Intel Foundry Update—A Plan to Secure Western AI Chips

Subscribe To Download This Insight

By Paul Schell | 1Q 2024 | IN-7263

Intel Foundry becomes a separate business unit aiming to become a leading fab by the latter part of this decade. An updated roadmap pivots the company toward the Artificial Intelligence (AI) era as a “systems foundry,” addressing all aspects of chip manufacture.

Registered users can unlock up to five pieces of premium content each month.

Log in or register to unlock this Insight.


Intel Repackages Its Foundry Business


Intel Foundry Direct Connect saw several significant announcements around Intel’s foundry business strategy, which will be operated as a separate business unit. The ramifications of the project extend beyond the commercial success of one company, hence why U.S. Secretary of Commerce, Gina Raimondo, tuned in to voice her approval of a more diversified semiconductor supply chain. Geopolitical tensions are influencing chip manufacture, and the surging demand for strategically essential AI chips only increases their importance. Intel is in the unique position of being the only Integrated Device Manufacturer (IDM) and foundry headquartered in the United States; however, it faces stiff competition from pure-play foundries like Taiwan’s TSMC, which it plans to rival, and has lost ground to competitors from issues arising from product slips, like the delay to its 4th Gen Xeon Scalable processors.

Under Gelsinger’s leadership, Intel has doubled down on its foundry business and invested billions of dollars to reclaim leadership. The “S” from IFS has been axed, and the company has launched what it brands the “world’s first systems foundry,” with Artificial Intelligence (AI) front and center. This involves establishing Intel Foundry as a separate business unit with its own Profit and Loss (P&L) statement that will need to compete for orders like any third-party supplier. The service will also open up its entire fabrication ecosystem to outside customers, from lithography and packaging to testing. Boosted by these changes, the company aims to reestablish U.S. semiconductor fabrication leadership in 2025 with Intel 18 Angstrom (18A) , and the plan to launch Five Nodes in Four Years (5N4Y) was reaffirmed. Investors were entertained by a recorded appearance from Microsoft’s Satya Nadella in which he revealed a chip deal with an expected lifetime value of over US$15 billion.

These ambitions apply across Intel node families, and will be realized with the help of an extensive line-up of ecosystem partners, including Synopsys, Cadence, Siemens, and Ansys, whose Electronics Design Automation (EDA) and Intellectual Property (IP) portfolio will be used to support foundry customers. The guest of honor, however, was Arm (already working with Intel’s foundry), which will further collaborate on supporting emerging System-on-Chip (SoC) designers using Arm IP, including startups that may also be entitled to financial support. The ambitious roadmap aims to carry customers through their long-term product evolutions and position Intel Foundry as a one-stop shop for fabless chip vendors of all stripes.

Intel Foundry Pivots to Capitalize on Industry Trends


The open systems foundry approach coincides with several industry trends. First, the shift toward systems of chips, aka chiplets or tiles, whereby individual dies with various Instruction Set Architectures (ISAs) produced on different semiconductor process nodes are packaged together. These packaged systems provide flexible solutions that can target the economics of individual markets. This has been implemented by AMD and Intel in their latest heterogenous x86 AI PC and server chipsets, as well as a host of fabless chip vendors like Ventana and Untether AI in the RISC-V space. By breaking the die down into individual processors—like Central Processing Units (CPUs), Graphics Processing Units (GPUs), and Neural Processing Units (NPUs)—with a standardized interconnect, elements can be produced on different nodes. For example, accelerators can be made on legacy nodes, with a simpler manufacturing process than for GPUs and CPUs. All component dies can then be “packaged” together using a standardized interconnect like Universal Chiplet Interconnect Express (UCIe), which Intel has promoted heavily. Heterogenous System in Packages (SiPs) can be brought to market faster than their monolithic counterparts, which is important going forward, given their applicability to rapidly changing AI workloads. Finally, chiplets make better use of the precious real estate on a chip, breathing new life into Moore’s Law, to allow more transistors to be packed effectively and economically into heterogenous packages.

Second, vendors of High-Performance Computing (HPC) data center servers are considering Arm Neoverse designs as an alternative to Intel’s own x86 architecture for certain market segments. Intel’s deeper collaboration with Arm on the production of the latter’s IP is arguably a strategic move to capitalize on this shift; for example, Arm processors are in a growing number of supercomputers, and feature in NVIDIA’s GH200 server platform for generative AI (not currently produced by Intel). A deal to make a Faraday Neoverse SoC using Intel 18A shows confidence in the process. So, even if Intel’s own processors lose market share, which is a possibility in both PC and server markets as more power-efficient RISC architectures gain traction, the company can still ride the wave of demand for Arm’s IP.

Third, geopolitical risk-averse semiconductor executives will appreciate Intel’s prospective position of having more capacity to produce advanced nodes in the United States. Intel Foundry can capture orders from fabless vendors looking to “reshore” their advanced process manufacturing and derisk their supply chains away from Taiwan. But one peculiarity of this business model is also the fact that, as an IDM, Intel’s own semiconductor IP will be manufactured at the same plants as designs produced by fabless competitors, so ensuring trust by constructing appropriate silos is vital for attracting new customers. Moreover, and as acknowledged by Intel, the status as an IDM means higher Capital Expenditure (CAPEX) and Research and Development (R&D) spending than fabless competitors, thanks to the investments needed to maintain said manufacturing capacity. But operating Intel Foundry as a separate P&L entity should enforce more efficiency in the manufacture of Intel’s own semiconductors for its product group and create a more agile business unit with clearer pricing, better able to compete with other foundries.

Orders, Not Subsidies, Are the Ultimate Bellwether


Semiconductor supply chains are among the most complex in the world. While world-class IP capable of handling the most demanding AI workloads still comes out of the United States, only a very small percentage of overall cutting-edge manufacturing is located at home. The Creating Helpful Incentives to Produce Semiconductors (CHIPS) Act contains US$50 billion to bolster semiconductor manufacturing, some of which is destined for Intel’s coffers. However, Intel’s commercial success will be borne out not by its eligibility for government subsidies, but its ability to convert its sizable and ambitious investments in the system’s foundry strategy into billions dollars of orders—all while staying on top of its own product development. There are several steps Intel can take to promote and improve the long-term chances of success in its foundry business:

  • Intel should focus on the delivery of the 5N4Y roadmap, not only for the direct commercial value, but also to win back the trust of its customers and the investor community, and make up for past failures. This message is a strong external marketing tool to demonstrate the capabilities of its foundry business, and inspire confidence in its strategy.
  • Intel should continue to promote and nurture the UCIe standard essential for the proliferation of systems of chips using chiplets, which it plans to produce and package.
  • Investments in new fabs require vast upfront CAPEX, but these new facilities also require a long-term plan to find and retain talented professionals with experience on leading-edge (as well as legacy) nodes. Intel’s community college program to train technicians in Ohio is a good example of this. Investments in electrical engineering courses at the bachelor’s, master’s, and PhD levels are also vital.
  • As well as tapping into the burgeoning Arm data center market, Intel should continue to explore RISC-V customers targeting AI workloads (ABI Research’s RISC-V for Edge AI Applications report (AN-6121) forecasts an increasing share of RISC-V into AI chipsets). It has invested capital in leading startups like Tenstorrent and SiFive, but it should also seek foundry customers implementing RISC-V architectures. The deal with Esperanto to manufacture its data center and edge AI chips is a good example of this.

The ongoing threat to market share from China’s homegrown processors is another consideration as the country invests heavily in its own domestic supply chains in response to U.S. sanctions. However, surging Western demand for AI chips somewhat offsets this trend, as do the incentives provided by the CHIPS Act and the appeal of “Made in USA” semiconductors. The potential efficiency improvements of operating Intel Foundry as a separate business unit, along with the high-profile deal with Microsoft to secure capacity on the 18A process, are positive developments.


Companies Mentioned