology companies that want to train their own large models have a huge demand for H100 GPUs. It is estimated that the H100 had a gap of 430,000 at the time.

With the AI boom, it is not only NVIDIA that is sold out, but also HBM, which matches high-performance AI chips. In February this year, SK hynix's management said that although it plans to double HBM's production capacity by 2024, the production quota for that year (2024) has been sold out, and the company has begun to actively prepare for orders for 2025. Its vice president of sales and marketing also said that Hynix foresaw the high demand for HBM's memory and was ready to increase production capacity in advance to meet market demand and protect its leading position.

In May, the CEO of SK hynix announced at a press conference that HBM products produced in 2025 were also sold out, according to the mass production plan.

In March this year, Micron CEO Sanjay Mehrotra said on an earnings call that Micron's HBM production capacity has been sold out in 2024. This is mainly due to the rapid growth of HBM driven by the demand for AI servers, which has led to extremely high demand for HBM. At that time, Sanjay Mehrotra mentioned that the vast majority of HBM's capacity for 2025 had also been booked. In June, Micron Technology once again said that all HBM memory chips for 2024-2025 had been sold.

Behind such a hot demand, there is a company that silently collects money - TSMC.

It is no exaggeration to say that TSMC is single-handedly stuck in the AI chip production capacity of AI chip leaders Nvidia, Broadcom and AMD, and controls almost all of the world's AI chip production capacity.

With industry-leading 2.5D and 3D chiplet advanced packaging, TSMC has eaten almost all orders for high-end chip packaging with 5nm and below processes in the market. And because the advanced packaging capacity is far from meeting the demand, Nvidia GPUs have been in short supply for a long time.

TSMC's production capacity continues to be tight. The legal person expects that by the end of this year, it is conservatively estimated that TSMC's monthly production capacity of 3 nanometers will reach at least 80,000 pieces, and it will be further expanded to 100,000 pieces in the future; In the future, another 15,000 pieces of 3nm production capacity will be built in the United States and Japan factories per month, which means that the total monthly production capacity of 3nm process plants at home and abroad will reach about 130,000 pieces in the future. Some institutions predict that TSMC's current 3nm process capacity utilization rate has reached 110%.

TSMC announced its third-quarter financial results today, with consolidated revenue of NT$759.69 billion in the third quarter, up 39.0% year-on-year and 12.8% quarter-on-quarter. Net profit for the third quarter was NT$325.26 billion, up 54.2% year-on-year and 31.2% quarter-on-quarter.

TSMC's gross margin in the third quarter was 57.8%, operating margin was 47.5%, and after-tax net profit margin was 42.8%.

In the third quarter, 3nm process shipments accounted for 20% of TSMC's wafer sales in 2024, and 5nm process shipments accounted for 32% of the quarterly wafer sales. 7nm process shipments accounted for 17% of the quarter's wafer sales. Overall, advanced processes, including 7nm and beyond, accounted for 69% of the quarter's wafer sales.

TSMC can be said to have made a lot of money.

Will the AI bubble be staged again?

Now Nvidia's market capitalization has risen to 3.39 trillion, and TSMC's market value has recently hit a trillion dollars.

The scary thing about Nvidia is that it has risen by almost 1,000% in less than two years, which means that a $10,000 investment at the end of 2022 will be worth nearly $100,000 today. With the stock price pushing up step by step, many people are worried: "Will next year's AI still be popular?" ”

A few days ago, a report about "$2/hour rental H100: the eve of the bursting of the GPU bubble" has attracted great attention in the domestic market. A related article points out that after the Nvidia H100 GPU hit the market in March 2023, its rental price soared from the initial $4.7 per hour to more than $8 per hour due to a surge in demand. But since the beginning of this year, the H100 has begun to be "oversupplied", and the hourly rental price has dropped to about $2.

Will the AI speculative bubble be staged again? This question tugs at the nerves of every AI investor. The answer to this question can be seen in combination with different forecasters, large manufacturers and market demand.

In terms of institutions, it is still in a state of optimism, and the optimistic period is this year and next year. Hyperscalers' capital spending will eventually peak, possibly as early as next year.

Citi: The demand for computing power is still huge

According to the latest forecast from Wall Street financial giant Citigroup, data center-related capital expenditures of the four largest tech giants in the United States are expected to grow by at least 40% year-over-year by 2025. The associated large capital expenditure is largely tied to AI, which means that Citi believes that the computing power demand for AI applications such as ChatGPT is still huge.

"We expect the growth opportunities for AI GPUs and AI networks to expand from connectivity between individual server systems within data centers to large AI computing platforms that connect multiple flagship data centers via DCI," Citi's analyst team said in an investor note. The team of analysts is led by Wall Street star analyst Atif Malik ·.

Goldman Sachs: Big tech companies will continue to invest large sums of money next year

"While these stocks (i.e., AI infrastructure-related stocks such as Arm, TSMC) may seem relatively expensive relative to history, the demand for AI could lead to large tech stocks spending more capital than investors and analysts currently expect in this space," Goldman Sachs analysts wrote in an Oct. 10 note. ”

Google, Microsoft, Amazon and Meta have all said they will continue to invest heavily in AI infrastructure by next year, which will benefit AI hardware companies led by Nvidia. Overall, big tech companies will spend $215 billion in 2024 and $250 billion in 2025 on AI capital expenditures, Goldman Sachs said.

JPMorgan: Earnings are expected to stabilize/rise in the second half of 2025

JPMorgan analyst Harlan Sur expects semiconductor industry revenue to grow by 6% to 8% in 2024. "We remain bullish on semiconductors and semiconductor equipment stocks as we believe equities should continue to move higher, expecting supply and demand to improve in the second half of 2025 and earnings to stabilize/rise in the second half of 2025," he said. ”

In terms of large manufacturers, Micron, SK hynix, and Nvidia are optimistic about AI demand next year.

Micron: Optimistic expectations

On September 26, Micron released its report for the fourth quarter of fiscal year 2024 as of August 29 this year after hours, and the overall performance far exceeded market expectations. Its revenue reached $7.75 billion, beating analysts' expectations of $7.65 billion, a year-on-year increase of nearly 93% and the biggest increase in a decade.

For the next quarter's revenue forecast, Micron also exceeded Wall Street's expectations, giving a figure of about $8.7 billion, more than double the year-on-year growth and higher than analysts' expectations of $8.21 billion.

SK hynix: Increasing investment in AI demand and continuing to expand production capacity

SK hynix is also actively expanding. SK Group emphasized, "With the advent of the AI era, the Group is expected to make historic large-scale investments in business areas closely related to the AI ecosystem, such as HBM, in the next two to three years. ”

To meet the rapidly growing demand for AI development, SK hynix plans to spend about $14.6 billion to build new memory chip production capacity in Korea. Its M15X fab in Cheongju, Korea, is expected to be completed in November 2025, and the company's HBM capacity is expected to increase significantly with the new facility coming online.

In addition, SK hynix has started mass production of the world's first 12-layer HBM3E product with a capacity of 36GB, which is the largest capacity of existing HBMs to date. Compared to the previous generation, the single DRAM chip in this new product is 40% thinner, maintaining the same overall thickness while increasing the capacity by 50%. This product will be available later the year.

Market demand

According to a report by DIGITIMES Research, shipments of high-end AI servers capable of developing and running ChatGPT-level generative AI are expected to account for only 3.9% of all servers by 2024. Analysts believe that this shipment simply cannot meet the needs of cloud service providers.

"So far, Nvidia's GPU craze has only just begun. A full-blown generative AI boom is coming. ”

He calculated that 34,000 high-end AI servers were shipped in 2022, so only one ChatGPT-level AI system could be built (which is ChatGPT). The following year, 2023, 200,000 high-end AI servers will be shipped, so 6 to 7 ChatGPT-level AI systems can be built. It is expected that this year, it is possible to build 18 to 19 ChatGPT-level AI systems. This shipment is not satisfactory for the foot cloud service providers in the United States.

Jensen Huang, the "helmsman" of Nvidia, also appeared a few days ago. In an interview with Altimeter Capital, he stressed that the continued bullishness on Nvidia is nothing like the frenzy surrounding Cisco at the peak of the dot-com bubble. Nvidia is "reinventing computing," and the future will be the era of "high machine learning."

"Moore's Law has basically come to an end," he said, adding that in order to provide the necessary computing power to keep up with the pace of future compute-intensive software, existing data centers will need about $1 trillion worth of GPUs in the next 4~5 years to upgrade.

epilogue

Will AI still be popular next year?

According to the current market situation, data centers and cloud service providers have snapped up AI chip production capacity next year, and capital investment is expected to peak next year.

For now, the situation seems to be very good. However, the time for investment slowdown will eventually come, and the key will be when it will appear. AI software usually comes in the form of a user subscription, while hardware is a one-time sale. In light of this, some analysts have warned that AI chip stocks are in a bubble and will eventually burst once big tech companies reduce their huge spending on AI infrastructure.

In fact, recent earnings reports from tech giants show that the gap between their huge investments in AI infrastructure and the return on investment they receive is widening, testing Wall Street's patience that is gradually being worn out. Shares of Google, Microsoft and Amazon all fell late this summer, after their quarterly financial reports showed that they were spending billions of dollars on artificial intelligence.