Cava is a quick-service restaurant chain that sells Mediterranean food, with a focus on strong value, quality ingredients and a warm, in-store ambiance.

My Cava Deep Dive can be found here.

For individual cost buckets, G&A was 12.6% of revenue vs. 13.9% Y/Y. Food, beverage and packaging costs were. 29.9% of revenue vs. 28.8% Y/Y due to steak. Labor and related costs were 27.3% of revenue vs. 27.8% Y/Y. Leverage was offset by 4% wage inflation. Occupancy & related costs were 7.6% of revenue vs. 8.3% Y/Y. It spent a little more than expected on other OpEx for store upkeep.

For the full year, same-store sales growth guidance of 7% missed 8% estimates by a point. Its 25% RL margin guide, which includes a 100 bps headwind from a full year of steak, missed 25.4% estimates. EBITDA missed by 6%. It continues to expect 17%+ store growth in 2025. They love to under-promise and over-deliver, and the most popular time to do that is when offering brand new annual guidance.

Cava trades for 71× 2025 EBITDA. EBITDA is expected to grow by 26% this year and by 29% next year. It also trades for 137× 2025 FCF. FCF is expected to grow by 60% this year and by 10% next year.

“Our guidance takes into consideration what we are currently seeing in the business and the fluidity of the macroeconomic policy environment. Our business continues to be and remains strong and resilient. That strength is appropriately reflected in our guidance.”

CFO Tricia Toliver

Cava’s value proposition is clearly resonating based on its masterful results since going public. What’s working? This gets back to what I talked through in the deep dive. Cava is combining elite service levels with strong operational rigor practiced for every single decision made.

What’s Working — Tech & Human-Driven Service:

Cava empowers its workers to have easier, enjoyable and more rewarding jobs. In turn, they feel inspired to take better care of customers, stick around for longer and drive a better omni-channel experience. Pairing this with quality food, a warm in-store ambiance and affordable prices has been a winning recipe.

For example, after careful testing of its new labor deployment model, it’s enjoying improved distribution and placement to create organized and less overwhelmed workers. As a result, store productivity and throughput are both rising. It thinks there’s even more to do here to improve speed of service, throughput and store average unit volume (AUV ) from already strong levels. It knows a lot of the high-volume stores still deal with long lines and this is one way it hopes to alleviate that service issue.

Next, as part of its Connected Kitchen initiative, its AI video technology is delivering expected objectives in 4 test stores. This will be added to more locations this year. As a reminder, the product tracks ingredient depletion and nudges employees to replenish needed items. This technology will be used for its in-store food prep lines, as well as the digital order make-line. As yet another example, its new kitchen display system (now in 25 stores) is improving order accuracy rates across channels, while improving customer service ratings.

“As we focus on running great restaurants in every location, every shift, we will integrate new technologies and tools to support our team members and enhance the guest experience.”

CEO Brett Brett Schulman

Lastly, I think its revamped loyalty program is a solid case of creating great, personalized customer service at scale. The new approach has been “warmly received” and raised the percentage of sales from the program by 230 bps since relaunching. Furthermore, 50%+ of redemptions are from the entry-level subscription tier, which is what Cava was hoping for. As of now, this is really just a bankable points model where customers are compensated for engagement. Looking ahead, 2025 will be the year where Cava fully leverages the power of its 1st party data to create more personal, granular and delightful experiences. This is how it will make the brand feel intimate while expanding across the nation.

As an important aside, internal development of this loyalty program is no small feat. It’s not normal for an $11B chain to do everything on their own. This is the byproduct of Cava carefully building its malleable, microservice-based foundation to enable faster learning and iterating. The deep dive covered this in more detail. For the sake of this piece, just know that this foundation was a prerequisite for all of the good work Cava is doing today. 

Lastly, I think price will become an increasingly strong differentiator for Cava service quality. Since 2019, Cava menu inflation has been 15% vs. about 35% for the sector (23% total CPI growth). This year, it plans to hike prices below the rate of inflation while its peers do the opposite. Strong operations and supply chains allow it to consistently find more margin and pass part of the relative savings onto customers. This is more of the same.

What’s Working — Disciplined Operational Rigor:

The other reason Cava is thriving is because of its obsessive operational rigor. The microservices foundation set the stage for them to closely track and optimize every tiny variable within their business, and they spend a lot of time doing so. They resemble Duolingo in this light.

The team doesn’t roll out a new ingredient until it has been intricately tested and they know it will work; they don’t implement new store technology until it obviously creates value; they don’t change loyalty program experiences until they’re certain it’s the right move. There is no guessing here. There is small-scale testing that expands to more stores as Cava learns and grows more confident in a given initiative. They always walk before they can run, always follow their playbook and always deliver product introductions on or ahead of schedule. Great team.

Cava stores are absolutely crushing it. In a year when most publicly traded competition struggled to maintain positive traffic, Cava had to cut marketing spend in key markets because lines were too long. Not normal. What an amazing problem to have. When looking at newer store vintages, 2024 is pacing to be its best year yet. This consistent outperformance led to Cava revising its new store financial targets it offered when it initially went public. For years 1 and 2, it now expects AUV of $2.3M and $2.5M, respectively. This represents a $200K boost for both figures. Furthermore, RL margin for year 2 is now expected to be 22% vs. 20% previously. Despite a small upward revision to restaurant CapEx estimates, it boosted its year 2 cash-on-cash returns estimate from 35%+ to 40%+. Very good. 

Clearly this means each store can contribute more to the financial engine than originally thought. And? There’s another more subtle perk here that I find equally enticing. More productive stores mean locations in fringe markets that weren’t quite attractive enough to pursue now are. This expands the addressable market for Cava stores and amplifies the opportunity for growth. That opportunity was already big, as it hasn’t even entered several major markets in the USA. Along these lines, it will enter Detroit (thank you), Pittsburgh and Indianapolis this year.

Cava is getting really good at marketing. Through influencer partnerships and strong word-of-mouth growth, its brand awareness is rising. This is leading to “people waiting for them” to open stores, which means a quicker AUV and revenue ramp.

The steak launch continues to outperform and drive incremental traffic. Despite this impacting RL margin a bit, management is quite excited about traction. In 2025, Cava will likely follow the same menu promotion cadence it did last year. There will be one major introduction and a few seasonal items.

Cava is progressing in its catering tests. It will expand from very limited access to debuting in a singular major market this year. A national rollout of this is probably coming at some point in 2026. Just like consumer packaged goods (CPG), this represents another opportunity to make its stores and its manufacturing even more productive.

Despite the modest 2025 guidance weakness, I think this was a good quarter. The company continues to outperform all peers in terms of store health, growth, and AUV expansion. It continues to effectively nurture its loyalty program and is thriving in every single new market it enters.

This is a special company with a special team and a sky-high valuation multiple that I’m not willing to pay. I love everything else about the investment. I would love to own this name at some point. For now, all I can do is admire the company from a distance. Getting closer… but not there yet.

Full review coming tomorrow alongside Duolingo.

Nvidia designs semiconductors for data center, gaming and other use cases. It’s unanimously considered the technology leader in chips meant for accelerated compute and Generative AI (GenAI) use cases. While it specializes in chips, it does a lot more than that too. Its toolkit includes chips, servers, switches, networking and cutting-edge software. It designs the entire next-gen data center layout with slick software integrations so customers can enjoy the best of accelerated compute. Nvidia calls these data centers “AI factories.”

The following items are important acronyms and definitions to know for this company:

DGX: Nvidia’s full-stack platform combining its chipsets and software services.

Chips:

Connectivity:

Software, Models & More:

The Blackwell ramp is going smoothly and the product exceeded internal expectations, with $11B in Q4 revenue. Production is in full gear for most configurations and supply is increasing steadily. There was no change to Nvidia’s previous guidance of Blackwell being supply constrained through the summer. Leadership acknowledged the mask design issue with Blackwell that led to production delays, but doesn’t see that as an issue for its upcoming Blackwell Ultra launch. This is because moving from Hopper to Blackwell entailed adding NVLink72 SpectrumX networking hardware to its menu to join NVLink8 Infiniband. Blackwell Ultra won’t have this same task.

For the networking (connectivity) piece of data center revenue, revenue fell 3% Q/Q. This was related to shifting from NVLink8 to NVLink72 Ethernet-based technology. SpectrumX revenue rose Q/Q and “represents a major new growth vector.” The segment is expected to resume Q/Q growth in Q1. Stargate data centers will use this technology, as Nvidia’s traction here grows. Notably, while it clearly leads in GPUs, other companies in networking (like Broadcom) are tougher competition. Good to see a win like this.

The longevity of the GenAI hardware demand boom is of the utmost importance for Nvidia. This debate has raged on for more than a year now, and that will continue to be top of mind. Within this topic, there are two things to consider.

The first is whether or not Nvidia continues to lead this technological boom. I think that’s highly probable at this point. The lead from the H200 chip is why demand growth has been so incredible and why margins have been so elevated. The tech lead fosters hefty interest and pricing power for Nvidia. And? Nvidia is sprinting to release brand new platforms on an annual cadence, with massive, incredible performance gains delivered every single year. For example, MLPerf (independent 3rd party) conducted a study showing Blackwell delivers 2.2x performance gains over the once-best-in-class H200 chip, with a 30x inference speed boost too. AMD has now moved to an annual platform launch cadence to try to match and catch this company. The issue? They’re pounding their chests about reaching Hopper-level performance while Nvidia moves onto its next platform with large incremental upgrades. This is not a matter of other companies being poorly run. AMD’s Lisa Su is a great CEO. Instead, this is a matter of Nvidia’s execution being ridiculously impressive.

The second part of the current debate is more uncertain in my view. While this chip cycle has dwarfed all predecessors, semiconductors will still always be cyclical. The intensity and length of this cycle will determine how long Nvidia’s growth can remain rapid.

But what determines the intensity and length? A few things. First is cost deflation and model improvements.

Cost Deflation and Model Improvements:

When DeepSeek came out with its new reasoning model that was trained at lower costs than competitors, panic struck. But? Just like in the cloud computing revolution, disinflation is a net positive for overall demand and revenue levels. The cost pressure is more than offset by an explosion in usage, as access to the latest and greatest technology becomes more feasible and rational to pursue. The real risk here is cost deflation and model improvements halting. If products aren’t getting better, there’s little reason to buy the latest version. If they are getting better and more efficient, companies can either pocket the savings from this or build better products. As soon as one competitor decides to build better products, all will be forced to follow suit.

We’re seeing the risk of slowing innovation play out in the CPU world, which is why GPUs have exploded so aggressively in demand. Lack of CPU innovation — as apps require more data processing — leads to rampant cost inflation which turns everyone away. DeepSeek isn’t hurting Nvidia… they’re ensuring GPUs don’t realize this same fate in the near-to-mid-term; they’re facilitating lower friction and another demand tailwind. Nvidia praised the development on the call, as DeepSeek is supporting its obsessive journey to drive lower total cost of ownership, higher performance, better customer results and so more company demand. Jensen pounded his chest about 200x inference cost reduction in 2 years to hammer home the idea of deflation being a very good thing in the eyes of Nvidia.

The 3 model scaling laws Jensen always talks about still have miles to go in delivering more model improvements. The 3 laws are pre-training (adding more data to models), post-training (retraining deployed models with reinforcement learning) and inference time scaling or reasoning (making models think harder for better answers). Pre-training is the only law of the three where the opportunity is at all mature. And even there, new developments like model distillation (smaller models made from a larger foundational models) provide great promise for more advancement, while also exponentially boosting pre-training compute needs. For the other two laws, we’re in the first inning and the first pitch hasn’t even been thrown yet. You still have time to go get those peanuts (man I miss baseball).

Blackwell was also built to be highly malleable. It’s offered within several different chips, with a plethora of networking equipment designs to fit any workload needs – from pre-training, to reasoning and everything in between. This creates yet another edge for using Nvidia’s hardware. Because everything in the hardware and software stacks work so well together, customers can easily shift their purchased GPUs to different types of workloads when need be. This should diminish concern associated with training revenue, as firms know they can reallocate this capacity elsewhere if advancements like DeepSeek’s make that necessary.

ROI:

Next is ROI. Last quarter Nvidia shared that $1 spent by hyperscalers yielded $5 in revenue over a 4-year period for these customers. That means a sub-1-year payback period for the mega-caps and is an obvious business decision for them to make. Aside from this, I think tangible ROI examples are the best form of evidence we have:

As a highly relevant aside, Blackwell was built specifically for DeepSeek’s inevitable disruption. Nvidia saw this training-to-inference shift coming. It was ready. Specifically, Blackwell boosts token throughput by 25x for reasoning models vs. H100. Higher token throughput means more revenue potential from the same amount of infrastructure cost to support ROI and elongate the runway. As Jensen explained, tokens (model outputs) produced by these data centers are in short supply and are highly coveted. So? 25x more throughput means 25x more revenue potential while these market conditions persist. From an expense perspective, Blackwell offers 20x lower inference cost vs. H100 — which was the leader before Blackwell. More revenue… lower cost… higher ROI.

DeepSeek is leading to budgets shifting from pre-training workloads to post-training and reasoning workloads. This leads us to the third and final major topic within this runway debate. How much overall compute demand will there be as DeepSeek and other innovators accelerate this shift? Boy were they ready to answer this question. Fortunately, long thinking and reasoning models require 100x more compute vs. one-shot inference. One-shot inference uses a single subset of data points to draw a very simple conclusion. Reasoning models combine data sets and scrape from different context to connect dots and provide deeper, more powerful insights. This will be vital for agentic AI, which is where the world is heading. And? 100x compute needs should mean the runway for workloads running on Nvidia GPUs remains long.

“Blackwell was architected for reasoning AI inference.”

CFO Colette Kress

All of this points to the demand runway remaining quite long, as we’re still early in the CPU to GPU infrastructure transformation and as newer developments like agentic AI and physical AI provide more promise. The only caveat I will add here is that semiconductor companies all tend to be a bit bold in their cycle forecasts in terms of length and size. I do think this evidence bodes well for the runway, but this still needs to be said. They’re a charismatic bunch and Jensen is certainly no exception.

Nvidia’s software continues to help it stand out in a few ways. First, its libraries of models and apps provide guardrails and inspiration to help companies embrace GenAI. The technology is basically a blank canvas for most firms right now. They need to be told what they can do and how to use it. NVDA’s software suite provides this education and also compelling cross-selling opportunities to create more growth and stickier customers. Once you’re using Nvidia NIMs and its libraries to run Cuda alongside your GPUs, it’s very hard to switch to other vendors. AMD is trying to make it easier, but vendor lock is still very real. So? Software should continue to mean higher retention.

But there’s another way in which software supports Nvidia that I think is often overlooked. Similarly to Meta or other companies with massive scale, Nvidia’s product suite and large install base motivate (5.9 million) developers to build within their ecosystem. A lot of this work involves GPU fine-tuning to maximize utilization rates and performance for Hopper and Blackwell. That inherently makes Nvidia tougher to compete with, as the world is helping contribute to its lead.

Nvidia thinks its automotive business will reach $5 billion in revenue this year vs. $1.69 billion in 2024. This is thanks to the explosion of physical AI demand in the world of robots and autonomous driving specifically. Its Cosmos World Foundation Model Platform for physical AI is expected to supplement this momentum, with Uber an early customer. Hyundai Motor Group (and basically every other autonomous vehicle (AV) program) is using Nvidia for several use cases to advance its AV program, which should 10x the compute that company is demanding from Nvidia. Company leadership is extremely excited about this opportunity, as most of the industrial economy has yet to embrace AI; the value Nvidia can deliver with digital twins (through its Omniverse software) to optimize and automate processes is immense. Nvidia Drive, its driverless platform, passed important safety assessments from two important German inspectors called TÜV SÜD & TÜV Rheinland. It’s the first AV platform to “receive a comprehensive set of third party assessments.”

China remains a materially lower percentage of revenue than it was before the export restrictions. Nvidia thinks it will remain around here as a percentage of revenue. They were not asked about large, over-indexing orders in Singapore that many think were redirected to China to power DeepSeek’s R1 model.

Gaming is supply constrained. Q/Q growth will resume in Q1 as this is resolved. Its new Blackwell laptop GPUs extend battery life by as much as 40%.

This was an excellent quarter. Nvidia clearly continues to lead the GenAI chip boom by a country mile and commentary points to this opportunity having plenty of room to run. That encouraging language was highly important to me, as the main risk here is when this cycle will finally begin to slow down. According to Nvidia, 2025 CapEx guidance from large customers and these great results… the answer is not yet. Its incredible pace of innovation should ensure it stays ahead of the pack to keep dominating and delivering historic results. As long as overall demand levels remain this robust, Nvidia will thrive. With agentic AI, physical AI and rapid model innovation still playing out, that likely won’t end any time soon.

If you’re wondering why the stock isn’t responding more positively to stellar results, it’s because Nvidia has spoiled all of us. They’ve trained us to expect monstrous beats and raises, while the level of outperformance was “only” great instead of record-setting. For the long-term Nvidia bull, the story remains firmly intact. What a company.