Join institutional investors, analysts and strategists from the world's largest banks: Subscribe today
Already have an account? log in
Amazon Asks For Patience Amid Slower Sales, Soaring AI Spend
Amazon said Thursday sales grew 10% to $187.8 billion during the holiday quarter. That was marginall
I’m impressed with how quickly you get these notes out. Is AI involved?
Anyway, thoughts post the call. +11% ex-FX topline growth is impressive for a Mega-sized company. But if we only care about AWS, then we see a little decel from 3Q and while EBIT margin was up +730bp YOY, it was down -120bp sequentially. Said growth will be “lumpy” and margins will “fluctuate” in coming years. Growth is capacity-constrained, both NVDA and AMZN chips, Trainium2 to ramp in coming months but AI is lower-margin than non-AI, so accel is a margin headwind. Investors may be wondering where margins are going, because the 2024 increase to server life added +200bp (!) to AWS 4Q margin and now they are decreasing the useful life for some hardware and retiring some hardware early; the hit to AWS margins sounds modest (<50bp?) but any hint of unexpected obsolescence is bad when AMZN is spending more on capex than any of the Megas and hence more than any company in history (not rigorously fact-checked, sounds reasonable). Trends and margin ramp in NorAm and International look good, but that’s not the investor focus. 1Q estimates should go down, 2025 maybe gets trimmed or backloaded.
To answer that first question you posed… I’ve been pushing out near-instantaneous analysis of quarterly reports and macro data points for longer than OpenAI (the company, not the chatbot) has existed. I have this down so cold after all these years that I can write most of this analysis before the numbers are released, which in a lot of cases is exactly what I do. I write it ahead of time, then just fill in the details after the numbers are out, both for earnings and macro data. Sometimes it doesn’t work (i.e., a given report or a given macro releases surprises enough that I have to start from scratch) but about 70% of the time, my pre-written “post”mortems are dead on. (See what I did there? Postmortem… “dead” on.)
Mortem or no, it’s very impressive.
I guess this approach is cheaper than extra staff. Higher productivity for sure.
The issue when contemplating amazon is the price of the stock. It’s an excellent company, with a wide moat protecting its business. The price assumes a long runway of high growth. The challenge in buying a growth stock is when to sell. So far buy and hold has worked pretty well. How long will the moat hold and the growth continue?
Jassy better hope the bull market continues. Patience is hard to find during a bear. The Street selectively forgets the meaning of ROI when convenient to the bull case.
So much craziness going on in the US and across the globe. Very few seem to notice or care. H and all or most of those reading his work notice and care. Rare group.
“Risk” is a word that used to have meaning. I can practice compartmentalization and outright denial w/ the best. But when I let myself actually envision what lies on the other side of this risk-free, overhyped market, it is gut-wrenching. Probably b/c I’ve lived through similar times before. But it won’t matter. It is unnerving at best even when you expect it and make money on the way down. Maybe this time is different. Death by a thousand cuts rather than a few machete slashes.
Yeah, Musk and a couple of his twenty-something year old pupils gain control over the payment system the US govt. uses and there’s no recognition of additional risk in the market. I don’t get it. That’s a great example of why I come here.
One thing all of that loquacious drivel implies is that AWS is positioning themselves as having a distinct cost advantage when people run or train their models on AWS.
Their argument, reiterated tirelessly at re:Invent this year, was that the Trainium chip and the new computer architecture that accompanies it can deliver a far lower cost per token generated (or round of training performed) than “commodity” server hardware with Nvidia hardware plugged into it.
I use scare quotes because the IT hyper scalers all figured out a decade ago that building one’s own servers drives a cost advantage due to power and cooling efficiency. Heretofore, however, the chips themselves and the basic layout of the chips and the buses, did not vary substantially between cloud providers.
Trainium’s architecture is so nonstandard that it can’t run an operating system; it’s more like a GPU the size of a small SUV with some absurd amount of “onboard” memory. Typically this would be a nonstarter for ML people, as all of their tools wouldn’t run on the new toy. Since AWS controls the tooling layer (SageMaker and Bedrock), compatibility is guaranteed for anyone who builds inside AWS’ walled garden.
I know I’m joining the ranks mentioned in your article by disregarding retail and hyper focusing on AWS, but assuming that Trainium is no paper tiger and that AWS passes the savings on to customers, I do believe that Nvidia (or someone) will need to pivot to more vertically integrated solutions to keep up with AWS’ efficiencies.
A new pick and shovel player may have entered the arena last quarter; we shall see.
Thanks! That is really fascinating. Another walled garden.
Tracking earnings season. As of today, 49% of the S&P500 has reported 4Q24, 63% by name count (68% by market cap) beat cons rev, 80% by count (88% by cap) beat cons EPS, but only 36% by count (29% by cap) saw next quarter’s cons rev go up, and only 28% by count (31% by cap) saw next quarter’s cons EPS go up. Average reaction to the reports was -0.5% by count (-0.7% by cap). I haven’t gone back and checked, but I think that is materially worse than I recall for the previous few quarters.
“When we look back on this quarter several years from now” – is that a thing people do?