Product Management for Engineers · Page 6
Business Viability for a Developer Product: Stakeholders, Strategy, Pricing and Positioning
Cagan says product managers too often under-estimate or avoid business viability1, and for a runtime or SDK it decides whether the product survives as a business. This page covers stakeholder constraints, deciding under disagreement, four definitions of strategy, commoditizing complements, the pricing mechanics of PostHog and Stripe, and positioning. Afterwards you can name the viability questions a developer product must answer, derive which price metric fits a cost structure, read PostHog's rules as one company's policy and not a consensus, and tell evidence from inference on this page.
1. Viability: the risk with the least engineering in it
1.1 What the source claims, and who inherits the risk
Cagan's claim is that business-viability risks "are too often under-appreciated and under-estimated (or simply avoided) by the product manager"1. His list is open-ended: fit with the go-to-market or sales channel, constraints in partner contracts and legal compliance, whether "we can afford to cost-effectively acquire customers", whether "we can effectively monetize the product", and consistency with the brand promise, "as just a few examples". He assigns the risk to the product manager, who is "responsible for the value and viability risks, and overall accountable for the product's outcomes"1. Two limits apply. The claim is one practitioner's assertion, with no count of how often teams skip viability. And it is about product managers: nothing fetched for this track measures what engineers weigh.
What follows for an engineer depends on who holds the risk when the team has no product manager, and PostHog is the fetched case. Its handbook says that in early-stage products "a team usually doesn't have a PM yet" and "product engineers own all product decisions", that "it's typically the team lead's job to decide" three things, "What to build", "When to release the first version" and "How to charge for it", and that a PM is usually added "after a product has launched and started generating revenue"2. Under that model, pricing is a decision an engineer lead makes by default. Page 5 compares the role models; the narrower point here is that under at least one fetched model viability is not something an engineer can leave to someone else.
The engineer-side guidance is a prompt and not a method. Orosz's advice to engineers on user-facing product teams begins: "Understand how and why your company is successful. What is the business model? How is money made? What parts are most profitable, what parts of the company are expanding the most?"3. How a team then tests a viability hypothesis is described by SVPG's stakeholder article: prototypes shown to the people who own business constraints are "how your product partners demonstrate they understand your business constraints", and "more generally, this is how the product team assesses whether the proposed solution would be viable for the business"4. Section 2 takes that exchange apart.
1.2 Viability questions for a runtime or SDK, and the evidence behind each
| Item | Question for a runtime or SDK | What a fetched source says | Status |
|---|---|---|---|
| Go-to-market fit | How does a developer first run it, and who approves it for production? | "Developers generally adopt from the bottom up, but when teams reach a certain size, production-level deployment usually requires a top-down decision", with single sign-on (SSO), audit logs, intellectual-property indemnification and data-handling guarantees needed "before they'll approve a contract"5 | Vendor guidance, about AI coding tools |
| Contracts, compliance | Which licences, for models, SDKs or hardware-vendor tooling, limit what we may ship and sell? | Stripe lists indemnification and data-handling commitments as procurement requirements5. Nothing fetched covers third-party licence terms for a runtime. | Question only |
| Cost of acquiring | What does winning one customer cost, against what one customer pays? | A 2004 essay ties the price level to the sales channel and gives an unsourced cost per sale (Section 1.4) | Attributed estimate |
| Monetization | Which unit is charged, and does it track the vendor's cost and the customer's value? | PostHog's published rules, Stripe's five models, and the derivation in Sections 6 to 8 | Company policy; vendor guidance |
| Brand promise | Does the promise survive the free tier and the early failure modes? | PostHog lists ways to "limit brand damage" when shipping early: "internal use first, a feature flag rollout, messaging a couple of friendly users, not marketing it or limiting the marketing, or shooting for Minimum Lovable Product instead of Minimum Viable Product"6 | Company handbook |
1.3 Letting the market answer: precedent and revealed prices
PostHog's engineering handbook says that before product-market fit "We only build things that already have successful competitors with real revenue", and draws the implication that "'just build it' works disproportionately well because other people have already figured out that new product X has product market fit"6. Its pricing rule for an established market has the same structure: "roughly match the pricing of the cheapest big competitor", counting only a competitor "making actual revenue at significant scale"7. Each rule swaps a forecast for an observed fact. The forecast being avoided is a hard one. Spolsky argues in a 2004 essay that you cannot learn the demand curve by asking, because the answers swing from day to day, and that "the only way to determine how much someone will pay for something is to put it up for sale, and see how many people actually buy it"; even then, he says, you would need "something like 1,000,000 customers" before price experiments mean anything8. A competitor earning real revenue at scale has run that experiment already. That is the writer's reading of why the two PostHog rules are the same rule; PostHog states neither reason.
The rule needs precedent. For "innovative products where there is no established market", PostHog says it is "also the innovator on price", should not be "afraid to experiment", and should change pricing early if "the usage-based unit we picked is unappealing to customers"7. Whether a given runtime has such precedent is a market fact no fetched source supplies. The same handbook gives one cheap interview probe aimed at willingness to pay: it says to "look for evidence that users have actually done anything about the problems they say they have", and suggests asking "Has your company spent money trying to solve this before?"6.
1.4 Price decides the channel
Go-to-market fit is not separable from price. Spolsky's 2004 argument is that software is priced "free, cheap, and dear". Cheap is "$10 – $1000, sold to a very large number of people at a low price without a salesforce"; dear is "$75,000 – $1,000,000, sold to a handful of rich big companies using a team of slick salespeople". Nothing is priced between, he says, because above $1,000 "you need to get serious corporate signoffs", and "the cost of making one successful sale is going to average about $50,000"8. That figure is his estimate in 2004 dollars with no data behind it, so treat it as an illustration of the mechanism and not as a number to plan with.
The mechanism is arithmetic. A sales-led channel carries a fixed cost per won deal, \( C_s \). With deal value \( D \) and the fraction \( m \) of it left after delivery cost, the channel pays only if
$$ m\,D \;\ge\; C_s \quad\Longleftrightarrow\quad D \;\ge\; D^{*} = \frac{C_s}{m}. $$With his \( C_s = \$50{,}000 \), the lower bound is \( D^{*} = \$50{,}000 \) at \( m = 1 \). His own boundary of $75,000 implies \( m \approx 50{,}000 / 75{,}000 = 0.67 \); he does not state \( m \), so that is inferred. Below \( D^{*} \) the only viable channel is self-serve, which forces a price a buyer can approve without procurement. Stripe's 2026 guidance makes the same split from the buyer's side: if the buyer is "an individual developer or a team lead with a credit card, per-seat simplicity wins", while a platform-engineering team or vice president with a software budget "might reach for predictable annual spending and a clean contract structure", and "trying to serve both customers with the same model usually won't work"5. PostHog's price list carries both channels: a pay-as-you-go plan with a "$0/mo base price", and annual credit commitments discounted 20% from $25,000 of credit up to 40% at $500,000 to $999,999, with "Contact us" above $1,000,0009. The arithmetic shows the failure to avoid: a price chosen for self-serve adoption, sold into a channel that needs \( D^{*} \) per deal, loses money on every contract.
2. Stakeholders as suppliers of constraints
In SVPG's account, a stakeholder is anyone "responsible for some key aspect of your company's business", whether that is operations, profit and loss for a business unit, or a service group such as legal, finance or HR; "most importantly", it is someone who is "not directly part of the product organization, but are still reliant on that group to create technology solutions that support your business needs and constraints"4. What stakeholders supply is the other half of the problem. Constraints range "from go-to-market considerations, to industry regulations, to financials (both costs and monetization), to business partnerships", and the reason product work is hard is that "solutions must solve for both customers and also for the various dimensions of the business"4. The categories line up with Cagan's viability list in Figure 1.
2.1 The exchange, step by step
- Constraints shared. The stakeholder explains the business context: go-to-market, regulation, financials, partnerships. The article says product leaders "rely on you" for recommended reading, introductions and specific guidance4.
- Request reframed. Requests for features and projects, "no matter where they originate", become "problems to solve, along with a definition of success". The most useful thing to share is "the specific problem you need solved, who specifically you need that problem solved for, and how you will measure success"4.
- Access granted. Teams "require direct, unencumbered access to customers, users and data"; where privacy or governance limit access, "this can all be accomplished with tooling and authorization controls"4.
- Prototype reviewed. Before anything is built, the stakeholder sees "very quick to create simulations" and gives feedback "while changes are still easy and inexpensive to make"; the same prototypes go to users for usability and value and to engineers for feasibility4.
- Built and instrumented. Engineers build "a production-quality solution, which includes instrumenting the product" so that stakeholders and team "can immediately see whether the new offering is generating the necessary business outcomes"4.
- Iterate until the result is met. If outcomes fall short, "the product team will immediately investigate the reasons, and proceed to iterate until the desired business results are achieved"4.
Three details matter for an engineer. Step 5 is engineering work: evidence that viability held after launch exists only if someone logged it (page 4 covers what to measure). The article exempts "keeping the lights on", meaning essential reporting, compliance changes and critical fixes, from reframing and discovery, and says that if that load "climbs too high" the company must decide whether it can also move the business forward4. And dates are where the old model returns: a "high-integrity commitment" is "a date you can count on, but it requires some product discovery work", and since "these come at a cost" it "should be used sparingly"4; page 3 (Section 8.4) covers the mechanism.
2.2 What the article specifies and what it leaves open
| Question | What the article says | What it leaves open |
|---|---|---|
| What does the team owe? | Prototypes before building, instrumented outcomes, dates only as high-integrity commitments | No cost, duration or effort figure for any of them |
| Two stakeholders' constraints conflict? | A product manager works "simultaneously with multiple stakeholders" on solutions that "satisfy the various, sometimes conflicting, constraints" | No method for resolving the conflict; the article defers to the book TRANSFORMED |
| An executive asks for a specific feature? | Requests are reframed as problems "no matter where they originate" | No method for when the requester rejects the reframing |
| Is there evidence it works? | "Most companies that have experienced the difference will tell you of the power of true collaboration" | No data. The page points readers to SVPG's workshops and book. |
The article is written from the product-team side, and no fetched source presents a stakeholder's rebuttal. Its silence on conflict and on executive demands is the gap that matters most for someone about to be on the receiving end of both; Section 3 gives what the sources do supply, and Section 10 records what they do not.
3. When people still disagree: who ends the argument
Section 2 describes how a constraint reaches a team. It does not say what happens when the people involved still disagree about the solution or the constraint. Four mechanisms in the fetched sources answer, and they answer differently: one ends the argument by commitment, one declines to end it by role, one assigns a single owner per decision, and one is the failure mode of the rest.
3.1 Four mechanisms
Amazon's 2016 shareholder letter has the sharpest statement of the first. It says to "never use a one-size-fits-all decision-making process", that "many decisions are reversible, two-way doors" suited to "a light-weight process", and that "most decisions should probably be made with somewhere around 70% of the information you wish you had"10; page 4 (Section 8.1) turns that into an inequality. The conflict mechanism is the next paragraph: "If you have conviction on a particular direction even though there's no consensus, it's helpful to say, 'Look, I know we disagree on this but will you gamble with me on it? Disagree and commit?' By the time you're at this point, no one can know the answer for sure, and you'll probably get a quick yes." It runs both ways ("If you're the boss, you should do this too"): the example is Bezos committing to an Amazon Studios original he doubted because the team "wanted to go ahead", saving what he calls the slower cycle "if the team had actually had to convince me rather than simply get my commitment". He also says what it is not: not "these guys are wrong and missing the point, but this isn't worth me chasing", but "a genuine disagreement of opinion, a candid expression of my view, a chance for the team to weigh my view, and a quick, sincere commitment to go their way"10.
Eriksson's 2017 essay is the opposite pole on authority. He argues that "unless you're the founder and the product manager at the same time, you are not the CEO of anything", because product managers "don't have any direct authority over most of the things needed to make their products successful", and quotes Ken Norton: "You need to gain authority through your actions and your leadership skills, not your role"11. The failure he names is the manager who believes "their teams should just do what they're told", and he says such managers often produce "disastrous results", an assertion without examples or counts. He also does not let the absence of authority excuse the absence of accountability: "Just because you can pass the buck to someone higher up the chain of command, it doesn't mean you should"11.
PostHog's handbook supplies the third mechanism for exactly the decisions this page is about. It says that "when major calls (like positioning, pricing, or launch scope) need to be made, the PM and team lead collaborate closely to select an informed captain: One person who makes the decisions and sees them through. The other can provide feedback when requested, but focuses elsewhere"2. The fourth item is Bezos's warning about the thing PostHog also uses: "the process becomes the proxy for the result you want. You stop looking at outcomes and just make sure you're doing the process right", with the diagnostic question "do we own the process or does the process own us?"10. PostHog runs a yearly pricing evaluation that is "submit[ted] as an RFC for others to review"7; whether that process owns the team depends on whether outcomes are checked, and the role page does assign the PM to "track the impact of pricing changes on revenue, conversion, and churn to understand what worked and what didn't"2. That is the writer's reading of how the two documents fit; neither makes the connection.
| Mechanism | Who ends the argument | What it assumes | What it costs or leaves open |
|---|---|---|---|
| Disagree and commit10 | The person with conviction asks for a gamble; the other commits sincerely, in either direction | Wrong is cheaper than slow; bad decisions are "quickly" recognized and corrected | Needs candid disagreement first. The one example given is a success. No data. |
| Earn authority11 | No one by role; influence settles it | Time, and a team that follows influence | No procedure for a standoff; the harms of the alternative are anecdotal |
| Informed captain2 | One person per major call, chosen jointly by the PM and the team lead | The pair can agree on who captains | Not stated how the choice is made if they disagree about it |
| Process as safeguard107 | A defined procedure, such as a yearly pricing RFC | Outcomes are checked, so the process stays a means | "Well, we followed the process" defends a bad outcome |
3.2 Opinions can be committed to, facts cannot
None of the four is specific to a viability constraint, and one distinction is missing from all of them. A constraint is either a fact about the world, such as a licence clause or a compliance rule, or a judgment about the business, such as how much acquisition cost is tolerable. Disagree-and-commit settles disagreements of opinion, which is what Bezos's example is; it has no purchase on a clause that says what may be redistributed. In Figure 2 the prototype review is where a fact-type constraint shows up as a pass or fail and a judgment-type constraint becomes a decision someone can commit to. This split is the writer's reading, not a claim in any source.
3.3 What no fetched source supplies
The stakeholder article says product managers work with several stakeholders "so that they can come up with solutions that satisfy the various, sometimes conflicting, constraints", and stops4. It gives no method for two stakeholders whose constraints cannot both hold and none for an executive who demands a specific feature, and it defers to a book that was not fetched. The nearest mechanisms elsewhere on this track are page 4's guardrails, which turn a constraint into a pass-or-fail gate, and its exchange-rate derivation of an overall evaluation criterion, which puts two goals in one currency. Neither source presents either as a method for stakeholder conflict, so that pointer is the writer's. How to refuse an executive without losing the argument or the working relationship is not in the evidence this page rests on.
4. Four things called strategy
PostHog's handbook lists "Can't explain our company strategy" among the marks of a bad product engineer, and "Understand the company strategy, and prioritize based on this and what they believe users want" among the marks of a good one6. An engineer asked to do either meets a problem: the fetched sources use the word for four different objects. Rumelt means a particular structure of argument. Cagan's team-facing statement of direction fills the slot a strategy fills, without using the word. Kohavi, Tang and Xu mean something that can be written as a metric. Spolsky means a position among the products around yours. They are set side by side under "Where experts disagree".
4.1 Rumelt: coherent action backed by an argument
The text here comes from Rumelt's own June 2011 McKinsey Quarterly article, "adapted from his forthcoming book"12. The book itself, Good Strategy/Bad Strategy, was not read: the fetched catalog page gives only the publisher's description and a table of contents, which state the thesis, "Strategy should not be equated with ambition, leadership, vision or planning", and show chapters titled "Bad Strategy" (p. 32), "The Kernel of Good Strategy" (p. 77), "Sources of Power" (p. 95) and "Proximate Objectives" (p. 106), with page numbers from the Profile Books UK edition13. Treat anything about the book beyond that as book not fetched. The article's version of the thesis matches the blurb: "strategy is none of these. Rather, it is coherent action backed by an argument"12.
The kernel has three parts: "a diagnosis: an explanation of the nature of the challenge", which "simplifies the often overwhelming complexity of reality by identifying certain aspects of the situation as being the critical ones"; "a guiding policy: an overall approach chosen to cope with or overcome the obstacles identified in the diagnosis"; and "coherent actions: steps that are coordinated with one another to support the accomplishment of the guiding policy"12. His worked example is Nvidia's. Its first product, a PC add-in board for video, audio and 3-D graphics, "was a commercial failure"; a rival, 3Dfx, took the lead with gamers, and Intel was rumored to be preparing its own chip. The diagnosis was "We are losing the performance race". The policy, from CEO Jen-Hsun Huang's insight that releasing a chip every 6 months instead of "the industry-standard rate of every 18 months" would matter, was to "release a faster, better chip three times faster than the industry norm". The coherent actions were three development teams on overlapping schedules, "massive simulation and emulation facilities" to avoid delays in fabrication and driver development, and, over time, regaining control of driver development from the board makers12. Rumelt adds that good strategies are not permanent: Nvidia "has been challenged in the fast-growing mobile-graphics business, where cost is often more important than performance"12. For an engineer whose product runs on devices, that caveat is the useful sentence. A diagnosis is indexed to a market and a date, and the performance race of one decade is not the cost race of the next.
He condenses bad strategy to four hallmarks: failure to face the challenge, mistaking goals for strategy, bad strategic objectives, and fluff12. Each has a case: International Harvester's 1979 plan, which never mentioned its "grossly inefficient production facilities"; a "20/20 plan" of 20% annual growth and a margin of 20% or higher, defended with "the will to win"; a city plan with "47 strategies and 178 action items", the 122nd being "create a strategic plan"; and "customer-centric intermediation", which meant being a bank12. Of the two roots he gives, "the inability to choose" connects to Section 3. At DEC in 1992, in Rumelt's simplified telling with three executives, each argued for a different strategy ("Boxes", "Solutions", "Chips"); a majority preferred Solutions to Boxes, Boxes to Chips and Chips to Solutions, so consensus had no solution, and the group issued a statement Rumelt calls "a political outcome reached by individuals who, forced to reach a consensus, could not agree on which interests and concepts to forego"12. He does not claim a captain or a commitment rule would have saved DEC; the case shows that consensus on a strategic choice can be impossible, which is what Section 3's mechanisms are for. The cases are the author's own, with no counts or controls.
4.2 Cagan: what the team is given in place of a roadmap
Cagan's 2015 article never uses the word "strategy"14. What it supplies is the context a team needs: "a deep understanding of where the company is heading" and knowledge of "how their particular team is supposed to contribute to the larger purpose". It offers two components: a product vision, "the holistic view of what the organization as a whole is trying to accomplish", and "specific, prioritized business objectives for each product team", managed by OKRs, "my current favorite"14. A roadmap, he says, serves two purposes, working "on the highest business value items first" and holding "date-based commitments", and an alternative has to meet both14. Calling vision plus objectives his answer to "what is the strategy" is the writer's mapping, made because the article fills the same slot.
Set beside Rumelt, the difference is where the work sits. Cagan's example objective, with a key result of "average on-boarding time less than 3 hours" against a current 30 days, is hypothetical; it names a target, and the article leaves the diagnosis of why onboarding takes 30 days and the policy for shortening it to the team, which is told to "figure out the best way"14. In Rumelt's terms the objective is neither good nor bad until someone has done the diagnosis. In Cagan's model the team does it; in Rumelt's the strategist does, and Rumelt's tests for bad objectives are the "blue sky" restatement of the desired state and the "dog's dinner of goals"12. The tension on vision is explicit: the blurb says strategy "should not be equated with" vision13, and Cagan puts vision at the head of what a team is given.
4.3 Kohavi, Tang and Xu: strategy as the choice of metric
The experimentation textbook's position is operational. "Having a strategy is critical for running experiments: the strategy is what drives the choice of OEC"; "by encapsulating strategy into an OEC, controlled experiments can provide a great feedback loop for the strategy", and the OEC is "the perfect mechanism to make the strategy explicit and to align what features ship with the strategy" (Ch. 1, pp. 21-22). Guardrail metrics state "what the organization is not willing to change, since a strategy also 'requires you to make tradeoffs in competing – to choose what not to do'", a line they quote from Porter15. The book also gives a test that a strategy can fail: one experiment that misses the OEC may indict only a tactic, but "if many tactics evaluated through controlled experiments fail, it may be time to think about Winston Churchill's saying: 'However beautiful the strategy, you should occasionally look at the results'" (p. 23). Their case is Bing's two-year social-media strategy, abandoned after "over $25 million" with "no significant impact to key metrics" (pp. 23-24)15; page 4 covers the sunk-cost argument. The book says nothing about where a strategy comes from, so it complements Rumelt without replacing him: it checks a strategy after someone has chosen one.
4.4 Spolsky: position
The fourth object is a position among substitutes and complements, argued in a 2002 essay16. Section 5 gives it a full treatment because it is the only one of the four with a mechanism that can be written as arithmetic.
5. Commoditize your complements
5.1 The argument, as written in 2002
Spolsky's "Strategy Letter V" starts from two definitions. "A substitute is another product you might buy if the first product is too expensive." "A complement is a product that you usually buy together with another product", with gas and cars, hardware and operating systems, and babysitters and dinners out as examples16. The claim follows: "All else being equal, demand for a product increases when the prices of its complements decrease". From it he derives a corporate rule: a company's "strategic interest is going to be to get the price of their complements as low as possible", the floor being "the commodity price", "the price that arises when you have a bunch of competitors offering indistinguishable goods". Hence "Smart companies try to commoditize their products' complements", and "if you can do this, demand for your product will increase and you will be able to charge more and make more"16. One refinement matters for anything free: economists "consider the total price, including some intangible things like the time it takes to set up, reeducate everyone, and convert existing processes", so a zero sticker price is not a zero price16.
The essay is a verbal argument. It gives no demand data and no elasticities, it is dated June 2002, and its author hedges the Netscape case himself: "My argument is even more shaky because it's pretty clear that Netscape in the days of Barksdale didn't exactly know what it was doing"16. Read it as a mechanism, not as a finding.
5.2 A derivation of the mechanism (writer's toy model)
The mechanism can be made exact under simple assumptions. Customers buy one unit each of product A and its complement B, and the number who buy depends on the total price: \( Q = 1000 - (p_A + p_B) \). The vendor of A has a constant marginal cost \( c_A = \$100 \) and chooses \( p_A \) to maximize \( \pi_A = (p_A - c_A)\,Q \), taking \( p_B \) as given. Setting the derivative to zero gives \( Q = p_A - c_A \), so
$$ p_A^{*} = \frac{1000 - p_B + c_A}{2}, \qquad Q^{*} = \frac{1000 - p_B - c_A}{2}, \qquad \pi_A^{*} = \left(\frac{900 - p_B}{2}\right)^{2}. $$Each dollar taken off \( p_B \) raises A's best price by $0.50 and its volume by 0.5 customers. At \( p_B = \$600 \) the vendor of A prices at $250, sells 150 and earns $22,500. At \( p_B = \$300 \) it prices at $400, sells 300 and earns $90,000, four times as much. At \( p_B = 0 \) it prices at $550, sells 450 and earns $202,500. That is the sentence "demand for your product will increase and you will be able to charge more and make more" with numbers: price up, volume up, profit up. Two caveats belong to the model, not to Spolsky. The parameters are illustrative; nothing here is measured. And the strategy has a cost the model leaves out: from $600 to $300 the gain is $67,500, so spending to push B's price down pays only if it costs less than that. The vendor of B loses what A gains, which is why B's vendor resists.
5.3 The examples, and where the author hedges
| Company | Sells | Wants cheap | Move | Reservation |
|---|---|---|---|---|
| IBM (PC design) | PCs | Add-in cards, drives, printers | Off-the-shelf parts, "carefully documented" interfaces so other manufacturers "could join the party" | None stated |
| Microsoft | The PC operating system | PCs | Licensed PC-DOS non-exclusively; the PC "was basically a commodity" | None stated |
| Netscape | Servers | Browsers | Gave the browser away as "freeware"; later released Mozilla as open source | "even more shaky"; management may have been "technologically inept" |
| IBM (2002) | IT consulting | Enterprise software | Pays for open-source development | None stated |
| Sun and HP | Desktop hardware | Windowing systems | Paid Ximian to work on Gnome | Sun's Java commoditizes the hardware Sun sells, "the last thing they want to do" |
One observation in the essay transfers to software that sits on top of hardware. "It's easy for software to commoditize hardware (you just write a little hardware abstraction layer, like Windows NT's HAL, which is a tiny piece of code), but it's incredibly hard for hardware to commoditize software", because "even when the price is zero, the cost of switching from Microsoft Office is non-zero" and "one tiny difference and you lose your commodity status"16. The asymmetry rests on switching costs, and the essay measures none.
5.4 Mapped onto an inference runtime
The mapping below is not in any source. It applies the essay's definitions to a runtime vendor and stops where the premises stop. A runtime and the models it executes are complements, since they are used together; two runtimes are substitutes. By the essay's logic the runtime's demand rises as the price of models falls and as device silicon becomes interchangeable, which is what a hardware-abstraction layer does. The same logic gives each complement's vendor, whether silicon or model provider, an interest in a cheap runtime, as it gave Sun and HP an interest in cheap windowing systems. The asymmetry favors the runtime only if switching away from it is costly for its users. Whether any of that holds depends on facts the sources do not contain: how large the switching costs of a runtime are, what silicon vendors charge for their own runtimes, and how elastic demand for on-device inference is.
6. PostHog's pricing rules as a mechanism
PostHog sells an open-source product and data tools platform, and publishes both its pricing principles and its price list79. That makes it the one fetched case where a developer-product company's pricing policy can be set beside the artifact it produced. Everything in this section is one company's stated policy, as of 2026-10-26. No page fetched gives revenue, margin, churn or conversion figures, so nothing here shows that the policy works; it shows what the policy is, whether the price list agrees with it, and what arithmetic follows from it.
6.1 The rules and the reason each one states
The handbook opens with five properties PostHog wants from its pricing: generous free allowances for hobbyists and pre-product-market-fit teams, the chance to try the product before paying, starting to pay "on their own, with few hurdles", paying "transparently" for the value received, and making it "a no-brainer to pick PostHog over other competitors"; its goals include keeping engineers "as close to our customers as possible" and tying "our success with that of our customers'"7. The runbook that follows is a list of rules. Table 5 states each with the reason the handbook gives and, in the last column, the assumption about cost or value that the rule needs. The last column is the writer's reading.
| Rule | Reason the handbook gives | Assumption it needs |
|---|---|---|
| Usage-based by default | "The more a customer uses a product, generally the higher their bill should be", because "as our costs for providing a product scale through increased usage, so does our revenue" | Cost and value both rise with the metered unit |
| Flat add-ons only where cost does not scale | Platform packages give larger customers "flexibility & controls (e.g. RBAC), without hindering individual user adoption (vs. a per-seat price)"; "extended retention should be charged usage-based, not as a flat-fee add-on" | The test for flat versus metered is whether vendor cost scales with use |
| No charge for extra users | "we shouldn't (and don't) charge for extra users in an organization because the more people we get inside PostHog, the better" | A seat count taxes the spread that drives word of mouth |
| Each product priced separately | Lets a one-product customer "compare each of our products to our competitors'"; fundamental dependencies are bundled into the product that needs them | Separate prices keep each product's value and cost visible |
| A path to pay for itself | "We might not fully understand our cost base"; "a good pricing structure allows us to bring costs down over time with pricing staying fixed, thus increasing our margins" | Cost knowledge is incomplete at launch; margin arrives as cost falls under a fixed price |
| Match the cheapest big competitor | "Roughly match" it "so long as the unit economics make sense"; the rival must be "making actual revenue at significant scale"; "the total ACV is higher even if the per-product ACV is lower" | A multi-product vendor recoups a thin per-product margin elsewhere |
| Core or sticky features free or near cost | An agentic assistant should be charged "close to cost, so we don't put users off"; cohorts are free because their cost is "partially covered through other products" | Some features exist to raise usage or switching cost, and are priced as that |
| Free allowance as a one-way door | It should let customers "experience the product before paying for it"; "it's easy to increase the free tier for existing customers, but it's very painful to decrease it" | Changes to the free tier are asymmetric in customer reaction |
| New features paid by default; costly ones priced in beta | "Full usage is only available on the paid plans"; features that must be experienced get "a reasonable limit"; "high costs or low margins (e.g. AI products)" are priced "even during beta" | Free is a cost with no ceiling unless it is bounded |
| Enterprise features gated | "extra security, permissioning, compliance, or other enterprise-style upgrades" go in the enterprise tier | Governance tracks buyer type, not usage |
| Grandfathering is time-boxed | "Grandfathering is expensive for us to maintain"; self-serve payers get "a limited transition period (e.g. three months)" | Each legacy plan is a maintenance cost |
| Yearly evaluation, notice on rises | Each product reviews "competitive market", "new features" and "costs & margins" as an RFC; rises need "plenty of advance warning (e.g. 60 days)" | Cost and competition drift, so over- and under-charging both need a scheduled check |
PostHog's usage documentation states the first rule's logic in one sentence: "The more you use PostHog, the more value you get, and the more it costs us to process and store your data. Thus, we charge based on usage"17. That sentence carries the whole framework of Section 8: the metered unit is chosen because both cost and value track it.
6.2 The rules as a decision flow
The handbook calls its list a "runbook for deciding which plans, and at what limits, features should be assigned to"7 and does not give an order. Figure 6 arranges the rules as a sequence of gates for a new feature, which is the writer's arrangement: each gate is a rule from Table 5, and the order is the one in which a question rules out the later ones. A feature that is core to the platform skips the rest; a governance feature goes to the enterprise tier whatever its cost; only a feature that is neither gets a metered or flat price, then a free allowance or paid-only access, then a price level.
6.3 Does the price list follow the policy?
A policy is easier to trust if the artifact it governs agrees with it. Table 6 sets each rule against the price list as fetched on 2026-10-26, and marks the ones no fetched page can check.
| Rule | What the price list shows | Check |
|---|---|---|
| No per-seat charge | "There are no per-seat charges. Your whole team can use PostHog." | consistent |
| Free allowance on each product | First 1,000,000 events; 5,000 recordings; 1,000,000 feature-flag requests; 100,000 exceptions; 1,500 survey responses; "every product has a generous free tier" | consistent |
| Dependencies bundled | Web Analytics is "included with Product Analytics at no extra cost"; Experiments is "bundled with Feature Flags — same billing, same free tier" | consistent in spirit; the handbook's own example is a different event |
| Flat add-ons only where cost does not scale | Boost $250/mo (SSO enforcement, HIPAA BAA, unlimited projects); Scale $750/mo (priority support, SAML); Enterprise $2000/mo (RBAC, dedicated support, training) | consistent; the handbook's example of RBAC sits in the top package |
| Retention charged by usage | Logs custom retention costs $0.05 per GB-month and "bills for the data you keep, not this month's usage" | consistent |
| Costly AI features near cost | PostHog AI: first 500 credits free, then $0.01 per credit. PostHog Desktop: "billed by AI token usage at cost" | stated, not checkable without cost data |
| Match the cheapest big competitor | No competitor prices appear in any fetched page | not checkable |
| Each product pays for itself | No margin or cost figure is published | not checkable |
PostHog's documentation adds one claim about the free allowance: "Our large free tier means more than 90% of companies use PostHog for free"18. It is the company's statement, with no method or date, so it is not evidence about customers; if true, it means the free tier is the product for most users and the paid brackets are paid by a minority.
6.4 A worked derivation: what the event brackets cost
Product analytics is billed per event in seven brackets: the first 1,000,000 free, then from $0.00005 per event up to 2,000,000 down to $0.000009 above 250,000,000, with brackets ending at 15, 50, 100 and 250 million9. The Product Analytics pricing page shows, for each bracket, an allocation, a price and a subtotal18, which means each event is charged at the price of the bracket it falls in: a graduated schedule in the sense Section 7.2 defines. With bracket bounds \( U_k \) and prices \( p_k \), the monthly bill for \( N \) events is
$$ \text{bill}(N) = \sum_{k} p_k \,\max\bigl(0,\ \min(N, U_k) - U_{k-1}\bigr), \qquad U_0 = 0. $$For \( N = 10{,}000{,}000 \): \( 1{,}000{,}000 \times \$0 + 1{,}000{,}000 \times \$0.00005 + 8{,}000{,}000 \times \$0.0000343 = \$50.00 + \$274.40 = \$324.40 \). Table 7 gives the bill, the average price per million events and the marginal price at ten volumes, computed from the list.
| Events per month | Bill (USD) | Average (USD per million) | Marginal bracket (USD per million) |
|---|---|---|---|
| 1,000,000 | 0.00 | 0.00 | 0.00 (free) |
| 2,000,000 | 50.00 | 25.00 | 50.00 |
| 5,000,000 | 152.90 | 30.58 | 34.30 |
| 10,000,000 | 324.40 | 32.44 | 34.30 |
| 15,000,000 | 495.90 | 33.06 | 34.30 |
| 50,000,000 | 1,528.40 | 30.57 | 29.50 |
| 100,000,000 | 2,618.40 | 26.18 | 21.80 |
| 250,000,000 | 4,868.40 | 19.47 | 15.00 |
| 500,000,000 | 7,118.40 | 14.24 | 9.00 |
| 1,000,000,000 | 11,618.40 | 11.62 | 9.00 |
One reading follows. The price list's tagline is "the more you use, the cheaper it gets"9, and it is exactly true of the marginal price: from $50.00 per million in the first paid bracket to $9.00 above 250,000,000, which is 82% lower. It is not true of the average price, which is what a customer experiences as cost per event. The average rises from $25.00 per million at 2,000,000 events to a peak of $33.06 at 15,000,000, 32% higher, and only then falls, because the free first million dilutes the average at low volumes and every bracket up to 15,000,000 costs more per event than the average so far. Counting only paid events, the average never rises. A team that doubles from 5,000,000 to 10,000,000 events pays 2.12 times as much (324.40 divided by 152.90), not twice and not less. PostHog states no reason for its brackets, but its contract rules give one for volume discounts on annual commitments: "Our cost to serve a $500k customer is not 10x a $50k customer, so we can share those savings"19. Applying that reasoning to the brackets is the writer's reading.
6.5 When the billable unit meets the code
A price metric is also a specification, and feature flags show it. Flag requests are billed per request, with 1,000,000 free per month9. PostHog's usage documentation says that with local evaluation "each poll for feature flag definitions is charged as 10 feature flag requests when the definitions changed, and as 1 request when they didn't", that polls are made "every 30 seconds by default", and that a constantly running server whose flags rarely change is charged "about 2 * 60 * 24 * 30 = 86,400 feature flag requests each month, plus 10 requests for each poll that returns new definitions". The one-request price depends on the SDK sending the ETag of the definitions it holds so that the server can answer "304 Not Modified"; "SDK versions without ETag support are charged 10 requests for every poll", which "adds up to 10 * 2 * 60 * 24 * 30 = 864,000 feature flag requests"17.
Set against the allowance, one server polling at the default interval uses 8.64% of the free million with ETag support and 86.4% without. Billing is organization-wide, "one bill with one free allowance"20, so two servers use 172,800 requests with ETag support, inside the allowance, and 1,728,000 without. The latter is 728,000 above the free million, which at the second bracket's $0.0001 per request9 is 728,000 × $0.0001 = $72.80 per month, by my arithmetic from the list. Identical flags and identical polling cost $0 or $72.80 depending on whether the client library supports a conditional request. For an SDK vendor the implication is that adding conditional-request support is a price cut, and that deciding what counts as one unit is a product decision made in code. Predictability, the cost Stripe attaches to usage pricing (Section 7), is handled here with two instruments: the free tier gives "an accurate projected volume in a few days"17 before any payment, and billing limits cap spend, with the price that "PostHog drops data after you go over your billing limit"20. Cost certainty is bought with the risk of lost data, and the customer carries that risk.
6.6 Who decides, and what is not shown
The pricing page names no owner. The role page does. A PM "owns" pricing for a product: reviewing "whether pricing aligns with the value users get", identifying when "pricing or packaging creates friction for adoption, retention, or expansion", designing any change "taking into account our pricing principles" with the billing and marketing teams, and tracking "the impact of pricing changes on revenue, conversion, and churn"2. For early products without a PM, the team lead decides "how to charge for it" (Section 1.1), and the handbook's example project for the stage at which a PM decides the pricing model is the pricing of PostHog AI2. The review is yearly and goes to an RFC7.
7. Stripe's five models and the arithmetic of tiers
7.1 Five models and their stated trade-offs
Stripe's guide to pricing AI coding tools names "five basic pricing models", though "many mature products blend elements of multiple models"5. It is a billing vendor's guidance for one product category, dated 19 April 2026. It gives no prices, margins, conversion or churn figures, and Stripe sells billing infrastructure. Table 8 records what it says of each model.
| Model | What is charged | Stated strength | Stated weakness |
|---|---|---|---|
| Per-seat | A fixed fee per developer, whatever the usage | "Simple to procure and easy to budget for" | "Doesn't reflect compute costs or value delivery"; "the lighter users are subsidised by the heavier ones" |
| Tiered subscriptions | Individual, team and enterprise plans | Team adds admin controls and central billing; enterprise adds governance (SSO, SAML, audit logging, data residency, IP indemnification) | Tiers that differ only in seat count and support SLA "don't give buyers a reason to upgrade" |
| Capability-based | More for better models, larger context windows, advanced workflows | Suits teams where output quality "directly affects downstream work" | Needs "obvious differentiation between tiers, or they risk upsell reversals or customer loss" |
| Usage-based | Requests, tokens or credits | "Matches costs with actual use"; "works best when usage varies widely across customers" | Predictability: the discomfort "often slows adoption, even when average costs are lower than a flat seat fee would be" |
| Hybrid | A flat seat for standard use plus a usage layer for compute-intensive or agentic work | "The seat price stays predictable for the customer, while the usage component keeps the vendor's economics intact" | Not stated |
The guide's selection questions are four: how buyers budget; the "defensible value metric", "the thing your customers would pay to keep if you raised prices"; whether the model can "withstand power usage", since with "users who might run 1,000 agentic tasks in a month, a flat seat fee will eventually create a margin problem" and the usage component should be built in "before you need it"; and whether each upgrade delivers "something the previous tier genuinely couldn't". It adds that "many pricing mistakes aren't apparent until 12 months in", an assertion with no data behind it. The survey statistic it opens with has no method attached and is not repeated here5.
7.2 Volume and graduated tiers differ in what they bill
Stripe's documentation sorts the same space differently: flat rate, per-seat, tiered ("volume-based pricing" or "graduated pricing"), and usage-based, whose forms include "fixed fee and overage, pay as you go and credit burndown"21. The two tiered modes differ in a way that matters to an engineer implementing a meter. In volume pricing "the entire quantity (or usage) is multiplied by the unit cost of the tier" reached; in graduated pricing "the totals for each tier are summed together". The documentation's example has 1 to 5 fonts at 7 USD, 6 to 10 at 6.5 USD and 11 or more at 6 USD: 5 fonts cost 35 USD under both, 6 fonts cost 39 USD under volume pricing (6 × 6.5) and 41.50 USD under graduated pricing (35 + 6.5)22. The documentation warns that under volume pricing "the total might decrease when calculating the final cost". The condition is simple. Moving from \( n_k \) to \( n_k + 1 \) units across a boundary lowers the total exactly when
$$ (n_k + 1)\, p_{k+1} \;\lt\; n_k\, p_k \quad\Longleftrightarrow\quad \frac{p_{k+1}}{p_k} \;\lt\; \frac{n_k}{n_k + 1}. $$In the documentation's example the ratio at the first boundary is \( 6.5 / 7 = 0.929 \), above \( 5/6 = 0.833 \), so the total rises. A steeper cut would not: with a hypothetical 5.5 USD for the sixth unit onward, 6 units would cost 33 USD against 35 for 5. Graduated pricing cannot do that, because each unit is charged at its own tier's price. A volume schedule with a steep cut therefore pays a customer to buy more, or to split or batch usage across the boundary. PostHog's per-bracket calculator18 is the graduated kind, which is the safer one for a metered developer product.
7.3 Where governance sits in three real price pages
| Source | Unit that scales the price | What bounds the free tier | Where identity and governance sit | Compute-heavy features |
|---|---|---|---|---|
| Stripe's pattern | Seat, plus usage for agentic work | "Limited completions or model access" | Enterprise tier: SSO, SAML, audit logs, IP indemnification | Seat plus agent-usage add-on, or capability tier |
| Linear | Per user per month: $10 Basic, $16 Business, billed yearly; Enterprise custom | 250 issues, 2 teams, 10MB uploads; members unlimited | SAML and SCIM, granular admin controls at Enterprise | "Coding sessions" and "Loops" footnoted "Requires AI credits" |
| PostHog | Per-unit usage; "no per-seat charges" | A free volume on every product | Flat packages: SSO enforcement $250/mo, SAML $750/mo, RBAC $2000/mo | PostHog AI at $0.01 per credit |
All three put identity and governance behind a higher fee and differ on the unit that scales. Linear is a seat-priced developer tool whose compute-heavy features carry a credit footnote, which is the hybrid shape Stripe describes; PostHog avoids seats altogether. Neither company's page says why it chose its unit, and the cost structures are not published, so which is better is not a question the sources can answer.
8. Choosing a price metric from the cost structure
8.1 Three questions for any candidate unit
The sources never state a procedure for choosing a metric, but their rules imply three questions to ask of any candidate unit \( u \): a seat, an event, a request, a token, a device. Does the vendor's cost scale with \( u \)? PostHog reserves flat fees "for functionality where cost doesn't scale with usage"7, and Stripe's margin warning is the case where cost scales and the price does not5. Does the customer's value scale with \( u \)? PostHog's documentation rejects monthly active users on this ground: "Some users are very valuable. They use your product a ton, they respond to your feedback surveys, and they store a lot of data in your platform. Other users hit your landing page and bounce immediately. From a value perspective, you shouldn't need to pay the same amount for both of these users"17. And does charging for \( u \) tax something you want more of? PostHog does not charge for extra users because "the more people we get inside PostHog, the better"7, while a per-event meter taxes instrumentation, which is why its documentation has a section on how "to reduce your PostHog costs" by capturing fewer events17. The third question is the writer's addition, drawn from those two statements: a metric is an incentive on the customer's behavior.
8.2 What a flat seat does to margin: a derivation and a calculator
Stripe's statement that a flat seat fee "will eventually create a margin problem" under heavy use5 follows from arithmetic. Let seat price be \( s \), vendor cost per task \( c \), and customer \( i \) use \( u_i \) tasks. Margin on that customer is \( s - c\,u_i \), negative whenever \( u_i > s/c \), and the blended margin is \( s - c\bar u \). Under a usage price \( p \), margin is \( (p - c)\,u_i \) and the margin rate \( (p - c)/p \) does not depend on the mix. A hybrid with base \( b \), \( q \) included tasks and overage \( o \) bills \( b + o\max(0, u_i - q) \). The calculator uses illustrative numbers, none from a source.
Seat, usage and hybrid pricing against the same customers (illustrative)
Formulas: flat \( (s - c\bar u)/s \); usage \( (p - c)/p \); hybrid \( \text{revenue} - c\bar u \) over revenue. At the defaults the flat seat earns 52.0% blended and loses 20.00 USD on each heavy user.
At the defaults the light majority hides the loss: 80% of customers cost 2.00 USD against a 20 USD seat. Double the heavy tier to 4,000 tasks and the blend falls from 52.0% to 12.0% (mean usage 880 tasks, cost 17.60 USD per seat) while the median customer has not changed. Usage pricing holds 60.0% but the heavy bill doubles to 200 USD, which is the predictability cost Stripe names. The hybrid holds 46.3% with the heavy bill at 124 USD. These are the three numbers behind "will eventually", and none is evidence about any real product.
8.3 The value side: Spolsky's toy numbers, recomputed
Cost fixes a floor; demand fixes the rest, and demand is the part that cannot be observed. Spolsky's 2004 essay builds a price from an assumed downward-sloping demand curve and an incremental cost of $35 per copy, treating the $250,000 of development as sunk. His own caveat is stated first: "These numbers are not, of course, real"8. His reported totals can be recomputed from the inputs he gives.
| Scenario | Customers | Contribution recomputed (USD) | Essay states (USD) |
|---|---|---|---|
| One price, $220 | 233 | 233 × 185 = 43,105 | 43,105 |
| Top 42 pay $349 | 233 | 43,105 + 42 × 129 = 48,523 | "about $48K" |
| Plus 217 more at $99 | 450 | 48,523 + 217 × 64 = 62,411 | "$62K" |
| Each pays own maximum | not stated | not recomputable | "almost $80K" |
The arithmetic matches, which shows only that the example is internally consistent. His cautions are the useful part, and each is an assertion without data. Segmentation breeds resentment, and "customers would rather pay $100 when everyone else is paying $100 than pay $79 if they know there's someone out there who got it for $78". An "unlimited" site license, he says, means "giving a gigantic gift of consumer surplus to the least price-sensitive customers". A price absent from the website pushes the biggest buyers to bargain at quarter-end. Most basically, "you can't really find out what the demand curve is", because focus-group answers vary and "the only way to determine how much someone will pay for something is to put it up for sale"8. PostHog sits on the opposite side of several of these: it publishes every bracket, and its contract rules say "discounts are earned, not given"19.
8.4 The on-device inversion
In every source here, the cost that makes usage pricing necessary is the vendor's own: PostHog processes and stores the customer's data17, and Stripe's examples are tools whose agentic workflows "can consume much more compute than AI autocomplete does"5. For a runtime that executes on the customer's device, the compute is the customer's. If the vendor's marginal cost per inference is near zero, the first question of Section 8.1 answers "no" for every candidate unit, the left column of Figure 9 opens, and the choice rests on the value question and on what the vendor can observe. Usage pricing also needs metering: PostHog bills the events it receives, while an on-device runtime sends the vendor nothing unless it is built to, so metering becomes an engineering and trust decision before it is a pricing one. None of the units one would try, per device, per app, per active user, per model shipped or per inference, appears in a fetched source for on-device software; PostHog's argument against monthly active users is the one fetched argument about a per-user unit. This page therefore gives no price, margin or licence figure for an on-device runtime, and none is implied.
9. Positioning, briefly
book not fetched The source assigned to this topic is the author's book page for Obviously Awesome, which is marketing copy. It promises "a step by step process that any startup can follow", says positioning breaks into "five component pieces" and "three styles of positioning", and names none of them; it says "newly updated and expanded edition", so at least two editions exist, and no year is stated here24. The component names below are from Dunford's own March 2021 post summarizing the book25, which is a blog post and not the book.
She defines positioning as "how your product is a leader at delivering something that a well-defined set of customers cares a lot about", built from five components: competitive alternatives, differentiated capabilities, value for customers, target customer segmentation and market category25. The components depend on each other, and the method fixes where to start: "the starting point had to be competitive alternatives", meaning "what would customers do if our solution didn't exist", which often is the status quo and not a named competitor. Then come "What do we have that the alternatives do not?", "So what for customers?", who cares most, and last the category. She warns against "phantom competitors", ones that "theoretically could compete with you; you just never actually see them or lose to them in deals"25. It matters on this page because the category carries price expectations: it "sets off a really powerful set of assumptions about who your product competes with, what features your product should have, who the product is intended for, and even things like what the product should cost"25. Matching the cheapest competitor, as in Section 6, amounts to accepting the category's price; that link is the writer's reading.
Her opening example is the nearest thing in the sources to on-device software. A product sold as a "Microsoft Access killer" had "barely sold 200 copies"; phoning 100 customers, she found that "94 didn't even know they had it. 6 had transformed their business with it"; repositioned as an "Embeddable Database for Mobile Devices", it "took off"25. It is one author's account of an event about twenty years before she wrote it, and the $100 per copy that her records show in the story is not evidence about runtime pricing. PostHog's handbook lists "shape initial positioning" ("who the product is for and why it matters") among typical PM work at stage 12. No fetched source applies Dunford's method to an infrastructure runtime, and her book page describes B2B startup products; the first step, asking what customers would do without the product, has an answer the sources do not contain.
10. Where the evidence ends
10.1 The infrastructure and platform-team gap
A runtime is the product type the fetched sources cover worst. Orosz's product-minded-engineer page says nothing about infrastructure or platform teams, and its tips begin "If you work on a user-facing product"3. The PostHog role page has the one sentence in any fetched source: "When a team is working on an infrastructure-heavy product, it's more important that the team lead has strong engineering and infra skills than product skills. In that case, it can make sense to add a PM early who can focus on product direction, positioning, and user research while the lead focuses on technical execution"2. That is one company's staffing rule for an externally sold product, with no data. book not fetched The first-edition table of contents of Cagan's Inspired lists "Chapter 39. Keys to Platform Products"26, a pointer to a 2008 chapter whose text was not fetched and which may not survive in the second edition. Whether any mechanism on this page transfers to an internal platform team, which has no price and captive customers, is not answered by any fetched source. The closest analogs the sources offer are developer-product companies with external customers: PostHog and Linear.
10.2 An evidence map for this page
Areas the sources here could not carry, and that need a course with a second research pass and not another page: willingness-to-pay research methods, inference unit economics and cost of goods, go-to-market motions for developer products, opportunity sizing and business cases, strategy frameworks beyond the four above, platform-product management, and product careers.
Where experts disagree
Two disagreements run through the sources: how to price a developer product, and what the word strategy refers to. No fetched source compares any position with another on outcomes, so each is stated by its mechanism and its own evidence.
Pricing a developer or infrastructure product
Meter it, publish it, match the rival
PostHog handbook and price list79
Usage-based by default, no per-seat charge, anchor to the cheapest big rival earning real revenue, a free allowance on every product, a yearly review.
- One unit carries both cost and value; the list is public
- The free tier doubles as a metering instrument and a growth channel
- One company's policy; no outcome data
- The rival rule leans on cross-selling; billing limits buy predictability by dropping data
A seat with a usage layer, governance behind the top tier
Stripe's guide5; Linear's plan page as a seat-priced example23
Per-seat is simple but cross-subsidizes heavy users, so a hybrid keeps the seat predictable and the vendor's economics intact; identity and governance features mark the move to enterprise.
- Fits buyers who need a predictable budget
- Tiers follow procurement stages
- Qualitative guidance from a billing vendor, no numbers; hybrid's downside unstated
- Linear's costs and results are not published
Price is a signal; the lever is the stack
Demand cannot be observed by asking, price sends a message, segmentation and unlisted prices backfire, and a company gains most by making its complements cheap.
- Honest about what cannot be known
- Points to actions that raise demand rather than to a price
- Essays without data, from the desktop-software era
- Silent on subscriptions and metering
The first two differ on where predictability comes from: PostHog meters and caps, Stripe puts a seat under the meter. Both default to cost-tracking and both gate governance behind a higher fee. The third answers a different question: the price level, and which lever raises demand. Which is right depends on three things no source measures: whether the vendor's cost scales with use (Section 8), how the buyer budgets (Section 1.4), and whether a rival with real revenue exists to anchor on. For an on-device runtime the cost condition may fail (Section 8.4, the writer's unverified inference), which would remove the main argument for the first two and leave the third. The evidence that would settle it is margin and churn by pricing model after the first renewal cycle, by cohort; none is in the sources.
What a strategy is, and where it comes from
Coherent action backed by an argument
A diagnosis, a guiding policy and coordinated actions; not ambition, vision or planning.
- A concrete test; names how strategies fail
- The cases are the author's; the book was not read
Vision plus ranked objectives per team
Cagan14
Not defined as strategy; what teams are given in place of a roadmap.
- Transmits direction to autonomous teams
- Example is hypothetical; leaves the diagnosis to the team
Strategy as the choice of metric
Kohavi, Tang and Xu15
The strategy drives the OEC; guardrails state what is not traded; results can falsify it.
- Measurable and falsifiable
- Presupposes a strategy exists
Position among complements
Spolsky16
Make your complements cheap so demand for your product rises.
- A mechanism that can be written as arithmetic
- 2002; no data; the author hedges
These answer different questions. Rumelt says what makes a strategy good, Cagan how a team receives direction, Kohavi et al. how a strategy is checked, Spolsky where an advantage comes from. They collide only on vision: the blurb says strategy is not vision, and Cagan puts vision first. An engineer can use them together: Rumelt's test on the claim, Kohavi's metric on the result. The evidence that would settle the vision question, a comparison of organizations with and without a stated diagnosis, does not exist in the sources.
What is changing
Every pricing source here is likely to change within a year. Stripe's guide, last updated 19 April 2026, argues that agentic use is straining flat seats5; Stripe's documentation now says "Metronome is now part of Stripe" and names it the recommended starting point for new usage-based billing27; Linear's page already footnotes AI credits on a seat-priced plan23. PostHog's product team lists, among its Q2 2026 themes, "tracking human vs agent actions, and defining success for each"28: if agents generate the events and requests that a meter counts, the meaning of a usage unit shifts.
References
- blogCagan, M. The Four Big Risks. Silicon Valley Product Group, 4 December 2017. https://www.svpg.com/four-big-risks/. Accessed 2026-10-26. Used for: the five viability examples, the product manager's ownership of value and viability risk, and the claim that viability is "under-appreciated and under-estimated".
- officialPostHog. What product managers do at PostHog (Handbook). Living page as accessed. https://posthog.com/handbook/product/product-manager-role. Accessed 2026-10-26. Used for: PM ownership of pricing, the team lead deciding "how to charge" on early products, the informed-captain rule, the infrastructure-heavy-product sentence, and positioning at stage 1. Added beyond the source map.
- blogOrosz, G. The Product-Minded Software Engineer. The Pragmatic Engineer, 5 September 2019, updated 1 February 2022. https://blog.pragmaticengineer.com/the-product-minded-engineer/. Accessed 2026-10-26. Used for: the "understand how and why your company is successful" tip, and the page's scope (user-facing products).
- blogJones, C. and Cagan, M. Stakeholders and the Product Model. Silicon Valley Product Group, 1 December 2025. https://www.svpg.com/stakeholders-and-the-product-model/. Accessed 2026-10-26. Used for: the stakeholder definition, the three foundations, prototypes as the viability check, instrumentation, keep-the-lights-on, high-integrity commitments, and the absence of a conflict method.
- blogStripe. Pricing models for AI coding companies: What works at each stage of tool adoption. Last updated 19 April 2026. https://stripe.com/resources/more/pricing-models-for-ai-coding-companies. Accessed 2026-10-26. Used for: the five models and their stated trade-offs (Table 8), the selection questions, two-stage buying, and the margin warning. A billing vendor's guidance with no prices or margins.
- officialPostHog. How to do product, as an engineer (Handbook). Living page as accessed. https://posthog.com/handbook/engineering/product-engineering. Accessed 2026-10-26. Used for: good and bad product engineers, brand-damage options, "only build things that already have successful competitors with real revenue", and the interview probes. Added beyond the source map.
- officialPostHog. Pricing principles (Handbook, feature-pricing). Living page as accessed. https://posthog.com/handbook/engineering/feature-pricing. Accessed 2026-10-26. Used for: the pricing rules (Table 5) and the decision flow (Figure 6). One company's published policy.
- blogSpolsky, J. Camels and Rubber Duckies. Joel on Software, 15 December 2004. https://www.joelonsoftware.com/2004/12/15/camels-and-rubber-duckies/. Accessed 2026-10-27. Used for: the claim that the demand curve cannot be observed, the free, cheap and dear price bands with the sales-cost estimate, and the segmentation example with its stated totals. Added beyond the source map; the researcher fetched it but did not list it.
- officialPostHog. Pricing (Markdown view of posthog.com/pricing). Living page as accessed. https://posthog.com/pricing.md. Accessed 2026-10-26. Used for: bracket prices and free tiers (Table 7, Figure 7), platform packages, the annual-commitment discount table, and "no per-seat charges". Added beyond the source map.
- officialBezos, J. 2016 Letter to Shareholders. Amazon. https://www.aboutamazon.com/news/company-news/2016-letter-to-shareholders. Accessed 2026-10-26. Used for: "High-Velocity Decision Making" (two-way doors, 70% of the information, disagree and commit) and "Resist Proxies" (process as proxy).
- blogEriksson, M. Product managers: you are not the CEO of anything. Mind the Product, 15 March 2017. https://www.mindtheproduct.com/product-managers-not-ceo-anything/. Accessed 2026-10-26. Used for: the argument that authority is earned, the Ken Norton quotation, and the mini-CEO failure mode.
- analysisRumelt, R. The perils of bad strategy. McKinsey Quarterly, 1 June 2011. https://www.mckinsey.com/capabilities/strategy-and-corporate-finance/our-insights/the-perils-of-bad-strategy, read through the Internet Archive copy https://web.archive.org/web/2023/https://www.mckinsey.com/capabilities/strategy-and-corporate-finance/our-insights/the-perils-of-bad-strategy because the live page timed out. Accessed 2026-10-27. Used for: the kernel of good strategy, the Nvidia, DEC and International Harvester cases, and the four hallmarks of bad strategy. The author's own article, adapted from the book; added beyond the source map.
- textbookRumelt, R. P. Good Strategy/Bad Strategy: The Difference and Why It Matters. Profile Books edition, 2011 (Google Books catalog page). https://books.google.com/books/about/Good_Strategy_Bad_Strategy.html?id=nTU65JgxgOAC. Accessed 2026-10-27. Used for: the publisher's description and the table of contents with Profile Books page numbers (pp. 32, 77, 95, 106). The book text was not read.
- blogCagan, M. The Alternative to Roadmaps. Silicon Valley Product Group, 7 September 2015. https://www.svpg.com/the-alternative-to-roadmaps/. Accessed 2026-10-26. Used for: the two purposes of a roadmap, product vision plus business objectives per team, the OKR endorsement, and the hypothetical onboarding objective.
- textbookKohavi, R., Tang, D. and Xu, Y. Trustworthy Online Controlled Experiments: A Practical Guide to A/B Testing. Cambridge University Press, 2020. Chapter 1 sample: https://experimentguide.com/wp-content/uploads/TrustworthyOnlineControlledExperiments_PracticalGuideToABTesting_Chapter1.pdf. Accessed 2026-10-26. Used for: "Strategy, Tactics, and Their Relationship to Experiments" (pp. 20-24): the strategy-to-OEC link, guardrails, the Churchill test, and the Bing social-media case.
- blogSpolsky, J. Strategy Letter V. Joel on Software, 12 June 2002. https://www.joelonsoftware.com/2002/06/12/strategy-letter-v/. Accessed 2026-10-26. Used for: substitutes and complements, the commoditize-your-complements argument, the company examples, and the hedges; the hardware-abstraction asymmetry.
- officialPostHog. Estimating usage and costs (Docs). Living page as accessed. https://posthog.com/docs/billing/estimating-usage-costs. Accessed 2026-10-26. Used for: why PostHog prices on usage, the argument against monthly active users, the local-evaluation request arithmetic, billing limits, and cost-reduction guidance. Added beyond the source map.
- officialPostHog. Product Analytics pricing (Docs). Living page as accessed. https://posthog.com/docs/product-analytics/pricing. Accessed 2026-10-26. Used for: the per-bracket allocation and subtotal in the calculator (graduated reading), and the company's claim that more than 90% of companies use PostHog for free. Added beyond the source map.
- officialPostHog. Contract rules (Handbook, growth/sales). Living page as accessed. https://posthog.com/handbook/growth/sales/contract-rules. Accessed 2026-10-27. Used for: the cost-to-serve rationale for volume discounts and "discounts are earned, not given". Added beyond the source map.
- officialPostHog. Common questions about billing (Docs). Living page as accessed. https://posthog.com/docs/billing/common-questions. Accessed 2026-10-26. Used for: organization-wide usage with one free allowance, and data being dropped after a billing limit. Added beyond the source map.
- officialStripe. Recurring pricing models (Docs). Living page as accessed. https://docs.stripe.com/products-prices/pricing-models. Accessed 2026-10-26. Used for: Stripe's four pricing-model families (flat rate, per-seat, tiered, usage-based) and the forms of usage-based pricing. Added beyond the source map.
- officialStripe. Set up tiered pricing (Docs). Living page as accessed. https://docs.stripe.com/subscriptions/pricing-models/tiered-pricing. Accessed 2026-10-27. Used for: volume-based versus graduated pricing and the 7 / 6.5 / 6 USD fonts example (Figure 8). Added beyond the source map.
- officialLinear. Pricing. Living page as accessed. https://linear.app/pricing. Accessed 2026-10-26. Used for: plan prices per user per month, the free-tier limits, the Enterprise governance features, and the "Requires AI credits" footnote (Table 9). Added beyond the source map.
- textbookDunford, A. Obviously Awesome: How to Nail Product Positioning so Customers Get it, Buy it, Love it (author's book page). Edition unresolved. https://www.aprildunford.com/books. Accessed 2026-10-26, re-read 2026-10-27. Used for: what the book promises (a five-component process, three styles of positioning), the "newly updated and expanded edition" statement, and the absence of component names on the page. The book text was not read and no publication year is stated.
- blogDunford, A. A Quickstart Guide to Positioning. aprildunford.com, 15 March 2021. https://www.aprildunford.com/post/a-quickstart-guide-to-positioning. Accessed 2026-10-27. Used for: the definition, the five components in order, the phantom-competitor warning, the price-assumption sentence, and the embeddable-database anecdote. The author's own summary of the book; added beyond the source map.
- textbookCagan, M. Inspired: How To Create Products Customers Love, first edition. SVPG Press, 18 June 2008 (Open Library catalog record with table of contents). https://openlibrary.org/books/OL25426571M/Inspired. Accessed 2026-10-27. Used for: the heading "Chapter 39. Keys to Platform Products" only. The second edition (Wiley, 2017) was not fetched, so the chapter's content and its survival are unknown.
- officialStripe. Basic usage-based billing (Docs). Living page as accessed. https://docs.stripe.com/billing/subscriptions/usage-based. Accessed 2026-10-27. Used for: the statement that Metronome is now part of Stripe and is the recommended starting point for new usage-based integrations. Added beyond the source map.
- officialPostHog. Product management at PostHog (Handbook, product-team page). Living page as accessed. https://posthog.com/handbook/product/product-team. Accessed 2026-10-26. Used for: the Q2 2026 theme of tracking human versus agent actions.