Good morning,
I apologize that this Update is late. I really had to get into the weeds to make sense of these earnings.
On to the update:
Nvidia Earnings and Inventory
Two weeks ago, from the Wall Street Journal:
Graphics chip maker Nvidia Corp. issued a muted outlook and reported a sharp decline in quarterly sales, driven by waning consumer demand for its videogaming chips after a pandemic-fueled boom and the onset of the cryptowinter. America’s largest chip company by value on Wednesday said revenue fell 17% to $5.93 billion after gaming-segment sales more than halved in its fiscal third quarter. Net profit was $680 million. The sales were above expectations in a survey of analysts by FactSet, but net profit fell short.
Gaming is straightforward: it was bad, but we knew it would be bad. CFO Colette Kress said on the earnings call:
Moving to gaming, revenue of $1.57 billion was down 23% sequentially and down 51% from a year ago, reflecting lower sell-in to partners to help align channel inventory levels with current demand expectations. We believe channel inventories are on track to approach normal levels as we exit Q4.
The problem for Nvidia, though, is that their own inventories are exploding:
| Inventories | October 30, 2022 | January 30, 2022 |
|---|---|---|
| Raw materials | $1,936 | $791 |
| Work in-process | $768 | $692 |
| Finished goods | $1,730 | $1,122 |
| Total | $4,454 | $2,605 |
Kress said on the earnings call:
Now looking at our inventory that we have on hand and the inventory that has increased, a lot of that is just due to our upcoming architectures coming to market, our Ada architecture, our Hopper architecture and even more in terms of our networking business. We have been building for those architectures to come to market and as such to say. We are always looking at our inventory levels at the end of each quarter for our expected demand going forward. But I think we’ve done a solid job, at least in this quarter, just based on that expectation going forward.
I read this with a bit of skepticism given the numbers above. Sure, I could buy that raw materials or work in-progress are up because Nvidia is ramping, but why would finished goods be so much higher? Isn’t the Occam’s Razor explanation that Nvidia still has previous generation inventory that it has yet to clear (or write-down)? To double-check my assumptions I went back to look at Nvidia’s inventory levels for Q3 2020, when the company launched its previous generation:
| Inventories | October 25, 2020 | January 26, 2020 |
|---|---|---|
| Raw materials | $455 | $249 |
| Work in-process | $380 | $265 |
| Finished goods | $660 | $465 |
| Total | $1,495 | $979 |
This is a different picture than I expected:
| Inventories Increase | Q3 2020 | Q3 2022 |
|---|---|---|
| Raw materials | 83% | 145% |
| Work in-process | 43% | 11% |
| Finished goods | 42% | 54% |
| Total | 53% | 71% |
Finished goods increased a bit more this cycle (while work in-process increased less), but it is raw materials that is out of line with the last cycle.
In retrospect, though, this actually makes sense. Go back to Nvidia’s inventory charge last quarter; from that quarter’s 10-Q:
Gross margin decreases were primarily due to a $1.34 billion charge, comprised of $1.22 billion for inventory and related reserves and $122 million for warranty reserves. The $1.22 billion charge for inventory and related reserves is based on revised expectations of future demand, primarily relating to Data Center and Gaming. The charge consists of approximately $570 million for inventory on hand and approximately $650 million for inventory purchase obligations in excess of our current demand projections, and cancellation and underutilization penalties.
Notice that that charge was not simply for existing inventory (which, I have to admit, I assumed it was); in fact, less than half of it was. The majority was a charge against Nvidia’s purchase obligations with TSMC for its upcoming chips. In other words, Nvidia didn’t simply overbuild for the previous generation; it over-ordered for the just-arriving one (what I missed was this was labeled an inventory charge).
This, by extension, explains the increase in raw materials inventory: Nvidia wouldn’t have just pre-ordered space with TSMC, but would have also ordered the materials to be used in that space. However, while Nvidia likely received some sort of discount from TSMC for giving unused space back — if Nvidia doesn’t need that space now, then TSMC would be happy to sell it to someone else — Nvidia will use those raw materials eventually, so they might as well hold onto them.
That’s not the end of the write-off story, though: Nvidia took another inventory charge this quarter. From this quarter’s 10-Q:
Gross margin for the third quarter was down 11.6% from a year earlier, primarily due to a $702 million inventory charge, largely relating to lower Data Center demand in China, partially offset by a warranty-related benefit of approximately $70 million. Sequentially, gross margin was up 10.1% primarily due to lower inventory charges compared with the second quarter. The $702 million inventory charge consists of approximately $354 million for inventory on hand and approximately $348 million for inventory purchase obligations in excess of our current demand projections.
It’s a bit of a tough scene when you have to admit that your sequential gross margins are higher because at least this quarter’s inventory charge was less than last quarter’s inventory charge! Once again, though, note the split between inventory on hand and purchase obligations: this isn’t just about having made too much, but having agreed to make too much in the future.
Nvidia and China
Whereas last quarter’s write-offs were about gaming, this quarter’s are about China, but once again, the driver was a bit different than expectations.
First, go back to August 26 when Nvidia reported that the US government had banned the sale of A100 and H100 AI chips to China; from the SEC filing announcing the ban and its impact:
The Company’s outlook for its third fiscal quarter provided on August 24, 2022 included approximately $400 million in potential sales to China which may be subject to the new license requirement if customers do not want to purchase the Company’s alternative product offerings or if the USG does not grant licenses in a timely manner or denies licenses to significant customers.
The good news for Nvidia is that this revenue shortfall did not, for the most part, materialize; from Kress’s prepared remarks:
During the quarter, the U.S. government announced new restrictions impacting exports of our A100 and H100 based products to China, and any product destined for certain systems or entities in China. These restrictions impacted third quarter revenue, largely offset by sales of alternative products into China. That said, demand in China more broadly remains soft, and we expect that to continue in the current quarter…
That “alternative product” appears to be the A800, which is an A100 but with lower memory bandwidth. I wrote about the A800 in an Update earlier this month:
There are a few interesting points about this announcement. First, notice that Nvidia wasn’t exactly broadcasting a press release about this new chip; Reuters found advertisements in China and reached out to Nvidia for comment. That seems like an implicit admission on Nvidia’s part that this announcement would not be particularly popular in the U.S.
Second…I don’t think it is a coincidence that the rules specify 600 Gb/s given the fact that Nvidia’s A100 chip has 600 Gb/s interconnects…Maximizing compute is increasingly a matter of system design, not just chips; this is particularly important in terms of AI training, which is an “embarrassingly parallel” task: the speed of processing increases basically linearly with the amount of processors available, but the trick is keeping all of those processors busy, which is where the memory bandwidth comes in. In other words, the A800 chip, which is limited to 400 Gb/s interconnects, will operate as fast as the A100, but it will more likely be sitting idle more often because the interconnect is saturated moving data around.
Third, I did find the phrasing from the Nvidia spokesperson interesting: “The A800 meets the U.S. Government’s clear test for reduced export control and cannot be programmed to exceed it.” I can’t decide if this sounds like a software limit or a hardware one; the former would be faster to implement — and Nvidia sure started producing these cards quickly — but cannot be programmed to exceed it sounds like there is no software solution, which, if the limitation were software, would by definition exist.
CEO Jensen Huang addressed all three points in response to a question as to whether or not Nvidia was working against the spirit of the U.S. government’s regulations:
The hardware of A800 ensures that it always meets U.S. government’s clear test for export control. And it cannot be customer reprogrammed or application reprogrammed to exceed it. It is hardware limited. It is in the hardware that determines A800’s capabilities. And so, it meets the clear test in letter and in spirit. We raised the concern about the $400 million of A100s because we were uncertain about whether we could execute, the introduction of A800 to our customers and through our supply chain in time. The company did remarkable feats to swarm this situation and make sure that our business was not affected and our customers were not affected. But A800 hardware surely ensures that it always meets U.S. government’s clear tests for export control.
First off, the fact that Nvidia went from warning on August 26 that it might face a $400 million shortfall, to reporting for the quarter ending October 30 that this revenue impact was “largely offset by sales of alternative products”, strongly suggests that A800s are not new cards but A100s with some sort of modification. That would, in a vacuum, suggest software, but Huang insisted the limitation was hardware.
The reason to believe him is that Nvidia is still writing off already-built A100s; Kress said in response to a question:
So, as we highlighted in our prepared remarks, we booked an entry of $702 million for inventory reserves within the quarter. Most of that, primarily all of it is related to our data center business, just due to the change in expected demand looking forward for China. So, when we look at the data center products, a good portion of this was also the A100, which we wrote down.
Even if China doesn’t want those A100s, it’s surprising to me that Nvidia can’t sell them elsewhere; to be fair the H100 is just rolling out, but given how hard it has been to get an A100 it would seem that Nvidia could dump them somewhere. Unless, of course, those A100s have already been modified to be A800s…which Chinese companies don’t want. This certainly fits the timeline: Nvidia rushes to convert a bunch of A100s, Chinese companies buy some of them, but not as many as expected, and Nvidia has to both write them off and reduce its forecasts (for which it already paid TSMC). This also suggests that there may be an H800 (i.e. a reason to not buy those A800s in the future): the chip ban rules specify 600 GB/s memory bandwidth and a certain level of performance; in other words, you can sell a very performant chip as long as the interconnects are slow enough (indeed, the A800 exceeds the performance rule, but because the memory bandwidth is limited it is allowed).
Nvidia and Public Clouds
Once more from Kress’s prepared remarks:
Earlier today, we announced a multiyear collaboration with Microsoft to build an advanced cloud-based AI supercomputer to help enterprises train, deploy and scale AI including large state-of-the-art models. Microsoft Azure will incorporate our complete AI stack, adding tens and thousands of A100 and H100 GPUs, Quantum-2 400 gigabit per second InfiniBand networking and the NVIDIA AI enterprise software suite to its platform.
Oracle and NVIDIA are also working together to offer AI training and inference at scale to thousands of enterprises. This includes bringing to Oracle cloud infrastructure the full NVIDIA accelerated computing stack and adding tens of thousands of NVIDIA GPUs, including the A100 and H100. Cloud-based high-performance [Technical Difficulty] is adopting NVIDIA AI enterprise and other software to address the industrial scientific communities’ rising demand for AI in the cloud.
Like I said, there is definitely demand for A100s!
Other than that, I am reminded of what Huang told me in a Stratechery Interview last spring when I asked him if Nvidia would build their own cloud service:
If we ever do services, we will run it all over the world on the GPUs that are in everybody’s clouds, in addition to building something ourselves, if we have to. One of the rules of our company is to not squander the resources of our company to do something that already exists. If something already exists, for example, an x86 CPU, we’ll just use it. If something already exists, we’ll partner with them, because let’s not squander our rare resources on that. And so if something already exists in the cloud, we just absolutely use that or let them do it, which is even better. However, if there’s something that makes sense for us to do and it doesn’t make for them to do, we even approach them to do it, other people don’t want to do it then we might decide to do it. We try to be very selective about the things that we do, we’re quite determined not to do things that other people do.
I came back to this quote when Nvidia announced its Omniverse Cloud, but it’s worth pointing out that Huang was being honest: Omniverse Cloud is not a GPU compute platform; it’s a bespoke Nvidia creation for its Omniverse (i.e. Nvidia’s metaverse), which no one else is going to build. On the other hand, Nvidia is, quite clearly, willing to let cloud providers do what they do best, which is sell cloud computing capacity, in this case capacity that runs on Nvidia chips.
This Update will be available as a podcast later today. To receive it in your podcast player, visit Stratechery.
The Stratechery Update is intended for a single recipient, but occasional forwarding is totally fine! If you would like to order multiple subscriptions for your team with a group discount (minimum 5), please contact me directly.
Thanks for being a subscriber, and have a great day!