Why Gelsinger was wrong for Intel

By all accounts, Pat Gelsinger is affable, technically sharp, hard-working, and decent. Those who have worked for him praise him as a singularly good manager. In January 2021, when Gelsinger was abruptly named the CEO of Intel, this is more or less all I knew of him — and I found myself urgently needing to learn much more.

To understand my urgency, it’s worth rewinding the clock back to late 2019 and the earliest days of Oxide. We knew what we wanted to build, but we had some big (consequential!) decisions to make — first among them our choice of host CPU. Based on what we had seen in AMD Naples and Rome, we felt that AMD was clearly outpacing Intel, but this was a big bet and we wanted to be rigorous about it. Even though we felt Xeon was unlikely to be our direction, we scheduled a meeting with Intel in Santa Clara (pre-COVID!) to understand their roadmap.

We wanted to take advantage of the all-day meeting to understand the other components that Intel was making that might be relevant for us, so we also asked for roadmap briefings on NIC silicon (Columbiaville) and switching silicon (Tofino). The host CPU discussion ended up confirming our beliefs (befitting our writing-intensive culture at Oxide we wrote up our findings in RFD 12 Host CPU Evaluation), and the NIC discussion similarly was a dead end. The switching silicon discussion, however, was interesting: Tofino was TSMC-fabbed (the only Intel part at the time fabbed outside of Intel) and we found the programmable nature of it via P4 to be really compelling. Up until that point, we had assumed that we would have to use Broadcom, and we weren’t thrilled at the prospects; the idea of having not just an alternative, but one with such a novel approach was hugely appealing. The Tofino team at Intel was enthused by our early interest, proving itself to be very responsive in the weeks that followed as we continued to explore the prospects of building around their Tofino 2. The more we learned, the more we liked!

As we got deeper with Tofino, we did, however have a single, substantial reservation: we didn’t trust Intel to not kill it. Intel has a long, long (long!) track record of fostering innovation outside of its mainstay x86 product — and then killing it. Making things more complicated, these untimely deaths are not without ambiguity: the efforts that Intel kills are often early and interesting — but need patience and more iterations to be able to win a broader market. And while the cullings are not always wrong (no amount of patience would have saved the iAPX 432), the biggest mistakes in the last two decades at Intel (namely, its failures in mobile CPUs and discrete GPUs) are a result of discarding a flawed effort entirely rather than learning from it and iterating.

Still, we loved the idea of P4 and programmable switch silicon, and we really liked the Tofino team at Intel: we felt our teams had a shared vision for the programmable data center — and they were very accommodating of our desire to build something very different from a commodity switch. So despite our trepidations, we stepped over the desiccated carcass of Red Rock Canyon, and engaged on building a switch with Tofino. And we didn’t (and don’t!) regret it: it was of course brutally hard to build our own switch, but the dividends of programmability were enormous — and thanks to Tofino and P4, we were able to build some important, unique functionality.

But by the time Gelsinger became CEO in early 2021, our finished product was still a bit in the future: we knew what we were building, but we were in the thick of building it. It was important for me to understand the incoming CEO, and if possible, to get an audience with Gelsinger to explain what we saw in Tofino (and to understand what his vision was for programmable networking silicon). So — and as I relayed in our recent podcast episode on Intel after Gelsinger — I tried to learn as much as possible about the guy in his own words. This would normally mean binge listening podcasts, talks, interviews, etc. — but in Gelsinger’s case, I hit an early and familiar jackpot: as part of their superlative series, the Computer History Museum had a four-hour oral history with Pat Gelsinger.

And in listening to Gelsinger, his affability clearly came across, along with his technical acumen and depth of experience. But something else came across too: an indisputable undercurrent of arrogance. Now, it’s a little hard to be critical, given the context: the whole point of an oral history after all is for people to speak about themselves. Still, these oral histories are most interesting when they take the opportunity to expand on failures, and Gelsinger’s narrative of his time at Intel seemingly just went from strength to strength. Worse, in his own story, the bad things at Intel always seemed to happen when he wasn’t in the room or otherwise over his objections — and the good things always when he was called in to save an effort from failure. I’m sure this retelling has truth to it, but I found it surprising that his narrative about AMD Opteron is of Intel conquering AMD with Nehalem under his leadership — with no mention of Intel’s gross architectural missteps with respect to 64-bit that allowed (and even demanded!) AMD Opteron in the first place.

But, fine. Again, it’s an oral history; there’s going to be a Rashomon effect. And so the guy was confident; so what? My feeling of his arrogance was still a bit amorphous when we got to Gelsinger retelling the end of his time at Intel. I was waiting for this, because I knew that Gelsinger had led the charge on Larrabee — and that Larrabee was infamously a wreck. He lamented leaving Intel because he knew that Larrabee would be killed after he departed (which it was), but I was shocked when he said that had Intel "stayed with it…​ NVIDIA would be a fourth the size they are today." I was washing the dishes as I was listening to this, and it stopped me in my tracks: I turned off the faucet, dried my hands, and backed up the recording. Had I heard correctly?! If one wanted to make this eyewatering claim, it must be loaded with riders and caveats: it must acknowledge that Larrabee itself was an unusable disaster; that NVIDIA had an indisputable lead, even in 2009; that for Intel to dominate NVIDIA it would have required conjuring software expertise and focus with which Intel has famously struggled; that Intel had no pattern for sustained success out of x86. On the one hand, there were enough qualifiers from Gelsinger to soften this claim a little (and at least some passing respect for NVIDIA’s Jenson Huang and Bill Dally), but on the other…​ yes, he actually claimed that NVIDIA had his departure from Intel to thank for (checks transcript) three quarters of its size. (Which, even in 2019 when the conversation was recorded, amounted to a whopping ~$75B!)

I think even the interviewers were a little taken aback, and asked the question directly: "Do you think that it was in part that you weren’t there to drive [Larrabee] forward that they decided to withdraw from that?" Gelsinger’s response: "Yeah, I’ll say that very directly."

This was beyond mere confidence, and was looking more like a total disregard for one’s own limitations. Another feeling was growing inside me: despite his many positive attributes, Gelsinger was the wrong person for the job. And to be clear, the job in 2021 was to lead a company in crisis, viz. the letter from Third Point that precipitated the firing of Bob Swan. Intel’s strategic mistakes were (in my opinion) symptomatic of an acute cultural problem: the company still carried with it the inherited arrogance from an earlier age. A concrete manifestation of the company’s arrogance is that it didn’t listen: it didn’t listen to its own people (and therefore struggled to correct course even when the rank-and-file knew that the trajectory is wrong) and it didn’t listen to its customers (and therefore built the wrong things for new markets — or missed out on those markets entirely). Intel needed a leader that could confront this cultural problem directly — who could work to undo an accretion of generations of entitlement — but if Gelsinger’s narrative for himself was any indicator, it felt like he would instead be feeding the company’s worst impulses about its own exceptionalism. As Gelsinger returned, I found no reassurance in the company’s narrative, which felt less like the arrival of Winston Wolf, and more like a cherished prince returning home to live out his destiny.

The concern that Gelsinger would be reliving his past more than navigating Intel’s future wasn’t exactly put to rest when, shortly after his arrival, Intel launched its cringy-as-hell anti-Apple "Go PC" ad campaign. (It tells you everything you need to know that Intel deleted the Go PC ads, but then forgot about Canada.) Fortunately for future historians of corporate dreck, the cringiest of these ads was immortalized when Marques Brownlee and David Imel absolutely ripped it apart. "Go PC" was an embodiment of the arrogance that I feared came from the top; how could anyone think that Intel’s biggest problem in 2021 was competing against…​ the Mac?! They knew that AMD also made x86 parts, right?! The whole campaign frankly felt juvenile, as if they were trying to just deride Apple for their decision to build their own silicon.

Much meatier than an ill-advised ad campaign was Gelsinger’s announcement that they would have 5 process nodes in 4 years. This was — and is — an aggressive bet for a company that had an entire process node (10nm) slip for years and then quietly fail to yield but a single product (Cannon Lake). From the perspective of an Intel customer, Intel had never come completely clean about the failing of 10nm (publicly or privately); how could Intel expect outsiders to trust them to learn from their mistakes if it wouldn’t even publicly acknowledge what those mistakes were? I don’t think that that bet was necessarily the wrong bet, but it was indisputably going to be expensive and risky. And this gets to Gelsinger’s first real, unequivocal mistake: he didn’t eliminate Intel’s dividend. Eliminating the dividend would have sent an early message to everyone — shareholders and employees — that the company’s survival was at stake, and it would have helped build the war chest for the battles to come. While I think his dichotomy is a bit reductive, Ben Horowitz’s nomenclature is useful here: Intel needed a wartime CEO — but in maintaining a dividend that it couldn’t afford, Gelsinger was committing the ultimate peacetime act. (I am not alone in this view; Ben Thompson of Stratechery dropped a must-listen episode on Intel in which he cites the paying out of billions of dollars of dividends over Gelsinger’s tenure as an early and serious blunder.)

Beyond being merely expensive and risky, the newly-christened Intel Foundry Services (IFS) had another, deeper problem: the culture at Intel didn’t really seem amenable to the level of customer engagement that foundry customers (rightfully) expect. And this was compounded by the trust issue that we at Oxide felt with Intel over Tofino: why would a customer trust their future to something chancy like IFS? Absent the availability of 18A, the only answer that Intel ever seemed to give on this was to wrap itself in the flag: that one should choose IFS solely because the fabs are on US soil. In this regard, it reminds me (ironically?) of a long-ago CPU startup, Transmeta. They had interesting technology, but unable to compete with Intel on absolute performance, they retreated to power efficiency. Transmeta was implicitly relying on the fact that Intel wouldn’t release a lower power part, but Intel promptly turned around and introduced the Pentium-M line — and that was more or less it for Transmeta. Similarly, is Intel assuming that TSMC is unable to build fabs in the US or that the US government wouldn’t see such fabs as addressing geopolitical concerns? If so, it feels like a bad assumption on both fronts.

Skepticism of Gelsinger’s plan for Intel aside, we at Oxide anxiously watched Tofino. At Intel, the team itself believed it was safe under Gelsinger, and things did indeed seem okay for a while. Fast-forward two years to 2023, and we got an urgent request for a call from the executive leading the Tofino effort. Fearing the worst, we were honestly somewhat relieved to learn that Tofino hadn’t been killed outright — but all future development of the part had been cancelled. We were on the cusp of shipping our first product, and the switch was working, so our own die was cast. And we were reassured by the Tofino team, who (to their great credit!) assured us that they would support us, even going to far as to draft a formal letter from Intel that we could share with our own customers expressing Intel’s ongoing support of Tofino.

While the situation for Tofino was obviously dire, I also felt we had some stoppage time; could we get in front of Gelsinger to explain why killing switching silicon in 2023 feels a lot like walking away from the GPGPU in 2011? No, as it turns out, we couldn’t. (And nor, apparently, could anyone else who was buying the part.) Everyone we dealt with at Intel agreed that the move was short-sighted, but as with so many things Intel (that culture problem again!), they felt powerless to change it.

Tofino was clearly living on borrowed time, and we were disappointed (though frankly not that surprised) when we were notified earlier this year that Tofino was being formally killed. It’s a credit to the Tofino team (or what was left of it at that point) that they continued to be very direct with us; this was (clearly) a decision that they disagreed with, and they were especially apologetic for the sloppy manner in which the end-of-life was being handled (which made a mockery of Intel’s own process for end-of-life management).

In the end, for all of the decisions that we made at Oxide — out of all of the companies and parts that we bet on, out of all the partners that we had sent RFD 68 Partnership as Shared Values to — only one had walked away from us, and it was the largest and best capitalized partner, who had repeatedly told us that they would not do exactly what they in fact did. How can Intel ever expected to be trusted when they treat partners this way?

For Oxide, fortunately, Intel might have done us a favor in the limit. While we were grateful to Tofino for allowing us to prove out our ideas on programmable networking, we had issues with it too. Specifically, Tofino’s software and instruction set architecture remained needlessly proprietary. We believe that switching silicon is awaiting its "x86 moment", when open source software can be implemented for a well-defined ISA — and we were never going to get there with Tofino. As we looked for our post-Tofino future, we were delighted to find the Xsight Labs team and their X2 ASIC. We are going to have a lot more to say about this part, but suffice it to say that we see in X2 both the strengths that we saw in Tofino and also the potential to be the open substrate for programmable networking writ large. Stay tuned!

I have great reverence for Intel and its extraordinary history, and I would never count them out (the resurrection of a clinically-dead AMD shows what is possible!), but I also won’t be integrating with any of their technology until their acute cultural issues are addressed. With regard to these cultural issues (and his other strengths aside), Pat Gelsinger was indisputably wrong for Intel.