This day always comes. It is the nature of monopoly and hubris.
It came for IBM. It came for Microsoft, and it is coming for Facebook. It will come for Google and, even though it is hard to believe, it will come for Amazon. And it is most assuredly coming for Intel right now, and it is probably going to get worse. Maybe not as bad as IBM in the early 1990s – because nothing was worse than that self-inflicted and self-described “near-death experience,” where Big Blue had the biggest write-offs in corporate history and eventually had to lay off half of its 400,000 workforce – but most assuredly for Intel, bad nonetheless.
Intel’s problems are not intractable, but they are incredibly expensive to fix, which is why the company is panhandling for money in the halls of the US Congress and in the corridors of power in Brussels, and why it is reportedly considering hooking up with private equity and hedge fund firms to raise the money to revitalize its foundries and therefore, it is hoped, its central compute engine business that, quite frankly, Intel’s former top brass took for granted. But it is worse than that. The chip design part of Intel took the foundry business as a given and the foundry part of Intel took the chip design business as a given, and AMD, the Arm collective (including homegrown Arm chips from cloud builders), and Nvidia have been doing all of the real taking – mostly of market share in compute. Given the roadmaps Intel has laid out and the substantial competition Intel is facing, they are going to keep taking it out of Intel’s hide as they ramp up new chippery and spend money to buy more manufacturing capacity from Taiwan Semiconductor Manufacturing Co.
And so, Intel finds itself in a real fight for its life, luckily with the best person possible, former chief technology officer and Data Center Group general manager and for the past year and a half, chief executive officer Pat Gelsinger, firmly in charge and knowing what must be done. The question is: Can Gelsinger afford to make the fixes that need too get done?
Not if you look at Intel’s current balance sheet and the prospects it has for a continued weak – and possibly progressively weakening – financial condition. Hence the panhandling and fast talking. And when this is all done, Intel might be saved, but there is no way in hell that it can be the same Intel it was before. It will be gun-shy, like IBM has been for decades, and it will be beholden, which is like having your guns taken away. At some point, if Intel’s stock gets hammered enough, it might even be tempted to pull a Michael Dell maneuver and take Intel private so it can try to heal without so many eyes watching the process. But we think that no private equity firm will believe that Intel’s historical profit margins will ever return.
Let’s get this straight: They won’t. Period.
Intel can get its foundry in order and make a reasonable profit being a second source for all kinds of vendors waving the American flag. It can invest in advanced processes and eventually be a rival to TSMC and keep pace with Samsung. It can architect complex compute engines with chiplets stacked in all dimensions and interconnected inside and out with gussied up PCI-Express links. But unless and until AMD, Nvidia, and the Arm collective screw up and unless TSMC screws up, Intel can never recreate the conditions that allowed it to have hegemony in the PC and in the datacenter ever again. This is like the United States after Europe and Asia were destroyed after World War II, or Rome after the fall of Carthage. These are once in a lifetime opportunities for monopoly, and they just don’t come around again. The Holy Roman Empire was a pale shadow of the glory that was Rome. But still effective, mind you, given the times many centuries after the fall of Rome.
In the quarter ended in June, which we expect to be a very good one for AMD and Nvidia, Intel did not do so good. Revenues were down just a smidgen under 22 percent to $15.32 billion, gross margins were off 50.1 percent to $5.58 billion, and even with a $455 million benefit from taxes that Intel was keeping in its back pocket for a rainy day it posted a $454 million net loss. This is real red ink, and if it didn’t have that benefit and it had to pay some taxes, it could have easily been over $1 billion in losses. The company ended the quarter with $4.39 billion in cash and $22.65 billion in short-term investments, so it has some financial maneuvering room. But as the OEM and ODM channels cut back on inventories due to the slowdown in the PC market and, we think, due to share gains by AMD and the Arm collective in CPUs, Intel’s accounts receivables are down by 35.9 percent to $6.06 billion. (The fact that we are even looking at this metric is telling.) The accounts payable is $7.95 billion, up $38.3 percent, in the June quarter. So Intel looks like it is stretching its revenue stream out a bit, which is to be expected given the massive investments it is making and its sharply curtailed business in the quarter.
In past times, the PC business sometimes masked issues in the datacenter business, and vice versa, but not this time around. Both PCs and servers have taken a dive at the same time.
The Client Computing group had $7.67 billion in sales, down 25.2 percent as PC spending has slowed dramatically now that we have all apparently done our COVID-19 desktop and notebook upgrades. And whatever sales Intel has here, it has to fight for as well as cover the expensive costs of its Intel 7 (SuperFIN 10 nanometer) process ramp, because Client Computing posted an operating profit of only $1.09 billion, down a staggering 73.1 percent.
The Datacenter and AI group, which is not precisely the same thing as the old Data Center Group we have been tracking for a decade because of the recent financial reorganization that was done in preparation for this terrible day, fell by 16.2 percent to $4.65 billion. And if Intel had left the networking adjacencies in the old Data Center Group and reported those, it would not have had as bad of a quarter as it looks. But it wanted a better story to tell for a fast-growing Network and Edge group, which had sales of $2.33 billion, up 10.83 percent.
Because of the intense pressure coming from AMD and the Arm collective and the high costs of ramping the Intel 7 and Intel 4 (5 nanometer) processes at Intel Foundry Servces, Datacenter and AI group stomached a huge hit to the middle line, with operating profits falling 89.8 percent to $214 million. That is a long, long way from the 45 percent to 50 percent operating income that Intel typically had in the prior decade with Data Center Group. Like an order of magnitude smaller as a percent of revenue. And while Network and Edge (which is called NEX for some strange reason) had a revenue boost, operating income fell by 60.2 percent to $241 million – for the same reasons. Intel is investing heavily in future chip designs and Network and Edge group has to shoulder its share of the IFS burdens as well as compete against staunch rivals who want their pieces of datacenter and edge networking and compute.
“Over the next couple of years as we rebuild our server product portfolio, we expect to grow slower than the overall data center market,” Gelsinger conceded in the conference call with Wall Street analysts going over the financial results “It’s not a fact we like but the forecast we see. We have a singular focus to regain performance and TCO leadership across all workloads and use cases from enterprise to cloud. The advantage of our incumbency position remains underappreciated and provides significant opportunity to drive outsized advantages to our customers.”
Hmmm. Intel can talk all it wants about the custom chip deals it has with Amazon Web Services and Meta Platforms (Facebook) to do custom compute engines, but do you really think these deals have any margin? We don’t. Ditto for the use of “Sapphire Rapids” Xeon SPs inside of the future Nvidia DGX-H100 systems using that company’s “Hopper” GPUs. So what? We hear Intel cut price like crazy, and if Sapphire Rapids is delayed to early next year, as we keep hearing, then you will see just how fast Nvidia can global replace that CPU. But, that said, Nvidia is in kind of a bind there. It has its own “Grace” CPUs coming early next year, and it doesn’t want to give any money to GPU rival AMD for its “Genoa” Epyc 7004s, either. Ampere Computing is a possibility, but it will be rivaling Grace, too. Intel is a no-show so far in datacenter GPUs, and eager for a design win – any win – so Nvidia no doubt got a great deal on those Xeon SPs. If it can get them, that is. Nvidia should just get Grace out the door and be done with it.
For all the talk about Intel Foundry Services and it Accelerated Computing Systems and Graphics groups, these are piddling businesses. IFS saw its revenue shrink by 53.8 percent to $122 million, with a swing from a $52 million gain a year ago to a loss of $155 million this time around. The graphics business (which is called AXG for another strange reason) brought in $186 million, up 5.1 percent, but losses expanded from $168 million in the year-ago quarter to $507 million in this June quarter. (Oooof!) And Gelsinger reiterated that AXG was on track to generate more than $1 billion in sales this year and that it would deliver more than 10,000 nodes – that would be 60,000 GPUs – to Argonne National Laboratory for the “Aurora” supercomputer before 2022 closes out. That will also be over 20,000 Sapphire Rapids CPUs, and our guess is that if Sapphire Rapids is delayed, it will be the commercial launch not these nodes. Aurora is a big chunk of that AXG revenue, but so is Alchemist, Arctic Sound M, and Blockscale GPU products alongside the “Ponte Vecchio” Xe HPC GPU accelerators in the Aurora system.
Given all of this, it is no wonder at all to us that Intel is going to shutter the Optane persistent memory business to focus on core compute and foundry. Optane is slower than expected, skinnier than expected, more expensive than expected, and harder to integrate into systems and applications than expected – and Intel held on to it too tightly for it to emerge as something that could go into all systems. This is the kind of maneuver IBM would pull in the old days, and Intel got the same result in the end. The truth is, we need something like Optane, as originally pitched, in the memory hierarchy, something that can be used like DRAM but which is not as expensive and which can be evolved at a quicker pace. To our way of thinking, system architects are going to start pooling DRAM with CXL to drive up memory usage, and this will be better and easier than integrating two different kinds of memory into systems as Intel was trying to do with DDR4 and 3D XPoint. Both approaches will save some money.
As things now stand, Intel has lowered its guidance for sales by $8 billion to $11 billion, and now expects for annual revenue to be in the region of $65 billion to $68 billion. That could prove to be optimistic, but even if Intel hits that top line, everyone will be watching that bottom line get thinner and thinner.
The Next – And More Profitable – 10 Percent Of Server Share For AMD
When this is all said and done, Intel will deserve some kind of award for keeping its 14 nanometer processes moving along enough as it gets its 10 nanometer and 7 nanometer processes knocked together to still, somehow, manage to retain dominant market share in the server space. Or, maybe …
AMD Rounds Out “Aldebaran” GPU Lineup With Instinct MI210
When the “Aldebaran” datacenter GPUs were launched by AMD last November for the HPC and AI crowd pushing up into the exascale stratosphere, only the two top-end models of the Instinct GPU accelerators –ones that use the Open Accelerator Module (OAM) form factor put forth by Facebook and Microsoft under …
Cerebras Shows Off Scale Up AI Performance For Big Pharma And Big Oil
Historically, the largest organizations in the world – the Global 2000 plus the biggest national government and academic research institutions – have always had the most complex data processing needs. And by virtue of the size of their IT budgets and the sophistication and scale of their workloads, they get …
This all will look a bit different if China invades Taiwan.
My contention all along is that boosting American and European fab independence actually makes China invade even faster than it might otherwise given the national security risks the US and the EU would not otherwise tolerate.
We are creating the condition we fear.
Extremely well done Tim… I saw the apologist for INTEL claiming all they would need is for Qualcomm and others to use INTEL FAB. No fabless company is going to share their proprietary designs with INTEL. PERIOD. Besides Intel already failed at 10nm on their own. They are just fishing for Gov money at this point.
Secondly US FAB makes every chip produced by these Fabricators, if they ever build a single chip, subject to dumb politician’s sanctions and protectionism. So Asia will never be using US FAB and ASIA is the biggest customer in the world. This is an enormous dead end.
Once the US followed Trump into the protectionist abyss followed by UNION JOE BIDEN, that was the end of INTEL and US Semiconductors. Qualcomm a truly great company has fought the US Government at every turn but it is a losing battle. Even Samsung started to build their own chips for fear of Snapdragon being rejected by China. Qualcomm is a great company being systematically tormented by right and left-wing politicians.
China doesn’t need to take TAIWAN. 46% of the population identifies as Mainlanders. China just needs to take TSMC. If China blew up TSMC the US would have no further use for TAIWAN. Taiwan’s largest customer is Mainland China. China can take Taiwan diplomatically once TSMC is removed. Whether China removes the EUV lithography machines first is irrelevant. They really only have to remove certain parts of the machines.
At 99 years of age, Henry Kissinger has told the US to stop antagonizing China by parochial unschooled sentiments of local US populist politicians. China is not going to disappear. This nancy Kerrigan kneecapping US foreign policy will backfire. Protectionism has always failed. The US is not big enough to sustain its monopoly of the US Dollar Reserve. Like all monopolies, they all fold.
One more point… A US Military General talking head said that if Taiwan is invaded by the Mainland,, they should destroy TSCM. Why on earth would Taiwan do that? If Taiwan were smart they would Unify with the Mainland and secure their local gov control, similar to Macau but their foreign military alliance would be over. China doesn’t need to take Taiwan, they only need to destroy TSMC.
It’s not the American and European fab independence that is pushing Chinas’s buttons, but the restrictions the US is putting up in tooling / technology. Those are the real threat to China
Intel’s bottom line has also likely been affected by trade restrictions with China (that were felt necessary for many reasons) which saw their HPC efforts move from the #1 Tianhe-2A (Xeon) to the #1 TaihuLight (homegrown Sunway). But on the own-goal side, they probably got too distracted with swatting the ARM collecitve fly in mobile and IOT, with such discontinued efforts as Intel Quark and Galileo. AMD had a more focused Zen attitude there it seems. I think that Aurora could be the proof in Intel’s pudding, with the hope of a working system delivered on time, at better than 20MJ/Exaflop efficiency … but your detailed analysis does raise concerns about the likelihood for this happening.
“Intel’s bottom line has also likely been affected by trade restrictions with China (that were felt necessary for many reasons)”
I can name a dozen examples of Political paranoia that cost Intel business. Obama blocked the sale of Intel Chips destined for a Chinese Supercomputer. No secrets. Obama blocked the sale. 15 months later China had their own fast RISC chips and built the fastest supercomputer of its time.
Over and over again the blocking of INTEL and US politician sanction have made INTEL products unappealing due to the potential for sanctions and arbitrary and capricious political protectionism. And what are you protecting? Bananas in a freight train in Needles California.
Tech products have a very short shelf life. So if you block a stream, you destroy the product. Intel could not secure any handheld business for fear their products would be sanctioned by brainless politicians. So the product sits on the shelf, is not used in products built in Asia, No revenues and no profits and no R&D. Intel didn’t raise a fuss or legal arguments. So tough.
I know you think you have the answers but let’s simplify the equation. China is the largest biggest buyer of semiconductors in the world and the largest producer of products that use semiconductors in the world. US firms are Chinese Customers. How does blocking US firms from selling products with short shelf lives to China benefit US companies or the US GOV?
“That were felt Necessary”? I am glad you used the word FELT and did not use the word THOUGHT. For if anyone had thought this through, they would never have embarked on protectionism for a train car load of bananas in Needles California destined to rot.
I tend to agree with the article that hubris trumped politics in this situation, especially since AMD (under the same policies) did very well. Looking at the top500, Xeon (and POWER9) with NVIDIA A100/V100 run at 60 MJ/ExaFlop (12-HPC5, 19-ABCI 2.0, 4-Summit, 21-Marconi-100). Switching to EPYC+A100 improves that to 40 MJ/EF (8-Selene, 11-Juwels), and moving to EPYC+MI250x yields 20 MJ/EF (1-Frontier). Meanwhile, 6-Sunway runs at 150 MJ/EF and 9-Tianhe-2A is at 300 MJ/EF. My hope is that upcoming Hoppers, Graces, Xeons, POWER10/11s, and A64FXs keep improving on that performance.
Intel was hit by three short term crises simultaneously.
The China shutdown hampered PC availability. Apple’s 2Q Mac revenue was down $1 billion YoY for the same reason.
The war, inflation and recession pressures lowered PC, data center and cloud demand even further.
Finally the backlog at TSMC delayed their GPUs. Had Gelsinger chosen Samsung’s superior 4nm and 3nm available capacity over TSMC 6nm’s superior yields, their GPUs would have launched on time and in much larger volumes. That alone would have made their 2Q profitable. Intel only expects to ship 4 million GPUs this year, but in 2023 they should at least get the $9 billion in revenue from AXG that AMD’s equivalent had in 2021.
Intel Foundry Services isn’t generating revenue because their Intel 4, 20A and 18A nodes aren’t online yet. They have Qualcomm, Nvidia and MediaTek lined up for when they are.
It is true that Intel will never again enjoy their monopoly but what does that matter? In 2007 Microsoft had 97% PC market share, a similar monopoly for their Office suite, PCs were largely the only end user computing devices and the XBox 360 was at parity with the PlayStation 3. Look at what Apple, Google, Samsung and Linux have done to the Microsoft hegemony since. Microsoft responded by diversifying into cloud and mobile. AXG and IFS will do the same for Intel.
Even the claims that Intel will not be able to generate enough revenue to pay for their foundry investments is overstated, as this thinking requires IFS being unable to draw business from TSMC. This ignores that Intel does not need to surpass or match TSMC’s cutting edge nodes to do this. Instead, IFS merely needs to get Qualcomm, Nvidia and MediaTek to use Intel 4, Intel 20A and Intel 18A instead of TSMC 7, TSMC 6, TSMC 5 and TSMC 4 for their midrange CPUs and other lesser expensive offerings.
Intel had $80 Billion in stock buybacks. That is money not going into products but into hedge fund traders’ pockets. Intel’s downslope began long before the COVID EXCUSES, or Apple. Apple makes their own chips now having said bye-bye to Intel.
Let me also clarify that packaging is now king, not narrow gates. Narrow gate semis are only 5% o the marketplace. Intel already failed at 10nm scale but you have them leading the way.
Further, China is not moving toward narrow gated just yet. By improving the backbone of 5G with an eye to 6G, with supercomputers doing the heavy computational lifting, a dumb terminal can do as much as the $2000 smartphone. So the idea of a smartphone with massive computational power may be entirely irrelevant. It’s not like you can connect a phone easily to a keyboard and large monitor and do any real work with it. Microsoft’s phone already failed.
There are diminishing profit returns with narrow gated fab, for what? Sccalor size alone is meaningless. A data centre can use faster chips or more chips and more attached math coprocessors. Anyone with a PC made in the last 10 years knows that if you use a good GPU, performance soars even on the old machine. I think this easily proves the virtue of packaging. Want a faster desktop computer… add a GPU or two. Alibaba’s new server chip combines DRAM and LOGIC, a feat thought impossible.
Intel has a populist born-again Carnival Barker running the show. He failed at the 486 chip. Most of his background is software NOT HARDWARE. Intel is at the breakwater of its demise. It reminds me of UNION CARBIDE’s born-again CEO that talked the board into going into the diaper business. DRYDEES was the trademark. Hope springs eternal I suppose. FAB is one of the most labour-intensive businesses in the world. Intel is looking to Ohio, a move incomprehensible. Ohio is a UNION State, not a right-to-work state. There is a reason OHIO is a Rust Belt State.
Just to comment on the gpu thing: the issue is in intel’s hardware/software, not with TSMCs late delivery. They’ve been ready since q1, but intel hasn’t done a single serious tech demo to generate hype, because they can’t get the things to work reliably, either because of an unexpected hw issue or because they’re simply not good enough to write game drivers (more likely the former, or both). So now they have $500m worth of inventory they don’t know they’ll be able to sell.
” Intel can talk all it wants about the custom chip deals it has with Amazon Web Services and Meta Platforms (Facebook) to do custom compute engines, but do you really think these deals have any margin? We don’t. ”
Why do you think this not matter? Intel was doing packaging service only for AWS graviton. But they have now expand the partnership to co-develop multi-generational data center solutions optimized for AWS infrastructure. This is clearly a win for IFS and its initial mission when Intel announced that they would license x86 to other fabless players. In the earning call, they have said the benefits to the fabless companies that custom Xeon solutions will bring greater levels of differentiation and a durable TCO advantage to AWS and their customers. When AWS get to launch custom silicon, they can offer it at a cheaper price with same margin, or same price with higher margin. Yet AWS’s custom x86 will differentiate itself from other x86 chips that’s available in the marketplace. Essentially Intel will use IFS to provide customizable x86 chips to take revenue from Intel’s DCAI and revenue from AMD’s data center. Can AMD do that? yes they can, but they will offer it at higher price because AMD doesn’t have cost advantage. Can TSMC do that? no they can’t because they don’t have x86 license, and if they do, they won’t provide x86 IP to their customers because TSMC only do manufacturing.
When AWS offer the x86 silicon to its customers at lower price and higher capability, it’s going to take more share from Google cloud and Azure. Small cloud players will not be able to compete with AWS effectively. But Google and Azure still stand a chance, and their chance is to copy what AWS has done, co-develop x86 with Intel which in turns hurts DCAI of Intel and AMD. But Intel will be able to recoup the loss through IFS and x86 IP licensing.
So going forward, DCAI revenues will rely more and more on the on-premise data center, private cloud, and small cloud players. DCAI will be able to offer more through their SaaS products which is an interesting and very profitable path under Pat Geisinger. At the same time, IFS will gain its customers through following two ways:
1. Provide a geographic-balanced supply chain network that span from US to Europe (Note these are more stable location than TSMC and Samsung, with TSMC being too close to China and Taiwan is short of electrical supplies, and Samsung being too close to North Korea and its risky bet into GAA and its toxic culture)
2. Provide x86 IP licensing through IFS. This will enable customers build their differentiated product while utilizing some of technology used in Xeon processor and other Intel IP
I would not count Intel out just because of one or more quarters of earning miss because Pat has laid out the roadmap. The more problems they find they go along, the better the outcome. People originally thought only process technology has issue under prior management. But now it turns out design team has serious problem as well with things happening on Sapphire Rapids and Alchemist. So Pat and his team need to do more, I agree, but they are also on a path to fix multi-years of mistake and transform the toxic culture under prior three CEOs
if they retained me as a consultant i could turn intel around and put them back on top.. but i’d start it by killing x86
Hacked cell transmitters can become microwave weapons. Guess what is causing Havana Syndrome? Civilians are also targeted everyday. Consider:
“It’s not technically hard to make a device that complies with the FCC that listens to nonpublic bands but then is quietly waiting for some activation trigger to listen to other bands,” said Eduardo Rojas, who leads the radio spectrum lab at Embry-Riddle Aeronautical University in Florida. “Technically, it’s feasible.”
To prove a device had clandestine capabilities, Rojas said, would require technical experts to strip down a device “to the semi-conductor level” and “reverse engineer the design.” But, he said, it can be done.
https://www.cnn.com/2022/07/23/politics/fbi-investigation-huawei-china-defense-department-communications-nuclear/index.html
And now we have this … Why the end of Optane is bad news for all IT. (https://www.msn.com/en-us/news/technology/why-the-end-of-optane-is-bad-news-for-all-it/ar-AA10bEvJ?cvid=fe3783ab72a54daca923d0133e38c18e) I am suddenly feeling deflated. Are we computer techies so locked into the notion of fast dynamic memory and slow persistent memory that we can’t see the advantages of investing in fast persistent memory?
Intel is moving into a broad range of products … AV, ADAS, Pluggable optics, discrete GPUs, AI ASICS, Xeon-D edge servers, E-core Servers, IPUs, P4 programmable switches.
The SPR-HBM released benchmarks show huge gains. The 47 tile Ponte Vecchio GPU in Argonne’s Aurora computer has been reassessed as a >2EF supercomputer.
not hard at all to see Intel’s TAM tripling as all they enter these new markets … and that is without even considering the IFS.
Yes. But right now, it is tough.