Intel chips will no longer pay all foundry costs

tech

This article is compiled by Semiconductor Industry Review (ID: ICVIEWS) from nextplatform.

Intel has just restructured its financial situation.

The biggest advantage of Intel's separation of its chip design and marketing business from its foundry business is that Intel's chip product groups no longer have to bear the full massive costs of its manufacturing operations. The biggest disadvantage is that these chip product groups will not be treated any better or worse than external customers who also use Intel's foundry services.

As promised by CEO Pat Gelsinger, Intel has just restructured its financial situation, spinning off Intel Foundry Services (formerly known as Intel Foundry Services) from the rest of the company. Intel Foundry Services is a temporary name for a part of Intel's business that allows the outside world to view the company as a commercial foundry similar to semiconductor industry giant TSMC. This is not a large part of Intel's business, and its profitability is growing. Due to the substantial investments made by chipmakers in processes and factories, and frankly, the business does not have enough volume to cover the costs.

Advertisement

We know that Intel's fabs have not been operating at anywhere near peak utilization rates, and the PC and server businesses have taken a huge hit in the past few years after the massive sales during the COVID-19 pandemic. However, we do not have a clear picture of how deep under water the now-called Intel Foundry Services is. Through the new financial details provided by Gelsinger and CFO Dave Zinsner, we can see the full picture of Intel Foundry Services.

Intel has only readjusted the financial data for the years 2021, 2022, and 2023, and has not provided quarterly details for that period. The data for Intel Foundry Services is expected to go back to the time of the Great Recession, or even to 2015, when the company was united in providing a full-stack of components for HPC systems, excluding main memory, but that never happened.

Interestingly, as the cost allocation for chip manufacturing shifts from "how Intel feels" to "fair market prices," the profitability of the remaining Intel product groups that have not yet been sold - the Xeon products for data centers and the edge, the core CPUs on the desktop - has actually been enhanced. This data comes from historical financial data from the old method and the 8-K filing submitted to the SEC under the new method:In 2021 and 2022, on separate income statements, the Accelerated Computing and Graphics Group (AXG, colloquially known as the old guard) was included in the other categories of the old financial accounting methods. When AXG was disbanded in 2023, a portion of its revenue was allocated to the Data Center and Artificial Intelligence Group (DCAI), the Network and Edge Group (NEX), and the Client Computing Group (CCG), with the names and acronyms of Intel's various departments being inconsistent.

In the old method of Intel's foundry business, Intel merely discussed the external sales of its manufacturing capacity. With the new method, the fair market price of chip manufacturing is extracted from each group, and the new operating profit for each group is calculated by combining other costs. As we can see, the profits of all major groups have increased, with Altera and Mobileye remaining essentially unchanged.

Another positive development is the removal of revenue from DCAI: $3.51 billion in 2021, $2.34 billion in 2022, and $2.89 billion in 2023. The total value for these three years is $8.73 billion, which seems significant to us and appears to be the reason why the Altera FPGA business was spun off. The good news is that DCAI has become slightly more profitable despite a substantial reduction in revenue. However, this is mainly because DCAI does not bear much of the foundry costs.

In technical business terms, Intel's foundry is a real "bugbear." It's no wonder that Intel only separated its financial data after the U.S. government provided funding for the CHIPS Act, under which Intel has received $8.5 billion in direct funding and a 25% income tax credit, as well as an additional $11 billion in federal loans, totaling $44.5 billion.

Part of the confusion stems from Intel's own mistakes, and part is due to the industry returning to a more competitive computing environment while Intel's 10nm, 7nm, and 5nm processes have been stuck.

The generative AI boom that began at the end of 2022 has exacerbated the server downturn, which also hasn't helped Intel's cause. Even though shipments have declined, overall server revenue is still rising, but this is only because AI servers equipped with 8 GPUs cost about $400,000 each. Designs for AI servers using Intel CPUs are becoming increasingly rare, and Intel does not sell main memory, flash memory, Ethernet, or InfiniBand networking, nor does it have a large surplus of "Ponte Vecchio" Max series GPUs.

Intel has a certain number of Gaudi2 matrix math accelerators in the barn, and Gaudi3 is also set to be launched this year, both of which will receive considerable acclaim. But to be frank, if you can manufacture a fairly powerful AI accelerator that can outperform Nvidia's A100 and compete with Nvidia's H100, and it can run PyTorch, you can sell everything you can make because Nvidia has likely already allocated for 2024 and 2025. However, the Gaudi2 and Gaudi3 chips are manufactured by TSMC, not Intel's foundry, so Intel can only get a fair share of TSMC's capacity in chip etching and interlayer stacking.

Over the past five years, Intel's market share in the CPU market has slipped from around 97% to around 75%, depending on the quarter. Intel has steadied its position in recent quarters but still lags in chip manufacturing processes. By 2024, the situation does not look to improve significantly, but Intel could even start mapping out 2025. But as always, Gelsinger is hopeful, and he is right that there are political and economic reasons for having a large chip foundry in the United States.

"When we went through the setbacks of not accepting EUV and Intel 7, the industry was catching up. But the cost and complexity of EUV have caused the industry's curve to flatten," Gelsinger said in a conference call with Wall Street explaining the new data. "We have now broken through the EUV barrier and have moved on to high numerical aperture. Importantly, we are bringing economics back into Moore's Law. For the early EUV technologies and the chips and fabless suppliers using these technologies, they will lose competitiveness if they do not adopt leading-edge technologies. Looking at the history of the industry, we have seen that every decade, the complexity of capital technology reduces the number of companies that can economically manufacture cutting-edge transistors. The adoption of EUV is the barrier for this decade, and AI is driving an explosive growth in computational demand. Physics is driving the development of density and packaging, bringing more chips closer together and packaging them in a 3D manner. The rack is becoming a system. The system is becoming a chip, and when we reach the other side of the EUV transition, as seen in this chart, Intel's position in the industry is very valuable."Here is the translation of the provided text into English:

This is the chart that Gelsinger referred to, which outlines the competitive landscape for Intel 7 (10nm) and Intel 3 (5nm) processes, without mentioning the 20A (2nm) process, and directly jumping to 1.8nm and 1.4nm processes:

Intel's profitability hinges on its ability to become a top-tier EUV foundry partner for its own products as well as for others. So far, Intel has secured $15 billion in deals from external sources, spanning many forward-looking years, most of which are on the 18A process, where Intel will be on par (or close) with TSMC.

Intel's transition to EUV manufacturing has only just begun, with its operating expenses as a percentage of revenue being very high, leading to significant losses for Intel's foundry business. However, as shown in the chart on the upper left, by 2027, EUV is expected to account for about two-thirds of Intel's foundry wafer structure, and by 2023, this proportion seems to be around 85%, with operating expenses reduced to half of revenue.

In the meantime, Intel will increase its foundry utilization rate from the current approximately 75% to over 95% by 2030, while increasingly relying less on TSMC for wafer production, but not entirely eliminating its own use.

As it stands, Intel's foundry business looks quite bleak, with revenue of $18.91 billion in 2023 and an operating loss of $6.96 billion, expected to peak this year as Intel completes its 5N4Y (five nodes in four years) work. Intel's foundry business is projected to break even by 2027, with Intel aiming for a non-GAAP gross margin of 40% and a non-GAAP operating profit margin of 30% by 2030. In 2023, the operating loss calculated on a GAAP basis represents 36.8% of revenue. This is a level of loss seen in some startup companies.

Looking ahead, however, Intel believes that by 2030, it can generate over $15 billion in annual revenue from external foundry customers, which is a rather ambitious business goal. We believe that Intel's product group will bring in at least $20 billion in revenue annually, although Intel has made no predictions here, and if its competitive position in key computing engine markets improves, the annual revenue could reach as high as $25 billion. Generously assuming that Intel could have a $40 billion foundry business by 2030, with substantial profitability.Then Kissinger could become the chairman, find a new CEO, catch his breath and take a temporary break, but there is still a long way to go until 2030.

*Statement: This article is originally created by the author. The content of the article represents their personal views. We republish it solely for sharing and discussion, and it does not mean that we endorse or agree with it. If you have any objections, please contact the backend.

tech
1655 62

Comment Box