In case you missed it, our detailed PayPal Earnings Review from this afternoon.

This was a very long 16-hour work day. If there’s an extra typo or two in the reviews, I apologize. Struggling to keep my eyes open and Uber + Disney both report tomorrow morning.

A full review of this report will come on Saturday or maybe early next week.

Results:

Balance Sheet:

Guidance & Valuation:

For the full year, 9% store growth and about 3% comparable sales growth guidance for 2025 leave us with about 12% revenue growth guidance for the year. This missed 13.3% growth estimates. If we instead assume “low-to-mid single-digit” growth means 4%, then it only slightly missed estimates.

Chipotle trades for 42× 2025 EPS estimates. EPS is expected to compound at a 19% clip for the next two years. Estimate revisions will likely fall a bit after this report.

A full review of this report will come on Saturday.

Results:

Balance Sheet:

Guidance & Valuation:

Spotify trades for 56x 2025 earnings estimates and likely somewhere at or a little over 50x after revisions. It also trades for about 37x 2025 FCF estimates and likely a few ticks lower after revisions. EPS is currently expected to grow by 100% Y/Y this year and by 8% Y/Y next year. FCF is expected to grow by 45% Y/Y this year and by 2.5% Y/Y next year.

EBIT and EPS both rose by 31% Y/Y as OpEx fell by 1%. This was helped by headcount and cost discipline, and also the absence of a $1.2 billion real estate impairment charge that was in last year’s results. Excluding this help, OpEx would have risen by 5% Y/Y. EBIT and EPS would have risen by 25% instead of 31%.

As is typically the case, the company did not offer much guidance. It called out leap year and ramping FX headwinds as growth headwinds. It also told us to expect $75 billion in 2025 CapEx, representing about 42% Y/Y growth. Most of this will be for short-lived assets like servers to meet more near-term cloud and AI demand. Full speed ahead on spending. It sees no issues with matching all of its 2025 CapEx with demand. More on this confidence later.

As a result of recent CapEx, it also expects depreciation expense growth to accelerate from 28% in 2024.

The company trades for 21x 2025 GAAP EPS (chart below has not yet been updated) and 27x 2025 FCF. EPS is expected to compound at a 14% clip for the next two years. FCF is expected to compound at a 17% clip for the next two years. Estimates should be rather stable following this report.

Full Stack Approach to GenAI and DeepSeek:

The majority of Sundar’s prepared remarks were dedicated to updating investors on the company’s full-stack AI progress. As a reminder, it’s this full-stack approach that it believes will allow it to differentiate vs. the field and create great products. It builds and designs its own custom data centers, offers cutting-edge Tensor Processing Units (TPUs) for machine learning workloads, boasts a world-class team of researchers driving rapid model innovation, and? Has 7 products with more than 2 billion users to offer a humongous and diverse data set to train all of its products. Owning more of the hardware, software and distribution means it can optimize every single component of the tech stack to drive more cost savings and incremental efficiency… from models, to hardware, to apps to databases. 

This powerful vertical integration means its data centers are more efficient than the competition on a performance-per-watt basis (per the team). It also makes the firm rather positive on the DeepSeek model training efficiency news. It wants models to get cheaper and invites compute shifting away from pre-training due to this welcomed cost disinflation. Why? Because it means the mega-cap can offer far better tools and far more compelling apps with the same budget and same compute. It also means more budget will shift to inference use cases, which is what its TPUs were purpose-built for and where their performance and efficiency shine brightest.

Cost disinflation has been the goal and the plan all along for this company. Pieces of its full-stack approach getting cheaper mean other pieces of it can get better. It can lean on unparalleled distribution, apps, budget and talent, along with competitive models to keep winning for a long time. And? Leadership was quick to remind us that it’s confident in its ability to compete in model efficiency:

“Gemini models shine in the Pareto frontier of cost performance and latency. And if you look at all three attributes, I think we are leading this frontier. And I would say both our 2.0 flash models or 2.0 flash thinking models are some of the most efficient models out there, including compared to DeepSeek's V3 and R1.”

CEO Sundar Pichai

More on AI Models:

Gemini 2.0 is the company’s foundational model series for the “Agentic Era.” Ranking from largest to smallest, this comes in Ulta, Pro and Flash versions. There’s also a Gemini 2.0 Flash thinking model, which works users through its contemplation and response processes to sharpen model reasoning. This is another reason why model training efficiency is so compelling for Alphabet, as it unlocks an ability to create better reasoning and Agentic models. Reviews early on for the thinking model have been “extremely positive” and it’s leaning into more work here.

Gemini Advanced is the company's paid subscription that comes with Google Deep Research. This functions as an “AI research assistant” to build and automate your research. This is a new example of this firm’s push into Agentic AI, as it intricately plans and delivers complex, multi-step, goal-oriented tasks. Project Astra will build on this Agentic progress as another “leap forward” for Agentic AI assistants. Much more on this throughout 2025. Project Mariner will help further as a “tool that can understand and reason across information on a browser to complete tasks.”

Finally, traffic and usage of its video generation model called Veo 2 was called strong. 

Whether it’s Gemini 2, Deep Research specifically or Veo 2, Alphabet is enjoying rapid developer adoption. Specifically, its base of Gemini developers has doubled over the last 6 months to 4.4 million. Why is it enjoying such strong momentum? It will tell you this is because its models “top industry leaderboards across industry benchmarks.” Pairing that with unmatched distribution and customer traffic is understandably enticing.

Search:

AI overviews continue to deliver incremental engagement for the company’s search business and drive more usage growth for users vs. non-users. Users give this format higher favorability ratings than traditional search too. Additionally, the product continues to monetize at a similar rate vs. its established advertising products – and there’s still a long runway for introducing more ad load. All of this progress should be augmented by the product integrating Gemini 2.0 later this year. Circle-to-search has proven quite popular with younger generations. Impressively, this interactive search tool, which turns your screen into a query interface, is already more than 10% of total search volume for users so far. Lens is its search tool to let consumers use their camera to initiate a search. This is already used for 20 billion searches per month, most of which are purely incremental to overall demand.

There has been significant concern about both cannibalization and margin associated with AI-powered search. Along with outperforming search results overall, the monetization parity reiteration is encouraging, and the continued language surrounding these use cases being incremental is too. And to me, this opportunity-expanding thesis makes sense. GenAI is unlocking entirely new query formats and extending the power of what search can actually provide. As it becomes more diverse and valuable, it makes sense that the range of popular use cases would too. Traffic always follows compelling value… and ad demand always follows traffic.

More on Cloud:

Cloud continues to turn into yet another elite business for this company. Its GenAI services netted customer wins with Mercedes Benz, Mercado Libre and Servier; its first time commitments in 2024 doubled vs. 2023. Deals over $250 million in value also doubled Y/Y. The firm’s “leading performance and cost” dynamics for GPU and TPU-related workloads are helping it stick out and land large clients such as Citadel and Wayfair. For Wayfair, it has already improved performance of that enterprise’s operations by 25%. Specifically, Trillium, its 6th generation TPU, “enjoyed strong uptake” thanks to 3x inference throughput gains vs. the 5th generation.

Its AI cloud developer platform, Vertex AI, enjoyed a 5x increase in customers Y/Y and a 20x increase in usage, as the new product enjoys more clients and sharply rising engagement. This product allows developers to create and ship highly configurable and powerful models and apps, with a roster of more than 200 foundational models to choose from. This directly competes with Azure’s Copilot Studio and AWS’s SageMaker. This quarter, it also added Agent Space to extend Vertex AI’s building support to Agentic AI applications, with deep data integrations to ensure these custom tools have the context they need to actually provide value.

Because this was the source of this miss, I think it’s important to call out the timing of CapEx order fulfillment. Cloud is still quite capacity constrained, so revenue is a byproduct of how much capacity it can secure in a given quarter. The miss was not a matter of softening demand… it was a matter of still playing catch-up to meet excellent demand. Supply should catch up to demand at some point towards the end of this year. This is also why the hefty CapEx number doesn’t bother me in the least. Again, most of it is earmarked for short-lived assets, meaning near-term revenue and profit generation.

YouTube:

YouTube remains the #1 streamer by market share, per Nielsen, with a new all-time high set this quarter. The platform enjoyed rapid adoption in newer content formats like podcasting. Impressively, according to Edison, YouTube is “now the most popular service for podcast listening in the USA.”

Its YouTube shopping affiliate program, which lets popular influencers match with brands to sell product, now has 250,000 creators in it. It’s expanding to 3 additional countries. This quarter, it also added the ability for ad buyers to promote YouTube videos as part of their advertising campaigns.

Demand Gen is a tool that lets advertisers extend audiences and sharpen targeting to uplift the number of relevant eyeballs a campaign can affordably reach. This quarter, Sephora used a YouTube Shorts-only campaign to deliver a whopping 82% lift in search volume during the holidays. Speaking of YouTube Shorts, the monetization gap many were concerned about between this and the rest of its ad formats is quickly closing. Per Chief Business Officer Phillipp Shindler, the monetization rate of “Shorts relative to in-stream viewing increased by more than 30% in the USA.” It expects more progress this year. This content format is also quickly proliferating within its Connected TV (CTV) branch, as 15% of all shorts are now watching via YouTube’s CTV offering. Louis Vuitton tested a campaign with both traditional ad placements within YouTube and Shorts. For both, performance significantly exceeded industry benchmarks.

For YouTube ad campaigns overall, Petco used demand generation for targeting, creative generation and bidding across YouTube to generate a 275% boost to return on ad spend and a 74% boost to click-through rate. Simply put, AI continues to make the firm’s campaign-building and ad-targeting tools better. This is where that shows.

More on Advertising:

The company added easier reporting tools for its Performance Max (PMax) campaign builder. This should mean more informed buyers with a better grasp on which channels are working best and where to lean in. For Event Tickets Center, this helped it 5x creative asset generation and raised conversions by 4x. Most recently, its “marketing mix model” called Meridian broadly debuted. Marketing mix models dissect every piece of the customer interaction chain to observe what is working, what precise impacts are and where to focus. This helps the mega-cap become much more exact in its ability to connect ad dollars spent to revenue generated.

More:

I don’t care what the stock is doing after hours. Good quarter. One may pick on a small revenue miss in the cloud segment, but I think that’s missing the forest for the trees and misinterpreting the softness. The revenue miss is related to capacity constraints in the cloud business, not demand. And? The all-important search business beat estimates yet again

You can also pick on the $75B CapEx number, but that budget is to meet more near-term demand for its cloud business, and is money well spent in my mind.

What’s important? The Search King’s unmatched ability to deliver full-stack AI at compelling cost and broad utility. And? Its ability to scrape data from 7 products with 2 billion users to train its products better than anyone else can. It’s that the firm has fended off countless new entrants in its bread-and-butter search business while continuing to deliver stellar results there. It’s that Cloud, YouTube, Android and Maps are all elite businesses and Waymo eventually will be too.

Financial trends remain excellent and leadership continues to explicitly talk about much more cost bloat to trim. People can feel free to fret over a tiny revenue miss… I’ll focus on all of the other overwhelmingly positive things in this report. That’s the forest… and this forest is full of mighty red oaks with decades upon decades of thriving left to do.

If there’s one thing the semiconductor industry loves, it’s constantly changing the names of products with a swarm of acronyms for us to juggle. Fun, fun. Those acronyms all fall into neat categories: chips, networking and connectivity, and software. It’s these ideas and AMD’s positioning within them that matter to investors. Not that they’ve memorized what an MI325 HBM3E chip stands for. That’s how we’ll frame this coverage, with an emphasis on data center results.

We got a lot more commentary on guidance for 2025 throughout the call. The company expects double-digit revenue growth in 2025, which compares to 25% estimates. When companies give guidance like this, it leads analysts to lean negative. They’ll think “double-digit” means “close to 10% growth” when that may not actually be the case. For this reason, I would have loved some more detail here, or for AMD to forgo any annual commentary altogether.

The company expects data center revenue to be flat during the 1st half of 2025 vs. the 2nd half of 2024 for both CPUs and GPUs. This is below the wide range of consensus estimates that I see and led some to worry that the runway for GPU growth was already beginning to slow for AMD. As leadership explained, it thinks this is simply a lull in demand as customers wait for its new MI350X chip and promised performance gains (more later). While that may be true, this is not something Nvidia has dealt with during its own GPU ramp. Furthermore, banking on faster 2nd half of year growth to meet annual guidance will always make people a bit nervous.

All of this should lead to modest downward revenue revisions and slightly sharper downward profit revisions, considering data center is its highest-margin product. 

“We've talked about a data center accelerator TAM being upwards of $500B by the time we get out to 2028. I think all of the recent data points would suggest that there is strong demand out there. Without guiding for a specific number in 2025, one of the comments that we made is we see this business growing to tens of billions as we go through the next couple of years. And that gives you a view of the confidence that we have in the business.”

CEO Lisa Su

Data Center – General Compute CPUs & EPYC:

EPYC is its lineup of data center CPUs for general compute. While GPUs are leaned on for complex inference, not all workloads here need this level of sophistication. CPUs work well for static, step-series instructions and tasks. They’re good building blocks to be accelerated by high-performance GPUs. 

Demand for EPYC this quarter was quite strong and broad-based, with cloud and on-premise deployments both healthy. It sees 2024 as a “major inflection point” for its CPU business as “market share gains accelerated” thanks to a successful launch of its 5th generation EPYC processors (called Turin). Its 4th generation EPYC product still enjoyed strong double-digit Y/Y growth as well, showing these products may have a bit more staying power than some think. For EPYC Cloud deployments, it crossed 50% market share with “the majority of its largest hyperscale customers.” These hyperscalers are leaning on AMD to provide needed efficient compute, power their internal infrastructure and enable scaled deployment of their lucrative apps. Cloud service providers (CSPs) more generally speaking also delivered strong EPYC adoption, with cloud instances (virtual cloud-based machines that run on AMD’s EPYC hardware) rising 27% Y/Y. AWS, Alibaba, Microsoft and many others launched 100+ EPYC instances in Q4. Enterprise clients overall doubled the number of AMD cloud instances activated for usage Y/Y too.

The EPYC on-premise business is also finding double-digit growth and secured wins with Verizon, Visa, Akamai, ServiceNow and LG during the quarter.

While everyone loves to focus on GPU demand (for very good reason), that does not mean CPU demand will vanish – as these results and 200% EPYC cloud growth for 2024 overall clearly depict. Growth might slow considerably as these products aren’t a good fit for running GenAI and Agentic AI workloads (too much data processing and too expensive for CPUs), but structural decay for this space is far from imminent.

“Turin is clearly the best server processor in the world… we see clear growth opportunities in 2025 across both cloud and enterprise based on our full portfolio of EPYC processors.”

CEO Lisa Su

Data Center – Accelerated Compute & AI:

Instinct MI300X GPU deployments “expanded” materially during the quarter. Meta is using these GPUs “exclusively” run one of its Llama models for Meta AI (the 405 billion parameter Llama 3 model) and also added MI300X compute to its Grand Teton AI Platform. Meta will use these within Grand Teton for “deep learning recommendation models and large-scale inferencing workloads.”  Microsoft is using MI300X to power Co-Pilot apps based on GPT4 and more cloud instances. IBM, DigitalOcean and Vulture all are deploying these chips too, with IBM using them for WatsonX and GenAI applications. Per Lisa Su, overall, Instinct platforms (cloud servers that use AMD hardware as building blocks) are running for more than 12 cloud service providers, and it expects “growth in 2025. A little more detail here would have been nice.

The newest MI325X chip began ramping production in Q4 with a few early contracts secured. This is delivering large total cost of ownership advantages over competition, but it’s unclear if that means Nvidia’s Hopper 100, Hopper 200 or the new Blackwell platform. I would think Hopper 100 or 200, considering Nvidia continues to dominate the GenAI GPU market and has enjoyed the lion’s share of demand from this multi-year opportunity thus far. Regardless, several AI Lighthouse participants (its program for AI collaboration) have begun using MI325X chips to run their training and inference workloads.

Looking ahead, AMD is sticking to its annual cadence of product launches to match Nvidia’s rapid innovation and avoid falling too far behind. The next iteration in the MI family (MI350 series) will run on its new cDNA4 hardware architecture/backbone. This will deliver the “biggest generational leap in AI performance in history, with a 35x increase in AI compute performance vs. cDNA3. Per Su, robust preliminary interest will lead them to pull forward samples for some clients, with deployments beginning at scale earlier than previously expected this year.

AI Software:

AI Software is a vital complement to AMD’s Instinct software. This is how it can create more “out-of-the-box” experiences for its hardware, by offering slick software tools and broad integrations to make sure companies have more of what they need in one place. It diminishes friction associated with using these chips and allows AMD to conduct consistent performance updates to 2.7x MI300X inferencing performance since its launch. It now has 1,000,000 models offered through Hugging Face that run out-of-the-box with AMD.

It just launched ROCm 6.3 as the latest iteration with more upgrades to drive more Instinct ease-of-use. The ramp of this product was called successful with “numerous customers including its lead hyperscaler partners.” This is how it plans to be the “open AI stack.” Nvidia has done marvelously well with software products like Cuda, and it has created some vendor lock in the process. This is AMD’s response to drive better interoperability for the whole ecosystem. Starting last month, it started a twice-per-month software release cadence to drive more regular enhancements.

Two Threats to the Growth Runway

Application Specific Integrated Circuit (ASICs) are custom AI accelerator chips for specific use cases. Some believe this will diminish the demand ceiling for GPUs. Lisa Su isn’t concerned.

“I have always been a believer in needing the right compute for the right workload. With AI, given the diversity of workloads, when you're talking about broad foundational models or very specific models, you're going to need all types of compute. And that includes CPUs, GPUs, ASICs etc. Relative to our $500B+ TAM, we've always had ASICs as a piece of that. But my belief is given how much change there is still going on in AI algorithms, ASICs will still be the smaller part of that TAM. GPUs will enable significant programmability and adjustments to all of these algorithm changes.”

CEO Lisa Su

On DeepSeek driving model training disinflation and if AMD views that positively or negatively in terms of demand:

“DeepSeek innovation on the models and the algorithms is good for AI adoption. The fact that there are new ways to bring about training and inference capabilities with less infrastructure is actually is a good thing… it allows more adoption.”

CEO Lisa Su

Other Segments & Notes:

For the Client segment, its Ryzen processors (mainly CPUs for gaming and PCs) will be used to power Dell’s new Pro notebook and desktop PCs. It expects 150 Ryzen AI platforms running on Ryzen AI processors to be available this year. AMD took share in this segment for the 4th straight quarter.

Embedded segment’s recovery has been “slower than expected.” Industrial and communications clients are lagging behind aerospace and defense. Vodafone and AMD announced a new partnership for “higher capacity AI and digital services.” It thinks it’s taking market share here; the market is just quite weak at the moment.

Its purchase of ZT Systems secured approval from Japan, Singapore and Taiwan. It expects to seamlessly divest ZT’s manufacturing business when the deal closes during the first half of this year.

AMD will combine its client and gaming revenue segments in 2025 to “align with how it manages the business.”

The quarter really was not that bad. The issue is that the weakness is within its most important, highest-growth, longest-runway GPU business. That issue is amplified by the Q4 miss coinciding with Q1 2025 and Q2 2025 misses as well. AMD remains one of the cheapest names that I cover based on forward estimates, but the risk to those estimates being accurate is perhaps larger than any other that I cover too. This leads us to a compelling bear/bull debate. Does Nvidia simply have too large of a lead for AMD to capture more aggressive GPU-related momentum and market share? Or… is MI350X a sleeping giant with a 2025 deployment helping close the performance gap between that competition? Will that secure forward estimates and lead to upward revisions? Will MI350X mark an inflection point in the GPU narrative, like AMD’s leadership seems to think? If that’s the case, there is probably quite a bit to like here.

I’m simply not able to confidently answer these questions. So? I’m not able to confidently invest in what is admittedly a compelling opportunity from a valuation point of view. Lisa Su is still a great CEO; AMD’s CPU success over the last decade has still been admirable. But its positioning in this new technological wave continues to be uncertain. “Too hard” pile for me.