Intel’s decision to tap pure-play foundries like TSMC for potentially any future project has sent shockwaves through the industry. One of the biggest questions raised in the wake of the announcement is just how temporary it truly is. Intel is definitely positioning the delay as a pragmatic matter of getting its house in order, and for now, TSMC seems to agree.
According to DigiTimes, TSMC has no plans to expand its fab capacity for Intel, viewing it as a temporary customer rather than a long-term win. If you’re a fan of Intel manufacturing (or would like to be again), this is a very good sign. While Intel isn’t one of the five largest semiconductor manufacturers any longer, all of which are capable of over 1,000,000 wafer starts per month, it’s still one of the largest at an estimated 817,000 wafer starts per month. And unlike TSMC, Samsung, or GlobalFoundries, Intel uses its own capacity almost entirely for its own products.
Further complicating the issue is the fact that Intel manufacturers the majority of its products on its leading-edge nodes, while pure-play foundry wafer starts will be spread across all of the nodes the foundry manufacturers, not concentrated in 1-2 processes. In short, the capacity gap between Intel and any foundry Intel wanted to tap to handle its manufacturing is likely to be large. TSMC is said to currently view itself as Intel’s “rescuer,” not its long-term manufacturing partner.
Intel Could Be Snared by the Same Trap That Once Caught AMD
Before AMD and Intel settled their antitrust lawsuit, AMD was under a much more restrictive x86 license than it currently holds today. One of those restrictions was the requirement to own its own fabs. AMD was not allowed to pay TSMC, UMC, or any other pure-play foundry to manufacture x86 CPUs. It had to use its own facilities. Intel, of course, is under no such restriction, but there’s a variation of this problem haunting the semiconductor stage nonetheless.
Intel’s foundries are explicitly optimized to build Intel x86 CPUs, and its CPU designs are intended for fabrication at its own foundries. This hand-in-glove engineering is the advantage of being an integrated device manufacturer (IDM), and it’s one of the reasons Intel used to give to explain its industry-leading performance. So long as Intel is building and filling its own fabs, this model works well. Shifting business to a different foundry, however, creates problems.
The fact that Intel’s fabs are so specialized means they principally have value to Intel. Every dollar that Intel invests in paying TSMC to implement a specialized product line is a dollar it isn’t investing in fixing its own fabrication technology. Every facility that TSMC brings online for Intel to use is a facility it’ll probably need to use for something else as soon as Intel can return to its own fabs.
What happens if Intel can’t return to its own fabs? This doesn’t actually fix anything in the near term. TSMC doesn’t have the capacity to absorb Intel’s entire business because 1). Intel’s business is huge, 2). Fabs take 3-5 years to build. Even if Intel wanted to dump all of its own manufacturing today and TSMC actively wanted to take over as Intel’s manufacturer, it would take years to bring up new factories or modify existing Intel fabs to meet TSMC’s new guidelines.
I wrote the other day that Intel had given itself a 24-36 month deadline to fix its manufacturing, but upon further reflection, I’m not sure that was the most accurate way to summarize the problem. Intel can’t afford to wait 24-36 months to begin making plans to move its manufacturing to TSMC or Samsung. Given the difficulty and complexity of such transitions, Intel would need to be making an announcement more like this:
“We intend to deploy our own Xnm node and will transition to TSMC for Ynm beginning in 2022 / 2023.”
It seems unlikely anything less would suffice. Intel’s foundry partner would want the assurance of a publicly announced roadmap and would need the lead time to either build or allocate capacity. Intel, in turn, would have an obligation to inform its investors and to develop plans for how it would dispose of its foundry business. Building on n-nanometer at Intel and n+1 at TSMC would also smooth the roadmap rather than forcing Intel into further delays.
Given these facts, I’d modify my own earlier statement. The fact that Intel is still committing to a late 2022 – early 2023 timeline for its 7nm CPUs shortens the window the company has to pull the trigger on a production change. Should it decide to shift to TSMC or Samsung, it would need to make that announcement within the next 12-18 months to avoid the likelihood of even larger delays. The bring-up time on new fabs is long enough that Intel would almost certainly need to either license a process node, pay TSMC directly to build a fab, or operate its fabs in tandem with TSMC to avoid severe production shortages.
I still think it’s more likely that Intel deploys its own 7nm node and moves ahead with its efforts to overtake the pure-play foundries and re-establish process leadership than that the company pulls the plug on its own IDM status. The timeline for making that decision just may be a little shorter than I initially implied.
- Intel, AMD Reportedly Fighting for Capacity at TSMC
- We’ve Never Seen Intel Struggle Like This
- Intel Announces 7nm Delays, May Use External Foundries For Future CPUs