I published an update on my portfolio and performance vs. the S&P 500 earlier in the week.

Welcome back to earnings season. I’ll have the preview article out before my first holding reports and 40+ reviews sent to you all over the coming weeks. Can’t wait. For now, a more abbreviated earnings snapshot on an iconic bank:

The economic fallout from the current trade war was not felt during the Q1 period. Results for this recent quarter still come from a relatively tranquil and robust economic environment. Q2 is when weakness would potentially begin to show up, which is why guidance commentary is so important.

The company beat revenue estimates by 2.7%

In terms of important credit lingo:

Some brief thoughts on the data below:

Granular NCO, Provision & Allowance Context:

Guidance:

JP Morgan reiterated expectations for $90B in net interest income excluding the volatile global markets business. Its expectations for global markets performance actually improved a bit Q/Q. It also reiterated card NCO rate expectations of 3.6% for 2025. The theme of this call was “things are still pretty good and going well but we are highly mindful of the volatile macro backdrop and approaching it cautiously.” They’re building cash reserves and boosting their CET1 ratio. Still, that caution wasn’t enough to cut guidance, which I found encouraging.

Valuation:

Based on current forward estimates (which will be based on unknown geopolitical outcomes and are inherently uncertain today) JPM trades for about 12x forward earnings with the quarterly beat. As of right now, EPS is expected to fall by 8% Y/Y and rise by 7% Y/Y next year. For more context, the company’s EPS compounded at a 28% clip during 2023 and 2024.

Important Quotes:

“We continue to believe it is prudent to maintain excess capital and ample liquidity in this environment… We have an extraordinary amount of liquidity, with $1.5 trillion of cash and marketable securities. The economy is facing considerable turbulence, with the potential positives of tax reform and deregulation and the potential negatives of trade wars, ongoing sticky inflation and high fiscal deficits. As always, we hope for the best but prepare the Firm for a wide range of scenarios.”

CEO Jamie Dimon opening remarks

“Our Q1 allowance is anchored on the relatively benign base case outlook. Due to the significantly elevated risks & uncertainties, we increased the probability weightings associated with the downside scenarios in our underwriting framework. As a result, the unemployment rate embedded in our allowance is 5.8% vs. 5.5% last quarter (unemployment currently at 4.2%). This drove a $973M rise in the allowance, with a consumer build of $441M driven by changes in the macro outlook.”

CFO Jeremy Barnum

“Consumer balances have stabilized.”

CFO Jeremy Barnum

“It's important to note that the increase in the allowance is not, to any meaningful degree, driven by deterioration in the actual credit performance in the portfolio, which remains largely in line with expectations.”

CFO Jeremy Barnum

Thank you to Bloomberg for publishing the Yipit chart this week. That company does great research, but has asked me not to share their private data anymore. Considering Bloomberg posted it, I think this is fair game.

So far so good for Waymo and Uber in Austin. Through the first 27 days of operation, ride volume is 80% higher than in San Francisco, with larger leads compared to Phoenix and LA. Austin is the only place where Waymo is only available on Uber, as that AV vendor experiments with several business models. The ultimate model selection will be based on maximum utilization rates more than anything else. Yes, route optimization, fleet management and payment orchestration are all positive features for Uber to offer partners. But still, utilization rate matters the most and this data is positive. 

This is how Uber can demonstrate proof of concept to win a closer relationship with Waymo and to keep rounding out its budding roster of other partners. And while many think Tesla and Google are the only two players in town, I would strongly respond with a “not so fast. Zoox is coming… AVRide is coming… Mobileye is coming… Wayve is coming… Oxa, Motional, Wasabi, Nuro and AutoBrains are coming… BYD and Pony AI are coming. Many of these companies have AV deployments set for 2025 and most of them are already partnering with Uber. ALL of these players are surely paying attention to Uber’s impact in Austin. As more of these vendors succeed in deployments, Uber’s positioning strengthens. It’s a lot easier to argue that you deliver better utilization with this tangible Waymo data AND in a fragmented market where no hardware vendor can mirror the network effect that Uber has as a demand aggregator.

The theme of Amazon’s shareholder letter was how its “Why Culture” has shaped decades of success. It constantly questions everything about the business with a perpetual goal of serving consumers better. Every piece of its logistics footprint… marketplace interface… return policy… everything. And in the eyes of leadership, it boldly tinkers with existing norms and tries to create entirely new norms as the “largest startup on the planet.”

Enabling the Why Culture:

There were various philosophical pillars Jassy worked through to explain how Amazon enables this approach and facilitates the company’s historically impressive long term compounding.

Initially, he spoke about various leadership principles. Amazon leaders have strong judgement and carefully formed opinions, but they seek out differing ideas to confirm or disconfirm their convictions. Amazon classifies this as leaders being “right a lot” but this doesn’t mean they always come up with the best ideas. It means they’re excellent at weighing all ideas, confidently selecting the best one and rallying their teams uniformly behind this decision. The company aims for leaders who are constant learners and have the “backbone to both disagree and commit.” This means they don’t simply go with the herd because it’s the easy thing to do. They challenge ideas whenever they think it’s time to do so… even if it’s not a popularly held view. They’re capable of fully embracing a team’s decision, even if it wouldn’t have been their own. There’s no room for “I told you so” in their groups.

Beyond these pillars, Amazon doesn’t use Powerpoints anymore. They see this format as “easy to prepare and hard for the audience to understand.” Conversely, written notes are “harder to create but much easier for an audience to engage with.” They don’t need to connect the data dots to figure out what we’re trying to say. They can just read it and that makes forming the right “why questions” much easier for collaborative meetings. I couldn’t agree more with this idea. Writing is a lost art.

Next, it creates “working backwards documents.” It starts projects with a written press release as if they’re already done. Why? Because oftentimes, putting the purpose of a product in writing makes it more clear if it’s actually valuable or not. It spells out which problems this is trying to solve, which helps guide how to solve them in the actual work. It also reveals when the problems being fixed just aren’t that valuable and time should be spent elsewhere. Along somewhat related lines, Amazon prioritizes teamwork. Ideas are typically formed by one person in a vacuum, but as a byproduct of collaboration. It thinks this process is far more effective in person than over Zoom.

And finally, as briefly alluded to, it is evolving  to act as the world’s largest startup. It’s fixated on solving hard problems and, last year, created a new org structure to prioritize builder accountability. It cut several layers of middle management that were creating more red tape, bureaucracy and confusion than value. It streamlined processes, fixated on pace of innovation and unleashed its developers. The company is constantly paranoid about competition and permanently fixated on outrunning them. And lastly, this culture entails “scrappiness,” or trying to do more with less whenever possible. Smaller teams with scarce resources are more productive at Amazon.

“We have this persistent feeling, throughout the company and in every business in which we operate, that there are closing windows all around us. We operate in fiercely competitive market segments, with highly talented, well-funded, ambitious companies at every turn. Customers are always looking for something better. We spend a lot of time identifying how to unlock these experiences for them as quickly as possible, and know if we don’t, somebody else will.”

CEO Andy Jassy

Current Why Questions Leading to the Following Conclusions:

AI will reshape every user experience and create new ones as well. Any company not working closely with these products will be left behind. This is why AWS is investing so heavily in capacity. There are near-term demand signals that support this CapEx as its GenAI business grows at a 100%+ clip and at a multi-billion dollar run rate. They need to make sure they have capacity to support demand and are still constrained in this regard (as of last quarter).

AI needs to get cheaper. This is why its Trainium2 chip is such a focus. And per the team, it offers 30%-40% better price performance than the other GPU products available today. That obviously means Nvidia, which was both encouraging and interesting to hear. Amazon will not replace Nvidia general purpose GPUs with its own GPUs. It continues to buy from that close partner aggressively. But? Some use cases don’t need the very best chip. Furthermore, some custom and company-specific use cases can be better served by purpose-built chips vs. Hopper GPUs. Meta is a great example of that, as it uses its new chips to power more of its ranking and discovery algorithms. It’s easier to deliver strong performance when building for 1 use case instead of countless. The majority of GPU demand will likely continue to be for Nvidia’s chips, but this hardware can supplement those semiconductors in some places.

More notes stemming from current why questions:

DraftKings reported a 9% hold rate for this past week in New York, which is slightly below its 10% target. Important to keep in mind this is one week in one state and is the byproduct of outcomes – so it’s nothing to worry about for long term investors like myself. I don’t care if bad luck leads to a quarterly miss.

More importantly, the company continues to close the market share gap with FanDuel. We’ve spoken over the last several weeks about how the March Madness volume for DraftKings vs. FanDuel was looking much better Y/Y, with DraftKings even taking a market share lead during the early rounds of the basketball tournament. That was new. We now have March Madness data for every game besides the championship, and it’s encouraging. FanDuel’s bet volume lead vs. DraftKings shrank from 28% to 5% in New York from 2024 to 2025 for Mid-March through the first week of April. While FanDuel’s handle grew by just 2% Y/Y during this period, DraftKings enjoyed 20% Y/Y growth; we’re more than 3 years into New York’s launch.

I got a question in the comments section of an article last week on the varying degrees of tariff impacts for certain countries and sectors. I worked through this idea a bit in a previous portfolio update, but wanted to expand on my thoughts more here.

Following all of this week’s developments, it looks highly likely that most of the trade war aggression will be directed at China. Tariffs everywhere else were reset to 10% for 90 days while negotiations were called advanced and productive. Not only is it important that countries are coming to the table, but it’s highly important that this new tariff policy is a negotiating tactic to win trade barrier concessions like lower levels of currency manipulation. That has always been intuitive to me, but many thought sky-high tariffs were the end goal and new normal. It’s great news that this isn’t true. With this positive change, things are looking more like China Trade War 2.0 vs. Global Trade War 1.0. That’s still a large headache, but it is easier to digest for companies partially reliant on China for their supply and/or demand. With Korea, Japan, Europe, Latin America and everywhere else taking a softer tone on trade negotiations, all of those countries should fare relatively better as the fallout from this altercation takes place. Readers saw me de-risk the small amount of exposure I had to China this week, while they’ve seen me add exposure to Latin America. This is why.

In terms of industries, there are two different impacts to consider from sky-high tariffs in China – direct and indirect. For direct, we simply have to ask how important cheap imports/exports between China and the USA are for the durability of a given business. Apple makes 90% of its iPhones in China (already announced production moves to India and Brazil), Nike produces 16% of its clothing there, Lululemon makes 20% of its goods there and Qualcomm generates nearly 50% of its total revenue from that market. And there are more examples of direct impacts. Amazon and Walmart rely on affordable inventory from Chinese sellers while electronic components, industrials and biotech companies routinely import ingredients from that nation. All of these companies will need to navigate an environment of higher operating costs and a potentially weakened global economy for as long as this persists. And the heavier weight and more entrenched the supply chain, the harder it will be to sidestep these costs without passing them onto the end customer and risking demand decay.

For companies that don’t do much importing or exporting – namely software which encompasses most of the coverage network – business models are not immune but they are more insulated. This is especially true now that we’ve seen the EU’s response and it does not include specific penalties against U.S. payment service providers (PayPal & card networks were a small risk) or large tech companies (which was a somewhat larger but still unlikely risk).

The impacts they would feel are indirect. Pure software & service-based companies, international marketplaces outside of the sphere of U.S/China influence and financial services don’t risk higher input costs from tariffs. They don’t do anything that requires paying them. For these companies to suffer, it would take global economic weakening to drive the type of budget scrutiny, consumer caution and enterprise sales cycle elongation to impact them materially. Despite this, software has been beaten up as badly as any other sector, and I’ve taken advantage. From the sell-side shops I’ve seen, the estimated hit to S&P 500 earnings from tariffs is right around 10%. That’s before better potential resolutions. That, paired with far more than 10% pullbacks (and multiple compression) for many of these firms, tells me the pain from tariffs would need to lead to global contagion for stock weakness to match fundamental weakness. If that doesn’t happen, or even if it doesn’t happen as much as expected, all companies in this bucket should fare relatively well as they have already been marked down in anticipation.

Tobi Lutke, Shopify’s founder, sent out a letter to his employees this week. In it, he discussed shifting from using AI as a suggestion to classifying it as a new mandate. With partnerships and the advent of agentic AI, Shopify is finding profoundly impactful use cases within not just advice and code automation, but completing large portions of entire projects. He spoke about wanting workers to experiment with this technology and talked about needing to frequently use AI to develop skills needed to extract its value. A lot of this learning will come from trial and error, which is why open communication of “wins and losses” was explicitly mentioned in the memo. 

The mandate includes AI usage becoming a new part of performance reviews and teams being required to prove they can’t do added work with an AI model before requesting more resources/people. When I hear this, all I can think of is “higher margin ceiling.” While there are costs associated with using models and building apps, the marginal expense of incremental usage pales in comparison to adding another software engineer for $250,000 a year. If Shopify can do more with less, fixed costs should scale more slowly and operating leverage should kick in more meaningfully. With Shopify’s already stellar margin trajectory, this will likely add another leg of upside. That leg will not be enjoyed all at once, but gradually and subtly over the coming years.

Cloudflare purchased Outerbase for an undisclosed price. Outerbase is a very small company and was built on Cloudflare’s serverless developer platform (called Workers). Outerbase specializes in helping developers access and infuse databases into next-gen applications. Per the presser, it makes this entire process “more approachable, thus enabling more teams to build and deploy applications on Cloudflare’s global network.” Apps and data go together like Marc Benioff and the word “Incredible,” and this is Cloudflare making itself a more attractive developer option for building in the modern era. The company is determined to attract as many developers to its platform as it can. That means more subscription revenue, more usage-based revenue and a larger community building in its ecosystem. Like for others such as Meta, a lot of this work will inevitably involve optimizing GPU utilization rates and overall network performance – thus creating more benefits for nurturing the Workers platform.

“At Outerbase, our mission has always been to make working with data easier for developers.”

Outerbase Co-Founder and CEO Brandon Strittmatter

Galileo debuted a compelling new deposit sweep product this week. It allows non-chartered fintechs to access a network of partner banks to direct excess consumer deposits to higher-yield or higher insurance level options. This allows companies to match the utility of high-yield savings accounts without adding the complexity of securing a charter or ruining their margin profile. Galileo allows clients to set their own deposit limits and “automate customer data exchange, reporting and fund flows beyond those limits. Bluevine (existing customer) is an early user of the product and has “already enhanced customer value” with it. Compelling launch.

In March, SoFi web traffic enjoyed its largest month-over-month growth in over a year. Traffic jumped higher by nearly 20%. Month-over-month growth from February 2024 to March 2024 was 5%, so this is not a byproduct of seasonality. Great news, and quarterly data tells the exact same story. My friend Futurenvesting on X brought this data from Similarweb to my attention. Thank you.

Meta launched the first models from its broadly anticipated and open-sourced Llama 4 series. The company used several new techniques to power better model performance and the results are impressive. Here, I’ll work through some of those new techniques, talk about how the models stack up vs. the competition and review why being the open-sourced leader is so compelling.

New Pre-Training Model Techniques & Features

Meta debuted what it calls a “Mixture of Experts” (MOE) for Llama 4 model pre-training. To this point, when a consumer or enterprise prompted a model with an input token, all parameters within that model were activated to work towards a quality output token. In reality, this creates a lot of waste, considering small subsets of parameters are actually used to form responses. Llama 4’s MOE strategy was created with this in mind, as the model will only activate what is needed. Nothing more… nothing less. This drives better compute efficiency and model performance as inputs are automatically routed to the optimal expert for a given question. 

Model distillation, like we’ve seen with virtually all of the other new models, was also heavily utilized. This Means using larger “teacher models” to accelerate the training and inference capabilities of smaller “student models.” This allows smaller LLMs to emulate the utility of bigger products with more parameters. MOE and distillation both create more efficient end products in their own ways.

The models also incorporate something called “early fusion.” This creates a tighter connection between various data modalities (text, image, video, sound etc.) to unlock unified pre-training with a more diverse and less structured base of data. Finally, Llama 4 powers more dynamic model tuning within the pre-training stage as well as mid-training, which sharpens models with extended context windows to facilitate incremental improvement beyond pre-training. Extended context windows allow models to process more information and insight as they arrive at a desired answer. This is highly important in agentic or reasoning models that rely on longer memory and more data processing to complete strings of multi-step tasks in a goal-oriented fashion. No longer do you tell the model “I need this done and here’s how to do it.” Reasoning models simply need to be told what the task objective is. They can reason from point a to point b in the best way possible.

New Post-Training Model Techniques & Features:

According to the white paper, the biggest hurdle for driving large Llama 4 performance gains was “maintaining balance” across input modalities, reasoning capabilities and conversational talents. It was tough to prioritize all three of these things in harmony without jeopardizing the quality of one or another. Meta found that supervised fine tuning (SFT) and direct preference optimization (DPO) (two post-training techniques) led to exploration bottlenecks during reinforcement learning. Essentially, it made models too shy to seek out improvement, which capped accuracy, reasoning techniques and performance. To address this, Meta actually reduced overall training data by 50% to cut out simple and redundant information. It geared the SFT function towards this subset of data to boost performance. It then (during reinforcement learning) was able to feed the model “harder prompts” that it responded to with lower hallucination rates and latency. 

The Actual Models:

Three models were part of this announcement: Scout, Maverick and Behemoth. Scout and Maverick are now live, while Behemoth is in public preview. Behemoth was the teacher model used to build both Scout and Maverick. While Behemoth is still a work in progress, performance surpasses GPT 4.5, Gemini 2.0 Pro and several other industry-leading models across several major benchmarks. It has 288 billion active parameters with 16 “experts” to segment parameters under the MOE approach (already defined). Scout can fit on a single H100 chip; Maverick can fit on a single H100-powered Nvidia DGX supercomputer.

Maverick & Scout are the two other models, which boast native multi-modality and double the context length of previous models. Maverick is a 17 billion active parameter model (400 billion total) with 128 experts while scout is a 17 billion active parameter model (109 billion total) with 16 experts. Both are smaller than the large Llama 3 models thanks to model distillation. Maverick outperforms GPT-4o and Gemini 2.0 Flash across most benchmarks and performs on par with DeepSeek v3 for reasoning despite half of the active parameters. Generally speaking, per the release, Maverick leads the planet in its category for performance to cost ratio and will be used to run general assistant apps. Scout outperforms the mini Gemma 3 model, Mistral 3.1 and many others across several benchmarks.

The highly technical construction of the models gets way too into the weeds for us. We’ve already arguably gotten too far into the weeds of what investors need to know. We don’t need to understand every acronym, proprietary strategy and piece of source code that goes into making these models shine. What we do need to know is that Meta is innovating rapidly and now thinks it has the best small, medium and soon-to-be large models on the market… including compared to leading closed-source players like OpenAI. Several research organizations like LM Arena AI ranked Maverick as the top open-sourced model on the market. According to them, Gemini 2.5 Pro is the only model that ranks above the medium-sized competitor, but Gemini 2.5 Pro should really be compared to Behemoth.

Open-Sourced:

For consistent readers, this will be a bit of a review. AI model cost deflation continues to rage and the push for optimal efficiency and performance is getting even more important. I don’t see models being entirely commoditized. Data, distribution, internal talent and cash on hand can all be differentiators. But still, I do think partial commoditization is inevitable. DeepSeek, Alphabet and Meta are all showing us how much more models can do without simply making them larger. New techniques are being discovered and implemented, and that will be a theme in the years to come.

Considering my hybrid commodity vision, Meta’s open-sourced approach is a good one. Developers will build with the best tools and where they can deliver their work to the most people. Llama offers best-in-class models while Meta pairs that with world-class distribution. This inherently motivates more developers to build on Meta and a lot of that work is meant to optimize model efficiency and other things like compute utilization to make its products stand out even more. Meta is inviting the world to do a lot of its work… developers are saying yes please… and that makes it a lower cost provider.

Unless you’re a full-stack player like Alphabet that can optimize across more pieces of a tech stack than others, I think open-source is the winning mentality here. I’d much rather bet on a company using all of this work to make its core product better and more engaging, rather than investing in a firm like OpenAI that is reliant on directly monetizing workloads.

Reuters published an article on SentinelOne losing federal security clearances due to a C-level employee that the current administration views as a political adversary. I don’t comment on politics, but the company has already come out and said the impact will be immaterial. Glad they responded to this news.

Dan Ives and Wedbush released a somewhat fiery note this week essentially calling for Telsa’s Musk to distance himself from politics and pointing out what they view as material demand disruption (10% of future customers) from this. Maybe we really are getting close to a local bottom.

Raymond James channel checks for digital advertising point to dollars flocking to the largest platforms. They noted demand “resilience” for Meta, Google and Amazon with “mixed results” for Snapchat, Pinterest and Reddit. Advantage+ was named as a highlight for Meta in fostering stronger return on ad spend (ROAS) and demand.

There are new lawsuits against The Trade Desk in California on impermissible data sharing within Unified ID 2.0. I think this will turn out to be noise. The company prides itself on being one of the only “opt-in” identifiers on the open internet. Consumers aren’t automatically opted-in to data sharing, they must “opt-out.” TTD’s default is to not include them in data sharing until permission is granted.

Morgan Stanley lowered its SoFi price target this week and downgraded Lemonade. Both notes were based on heightened macro concerns stemming from the trade war.

Barclays sees Shopify as somewhat vulnerable to tariffs, as its large base of merchants routinely sources from China. They’re right. Shopify isn’t immune. Still, I do think it is insulated thanks to relative geographic diversity within its merchant base and how sticky of a platform it has. They’re not just building website for merchants… they’re running their entire business. Good luck cutting this cost without operational headaches. That won’t happen, but merchant volumes may slow a tad in the higher price environment, which would impact Shopify. Affirm and Shopify are deepening their partnership with a Shop Pay installments launch in Canada. More countries are planned. Affirm runs Shopify’s buy now, pay later backend.

Cava’s CEO spoke with Wall Street Journal this week on tariffs fueling menu uncertainty. They source paper bowls from Canada, olive oil from Greece and a few other ingredients from various parts of the globe.

Mercado Libre plans to materially boost Argentinian investment levels to $2.6B annually. That economy continues to encouragingly recover from its lengthy dose of hyperinflation. Morgan Stanley also came out with a note on the firm being quite insulated from tariff impacts.

Baird downgraded Starbucks due to near-term same-store sales challenges as Niccol implements his turnaround plan. $85 target. Jefferies upgraded the name despite what it also views as near-term challenges (everyone knows there are near-term challenges) due to valuation support. $76 target.

JP Morgan sees strong market share gains for Duolingo as part of its expansion into new GenAI tools. I’m assuming they’re talking about the video chat product.

JP Morgan upgraded Nu to overweight following the stock’s decline fostering what it sees as a more compelling entry. They actually lowered their price target from $14 to $13.

Bank of America came out with a note talking about some current digital advertising weakness (as expected amid global chaos).

Broadcom announced a new $10B buyback. Small percent on the market cap, but a show of confidence during a time when that’s especially craved.

Wild week in the bond market. The 10-year yield looked like a growth stock, as the Japanese government offloaded some of their U.S. treasuries amid heightened geopolitical tension. 3-year secured overnight financing rate (SOFR) spreads (premium between SOFR and treasury yields) and high-yield corporate credit spreads showed more fragility too. Liquidity and calmness levels can be directly observed in the bond market, and this points to both of those levels waning. The administration directly acknowledged that this contributed to their tariff pause decision, while several fed board members committed to being there for markets if needed. I think quantitative easing will soon come… I think several rate cuts are coming (see this week’s inflation data)... and I think this clearly shows us the federal government and Federal Reserve puts are firmly in place. They can try to tell us they don’t look at markets. I don’t believe them or any other administration that has ever tried to say that.

One more note on the ice cold inflation readings from this week. This doesn’t yet include the full impacts of tariffs, but it’s still encouraging. I’d much rather have a Fed cutting to address growth, liquidity or employment issues when inflation is racing to 2% instead of 8%. It gives them so much more flexibility to use their tools if needed. Some say deflation means we’re heading into a recession. Those same people would have said more inflation will lead to more rate hikes which will lead to… you guessed it… a recession. These people have made up their minds that recession is inevitable. They’re not using data to challenge their convictions. They’re framing it to reiterate them no matter how it looks.

Inflation Data: