Intel's New ASIC Business
Analyzing the secret weapon to help Intel Foundry
Of all the things that came out of this quarter’s earnings, Intel’s new ASIC design business is likely the most underrated news.
Why?
Well, let’s look at the ASIC design players in a three-tier list:
A) Broadcom
B) Marvell
C) MediaTek, Alchip, GUC, etc.
The cost-saving proposition of the ASIC design business is pretty simple to see for a customer like Google. Nvidia charges you a ~75% gross margin on every part of the chip. You pay 75% on the HBM they bought from SK Hynix; you pay 75% on every part of the chip’s design, whether or not it involved their “secret sauce” or was just a cheap component.
In contrast, Broadcom comes to you and says, “We’ll just sell you some IP and help you with the stuff you can’t do.” (There is also the benefit of an ASIC directly tying the hardware to your specific software demands.)
Broadcom is obviously the main game in town. They offer the best IP and services and command substantial margins in the mid-60s for their AI business. The rest of the players aren’t even close. However, Broadcom’s main business is being a middleman for hyperscale customers. That’s always been a dangerous game when your customers are big, wealthy players. Broadcom’s main objective is to keep its customers in a box and dependent on them. So far, that has worked out pretty well.
But how long will that last? Idk lol.
Anyone who has worked at a big corporation knows how slow execution can be when working across subcontractors, let alone a chain like Company A → Company B → Company C that then has to loop back Company C → Company B → Company A. This is mostly how it works for hyperscalers at TSMC. Sometimes they might go directly to TSMC, but then they still have to loop in their ASIC partner. Hyperscalers have an inherent interest in going directly to the foundry: fewer layers, faster execution, and aligned interests.
With Broadcom, they are your partner, but certainly not a cooperative one. Hock’s goal is to extract as much money from you as possible, leaving you dependent on him. At the end of the day, you’re still better off versus Nvidia... but he still stole your wallet.
This is where Intel comes in.
Intel’s main objective for its ASIC business isn’t really to be an ASIC business; it’s to get you to Intel Foundry. If Google slowly starts to “eat into” some of the IP they buy from Broadcom, Broadcom starts to lose its whole business. That’s a problem for Broadcom. For Intel, if a customer begins to move up the value chain, that’s fine. The hyperscaler just ends up buying a 40% gross margin wafer + packaging. This relationship is fundamentally more aligned with the customer’s interest. As of right now Intel has no ‘AI’ revenue, so really any profits from the ASIC business is a cherry on top of a sundae. Plus Intel IP developed for ASICs can help Intels GPU efforts as well.
So you might ask: why didn’t Intel do this earlier? I think we can see the answer by looking at the 18A experience. Intel went on an IP spending spree across EDA companies and IP design firms. Intel needs Broadcom to have its SerDes on its node; it needs the Alchips and Marvells of the world to offer IP on its process node. This is beyond the “standard” IP that EDA players might offer and develop.
In my opinion, this is likely why Intel has struggled to land hyperscale customers. Intel was dealing with intermediary ASIC houses to get custom IP to their actual end customers (who have no experience using Intel’s process nodes). This hurt execution and time-to-market, which is key for a new entrant in the foundry business.
Lip-Bu Tan has realized his best bet is going direct and essentially offering hyperscale customers a clean off-ramp. If they choose Intel ASIC, they get the benefit of dealing with the foundry on a 1:1 basis. They know Intel has no interest in putting them in a box to extract a toll like Broadcom. In fact, if hyperscale customers use less IP and services from Intel ASIC over time, it just means they are developing more on Intel’s process nodes. This DTCO (Design-Technology Co-Optimization) is a key reason for TSMC’s success, and Intel needs this virtuous cycle for its foundry.
Right now, hyperscale customers are peanuts in volume compared to Apple, Nvidia, Qualcomm, and AMD. They are likely to be last in queues and get worse pricing. They also have less input on process node definition. If Google wants a specific metal stack for its TPU or to target a specific voltage range, TSMC is not going to prioritize them over what Nvidia wants. But at Intel, the hyperscale customers are their first priority.
Sam Altman got called a “podcasting bro” by TSMC and wants more capacity. Well, he’s got better luck getting that volume at Intel.
Other thoughts on the quarter
Intel delivered solid results Thursday, but much like last quarter, analysts seem to have misunderstood the story, which hurt the share price today. This doesn’t change where we are going over the next year.
Dave Zinsner listened to my Twitter posts and refused to guide for 2026 CAPEX. This is a very good development compared to the “maintenance CAPEX” talk during the Q2 call. When Intel announces Foundry wins, CAPEX will go up a lot and the street will love it. Will analsysts upgrade stock to ‘buy’ idk lol.
While I would like Intel to announce a foundry customer next week, this rollout is likely going to take the next few quarters. Intel’s 14A 0.5 PDK recently dropped, test chips are being made, and commitments will have to firm up by roughly Q2 2026. Of course, an announcement could come sooner, and if it does, it’s likely due to Lutnick’s chip tariff proposal.
Intel might announce some advanced packaging wins, as hinted at on the earnings call. Anything like that would be a pleasant surprise before the 2026 timeframe.
I recommend Lip-Bu get Trump to just announce the tariff proposal so most investors can begin to factor it in. The same is true of the $550B fund to expand chip manufacturing in the US. Lutnick is wrapping up negotiations in Korea, so the first deals should start rolling out.
Intel is likely to use the fund to help lower the CAPEX for Fab 27 in Ohio and hopefully announce three more shells next year (which would be for 2030+ HVM).
(Lutnick clip about $550B Fund)
The key question for Intel Foundry now finally getting external customers is just how good their PDK for 14 Angstrom is? This was (apparently) an issue in the past, it apparently showed that Intel's foundry was simply not used to dealing with external customers. In contrast, TSMC's PDKs have a great reputation.
Curious if intel’s crossover deigning their own products would deter customers or dtco? Anecdotally I have heard TSMC gets a trust advantage for being a pure play foundry, does intel suffer this problem vs tsmc as well as against more pure asic houses?