Understanding Customer Lifetime Value
with case studies on OpenAI and Hims & Hers
I was recently picking up coffee when someone spilled their drink at the counter. The staff saw it happen and immediately replaced it, free of charge. The marginal cost of making a new one is almost zero, and if they were asked to pay they would feel unvalued and go elsewhere. Coffee is a high-frequency purchase driven by strong customer loyalty. What’s really at stake is their annual and lifetime spend, not the cost of a single drink. Good companies recognise that all it takes is one bad experience to wipe out customer loyalty, whilst good experiences create lasting impressions which keep you coming back for more. Replacing the coffee wasn’t an act of generosity; it was a strategic investment in customer lifetime value.
It’s intuitive, but it’s still common to treat customers as transactions rather than relationships. Even at large companies that should know better, there can be a tendency to go from campaign to campaign, racking up short term wins that lack sustained returns over time. The next set of quarterly earnings look good, but long term performance suffers. In response, investors increasingly demand greater transparency through the public disclosure of customer health metrics. But these metrics often lack standardisation, giving companies significant discretion in how they define and report them—sometimes in ways that mask underlying weaknesses in customer health.
This post serves as a primer on customer lifetime value (CLV), explaining what it measures, why it matters, and looking at two cases - OpenAI and Hims & Hers Health - to see how it can be applied in practice to provide a better understanding of the business model. They are different types of consumer businesses: OpenAI’s cost structure is product-heavy, while Hims & Hers is marketing-heavy. Yet their success hinges on the same principle: how efficiently they turn customer relationships into long-term profit relative to the cost of acquiring and serving those customers.
What is CLV?
There are multiple variations of CLV depending on the nature of the business (e.g. do customers churn or repurchase?) and it is more of a concept than a formula, but it is essentially the present value of customer future cash flow over a time horizon, which can be either a defined period or the average retention time. Here are two example calculations.
E-Commerce (repeated purchase, average retention time):
average order revenue x average order frequency per year x average retention time in years.
SaaS (churn-based, defined period of a year):
average subscription revenue per account per year × gross margin × average retention time in years (inverse of the churn rate!)
Note that additional adjustments are often made to this crude measure, such as discounting future profits (time value of money) or to account for dynamics in churn, i.e. if the churn rate were 10% in a year, it obviously matters when the churn happens in the year. It can be further developed into a net figure by subtracting the fixed customer acquisition cost (CAC) and the yearly cost of customer retention (CRC). Beyond this ‘descriptive’ approach, we can deploy statistical modelling which enables us to use other leading indicators - think acquisition channel, demographics, new product launches - to generate a more nuanced forecast of CLV.
Applications of CLV
Fundamentally, you want to spend less on acquiring customers than you think they are worth over their lifetime. How much less? A common heuristic is a minimum 3:1 CLV:CAC ratio, which suggests a good product-market fit and a company that is growing sustainably, whilst possessing the potential to scale with minimal outside investment. However, the “right” ratio depends heavily on the business model, margin structure, competitive dynamics, and growth stage. For example, this ratio typically refers to gross CLV, but in industries such as retail where the gross profit margin is relatively low, this ratio should be higher. Likewise, mobile games often have CLV:CAC ratios below 1:1, as more users enable them to scale app charts and acquire future customers at very little cost. Many readers may have benefited from the millennial lifestyle subsidy, when the ‘era of free money’ enabled tech companies to ruthlessly acquire customers at almost any cost to expand market share1. Ultimately, the 3:1 heuristic rarely holds once a business reaches any meaningful level of complexity.
Whatever CLV:CAC buffer is appropriate for your industry, if your profit margins are high it may be tempting to cut it closer. However, there are drawbacks to this. Firstly, CLV is a forecast with inherent risk. It is almost certainly wrong; you just have to hope it isn’t too wrong. Secondly, more aggressive campaigns in a finite market will increase the cost of acquisition and likely attract worse customers, and you might be surprised how quickly CLV and CAC start to converge toward each other. And then you’ve thrown half your budget into marketing, whilst your competitor has been ruthlessly improving their customer experience and acquiring customers that way. Your customers probably - the ones you were banking on to deliver value in the future. Put simply, efforts may be better placed increasing the ratio by improving CLV, rather than trying to make money through arbitrage. CLV should not be thought of as solely a marketing KPI for the purpose of keeping CAC in check, but rather something that energises the whole business to deliver value for customers.
For example, one application of CLV is in informing customer retention efforts. Low CLV customers might be one-off customers that didn’t spend again, and/or markdown chasers. They tend not to give you much return for marketing spend (although shouldn’t be discounted entirely with the right tailored reactivation campaign). High CLV customers, meanwhile, can generate a lot of additional value from loyalty schemes and personalised campaigns. With a strong understanding of your customers’ CLV, you’ll be able to run experimentation to see how it changes with product improvements, discount strategy and so on. As with looking at CAC, it is best to segment CLV across dimensions like demographics and acquisition channel to create holistic customer profiles. CLV:CAC ratios can vary immensely by segment, and averaging across them may not be useful.
But all of this depends on getting the fundamentals right. And there’s no need to rush, since meaningful cohort insights only emerge with time. Focus first on building a historical view of customer value from your transaction data, layering in sophistication as your data matures (many companies may begin with a poor understanding of their variable costs and marketing spend, for example). Over time, you can introduce segmentation, experimentation, and predictive modelling to uncover the drivers of retention and profitability.
CLV Analysis in Practice
Case Study: OpenAI
ChatGPT is a phenomenon. By July 2025, less than three years after launch, it had ~700 million weekly active users, or roughly 10% of the adult world population. According to The Information, there were 20 million paid consumer users as of April 2025, and 5 million paying business users as of August 2025. Very crudely extrapolating on trends, it is likely that paid consumer users were around 24 million by July 2025. This implies that around 4% of total users are paying.
Subscriptions are globally priced at $20, so monthly revenue from consumer users would be $20 multiplied by 24 million, which is $480 million. Annualised, this comes to $5.8 billion. If ~0.5–1% of paid users are on the $200 Pro tier, the incremental uplift is ~$22–43 million/month, for a total of ~$502–523 million/month, or ~$6.0–6.3 billion annualised. OpenAI’s revenue projection for 2025 is $12.7 billion, with the delta presumably coming from enterprise and API monetisation.
ChatGPT’s compute costs are highly variable with usage, and paid users are by revealed preference likely to use much more than free users. But paid users also have access to more compute than free users. If I assume that paid users have 10x the usage of free users, they are then responsible for 29% of consumer variable costs. Including business and API users changes the mix: assuming 5 million business/API users who each consume roughly 50× the compute of a free user, total usage becomes 676m (free) + 240m (paid consumer) + 250m (business/API). That would then imply that paid consumers account for ~21% of total compute, and business/API users for another ~21%, with the remaining 58% coming from free users.
Note that these figures are purely illustrative, this information is not publicly available.
Usage is highly relevant because of the high variable costs OpenAI incurs from inference. This sits on top of the usual variable costs a company incurs with more users, such as staffing and infrastructure. Here is my attempt at breaking down likely costs, based on extrapolating available public information.
Compute Costs
Per The Information, OpenAI spent $3 billion on training models and $4 billion on inference costs in 2024. However, a research note from Epoch AI estimates a lower $2 billion in inference costs, with a larger $5 billion allocated to R&D, which includes training. Assuming inference costs are fully variable, and training is mostly variable (roughly 70%), I can scale these Epoch figures by the growth in usage — from about 200 million users mid-2024 to 700 million mid-2025 (a 3.5× increase). This implies inference costs of roughly $7 billion ($2 billion × 3.5) and variable training costs of around $12 billion ($5 billion × 0.7 × 3.5). The Financial Times reported a cost base of approximately $12 billion in H1 2025, translating to a net loss of $8 billion after revenue. So $19 billion in variable training and inference costs across the full year seems plausible.
Staffing Costs
In 2024, The Information estimated that OpenAI employed around 1,500 people, with total staffing costs of approximately $1.5 billion including equity compensation. By August 2025, The Verge reported that headcount had reached 3,000. While simply doubling the 2024 estimate might appear reasonable, the Financial Times reported $2.5 billion in stock-based compensation in H1 2025, indicating that total personnel costs are likely closer to $4–5 billion for the year.
Since the real bottleneck for OpenAI is compute, not people, the majority of staffing costs are not variable. If I assume that only about 20% of staffing spend varies with user growth, variable staffing costs could come to $1 billion annually.
Calculating CLV
Annualised variable costs = $19 billion + $1 billion = $20 billion
Annualised revenue = $6.3 billion, or $12.7 billion if inclusive of enterprise/API.
With 24 million paid users responsible for 21% of the usage, paid users are responsible for $4.2 billion of the variable cost. Per user, that is $175 in annual variable costs, with annual subscription revenue of at least $20 * 12 = $240. The monthly marginal contribution of a paid user is then ($240 - $175) / 12 = $5.42. If I had an average retention time or churn rate, it would then be fairly trivial to extrapolate this figure to obtain CLV. For example, if the average retention time is expected to be five years, the Customer Lifetime Value would be, without time discounting, $5.42 * 12 * 5 = $325.202.
A few crucial caveats:
If I had made different assumptions along the way, the unit economics would be very different. It is hard to predict, with certainty, what inference costs will be given rapid technological progress. This analysis assumes essentially no optimisation of inference, nor any evolution in the product offering.
Given my set of assumptions, most usage comes from the free tier. Revenue from paid consumer and enterprise plans does not cover total costs, so OpenAI records a net loss. OpenAI’s customer acquisition has been mostly organic and product led, and so this free tier is likely to function as the largest component of the customer acquisition cost, which makes net CLV negative. There has been some speculation that OpenAI may try to monetise this tier via advertising.
Case Study: Hims & Hers
Hims & Hers (Hims) is a rapidly growing direct-to-consumer healthcare company, established in 2017 and publicly listed in 2021. The company is expecting full-year 2025 revenue of $2.4 billion, per guidance issued in its Q2 earnings. Its core business model is subscriptions, with 82 percent of revenue coming from recurring subscriptions. The monthly online revenue per average subscriber was $79 in the first six months of 2025, and they had 2.4 million subscribers at the end of Q2 2025.
Churn
Unlike the analysis for OpenAI, I will explicitly include churn in this analysis. Hims does not publish an overall churn rate, but the company has said that subscriptions lasting longer than a year churn at an annual rate of 85%. Since Hims is growing quickly, most subscriptions are likely to be less than a year old, and churn is usually higher during that first year. Prior analysis, derived speculatively from their quarterly earnings, suggests that Hims subscriptions may have an annual churn rate of 34% and a quarterly churn rate of 10%. These are the churn rates I will use for this analysis.
Variable Costs
Variable costs for Hims are likely minimal. They primarily comprise the cost of goods sold —the expense of preparing medications—which remains low given most drugs are generic and off-patent. Additional variable costs stem from scalable staffing needs, such as customer support during demand spikes. As a public company, Hims reported a Cost of Revenue of $284 million in H1 2025. This implies a gross profit of ($79 * 6 * 2.4 million) - $284 million = $853m, or a gross margin of 75%.
Calculating CLV
This is sufficient to calculate a gross CLV figure. If I decompose quarterly churn of 10% as a monthly 3.4% churn rate, the average lifetime of a subscriber is 1 / 0.034 = 29 months (2.45 years). Gross profit per subscriber over that time horizon is then $79 * 0.75 * 29 = $1,718.
Calculating Customer Acquisition Cost
Unlike OpenAI, most of Hims’ costs are not variable but instead tied to customer acquisition. The company relies heavily on marketing rather than organic growth, with 56 percent of total operating expenses in Q2 2025 spent on marketing. For a business like Hims, CAC is therefore a critical metric to watch. A rising CAC could indicate that Hims is running out of easily reachable customers or that the direct-to-consumer health market is becoming more competitive.
In Q2 2025, Hims spent $218 million on marketing and added 73,000 net subscribers. It is a common error to estimate CAC by dividing marketing spend by this net figure. Since churn is already accounted for in the CLV calculation, the correct approach is to use gross subscriber additions to avoid double counting.
Hims does not disclose gross subscriber additions, but they can be approximated. With 2.4 million subscribers at the end of Q1 and a quarterly churn rate of 10 percent, roughly 240,000 customers would have churned during Q2. Adding back these churned customers to the 73,000 net additions implies around 313,000 gross additions. Dividing marketing spend by this figure gives an estimated CAC of $700 per new subscriber in Q2 2025.
Calculating Net CLV
After accounting for acquisition costs, Hims’ net customer lifetime value (CLV) is estimated at roughly $1,000 per customer. While a CLV-to-CAC ratio of 2.5:1 falls short of the conventional 3x benchmark, it’s fairly strong for a direct-to-consumer subscription model, underlining the importance of interpreting CLV within its industry context.
That said, this estimate is pretty basic. In practice, CLV tends to evolve as the business matures and customer behaviour becomes more complex. For instance, upselling is a big part of the business model at Hims, with new customers entering through a single product category before expanding across multiple SKUs. As such, ongoing marketing investment also functions as a form of “re-acquisition” to deepen engagement and could be thus understood as, partially, a variable cost. Another simplifying assumption I have made is that churn rates will stay at the same. But if the customer base stabilises and the share of repeat subscribers increases, churn dynamics typically improve, further enhancing realised CLV.
In 2021, I would routinely travel in Ubers and pay single-digit fares, which is inconceivable today. These subsidies were not just demand side, as Uber used billions of investor dollars to attract drivers, and with it a large enough car pool to offer an instant service to customers and dominate regional taxi markets. Some drivers report making as little as a third of what they made when subsidies were in place.
ChatGPT retention is likely very strong. Publicly available cohort charts show that the first cohort of users (those who joined at or near launch) saw usage decline during 2023 but rebound in late 2024, reaching higher usage than ever by mid-2025. This means that early adopters not only stayed but became more active over time. This is very rare.