News of the Week (April 8 - 12)

News of the Week (April 8 - 12)

21 min read

Table of Contents

1. Amazon (AMZN) – Shareholder Letter, India & Piper Sandler

Subscribe now

a. CEO Andy Jassy Shareholder Letter:

Amazon CEO Andy Jassy published his latest shareholder letter this past week. It dove into how Amazon thinks and how it has configured its business to drive historic success. It was a lengthy, yet valuable piece for understanding why you own what you own. Here are the highlights:

On Primitive Services:

The main theme of this letter revisits a 20-year-old Amazon idea about “Primitive Services.” In the firm’s mind, these are “discrete foundational building blocks woven together in whatever combination desired.” It famously introduced this idea in 2003 in its AWS vision letter. These services fixate on unbundling use cases to perfect them one at a time. The roadmap of development is guided solely by customer pain-points, and each service is beautifully tied together via slick application programming interfaces (APIs) to ensure the pieces are cohesively, not disparately, conjoined. This allows Amazon to emulate service customization without needing to build from scratch each time a merchant wants something slightly different. It creates a la carte style client purchasing. Why is this important? Because merchants are all unique, as are their needs. Think of primitive services like atoms. They’re indivisible, but can be combined in nearly-endless ways into different molecules to create a plethora of end products.

Jassy took us through how Amazon created this approach and how it’s implemented. Looking backwards, Amazon’s fulfillment network was the example used. The firm first built a primitive service to perfectly handle and distribute inventory for Amazon only. It then leveraged this service and opened it up to 3rd party merchants. This created operating leverage on existing capital infrastructure while giving its marketplace a broader array of inventory. He spoke about how Amazon applied the exact same idea to its payments, search, and other microservices like last-mile delivery. These needed to exist separately from one another, rather than being inherently stapled together, which Amazon learned the hard way through early failure. At the very beginning, it let Target.com use its tech to power their storefront and “jumbled” services together with some customization built-in. Instead, that design required partitioning of products into primitive services, which eventually laid the groundwork for AWS.

This idea seems to stand in stark contrast to the “vendor consolidation” and “platformization” software themes of 2023 where combining products is preferred… but it doesn’t. Amazon is still offering an overarching platform, just a malleable platform that looks different for different clients. This ensures that every single ingredient in the overall platform is best-in-class and used only when needed. Primitive services can be enjoyed by Amazon, leveraged by clients in isolation (like with Amazon Shipping for last-mile delivery for example), or weaved together to create something commonly needed by client needs like Buy with Prime (BwP) or Fulfillment by Amazon (FBA). Per Jassy, building with this mindset amplifies “freedom” of creation and fosters many more ways to drive profitable growth. Jassy discussed how Amazon failed early on in healthcare and robotics. It didn’t begin to succeed until it began to embrace this exact same philosophy of building one hyper-localized service at a time to solve one, specific pain-point.

The Next Primitive Service Focus Areas:

Generative AI (GenAI) is the next pillar of primitive services for Amazon. I’ve previously shared the three layers of GenAI. Layer one encompasses foundational models (FMs) and involves the chip and hardware infrastructure that goes into creating them. Jassy talked about how Nvidia’s graphics processing units (GPUs) have been used to create most FMs to date. However, advancements in Amazon’s Trainium (for model training) and Inferentia (for model inference/prediction) chipsets are powering new models of best-in-breed players like Anthropic and Hugging Face. He mentioned significant performance, cost and memory bandwidth upgrades for the new versions of these chips, as well as its Graviton central processing units (CPUs) for general compute. Amazon’s SageMaker is its managed means of guiding developers through the model creation journey in a more intuitive manner. It’s sort of like an instruction template. SageMaker is helping Perplexity AI train models 40% faster and Workday reduce inference latency by 80%.

Layer two is where companies leverage FMs to create niche, localized, and customized models. Amazon Bedrock is its big product here, which already has tens of thousands of customers and a broad set of available foundational models to utilize. 

Layer three is the consumer application layer, where its shopping assistant (Rufus), automated AWS coding (Amazon Q) and Alexa upgrades will come into play. The first spending wave has been within layer one as hyper scalers and consumer internet giants race to put infrastructure in place so they can enjoy monetization in the other two layers down the road.

On E-Commerce & Logistics:

Jassy essentially took a victory lap here on how rapidly Amazon has been able to optimize its supply chain via regionalization and same-day facility expansion. He reviewed how these factors have cut Amazon’s per unit costs, with improved delivery speed and customer service, for the first time in six years. The new layout and its last-mile delivery investments are also freeing Amazon to be a bigger player in perishable goods as well as everyday essentials. This category grew by 20% Y/Y in Q4 2023, which was a new disclosure and is encouraging. He also told investors that more reductions in per unit fulfillment costs were coming in 2024. Focus areas will evolve from localization and better supply chain monetization to inbound fulfillment optimization and inventory placement. Its newer same-day facilities and inventory algorithms are placing inventory closer to the final predicted destination more often. That drives down cost.

In other e-commerce news, Mexico became its latest country to turn profitable. It generally takes Amazon about a decade to deliver this inflection in new markets, and it remains confident that every single newer market is on that trajectory.

“There has never been a time in Amazon’s history where we’ve felt there is so much opportunity to make our customers’ lives better and easier. We’re incredibly excited about what’s possible, focused on inventing the future, and look forward to working together to make it so.” – CEO Andy Jassy

I was already a big fan of Jassy’s. The more I hear him talk and read his thoughts, the more I like him.

b. India

A few years ago, Amazon lost a bidding war to Walmart in its quest to buy a majority stake in Flipkart – the Indian e-commerce giant. Following that development, Amazon pushed aggressively into the market and claimed the top site traffic spot as of January 2024. Still, Flipkart leads in overall e-commerce market share… but Amazon has its ambitions set on a larger piece of the pie. This week, it announced “Bazaar” as a new, low-priced e-commerce site in that market. It will sell fashion and home goods for prices ranging from $2.39 to $7.21.

India is an incredibly compelling international growth market. Its population is young, giant and growing while its middle class continues to expand too. Its government is both business and US friendly; its economy is poised for 6.5% growth this year as it races to become the 3rd largest in the world. This news makes a ton of sense as Amazon gears up to invest $26 billion in India over the next 6 years.

c. Piper Sandler Data

Amazon remains the most widely used e-commerce site among teens. 61% of respondents surveyed by Piper Sandler use it vs. 57% Y/Y. This just keeps ticking higher & higher for every new survey released.

2. Alphabet (GOOG), Nvidia (NVDA) & Intel (INTC) – Semiconductor Market News

a.  Alphabet’s Chip Announcement

This week, Google announced its latest chipset called Axion. Axion is based on Arm Holdings architecture and is a CPU product. Google has been hard at work on custom chipsets for about a decade, and Axion is the latest result of that work. 

Importantly, Axion will be used for general compute (GP) purposes. This isn’t like Amazon’s Trainium and Inferentia chips that will be for high-performance compute (HPC) apps. That makes this new release more of a competitive threat for Intel and AMD than for Nvidia. HPC applications, powered largely by Nvidia’s GPUs today, have unlocked more data center flexibility, utility and potential. But? GP is not going to go away according to Google. Google sees it as “remaining a critical portion of cloud-based workloads” and Axion is meant to claim a piece of that specific market. Axion can greatly help with overall cloud cost optimization; HPC is very expensive and it doesn’t make sense to use it unless you need to.

In the release, Google wrote about GP CPU innovation slowing and falling too far behind HPC. If the gap widens too much, GP becomes a performance and efficiency bottleneck and cannot be effectively utilized. Axion is Google’s strategy for ensuring that the gap doesn’t widen too much, while also creating new revenue opportunities.

Per Google’s internal data, Axion delivers 30% better performance than the next-best Arm-based GPU workloads and 50% better performance than other current generation competition. It’s also 60% more energy efficient than those competitors. Amazon says similar things about its Graviton CPUs. Meta, which delivered more AI semiconductor releases covered the next section, and other chipmakers do too. As these claims are largely internally-derived, take them with a grain of salt.

Axion was built to work seamlessly with Arm’s infrastructure and workloads. This will make it easier for Arm’s “tens of thousands of cloud customers” to deploy their assets in Google Cloud. Early on, CrowdStrike, Datadog and Elastic have been named as early users.

b. Competition Heating Up

Intel released its latest AI chip called Gaudi 3. It will directly compete with Nvidia’s Hopper, Grace and Blackwell systems. Per Intel, it’s 2x more energy efficient and can run models 50% faster than the Hopper 100 (H1000) GPU. Nvidia has already released two higher-performing chips for these use cases since the H100, but this is still considered a large leap forward for Intel.

3. Meta Platforms (META) – More Semiconductor News & Piper Sandler Survey Data

Subscribe now

a. More Semiconductor News

While Nvidia’s chipsets will power a large chunk of Meta’s next-gen data center and GenAI infrastructure, Meta’s Meta Training & Inference Accelerator (MTIA) AI will help vertically integrate semiconductor creation for certain use cases. Just like Amazon… just like Google… just like Apple. These use cases for Meta include its advertising and discovery algorithms, which are quite unique workloads. That uniqueness makes internal design and manufacturing a more rational move, given Meta can create them precisely for the unique features it needs. Per the team, the process is more efficient than delegating to a 3rd party.

The potential use cases are expanding with the second version of MTIA AI chips announced this past week. The new chipset can handle more advanced recommendation and targeting workloads, along with unlocking “new experiences,” with better efficiency and performance (so lower cost). It plans to continue investing more aggressively here in the coming years to optimize efficiency of its tech stack and enable more seamless, flexible scaling. Early success is emboldening its approach despite Meta’s massive orders for Nvidia’s advanced chips. MTIA version 2 is now live in Meta’s data centers and is being used by developers to build new models and applications with the firm.

b. Piper Sandler Data

So teens in the U.S. don’t use Facebook anymore, right? Right? Not so fast. Facebook monthly usage among teens rose to 32% and set a 4 year high in Piper’s data release from this week. It attributes this to improving advertising products and discovery algorithms.

Instagram saw its largest rise in favorite app preference since TikTok debuted with 7 points of additional share Y/Y. It’s now the 2nd favorite teen app behind TikTok as it passed a struggling Snapchat. Piper called time spent trends there “encouraging.” It remains the most used app in the U.S. with its lead growing from 5 pts to 8 pts Y/Y.

In other Meta news, Piper Sandler’s advertising channel checks were quite positive and led to it raising estimates from Q2-Q4 by about 1%.

4. Lululemon (LULU) – Valuable Brand & Company Data

a. Store Traffic Before & After Alo Store Openings