“If you take one thing from today's call, let it be this: Despite near-term challenges, we have significant strengths and a clear plan… We are where we want to be one quarter in, but much of our work is just beginning. While we’re only one quarter into our turnaround, we’re moving quickly to act on the 'Back to Starbucks' efforts and we’ve seen a positive response.”

CEO Brian Niccol

North America revenue fell 1% Y/Y and beat estimates by 1%. This was driven by -4% comparable store sales (CSS) growth, which beat -5% Y/Y estimates. Within that CSS growth, transactions declined by 8% Y/Y and ticket size rose by 4% Y/Y. Things like cutting some up-charges and a mix-shift to cheaper drinks were more than offset by price hikes last year and lower discounting. Much more on all of this later. While revenue growth continues to be lackluster, this does mark an improvement in the trend and Starbucks thinks it has a “clear path forward.”

Outside of the USA, revenue rose by 1% Y/Y and beat estimates by 2.5%. This was thanks to -4% international CSS growth, which beat -6% Y/Y estimates. China CSS growth was -6% Y/Y vs. -9% Y/Y expected. 9% store growth for this segment helped offset 2% declines in traffic and ticket size, while more revenue from purchasing its U.K. franchise partner helped here too.

The channel development segment saw -3% Y/Y revenue growth due to lower revenue from the Global Coffee Alliance (Nestlé partnership) and category weakness.

North American EBIT margin sharply contracted from 21.4% to 16.7% Y/Y. This was as planned and as a result of its “Back to Starbucks” Investments. As a reminder, this Brian Niccol-inspired campaign aims to fix crippling throughput issues, reduce menu clutter, revamp the food offering (with actual testing), improve in-store customer experience and focus on serving consistently great coffee. This quarter, the main margin headwinds stemming from this were via more marketing activity, wage and benefits investments and eliminating the up-charge for non-dairy milks. It was able to recover a bit of this decline with supply chain and brick-and-mortar efficiency gains.

Internationally, the EBIT margin decline was much more modest. Issues outside of the USA are more to do with macro than terrible execution, so things like more marketing and free premium add-ons aren’t as necessary as they are here. As a result, the margin fell from 13.1% to 12.7%.

“Although we are in the beginning chapter, and have much more work ahead of us, we will continue to prioritize shareholder value through dividends, providing a predictable return of capital while we turn around our business.”

CFO Rachel Ruggeri

We got very little color on guidance, as expected. Organizational changes discussed below will lead to EPS troughing in Q2 2025, with positive growth resuming in Q3 and Q4. The Q2 weakness is related to restructuring and lapping lower performance-based compensation in the prior year period. Finally, it walked back the old team’s commitment to $4 billion in operating expense (OpEx) savings.

EPS is expected to compound at a 21% clip for the next two years. Estimates should be mostly stable (might move a tick higher) following this report.

Back to Starbucks:

As a reminder, like when he led Chipotle, new CEO Brian Niccol has instituted a “Back to Starbucks” plan to turn things around. The key focus areas within this plan include: Reintroducing Starbucks to the world, greatly improving customer experiences, revamping the in-store environment and creating a better work environment for internal career mobility. Candidly, this company was a complete mess before he took over. Old leadership had made inexplicably poor decisions like abandoning food testing and stage-gate processes, while turning menu introductions into trial and error. Gigantic companies that are fundamentally lost cannot be turned around overnight. This will take time, but Niccol’s track record is elite and progress is already beginning to pop up. 

“To be the premier purveyor of the finest coffee in the world, inspiring and nurturing the human spirit — one person, one cup and one neighborhood at a time.”

New Starbucks Mission Statement

Reintroduce Starbucks to the World:

Under the old team, broad-based marketing campaigns were largely abandoned. The company fixated on marketing to its most loyal rewards members, but ignored the demand and preferences of everyone else. Under Niccol, frequent discounting has been greatly slashed and a 40% Y/Y decline in discounted transactions is the early result. Starbucks is not letting these added profit dollars flow down the income statement just yet. Instead, it’s re-earmarking this budget to national marketing campaigns to take back control of the Starbucks brand story and perception. I’ve personally started to see a lot more ad placements from them, and they do seem to be well done in my view. Directly following the debut of these campaigns, customers responded positively. Specifically, Starbucks comparable store sales trends improved throughout the quarter, with this being one of the main contributors.

Improve the Customer Experience & In-Store Environment:

There are many facets to this core initiative, with throughput perhaps being the most important piece. Through intricate testing, it has become apparent that mobile order sequencing is the biggest issue as Starbucks works towards its goal of fulfilling all orders in 4 minutes or less. It’s this poor sequencing process that leads to a mountain of cluttered, cold drinks at the end of the ordering line. Capacity is also an issue, which is why the team has invested in more labor in 3,000 U.S. stores (more initiatives coming to support staff in 700 more stores), upgraded scheduling and changed assembly routines, but throughput comes with the most low-hanging fruit to devour.

It’s now testing a new “in-store prioritization algorithm” to better organize workflows and take the guesswork out of which drink to make first. As of now, there is still a lot of that guesswork involved. During hectic peak hours, baristas are expected to know which order to fulfill first – across several channels and with little guidance. This is meant to fix that and take this massive headache out of their day-to-day work.

Other customer experience changes, like eliminating non-dairy up-charges, simplifying customization and ordering processes and reducing menu clutter, have all also been well received so far. This year, Starbucks plans to cut 30% of its overall menu (food and drink) to “free it up to make sure it has the right offerings.” All menu boards will become digital through 2026 to improve flexibility. It’s starting to lean more heavily on its baristas for feedback on what customers actually want to better guide testing. What a concept. This worked extremely well with its lavender drink lineup. 

To recapture the “community coffeehouse reputation,” in-store drinks are now served in ceramic mugs, barista sharpies are back to deepen the personal connection and it expanded access to free refills for some drinks. It also debuted a new code of conduct to operationalize in-store standards.

All of these changes help the customer experience, but also the employee experience. It’s items like these and reintroducing its coffee condiment bars that make working for Starbucks more enjoyable and create a compelling win-win-win: happier customers, happier workers and happier shareholders.

Better Place to Work:

As previously discussed, Starbucks doubled paid parental leave and is committed to 90% of retail leadership roles being internally filled over the next 3 years. Based on its commitment to employees, “shift completion, hours per partner, retention and hourly partner engagement” all improved.

More Changes:

Starbucks is reorganizing its support teams to “improve efficiency and accountability.” As part of this, it hired a new Chief Store Officer to “drive store excellence” and a Chief Development Officer to help with real-estate pipelines. This will lead to a spike in Q2 G&A, but some savings beginning in Q4.

“Similar to what we're trying to do for driving store accountability, we want to make sure that we've got the support center also focused on supporting the stores in an efficient manner and an accountable manner to where the business happens.”

CEO Brian Niccol

The Starbucks app is adding an option for customers to schedule mobile orders, while gearing up for another update this year to “improve customization and pricing.” On the scheduling, SBUX has found that making people wait more than 15 minutes leads to high abandoned order rates. This item is meant to improve proportion of mobile orders ready in under 15 minutes.

Starbucks is no longer eager to deploy Siren Craft in all stores. This marks the final strategic priority from the old team that has now been abandoned by Niccol. Siren Craft simply refers to upgraded workflows and tools (like a new blender and food warming systems) that make service more efficient. Niccol thinks these ideas are good, and the implementation has just been too slow and disruptive to operations. The idea here is to improve throughput, but most of its stores don’t need help getting to the sub-4 minute goal. Now, they will exclusively focus this system on the stores that actually do need the help. This should minimize disruption and cost.

Early Signs of Progress:

Starbucks quick service market share stabilized in Q1 after two consecutive negative quarters. It maintained its #2 share position in the gift card space. Non-rewards members delivered positive traffic growth Q/Q and rewards membership resumed positive growth. The decision to eliminate non-dairy up-charges also “brought back lapsed Starbucks Rewards members, while bringing back condiment bars and adding more free refills were the “top drivers of purchase intent.” It’s important to keep in mind that these expenses aren’t all incremental, as again Starbucks is greatly cutting back on discounting to make room for providing value through compelling perks such as these.

“Progress like this shows me that the Starbucks brand is still resilient and strong and that we have significant future potential.”

CEO Brian Niccol

Store Footprint:

Starbucks sees a path to doubling USA store count “while improving portfolio health.” Strong debuts in newer markets across the Southeast build on this confidence, while new stores being 90% incremental to results does too. 

The new mobile order sequencing plan is a big piece of its better-than-expected total addressable market (TAM) assumption. Per Niccol, this “frees up another degree that it hasn’t totally comprehended yet.” That made it sound like the doubling estimate may not even be the ceiling here. Large efficiency gains unlock significant potential for once-fringe markets to become more compelling from a unit economics standpoint.

International:

Niccol toured facilities across Italy, Japan and Korea. He sees many of these markets as “setting an example for the experience we aim to deliver in the USA.”  He also visited China for the first time, and sees several opportunities to address near-term challenges to “stabilize and strengthen the business.” It’s also exploring more strategic partnerships and borrowing supply chain efficiencies from the Chinese market to implement in other countries. There were rumors of Starbucks selling this business once and for all, but this commentary (in my mind) makes that a bit less likely.

Dissecting Sources of EBIT Margin Contraction:

All of the investments we’ve spoken about thus far are currently weighing on margins. Labor investments hit the EBIT margin by 180 bps; cutting non-dairy upcharges hit it by another 60 bps (while delivering the intended engagement uplift). More marketing actually didn’t impact things all that much, because those dollars came from fewer discounts. On the other hand, supply chain efficiency gains offset these declines by about 150 bps. While coffee commodity prices do continue to zoom, Starbucks does a fantastic job of hedging away this risk. This wasn’t a material source of margin weakness.

Comments on Current Labor Disputes:

“In December, the company shared an update on contract negotiations with Workers United and remains committed to engaging constructively and in good faith to reach collective bargaining agreements for represented stores and partners.”

Press Release

The quarter was very bad, but a bit better than feared. What’s more important is that Niccol’s plan of action is actually a good one and his track record clearly points to him being capable of executing a turnaround. The changes implemented will need time to fully take hold, but early progress is palpable and shows the brand is entirely fine. They just need to execute, and I’m confident they will with Niccol running the show. I still think we have a few more bad quarters ahead of us, before double-digit EPS growth resumes and Starbucks reclaims its rightful place as a blue-chip fundamental darling.

Headlines swirled over the last week about a Chinese AI firm called DeepSeek. It’s supposedly able to train high-quality models, like its R1 model, at a fraction of the cost of other leading options. That led to many arriving at a broad range of conclusions about overall market implications.

There are many conflicting notes on how expensive R1 actually was to build and how many H100 chips it illegally had access to. While the $5.6 million training figure floating around is likely way off, we just don’t know the actual price tag. There are still many, many moving pieces here (we don’t even know if this will eventually be banned by the west), but I did want to work through my thoughts. Here, I will assume there were actually large training efficiency gains discovered by DeepSeek and this piece will be highly opinionated.

Mega-Caps Besides Nvidia:

I do not think DeepSeek cutting model training costs is a large threat to these businesses. None of them struggle with competitive moats. All of them have world-class network effects, databases, infrastructure, product bundles and/or ecosystems. They do not need models to cost two fortunes to build to prevent others from being able to make them. They don’t need this to be yet another competitive differentiator.

Meta:

For Meta, the company has intentionally made itself an open-source player in this market. All of the work DeepSeek has done here is open source and a lot of it was built with the help of Llama. So? Meta can use all of this work for its own improvement (as Yann LeCun told us) to drive product parity and pocket a fortune in forgone CapEx. Or? It can just build way better models more efficiently. That choice will be a key theme throughout this piece for most companies mentioned.

Furthermore, it’s not like it’s meaningfully monetizing Llama directly. So? Workload pricing power isn’t a relevant debate like it is for OpenAI.

Meta can lean on having half of the world on its apps and all of the data that entails to ensure it stays competitively differentiated. It can also gate some of this data for its most advanced, proprietary Llama models to ensure they stay better than others if needed. At the end of the day, all of these models are complex algorithms that are only as good as the data they’re trained on. App quality also reflects this reality. Really what Meta wants to do here is use these models to create more apps to maximize engagement and create new ways for us to interact with agents and chatbots on its products. That will mean more monetization.

It also wants to use the models to create best-in-class AI hardware; I see no reason why LLMs being less expensive would do anything but help glasses development. To me, model cost reductions simply mean Meta can immediately do way more cool things with app building to accelerate these objectives.

And if models are truly becoming commodities (to be determined), what will matter? Cost advantages. How do you create those cost advantages? By open sourcing your products and inviting developers to do your optimization work for you. What do developers want to inspire them to build for you? Great models paired with more consumer traffic to give their work a better shot of being monetized through a platform they don’t hate building on. Who has that traffic and model combo? Meta. In the mind of Meta leadership, Deepseek isn’t a threat… it’s validation of open source being the correct approach. Maybe that’s why Meta reiterated its CapEx guide after all of this news broke. Full speed ahead.

Amazon:

For Amazon, if cost-to-train figures are remotely accurate, I think they will have overpaid for their Anthropic stake in hindsight. That will turn out to be a financial blip on the radar.

I think this is also mostly positive for the e-commerce king. While they are building their own foundational model and could potentially pocket some costs there, the main focus for AWS is creating best-in-class model choice. They just want Bedrock to be the place developers can go to use whatever model they want. They don’t care which model is chosen. If DeepSeek leads to other models becoming better and more efficient, AWS will offer those too. Alternatively, this could diminish the amount it could charge for renting GPU capacity, which is the biggest potential headwind here. At the same time, this headwind (if it manifests) will lead to lower CapEx to greatly offset profit obstacles. Finally, this will likely diminish workload pricing power for work on Anthropic and other models through AWS… but? If these GPU renting and workload pricing power challenges surface… it will mean a new tailwind for everything else AWS sells. It will mean far more data storing and processing demand, far more GenAI app-building that is hosted on AWS and far more usage of other supporting tools AWS provides such as SageMaker. It would also mean Amazon can update its own apps (that will hopefully now be way better than Alexa) with more utility. For example, it can do things like make its discovery agents far more powerful to improve conversion rates on its marketplace. That means real financial impact today.

Alphabet:

I also think Google will be entirely fine. They also overpaid for Anthropic and also could lose some GPU renting demand and pricing power (and so maybe shed some CapEx). But the full-stack AI approach here cannot be disrupted by one measly piece of that stack being cheaper to build. Google still has more search and video data than anyone else; it still has a fantastic team of AI researchers; it still is a leader in quantum computing and physical AI; it still has elite, scaled infrastructure; it still makes great machine learning chips; it still boasts a GenAI bundle ecosystem that is very difficult to disrupt. Vertical integration within Alphabet’s GenAI arm and its gigantic datasets make everything better together and should help to preserve GenAI workload pricing power I think better than AWS or Azure can. And again, any model-related workload pricing compression will support demand for most of the rest of its stack. When customers are using you for many more things… they quickly become dependent. That’s why Salesforce can pass on price hikes for average products with no pushback every year.

The full-stack fosters vendor consolidation and higher customer retention. That formula will be entirely fine regardless of where this DeepSeek news goes. And Gemini can take all of this DeepSeek work to use themselves just like Llama will.

For these three, simply put, the moats are fine and come from many facets of their operations. They have the assets to outcompete the field and the balance sheets to invest whenever that’s needed. Amazon and Google headwinds also feature fortunate offsets and all three will benefit from this in various ways.

Tesla:

I think Tesla will also be completely fine. It separates itself from its pack via building better cars more cheaply through unmatched vertical integration and software add-ons. Its work in AI is purpose-built for apps and physical AI applications like Optimus. Its supercomputer work, even if it overpaid for it, won’t suddenly become worthless if models are cheaper to train. Scaling these supercomputers will instead become much easier.

This work getting cheaper doesn’t mean companies can suddenly collect the vast data Tesla has scraped and build their own Cybercab or Optimus bot. It probably just means Tesla will either get more profitable or (more likely) realize its AV ambitions more quickly. Leading in EVs (and maybe AVs) doesn’t become worthless because tools used to augment current and future software offerings are cheaper. The hardware moat is real here.

Microsoft:

For Microsoft, while I don’t think this is an existential threat, I do think this is a bit more negative. They’ve made that company an instrumental, ingrained piece of their GenAI work. Furthermore, GenAI (Copilot) monetization has had a larger impact on Microsoft’s growth than any other enterprise software name I follow. Reviews for Copilot are not all that good, and willingness to pay for it could diminish. That would have a material impact on this business. At the same time, Microsoft has many competitive differentiators of its own outside of OpenAI models. It has more data than competition and a broader enterprise software bundle than them too. Azure would surely benefit from an uptick in data processing and GenAI app demand, and that would help offset some of the GPU renting and Copilot monetizing headwinds. Puts and takes are somewhat similar to Google and Amazon, but I think the app monetization risk is much larger here than it is for them. The other two are still working on building that revenue stream while Azure already has.

Apple:

Finally, on Apple, I don’t think this news is all that material. The company has mostly partnered in the realm of LLMs and can partner with others who build better models. They don’t rent GPU capacity to customers so won’t deal with that headwind.

Nvidia & AI Hardware:

I go back and forth here on what this means for Nvidia and see two potential outcomes.

First, if model training costs plummet and the world has more compute than it needs, that could sharply hurt forward demand for the company. If we need fewer GPUs to build the future, Nvidia suffers. No way around it. Secondly, and I think more probably, efficiency gains could simply pull forward model and app development by several years. It will allow current budgets to create far more advanced products than we otherwise would have in 2025. If models cost $10 million to build instead of $1 billion, that means we can build 100x larger and more complex models at the same cost. Again… There's a choice here for its customers. Either save a lot of money or build way better products with your existing budget. I think it is inevitable that some will see maintaining budgets as a way to leapfrog competition. It will only take one prominent player to make that choice… the rest will likely feel forced to follow to avoid being left behind. I think it would be a domino effect, with the overall impact to Nvidia being somewhat modest. Still negative… but not alarming.

Candidly, my confidence in implications for Nvidia and AI hardware names is lower than for the rest of Mag7 or enterprise software (next section). This is why quantifying DeepSeek’s impact today is a crapshoot. It’s a guessing game. Nobody has a clue how much of the potential efficiency gains will be pocketed and how much will be leveraged in development of more advanced products.

The risk for Nvidia and other hardware leaders like Broadcom is not that they suddenly won’t lead in their respective areas. It’s that the demand runway for this current hardware cycle was greatly shrunk from these advancements. Whether that risk manifests or not is to be determined.

Enterprise Software:

This is extremely positive for enterprise software names. They don’t deal with GPU renting headwinds and they have a choice too. They can rent model workloads and GPU capacity at lower levels to pocket savings… or they can build better products for their customers and give them the tools to build better products for themselves too.

Infrastructure monetization for all tech waves always comes before app monetization. One of the main bottlenecks for supporting app monetization was cost. If that bottleneck goes away, GenAI apps should explode. That’s great news for firms like Snowflake, Palantir, Salesforce and MongoDB, which all house and facilitate the building of these apps. More consumption of the platforms to use data and build apps means more revenue. It also points to the databases and software tools from these companies being more defensible niches than trying to build the best model.

It’s also fantastic news for Shopify, Servicenow, CrowdStrike and any other company trying to build and sell their own value-creating GenAI apps. All of them have AI products and basically none of them have turned those products into material revenue contributors (besides Palantir). There’s less risk associated with losing GenAI app pricing power for existing products than there is opportunity for building better apps that get these firms paid. This is great news for pretty much the entire sector.

6% impression growth missed estimates by 4 points. 11% Y/Y price per impression beat estimates by 8 points. Price strength was powered by improving ad performance.

Meta’s revenue estimate missed by 2.2%. This was related to a much larger than expected 3 point foreign exchange (FX) headwind. On the expense side of things, it reiterated the $62.5 billion CapEx number and set a $116.5 billion operating expense (OpEx) guide. Headcount growth will continue to be mainly via hiring for highly technical roles across its AI, advertising and reality labs work. It expects “strong growth” throughout the year.

Meta trades for 27x 2025 earnings estimates. EPS is expected to compound at a 13% clip for the next two years. I think estimates will be stable following this report. While the revenue miss likely brings down annual revenue targets a tad and the expense guidance is ahead, the decision to extend the useful life of some assets (more later) should add nearly 4 points to 2025 EPS growth. That should offset the other items here, along with the massive beat this quarter on a smaller revenue beat. Furthermore, while I think it’s possible that revenue numbers fall a bit for the year, I don’t think that will end up being correct. Meta loves to sandbag, is lapping leap year in Q1 and dealing with the large FX headwind.

2025 Vision:

Zuck was noticeably excited about 2025 on the call. To him, this is the year when “trajectories of long-term initiatives will get way more clear” and financial applications become more apparent. 

2025 Vision – Meta AI:

He’s exceedingly confident in a highly intelligent AI assistant reaching 1 billion people in 2025, and equally confident in that product being Meta AI, which already has over 700 million users vs. 500 million Q/Q. Updates to Meta AI this year will include query memory, so the product can pull from previous conversations and a complete set of customer app interaction history to sharpen outputs.

As he explained, the wildly valuable data that coincides with this scale leads to better post-training, better inference and better results. It’s a true network effect that gets harder to compete with as it grows, just like Meta’s Family of Apps (FOA). He sees Meta AI getting a lot more “personal.” As he has said time and time again, there will not be one massive AI model to rule them all, but more granular, purpose-built models that reflect the diverse needs, cultures and personalities of various users.

2025 Vision – Llama 4:

On Llama 4, there are no changes to his bullishness stemming from the DeepSeek news (much more later). Llama 3 was built to create product parity with elite closed-source models like OpenAI’s and Anthropic’s. And now? The goal for Llama 4 is to be better than any closed-source model one can find. Between Meta’s gigantic user base, datasets and distribution, as well as its world-class research teams and already deeply popular models, I think that’s very possible. We have to remember that Meta open sourcing this work means it’s not just up to it to drive improvement. Developers across the globe are eager to build for Llama and Meta’s consumer base, and a lot of that work goes towards optimizing cost and accessibility of Llama. This is the beauty of open source and the true power of combining elite internal teams and an innate incentive for elite 3rd party developers to work with you. Llama 4 will unlock significantly more complex agentic (or goal oriented) use cases vs. the predecessor. One of these will be Meta internally creating an AI engineering agent that will emulate the skills of mid-level developers. He called this a “profound milestone.”

“We expect that the continuous advancements in Llama's coding capabilities will provide even greater leverage to our engineers and we are focused on expanding its capabilities to not only assist our engineers in writing and reviewing our code, but to also begin generating code changes, automating tool updates and improving the quality of our code base.”

CFO Susan Li

2025 Vision – Ray-Ban:

As leadership reminded us, history points to 5 to 10 million units sold for 3rd generation hardware being a key benchmark to clear. This is when scaling to consumer ubiquity becomes a lot more likely, as early traction serves as a sort of viral domino effect to spread the word. This will be the year where it becomes clear if Ray-Ban is rapidly heading for that ubiquity, or if it will be a “longer grind.” My money is on this going well. These units have been flying off the shelves since the introduction, and consumers clearly don’t care about a full holographic AR display to want to own these things. The Meta AI tools being infused into it seem to be more than enough.

2025 Vision – Investing Philosophy:

The $62.5 billion CapEx figure includes bringing almost a full gigawatt of compute online and a truly gigantic data center. These things are expensive. But? Meta is determined to fund these aggressive investments while using the work to advance revenue growth. This is the first time we heard about AI investments actually moving the overall financial needle – beyond improving ad conversion rates by a few points here and there. This is becoming more material… and quickly.

App Updates:

Threads crossed 320 million monthly active users (MAUs) vs. 275 million Q/Q and is now adding 1 million new users per day. As previously noted, ads are being introduced to the app, but that ramp will be slow and will take time to become a large driver of top line growth. At the same time, all of this revenue should come with sky-high margins, as this is simply Meta monetizing existing traffic that it has already built. Meta plans to introduce several updates to its recommendation and content prioritization systems to keep improving engagement levels on that newer product. It also continues to work on custom feeds for even more personalization.

WhatsApp keeps gaining share and has well over 100 million MAUs in the USA.

On Facebook, Zuck wants it to become more of a cultural centerpiece for its users. The plan is to keep updating the video offering there, with a full screen product and revamped recommendation algorithms. He also wants to “get back to OG Facebook” but wanted to keep a lid on what this means for now. In terms of changes to content moderation policies on Facebook, there has been no material impact on advertiser demand or engagement. Those engagement levels remain quite strong, and that has a lot to do with improvements to the video offering there. Time spent on that form of content rose by more than 10% Y/Y and it sees this momentum continue to roll.

Creator Support & Tools:

Meta continues to prioritize original content on its apps to help “the little guys.” It also added the ability for them to share their reels with non-followers first, in order to test content before sharing it with your supporters. In other related news, it’s launching an app called “Edits,” which will offer end-to-end creative tools for mobile reel creations. Speaking of Reels, 4.5 billion are shared daily and introductions such as curating feeds based on what’s being liked by who you follow should help more.

Present Monetization Work:

Meta is hard at work on optimizing ad load and supply across its apps. It has been working on this for decades… and hopefully will be working on it for many more decades. There’s always a split test to learn from and a tiny, subtle tweak to make to squeeze a little more engagement out of its products. That work will not slow down. This year, it’s focused on improving when and where ads show up on a person's feed. Per CFO Susan Li, this is unlocking more efficient impression growth.

From a marketing performance point of view, relevancy and ad ranking continue to be top of mind. Last year, Nvidia and Meta partnered to create “Andromeda.” This 10,000Xed Meta’s ability to conduct ad retrieval, which is another way of saying it vastly improved its ability to screen, sort and match ads with consumers more granularly. More personalization means better targeting, better conversion rates and higher willingness to pay from buyers. And? This change has already improved ad quality on the apps by a full 8%.

Its “Advantage Plus” campaign builder also continues to thrive. Annualized revenue crossed $20 billion with 70% Y/Y growth. It’s now iterating on this campaign tool to build a campaign service that lets advertisers more selectively choose between Advantage Plus or manual campaigns. All performance campaigns will “have Advantage Plus turned on from the beginning, with more manual controls to use for more advertiser customization. Now, campaigns can freely move between Advantage + automation and manual curation. This is going well in beta testing and will roll out to more advertisers in 2025.

Within the intersection of GenAI and advertising, 4 million advertisers are using one or more GenAI tool vs. 1 million as of Q2 2204. Its video generation tool, called Image Animation, also now already has 100,000s of customers.

GenAI CapEx:

Meta plans to extend the useful life of its AI and non-AI servers to reduce depreciation expense and improve profitability. It thinks it can use these assets for longer than it previously thought.

It’s also looking to use its custom chips (MTIA) in areas where this can drive down compute costs. MTIA can let Meta more flexibly allocate capacity to “a different mix of memory vs. network or bandwidth vs. compute to make sure these chips are perfected and optimized for its own workloads. It will ramp usage of these chips throughout 2025 to more core inference, training and ranking workloads. It thinks it can get to a point of these chips replacing some GPU needs and supporting GenAI use cases (core AI is current AI, GenAI is more speculative and future AI).

DeepSeek:

I’d recommend reading the Meta piece in section 2 of this article for my views on how DeepSeek impacts Meta (and what DeepSeek is). Per Meta, there are a “number of novel things DeepSeek did well that they’re digesting” to use in their own work. Again, DeepSeek is open-sourced and built partially with Llama; Llama can use all of this for itself. As Zuck put it, “every new firm with a launch will have some new advances to learn from. That’s how the tech industry goes.” DeepSeek changes nothing about Meta’s current CapEx plans and, if anything, “the recent news has strengthened conviction that we are focused on the right things.” Right things meaning models and open sourcing those models. Meta can rely on its traffic and distribution for competitive differentiation while pushing to be the open source standard for models. Being that standard also comes with a plethora of advantages like developer platforms standardizing to build on Llama. That is the goal. And if it happens… it means lower optimization costs vs. the field. Per Zuck, this is part of why Meta thinks Llama will leap frog closed source competition in 2025.

In terms of what this does to multi-year CapEx assumptions, which is highly relevant to hardware players, Zuck doesn’t know what implications are just yet. Before DeepSeek, the industry was already heading towards shifting expenses from pre-training to post-training, inference and reasoning models within the realm of agentic, goal-oriented AI. This probably just sped that process up a bit, but “doesn’t mean companies will need less compute.” Regardless of where in the model and app building process that compute is deployed, he sees a large competitive advantage to building out this infrastructure, as more compute still means better products and higher quality service. And if that compute is cheaper? Products will be even better. Building out this footprint is a “strategic advantage.” If that changes, Meta will reallocate these CapEx dollars to other high-return areas.

“We are not sure where we are in the CapEx cycle.”

CEO Mark Zuckerberg

Great quarter. The revenue miss for Q1 is because of FX and everything else here looks fantastic. Threads keeps chugging along while WhatsApp monetization has begun and optimization work across Facebook and Instagram has miles left. The futuristic AI work is becoming more tangible and real by the moment, and 2025 should mark the year when these hefty investments begin to yield preliminary fruit. The company’s mix of an unmatched network effect, distribution, great models and a fortress balance sheet make this one of the easiest names for me to own. The recipe of scale, growth, margin, liquidity, optionality and positioning here is second to none. Meta is second to none. Great quarter.

Detailed coverage of the 2024 We Robot event can be found in section 4 of this article.

“During Q4, our focus was to reduce inventory levels in the automotive business. And we accomplished that by ending the quarter with the lowest finished goods inventory in the last 2 years.”

CFO Vaibhav Taneja

Tesla guided to auto growth resuming in 2025 and 50% energy storage deployment growth. Musk did not reiterate last quarter’s guidance calling for 20%-30% car vehicle growth for 2025. He could’ve simply forgotten, as his prepared remarks are usually mostly unscripted and he went on a tangent about future products. I really wanted someone to ask them about this in the Q&A, but no luck. 

New vehicle plans are on track. As a reminder, these cars serve as a transition between current models and its next-gen vehicle and manufacturing processes. They borrow some things from old lineups and some things from new approaches. They will use existing manufacturing capacity, which is informing confidence in a 60% production growth rate in 2025 (vs. -4% growth in 2024). Cybercab is on pace to begin volume production, with its entirely new manufacturing techniques, next year. 

Heading into this report, EPS was expected to rise by 44% this year and by 24% next year. FCF was expected to rise by 124% this year and by 26% next year. Estimates will fall following these results.

Automotive Puts & Takes:

Energy storage and services, vehicle delivery growth and regulatory credits were the demand tailwinds this quarter. ASPs across all vehicles, partially related to flexible financing, was the main headwind. This headwind will surely keep an important Tesla bear-bull debate going. Are consistent ASP declines because competition is catching up. Or… is this Tesla reacting to a higher cost of capital environment, leveraging its best-in-class electric vehicle gross margin, and absorbing some of the pain to more quickly grow market share? It does constantly tell us that it’s willing to accept lower up-front costs and lean on expected software upsells to harvest more margin down the road. At the same time, market share trends do not look amazing despite Model Y being the best-selling car on the planet in 2024. I go back and forth on my point of view here. I’ve never driven an electric car that resembled the quality of a Tesla. It’s simply better. But? As products become more and more similar and the large gap becomes a bit smaller, pricing power can be affected.

From an EBIT point of view, Tesla continues to reduce input cost per vehicle, which is helping with margin preservation. Average cost fell below $35,000 per car for the first time ever. While this hasn’t translated into improving gross margin just yet, this positions them to see a very nice margin pop if the exogenous backdrop improves and it executes. It can easily cut financing options as consumers start to feel more confident… and these input cost gains should be durable. More investments in AI R&D were also a large margin headwind this quarter.

Energy Generation & Storage:

The energy generation and storage business is rocking and rolling. This product is also interest rate sensitive, so it’s interesting to see this continue to thrive while its auto momentum cools off. That could just be a byproduct of being earlier on in the growth curve and market share trajectory here. Construction of its Shanghai factory is now complete and megapack (battery storage) production should now quickly ramp there. Within Powerwall (rechargeable home battery), Tesla also enjoyed a new quarterly volume record while version 3 of the product is now scaling before inevitable international expansion.

The growth here should remain solid, based on guidance and still being supply constrained across megapack and powerwall. It sells everything it can make and is actively boosting production capacity in Shanghai and now a 3rd planned factory. Batteries are the current production bottleneck and it’s hard at work on addressing that. This growth should also continue to come with increasingly compelling margins, as input costs keep falling more meaningfully than selling prices do. Interestingly, the team spoke about the current energy grid having very sparse and ineffective energy storage. This inherently makes it more inefficient, and Tesla can greatly help with these products. That should be an extremely durable structural tailwind.

AI & Full Self Driving (FSD):

Tesla finished construction of Cortex, which is a 50,000 H100 chip compute cluster in Texas. This is helping train the new FSD software and comes with a 4.2x boost to data capacity, better video inputs and more key performance indicator gains vs. Tesla’s previous capabilities.

FSD rolled out new parking and unparking features, while the Autopilot bridge technology continues to work nicely. Specifically, users of this drove nearly 6 million miles per accident vs. 700,000 for non-users. This is a material piece of validation and proof of concept as Tesla continues to convince regulators that its driverless technology is safer than people. It aims to launch unsupervised FSD and its Robotaxi service by the end of the year in Austin, Texas, as well as maybe California and other states. To start, Tesla will run its own fleets and go very slowly to make sure everything is working. It expects to begin allowing customers to add their own cars to these fleets in 2026. There was no mention of partnerships here, although I do still think that’s inevitable. International expansion should come thereafter, although the team spoke frustratedly about slow EU regulatory progress and not being allowed to do any AI training in China.

“We made many critical investments in 2024 in manufacturing AI and robotics that will bear immense fruit in the future. Immense. To such a scale that it is difficult to comprehend. I've said this before and I'll stand by it.”

An Always Charismatic & Flashy Elon Musk

Optimus:

Optimus is ahead of schedule for pilot production this year. Tesla wants to build 10,000 of them this year, but will settle for a few thousand. It expects these to be doing useful, tedious tasks in Tesla factories by the end of this year. It doesn't plan to sell these externally until version 2, when production lines are expected to be 10x in scale. He admitted this was a very rough estimate, but he did tell us he thinks they’ll be shipping robots next year. It’s very hard to forecast timelines when you’re building brand new technology, and Tesla basically had to build every single component of this by itself. The cost to produce one of these should also be about 42% lower than its average car. And with the massive potential utility of these units, selling price should be very high and so should the margins.

Car & Factory Production Schedules:

International Updates:

4680 Cell Batteries:

As a reminder, 4680 uses dry cathode technology. When used, this forgoes moist cell coatings to “direct the electrode material to the battery without the need for liquid.” In practice, that means higher density, faster charging, lower cost and lower environmental footprint. In-house production now equates to more than 2,500 Cybertruck batteries per week. It also processed its first lithium byproduct in its new refinery. “The intermediate material was on-spec” and it expects to ramp production through this year.

DeepSeek Implications:

While DeepSeek didn’t explicitly come up in the call, Elon did discuss the prospect of model training costs plummeting. Lower cost to train has been playing out for months now, and that trend is fully expected to continue. In his mind, this doesn’t diminish the benefits of having more compute, it just allows Tesla to more rapidly leap forward in its app and physical AI development to do more with the same amount of compute. He’s excited about this trend, as Tesla does not rent out GPUs and does not rely on model monetization to power its results. Tesla is firmly entrenched in the application layer of AI, and that layer benefits greatly from training costs tanking.

Tariffs:

“Over the years, we've tried to localize our supply chain in every market, but we are still very reliant on parts from across the world for all our businesses. Therefore, the imposition of tariffs, which is very likely, and any of the super capacity will have an impact on our business and profitability. “

CFO Vaibhav Taneja

Results were not good outside of the thriving energy business. Growth has stagnated and margins are challenged. Market share gains have flattened or reversed. The interesting thing here is that nobody seems to care about current financials. They’re willing to give a lofty multiple for this performance because they’re banking on FSD and Optimus carrying the financial burden on their backs down the road. Is that a good decision? Time will tell.

This is why Tesla remains firmly in my too hard pile. If they’re right about product timing and demand, this stock should do extremely well for a long time. If they’re not right? I think you know what will happen. It’s extremely hard to value this company based on business that should flow in over the near term. At the same time, Tesla remains the best company in its space with fantastic products and has a CEO who has supervised more innovation than any other in our lifetimes. I certainly wouldn’t bet against him. I just also wouldn’t bet on him.