Intel needs a whole lot of big wins for both its chip design and selling business and its increasingly arm’s length foundry business if it is going to turn itself around and be competitive with AMD and Nvidia on the left hand and with Taiwan Semiconductor Manufacturing Co on its right.
So far, what Intel is able to demonstrate is a bunch of small wins, which are steps in the right direction but about all both the company’s top brass and Wall Street can expect based on how far behind Intel fell in chip manufacturing techniques and therefore with its server and client CPUs, GPUs, and other kinds of compute engines.
We think that Pat Gelsinger, Intel’s prodigal son from EMC and VMware who has come home to save the company, has been doing a some pretty effective fast talking to the hyperscalers and cloud builders as well as to interested parties in the political sphere. And Gelsinger can be thankful that he didn’t get the chief executive officer job back in 2009 because the mess he has been cleaning up since he returned to the company in February 2021 is largely not his own, but rather those of former CEOs Paul Otellini, who could have been replaced by Renee James (now CEO of archrival Arm CPU provider Ampere Computing) but was not. Intel instead opted for Brian Krzanich in 2013, and the company was making money hand over until 2019 and it let its whole chip making business go to hell because of it, getting stuck in the La Brea 14 nanometer process pits in the process.
Roadmap after roadmap had to be torn up and redrawn, leaving the market wide open for AMD to take the pole position in CPUs and the Arm collective to coalesce – thanks in large measure to the desires of the hyperscalers and cloud builders to be rid of the X86 pricing tyranny Intel exhibited from 2010 through 2019 and to be in control of their own CPU architectures. A distracted but rich Intel fell far short with its many-core “Knights” family of accelerators and its “Ponte Vecchio” and “Rialto Bridge” GPUs, the former not really seeing the AI wave and the latter using a chiplet design and associated packaging that proved difficult to yield.
And to top it all off, Intel bought Nervana Systems, and then bought Habana Labs to replace them with the Gaudi line, and now Intel can’t make enough of the Gaudi matrix engines to create enough revenue to matter – $500 million in Gaudi 3 accelerator sales are expected this year, which is 32,000 devices, against more than $100 billion in datacenter sales for Nvidia in calendar 2024 and $5 billion in GPU sales and maybe $7 billion in CPU sales for AMD this year.
Don’t get us started on the past decade and a half of Intel’s attempts to build a datacenter networking business. . . . This obviously has not worked out, much to the detriment to the company. Networking represented 16.8 percent of Nvidia’s datacenter revenues in the trailing ten quarters the represent the GenAI boom. The share is coming down as Ethernet is gaining steam for AI clusters.
We believe in redemption, as we said as much in early August when going over Intel’s shocking financial results in The Resurrection Of Intel Will Take More Than Three Days. And the partnership just announced with Amazon Web Services is a step in the right direction. But we do not think it will lead to very much in the way of profits, even if it does result in billions of dollars in revenues. We are not sure how much of that purported and unspecified amount of revenue talked about in the deal this week is actually incremental.
Much is being made of the deal with AWS, so let’s break this down a bit.
Intel said in its statement that it is producing a custom Xeon 6 processor based on the Intel 3 process for the world’s largest cloud builder. Intel 3 is a refined version of the Intel 4 process, which is based on 7 nanometer geometries with some tweaks that make it more like a 5 nanometer process. The “Sierra Forest” Xeon 6700 and 6900 series announced back in June, the first of the Xeon 6 devices to come to market, use the Intel 3 process. And so will the impending “Granite Rapids” Xeon 6 chips, which have a smaller number of beefier X86 cores than the Sierra Rapids. We do not know if AWS is getting custom Sierra Forest or Granite Rapids chips – or both. We think the latter, and very likely with some modest tweaks to boost AI serving and database serving performance.
The nature of the customization is not clear, but we strongly suspect that it has more to do with the price of those X86 server CPUs than it does for any specific technical features. Yes, Intel has made custom processors for the big clouds for more than a decade, we know. Sometimes unneeded features are deactivated, sometimes the clock speeds are capped or boosted – the clouds were getting all-core boosted CPUs for years, and before everyone else could get their hands on these and also get more deterministic performance for their workloads. And, because the hyperscalers and cloud builders represented 30 percent of X86 CPU shipments worldwide and now represent in excess of 50 percent, they not only got better processors, but they got much better prices.
We do not think this is a change in the behavior of AWS at all, which has hewed closer to the Intel line than either Microsoft or Google ever did with their clouds. Google is well known for its enthusiasm for AMD server CPUs, and Microsoft has learned the same. All three are much more enamored of their own Arm server CPUs these days – which is without a doubt due in part to Intel not keeping its Xeon chips on a proper process roadmap, thus making them less competitive by necessity.
It is interesting that the AWS announcement did not say that Intel was being chosen as the foundry for future Trainium and Inferentia AI compute engines, but rather that it was “co-developing next-generation AI fabric chips on Intel 18A.”
Presumably, what Intel and AWS are referring to is breaking the Inferentia and Trainium processors into chiplets and having Intel etch the NeuronLink interconnect chips that were formerly on the Inferentia and Trainium monolithic dies free so they can be etched separately and, usually, on less expensive processes. A lot of the general business press and some of the IT trade press read what Intel and AWS said and mistakenly, we think, reported that Intel got the gig to co-design and etch AWS Trainium or Inferentia chips. If Intel is etching the whole Trainium and/or Inferentia chips, it did not describe this properly – and it did not give interviews on these announcements for further clarification.
Anyway, this AI fabric chip will be implemented in Intel’s 18A process, the first process to go into wide production using its RibbonFET transistor design and more or less at parity to TSMC’s 2N processes. (The 20A process that was dropped for server chips was more like TSMC’s 3N process.)
Intel also said the following in its announcement: “Intel and AWS intend to explore the potential for further designs to be produced by Intel based on Intel 18A and future process nodes including Intel 18AP and Intel 14A, which is expected to be produced in Intel’s Ohio facilities, as well as the migration of existing Intel designs to these platforms.”
So as we read this, there is a chance for Intel to get more chip fabbing deals with AWS at the 14A (or 1.4 nanometer) node, which is being brought up in the Intel Foundry operations in New Albany, Ohio.
As part of the announcements, Gelsinger also announced that Intel Foundry was not just having its financials broken out separately now, as happened a few quarters ago, but would be broken out as a independent subsidiary of Intel, much as the Altera FPGA business has already been. By being a subsidiary, Intel Foundry will have its own profit and loss statements and its own balance sheet, and it will be able to seek more independent investment from outside parties. (Hey, private equity firms can invest in NFL football teams now, so why not chip factories?)
Nothing was stopping Intel from seeking more equity investment to build out its foundry business, and it already has before making it a subsidiary. The US government gave Intel $8.5 billion in cash and incentives as part of the CHIPS Act. Apollo Global Management invested $11 billion in Intel’s Ireland fabs and Brookfield Asset Management invested $15 billion in Intel’s Arizona fabs, to get 49 percent stakes in new facilities being built there.
Making Intel Foundry a separate subsidiary might place stronger firewalls between the foundry and the rest of Intel, which very likely means seeding it with cash as well as keeping stuff confidential that foundry customers are doing. We think potential foundry customers want both, particularly with Intel Foundry hemorrhaging $8 billion in operating losses in the trailing twelve months and no reason to believe it will not continue.
Intel Foundry will be helped by the $3 billion in direct funding to make chips for the military that was announced as part of the Secure Enclave program within the CHIPS Act. This funding includes dough for the Rapid Assured Microelectronics Prototypes – Commercial (RAMP-C) and State-of-the-Art Heterogeneous Integration Prototype (SHIP) programs, and it is meant to ensure that chips of military significance are made indigenously. This $3 billion was not a surprise, given that Intel is the only indigenous chip maker. (IBM Microelectronics used to build a lot of chips for the US military, but it sold out to GlobalFoundries a decade ago; it still operates foundries in New York and Vermont, and they may also still make such chips for the Department of Defense.)
Intel is looking to shed 15 percent of its 116,500-strong work force, to cut its costs by $10 billion, and cutting its dividend payments to shareholders rocked its stock in early August, but the news of some of its actions happening helped the stock a little this week. Intel is about halfway towards it goal of shedding 15,000 workers by the end of the year. Intel is looking to consolidate its office space and will exit two-third of its office space by the end of 2025. (This sounds like a difficult thing to do.)
Gelsinger said in a memo to Intel employees that the Ireland fab will be its main hub in Europe and that it was pausing its manufacturing projects in Poland and Germany for two years, and that chip design and packaging operations in Malaysia would continue, but other manufacturing operations would halt.
“There are no changes to our other manufacturing locations,” Gelsinger said. “We remain committed to our U.S. manufacturing investments and are moving forward with our projects in Arizona, Oregon, New Mexico, and Ohio. We remain well-positioned to scale up production around the world based on market demand as we grow our foundry business.”
Ah, but will that market demand materialize?
There was some other re-org housekeeping news items. The edge and automotive business will be moved into the Client Computing Group, which makes chips for PCs and, in theory, other clients. The Network and Edge business will be focusing on “networking and telco” and presumably should just be called the Networking Group. Integrated photonics, which had presumably been part of NEX, is being moved into the Data Center & AI group and Intel is working on a roadmap to better serve the datacenter. Whatever that means. (Paring it down, more like.)
In the end, the hyperscalers and cloud builders were Intel’s biggest customers, but the malleability of the Arm architecture and the volume economics of the situation are going to make it very hard to Intel to win revenues at a reasonable profitability for its Xeons, and its process roadmaps are going to have to be successful for many years before they will trust Intel Foundry to be their etcher and packager. Intel’s foundry costs are so much higher than TSMC’s for any given process that it is hard to believe anyone would do it except for patriotic reasons – even Intel has to go to TSMC right now to get the best etching.