EPS is expected to compound at a 13% clip during the next two years.

Oracle provides a slew of software and hardware tools for on-premise and cloud environments… with an understandable focus on a continued shift towards cloud deployments. It has 3 main segments that tie very closely together. 

Oracle Cloud Infrastructure (OCI) is its fully managed business for infrastructure services (virtual machines, storage, managed high-performance compute data centers etc.). This segment also includes platform services to build apps in its safe, controlled environment (serverless and container-based).

Strategic software as a service (SaaS) includes Oracle NetSuite. This is a set of applications for enterprise resource planning (ERP), customer relationship management (CRM), human capital management (HCM), e-commerce and more. It’s hard at work on launching more industry-specific software apps across areas like Healthcare. It has a more customizable, feature-rich version of this product suite called Oracle Fusion geared towards larger customers.

The last segment is Oracle Database (OD). Creating valuable apps from GenAI infrastructure requires great models and great data products to properly season those models. That’s where its Oracle Database (OD) product comes in. It provides a not only structured query language (NoSQL) database for unstructured data, which is highly important in the age of GenAI. Oracle closely integrates with the 3 big hyperscalers to allow its OD database products to run anywhere. This also means that customers can migrate their on-premise databases to the cloud via OCI or through any of these hyperscalers, diminishing the friction associated with using OD. Oracle believes that this data cloud interoperability provides inherent cost advantages with data transferring. Cost benefits are estimated to be “several times cheaper” for model training than any competitive product, according to leadership.

Oracle is (re)-emerging as a digital infrastructure titan. While the company did take longer to roll out its high-performance compute product suite, that has since achieved fantastic traction. The results you see below are the byproduct of it taking its fair share of this infrastructure boom.

Oracle added that its pipeline growth is actually above the 50% RPO growth while its win rates are rising. Meta was the big win for this quarter to collaborate on Llama and AI agents; all of those bookings and the RPO benefit will come in Q3, not for this Q2 report. This will lead to a Q3 RPO “spike.”

“Today we’re telling you again that revenue growth will accelerate further in the coming quarters.”

CEO Safra Catz

EPS is expected to compound at a 13% clip for the next two years.

OCI:

OCI revenue rose by 52% Y/Y (GPU demand up over 300% Y/Y), with OCI consumption up 58% Y/Y as “demand continued to outpace supply.” OCI now is running in 98 regions vs. 85 Q/Q. There are many more coming. The seamless micro-expansion of Oracle’s data center footprint is a rare trait within public cloud computing. Its “Gen 2 architecture” allows it to debut brand new cloud regions with just 10 racks (“soon to be 3”) and to flexibly scale its own capacity to match customer demand growth. Racks are essentially exoskeletons or physical frames that hold servers/compute capacity. It does not need to build massive footprints ahead of demand like Azure, AWS and Google Cloud all need to do.

How does it uniquely position itself to provide micro-regions while others can’t? How has it figured out profitably servicing this demand? A few ways.

It has a single next-gen data center layout. There aren’t several custom configurations to add complexity, while deep layers of automation augment its ability to deliver low cost scaling for its customers and its own financial benefit. As leadership puts it, this simplifies manufacturing, lowers cost and creates one stock of technically-identical inventory to immediately add new racks to. It’s very efficient.

“When all of our racks are the same and all of our services are the same, they become easier to automate. They're all identical. We have 1 suite of automation tools that works in all 100 of our current regions.”

Founder/Chairman/Michigan Football Savior Larry Ellison

It also offers highly dense servers to pack more punch into the same amount of physical space, while featuring high-quality liquid cooling tech to reduce waste and cost inflation. Furthermore, while everyone has focused on building bigger and bigger GPU clusters, Oracle has paired that priority with heavy investments in “building networks that rapidly move data in and out of GPU clusters.” Data transference latency & scale bottlenecks are highly costly when, as Ellison reminded us, customers are renting out space by the minute.

All in all, added efficiency allows vendors like Oracle to service smaller projects and still generate enough margin to overcome smaller economies of scale. The modular-based, uniform design of its data centers allows building for what it needs. I will say that Oracle has been capacity constrained so far this year, which has held back growth and does support the argument that it often makes sense to pre-build capacity. Still, based on its assumption that RPO and revenue growth will continue to accelerate, this approach is working.

The unique strategy here allows it to go where the big 3 aren’t willing to go. This is Oracle carving out a promising niche within public cloud computing.

OCI CapEx:

Flexible capacity scaling means flexible CapEx scaling in tandem with actual bookings trends. That means relatively more stable multi-year depreciation schedules. That helps offset some of the relative margin pressures of being a smaller, newer entrant and helps power the 43% EBIT margin seen this quarter. This business is still just 8 years old. Azure is 14; AWS is 18. Oracle thinks cloud gross margin will keep expanding with scale (last update was 30% a few quarters ago) in the years to come.

“Our competitors always have to land with extremely large footprints before they can even get started. We can land with smaller footprints, have consumption and expand as customers need it.”

CEO Safra Catz

OCI + Oracle Database:

Rapidly proliferating demand for cloud-based architecture means a ton of data and insight moving from on-premise to publicly-managed environments. This is creating, as leadership puts it, a third promising growth lever in OD, joining OCI and its app business. They have an “enormous pipeline of customers” wanting to bring their data to the cloud. Many want to do so with OCI while many want to do so with other cloud infrastructure. Thanks to strong AWS, Azure and Google Cloud (GCP) partnerships, it services all of this demand and creates a truly interoperable environment for lower-cost cloud-based data storage, querying and analytics. This makes model training, app creation and all other GenAI work more compelling; it extends the possibilities of where GenAI can take us and how big of a role Oracle can play.

ORCL’s database services enjoyed 28% Y/Y revenue growth to reach an annualized run rate of $2.2 billion. Almost all of this is from OCI architecture cross-selling, but the big 3 are starting to contribute. Between Azure, AWS and GCP architecture used for Oracle Database services, that business crossed a $100 million run rate. All of these contracts are new and a large chunk of capacity is about to come online. All of this means that hyperscaler-related business for Oracle is poised to enjoy a period of explosive growth.

“We’re at the very beginning of multi-cloud… it will be a multibillion dollar business.”

Founder/Chairman Larry Ellison

To add more AI tools to its overarching database service, Oracle debuted Database 23AI. Per the company, this makes it even easier to “use existing data to augment and specialize training GenAI models.” It offers theme-based vector search to deepen querying breadth and retrieval-augmented generation (RAG) to push vector search results into associated large language models (LLMs)

Quick Note on Strategic SaaS Apps:

“Strategic SaaS apps are growing rapidly. We’re enjoying more industry-based traction for apps coming online and “immediately contributing to revenue growth.”

CEO Safra Catz

Strategic SaaS revenue rose 18% Y/Y to reach an annualized rate of $8.4 billion.

I thought this was a good quarter. RPO could have been a large beat if the Meta deal closed a bit earlier and it continues to sell as much OCI capacity as it can build. The expectation of more Y/Y acceleration in RPO/revenue growth is notably impressive, and is based on observations rather than hope. It’s delivering this stellar growth while enjoying strong OD and app-based demand. Some will pick on the next quarter revenue guide, but weakness is based on strengthening FX headwinds more than anything. Margins remain stellar and OCI leverage should have a long, long runway (like GCP a few years ago).

Oracle has positioned itself to cater to cloud demand that the big three just aren’t interested in pursuing. And? It has clearly demonstrated that this piece of the pie can be morphed into a great business. Excellent execution amid a multi-year vision to recapture this company’s technological edge. More growth… more leverage… more success. The stock volatility after a historic run is noise in my view.

MongoDB is a key player in data storage and analytics with a document-oriented setup. This differs from legacy relational-style databases and next-gen versions like Snowflake’s. How so? Legacy relational databases store data in static rows and columns linked by implemented formulas. These databases look like giant Excel spreadsheets and use structured query language (SQL) to work. Legacy relational databases cannot seamlessly handle unstructured data like MongoDB’s data lake can. This is a large limitation, considering how important unstructured data is for GenAI use cases. Legacy relational databases struggle to scale and unlock the most advanced querying. The datasets are fixed, with formatting and filtering more limited. The inability to provide “not only SQL” (NOSQL) can slow performance and diminish value. MongoDB’s NOSQL database and document-style storage fix these issues, which is why the pace of migration continues to ramp up.

The firm’s most exciting product is called MongoDB Atlas. This is a cloud-native database service that uses a group of servers (or a cluster) to actually store data for app creation within its platform. The nature of MongoDB’s product allows clusters to be easily added to or subtracted for easier tweaking as needs fluctuate. It also offers MongoDB Realm as a mobile environment for app creation and MongoDB Search for data querying. Finally, it offers MongoDB Data Lake specifically to house unstructured data.

Some more products & terms to know:

Reminder:

MongoDB pocketed $40 million in unused Atlas revenue and another $40 million in multi-year non-Atlas licensing business last year. That $80 million in pure margin revenue led to an $80 million revenue headwind and Y/Y profit declines assumed in its 2024 guidance. Keep this in mind as we go through its financials and why they look underwhelming.

MongoDB beat revenue estimates by 6.8% & beat guidance by 7%. Its 26% 2-year revenue CAGR compares to 25.7% Q/Q & 38.1% 2 quarters ago. Net annual recurring revenue (ARR) expansion rate was about 120% vs. 119% Q/Q and over 120% Y/Y.

MDB trades for 120x forward EPS. EPS is expected to be greatly challenged for the next few quarters due to continued comp headwinds. EPS is expected to be flat from 2024 to 2026 and grow by 32% during the following year. It also trades for 145x forward FCF. FCF is expected to grow by 12% this year and by 83% next year.

Focusing Go-To-Market on Larger Customers:

MongoDB is focusing more of its go-to-market (GTM) efforts on the large enterprise segment. It continues to find its strongest investment returns there, and so is leaning in. To nurture momentum here, it’s investing in support for a number of large accounts it thinks “have high upside.” When it focuses more on these customers, incremental, margin-accretive revenue follows. The firm is also investing more time and energy in educating large customer developer bases. As leadership told us, the massive groups of developers that giant corporations routinely have generally boast more experience building SQL apps. Educating them on the flexibility & diverse querying power of NOSQL “drives significant incremental adoption of the platform,” per CEO Dev Ittycheria. It’s pulling resources away from mid-market selling to nurture this more compelling opportunity. Leadership thinks it can shift some of the selling focus there to its self-serve channel, which should help buffer the headwind from abandoning some mid-market focus. And again, it thinks the incrementally better return opportunity within the large enterprise bucket makes this decision an easy one.

Accelerate Legacy App Modernization:

As a reminder, MongoDB thinks it’s great at automating the preparation of data movement for app migrations but not great at rewriting code as it enters their ecosystem. This creates app modernization friction and slows the evolutionary process to a certain extent. Its Relational Migrator Product is its full service response to this; it’s meant to make migration as painless as possible. Paired with its professional services work, this product, early on, is showing an ability to reduce cost of app modernization by 50%. That is how you help your customers ready themselves to embrace and use your technology.

“Not only are customers excited to engage with us, they also want to focus on some of the most important applications in their enterprise, further demonstrating the level of interest and size of the long-term opportunity.”

CEO Dev Ittycheria

Based on the early value creation, “customer interest is exceeding expectations” and MongoDB is allocating more investments to amplify this opportunity. Part of this will entail more professional services support, as relational apps pull from a wide series of fragmented languages and can be complex to migrate alone. While that revenue is not as exciting as software-based and recurring business, professional services are a needed tool to maximize Relational Migrator’s ability to accelerate adoption. It’s also trying to automate and operationalize some of these service engagements to make them more efficient.

During the quarter, Allianz, TealBook and Paylocity were among Relational Migrator customer wins. For TealBook, MDB drove workload interoperability, cost savings and easier scaling than its previous data vendor. For Paylocity, it cut database costs by 80% and time to app creation from weeks to minutes.

Another way MongoDB is accelerating app modernization beyond migration help is via MAAP (defined in 101 section). By providing hands-on support for actually creating GenAI apps within its platform, it puts more of the onus on itself to help clients as they adopt this new technology. MAAP added McKinsey, Confluent and more integrations during the quarter… including a new partnership with Meta to “enable developers to build GenAI apps on MongoDB using Llama.

GenAI Opportunity:

Relational Migrator and MAAP support MongoDB’s GenAI opportunity. The two concepts are closely related to each other. GenAI directly amplifies app modernization urgency for Relational Migrator as firms fall behind more quickly by not evolving. GenAI also helps automate portions of migrations to make them easier.

As discussed in the 101 section, MDB’s diverse NoSQL setup is perfect for the GenAI world. Apps and models devour unstructured data, and NoSQL is the perfect candidate to organize and unleash this unstructured data for commercial use. It “unifies source data, metadata, operational data and vector-based data” on one powerful querying platform for cohesive insight gleaning. More relevant data… means more training signal… means better apps. This also simplifies the life of developers by allowing them to use fewer tools and interfaces.  All of this means that as more of the apps we use become GenAI apps, MongoDB’s niche becomes more compelling. 

“We’re not just an online transaction processing database. We do text and vector search in one experience. No other platform offers that and we think we have a real advantage… MongoDB is uniquely equipped for query-rich and complex data structures typical of AI applications.”

CEO Dev Ittycheria

Today, the GenAI app opportunity still remains nascent, with most of the monetization happening within the infrastructure layer. MongoDB’s consumption-based revenue happens based on customer app traffic. If GenAI apps are largely not ready for ubiquitous product-market fit and usage, that means MongoDB’s main GenAI tailwind has yet to be enjoyed. The same is true for pretty much every other software name developing GenAI tools for other enterprises (or for internal use). It has a few examples of next-gen apps building meaningful traction, such as one with over 1 million total workloads and 10x year-to-date growth, but for the most part, it’s still early.

“We remain confident that we will capture our fair share of these successful AI applications as we see that our platform is popular with developers building more sophisticated AI use cases.”

CEO Dev Ittycheria

Product Sunsetting:

Between focusing more on MAAP, Relational Migrator and large enterprise selling, some other things needed to be deprioritized. Along those lines, it will consolidate the Atlas serverless product offering and its smallest server-based subscription tiers. The newly formed product will be called Atlas Flex clusters. It thinks this will lead to about 4,000 customers churning; these were very small users, so the revenue impact will be immaterial.

Demand Trends:

“We had a strong new business quarter & we’re happy with new workload acquisition for Atlas.”

CEO Dev Ittycheria

As outgoing CFO Michael Gordon told us, the seasonal improvement in Atlas consumption was “more muted” than in prior years. That was expected and results were actually a bit better than it assumed for this segment. Still, Y/Y consumption growth continues to slow and that will be the case in Q4, based on guidance. Much of that is due to holidays and the coinciding app usage headwind… but also 2024 growth levels remaining slower than during previous years.

Non-Atlas revenue was the largest source of outperformance. EA results were called “strong” and it signed a handful of large, long term licensing contracts. For EA deals, all revenue is recognized at the very beginning of the term, which leads to quarterly revenue lumpiness. As a reminder, there’s a $40 million revenue headwind in its guidance for this year from an expectation of lower multi-year licensing revenue. For this quarter, this piece of revenue was actually a $15 million tailwind, vs. the expected headwind. This revenue is lumpy. Meaning? A beat stemming from this source does diminish the quality of the outperformance a tad. But still, Atlas consumption did also modestly outperform. That’s what is most important.

Along this large contract timing note, MDB expects non-Atlas revenue to fall Q/Q in Q4, which bucks normal patterns. That’s due to the tough comp related to signing the large Q3 deals.

As a related reminder, MongoDB tweaked its GTM earlier in the year to focus more on existing client workload growth. Leadership said it was too early to know whether or not these changes are what drove the Atlas outperformance. At the same time, it also said macro was stable Q/Q, so something execution-related powered the Atlas upside.

“It’s too early to declare victory but we’re pleased with the results so far.”

CEO Dev Ittycheria

Final Notes:

It was named the AWS tech partner of the year for North America at the 2024 re:Invent conference. It closed a “ton” of AWS-related deals this quarter and is integrating with its new tools like Amazon Q and Amazon Bedrock. It also closed several large Azure-related deals as the “relationship has never been stronger.” Finally, Google Cloud has made some compensation structure changes with partners that have helped justify more focus and attention with that large partner.

Debuted quantization for Atlas VectorSearch. For less sensitive and pressing use cases, this slightly diminishes number-based data accuracy. In turn, customers can vastly lower data storage costs by up to 96%, which is an easy trade-off to make when clients can sacrifice some querying precision.

CFO & COO Michael Gordon is stepping down after 9 years with the company. A search for a replacement is underway and Gordon will stay on until the end of January. He will serve as an advisor thereafter.

This was a good quarter. Sharp outperformance for Q3 was related to lower-quality licensing revenue timing, but Atlas consumption still outperformed. Commentary surrounding Relational Migrator was highly encouraging and MongoDB’s positioning within GenAI may finally be entering the early phases of prime time. This is a great company with great products and a compelling data niche that should serve it well for years to come. At the same time, it’s very expensive… even when ignoring 18 months of flat profit growth due to tough unused revenue and licensing-related comps. I love the firm, but think there are similarly great companies in software that carry valuations and profitable growth profiles that I find more compelling.

The Search Giant also announced the next series of its Gemini foundational models. I’ll cover that news on Saturday.

The News & Quantum Computing Basics:

The Alphabet Quantum AI division announced a large technological breakthrough this week that has nerds like me excited. It’s in the realm of quantum computing. While traditional computers use “bits” as their building blocks, quantum computers use Qubits. This next-gen method of computing deploys “quantum mechanics” to exponentially accelerate ability to solve problems. Quantum mechanics includes concepts such as “superposition,” which means a qubit can exist in many different states at the same time (not just 0 or 1). For example, if I’m running a calculation on the best 9-iron to buy, this method of computing could theoretically run score simulations using all different versions at the same time… based on any weather outcome or if I decided to have a high noon at the turn etc… to give me the best answer more quickly. That’s a simplified example, but you can imagine how more complex challenges could greatly benefit from this.

I still won’t break 80, but at least I won’t break 80 more quickly. Quantum mechanics also includes a concept called “entanglement” which can link qubits from remote locations and lead them to effectively act as one. This enables complex problems to be solved with more compute power. The concept is somewhat similar to how networking equipment from Broadcom or Nvidia today allows more GPUs to connect to one another to bolster compute capacity.

This week, Alphabet introduced a quantum computing chip called Willow. It’s potentially the most advanced processor the world has ever seen. The chip is fabricated at its facility in Santa Barbara so it can control every single piece of the manufacturing and distribution process. If anything is off for any component of this hardware, there will be crippling performance bottlenecks.

Why Willow Matters:

There are two fundamental issues facing quantum computing that contribute to why commercially viable applications of this technology are still theoretical. Those problems are error rates and complex problem solving beyond the capabilities of traditional computers (without soaring error rates). As the head of Google Quantum AI (Hartmut Neven) puts it, “qubits rapidly exchange info with their environment, making it hard to protect information needed for computation.” This essentially leads to unstructured data chaos, which makes it likely for qubits to use the wrong context to solve a problem incorrectly… and as more qubits are added, more errors proliferate. For the last 30 years, researchers have tried to remove the negative correlation between qubit scaling and algorithm accuracy. This is called the pursuit of “below threshold.”

Now, with Willow, more qubits means lower error rates. Specifically, as it tested larger and larger grids of qubits, error rates fell by 50%. An ability to get more accurate answers as you add more qubits is simply massive. It removes the current cap on how large quantum clusters can become before becoming useless, and pushes the mega-cap much closer to technology with commercial viability.

This is how The Search King can pack 105 qubits into Willow (vs. 53 for the predecessor) while maintaining excellent accuracy rates.

The second issue is complex problem solving beyond traditional computers while still maintaining those lower error rates. Willow completed a standardized random circuit samping (RCS) benchmark test in less than 5 minutes. It compared this to Frontier, one of the most powerful supercomputers on the planet, and how that same problem would take it 10 septillion years to finish. Per the team, RCS is the “hardest benchmark test” for quantum computers today” and is a key hurdle to clear to demonstrate complexity beyond normal computers. In this light, it’s almost like a Turing Test 2.0. The Turing test measured if a computer was as powerful as a human brain; this testing measures if Willow is more powerful than a computer. The Alphabet team also made “conservative assumptions” for Frontier to make the comparison even more difficult for Willow.

Other Cool Willow Features:

Beyond fixing these two issues, Willow debuted new innovation that makes this technology more commercially compelling. It can now fix errors in real-time. It also offers “tunable qubits” to rapidly uncover underperforming “outlier qubits” to remove that performance headwind. Researchers can openly interact with this tunable architecture to constantly toy with algorithms for consistent error rate and performance improvements.  This is also the first quantum chip to extend the life of qubits as more are connected to one another via entanglement. This is yet another great feature that can make Willow more affordable for commercial deployment down the road.

The company acknowledged that normal computers will continue to improve, but it expects pace of improvement within its quantum branch to greatly outpace that progress.

Looking Ahead:

The company sees Willow as a massive step towards quantum computing being a real financial driver for the company and the deployment of relevant quantum-based apps. It will need to scale and perfect manufacturing processes first (to ensure this business comes with strong margins) but progress is palpable and this is looking less science-fiction-y by the year.

“Willow lends credence to the notion that quantum computation occurs in many parallel universes, in line with the idea that we live in a multiverse, a prediction first made by David Deutsch.” –

Vice President of Engineering & Founder of the firm’s Quantum Artificial Intelligence lab Hartmut Neven

High performance GPUs from Nvidia have vastly accelerated velocity of compute and data processing. It has made building complex, more data-intensive apps actually rational to pursue. This innovation could theoretically expedite that trend. Performance and efficiency are like oil and gas in a car. The driver is just going to pick whatever makes the car (or model or app) run more smoothly. These chips theoretically should do that. I think it’s hyperbole to say this will be the end of Nvidia GPU demand (I’ve seen others stake that claim), but if Alphabet keeps making massive strides here, it could find itself commanding a strong market share position in this large market. For now, this all remains “theoretical” which is now the 5th time I’ve used that word in this piece. That’s intentional. There remains a lot to prove to turn these advancements into ubiquitous apps. For now, this accomplishment will not benefit financials in the least.

Finally, many think this speed and power of compute could potentially crack Bitcoin encryption and risk the security of that asset. Bitcoin is not considered a “quantum safe” algorithm. Still, there are things holders can do to combat this risk, including using Segregated Witness addresses (added layer of security) and cold storage. Furthermore, the power of these chips (many think) would have to 10,000X to be able to crack those algorithms. Long, long way to go before that’s a concern.

General Motors (GM) is discontinuing its robotaxi project called Cruise; many thought of this news as a when not if. Importantly, GM is still pursuing the development of autonomous vehicles (AVs) for personal use. Cruise was one of many Uber partners and many think this adds to the risk of there being a robotaxi monopoly. In that case, as a reminder, Uber’s network effect would become immaterial and its value proposition would wane. Network effects don’t matter when you’re partnering with other monopolies. Still, I think this headline has been vastly overblown.

GM is still going to build autonomous cars. These cars, if they’re built, will be added to autonomous fleets and rider-sharing networks like Uber. To me, it doesn't really matter if these cars are added to fleets by individuals, smaller professional managers or GM directly. Those cars making their way into AV networks is what diminishes the risk of Waymo or Tesla (or Amazon or someone else) owning the entire fleet and eliminating the value of Uber’s value proposition.

Owning all of the hardware is the best way to own all of the software (the apps). And again, GM still plans on producing its own hardware.

The things to worry about are Tesla and Waymo winning this entire market and/or cutting Uber out of the business model. Those risks are real and are why I’ve trimmed so aggressively this year. And while I think these risks are very real, the Cruise development does not amplify them.

I find it ironic that Uber fell when the Cruise partnership was first announced… and fell again when this news came out. Just tells you where sentiment on this name currently sits.

Most of this interview was spent on the robotaxi risk. This was an extremely timely and valuable chat, considering recent AV regulation, Waymo and GM Cruise news. Thank you to Ross Sandler at Barclays for asking great questions.

New Yipit Data:

(To Yipit employees reading this, the Barclays analyst explicitly cited your data in a public forum. If you take issue with me sharing this data, you can take it up with him. I explicitly refrained from sharing this information until someone else publicly did so.)

This morning, Yipit published some data on market share for Lyft, Waymo and Uber in San Francisco. Its data showed that Waymo’s market share (for its own app, not rides through Uber) was approaching 20% and on par with Lyft. It also pointed out that Uber’s market share there was either stable or slightly declining. San Fran is the most mature city for AV deployment, so it’s a great case study for potential Uber implications. Rajah was ready for this question, and even provided brand new November data from that city showing stable Y/Y and Q/Q growth rates compared to previous years. This supports Uber’s thesis that AVs will expand ride-hailing use cases and grow the overall pie. In this light, Uber does not need to take any market share to enjoy strong growth.

Waymo:

The structure of Waymo’s expansion into Miami is the recent development that has me most cautious on this currently small holding. This interview assuaged those concerns a bit. Rajah reminded us that Uber has delivered higher-than-expected utilization rates for Waymo in Phoenix, which prompted Waymo to pick Uber as its exclusive app and fleet manager in Austin and Atlanta. That was then followed up by Waymo selecting Moove to manage its fleet there, which was disappointing at the time. Uber is an investor in Moove and has two board seats, but I still didn’t love hearing this.

Tonight, Rajah told us that Moove would likely be using Uber’s Enterprise Resource Planning (ERP) software to conduct this fleet management. Great to hear.

Waymo will likely continue experimenting with different business models in different cities to discover what works best. As long as Uber is providing higher utilization rates… with its stellar routing algorithms trained on more relevant data than anyone else has… there’s every reason to continue deepening this partnership. That’s what leadership expects.

Between higher utilization contributions, its seasoned routing processes, and tedious things like ID verification and payment fraud prevention… Uber remains adamant that the transition to AVs will be very positive for its overall base of supply and its business. While all of these small perks are nice, it’s maximizing occupancy rates that matter the most.

AV Market Outside of the USA

Outside of the USA, Uber sees several deployment-ready vendors in China with hardware ready for scaling. One of those is BYD, which is a close Uber partner. The regulatory landscape in China and Europe are the main things holding back the sector there.

2025 Guidance Commentary:

Uber expects mobility bookings growth to remain slightly below or above 20% for most of 2025. Rajah wasn’t ready to offer a full year guide, but this disclosure was still a pleasant surprise. Uber also sees insurance cost growth greatly lagging revenue growth next year, paving the way for yet another source of operating leverage.

I am sticking with my remaining piece of this name for now. I would love to add more shares as the AV risk picture becomes a bit clearer. Tonight helps a little, but I need more evidence. I may also have to add a bit if the recent multiple contraction gets much harsher. This thing is so cheap. 19x free cash flow; 24% 2-year free cash flow CAGR. 20x EPS; 20% 2-year EPS CAGR. We’re a Tesla partnership or a positive Waymo announcement away from this thing re-rating higher and turning back into a Wall Street darling. I’m happy to risk about 2.8% of my portfolio on this getting cut out of the AV equation, as the potential upside from that not happening is (in my mind) massive.

Bank of America downgraded SoFi and Duolingo based on valuation concerns. For SoFi, I don’t think those concerns are warranted and are based on using legacy valuation metrics to judge this disruptor bank.

Analysts see its 3.7x tangible book multiple and 75x GAAP EPS multiple and immediately conclude it’s crazy expensive. What they ignore is the expectations for rapid TBV and EPS growth in the coming years. Rate of profitable growth is the single most important variable for determining what a fair multiple is. Analysts continue to completely overlook that reality.

Specifically for EPS (which feeds TBV), SoFi is expected to compound that metric at a 90% pace through 2025 and 2026. This ignores faster growth in 2024… ignores SoFi’s far more optimistic guidance… and ignores that its optimistic guidance doesn’t even include products like its credit cards or the new piece of its thriving lending platform. But even if we refrain from assuming any upside, the PEG ratio here is 0.83X. To call that priced for perfection, like Bank of America did, is simply not accurate in my mind. Furthermore, the analyst thinks this recent move is related to the election and student loan refinancing. I would just remind them how powerful of a macro tailwind rate cuts are for all of its business segments; I think it’s just inaccurate to think the election is having a bigger impact than the Fed and SoFi’s overall execution right now. Bank of America’s estimates are far too low… their arguments are incomplete… their understanding of SoFi’s hedging program and fair value accounting is off… I don’t find this note compelling at all.

On the Duolingo downgrade, I think this is very fair. The company has more than 4Xed from the lows. It has gone from stupidly cheap, to more fully priced and I take no issue with the analyst acknowledging that. This is why I’ve done a bit of trimming here in recent months (nothing all that recent). I continue to view this as one of the highest quality software names on the planet, and am willing to own it at what I now view as a justified valuation. I think 50x forward free cash flow can be maintained, with profit growth and share price growth tracking each other rather tightly in the coming years. I don’t think there’s a lot more multiple expansion left to be enjoyed, and I don’t think that’s necessary for this to work.