The Loyalty Deficit: Is Developer Community a Retention Moat in the AI Race?

A testable case, with a CFO-ready scorecard and a decision gate

Author: Stevan Lohja, Developer Relations Leader and founder of a ~50k-member community.

---

Who this paper is about

"The AI race" covers three different companies, and community plays a different role in each.

• Application companies build products on someone else's model. Their exposure is the thinnest: differentiation can sit in a prompt pipeline or interface that a model vendor can replicate. This is the main subject of the paper.
• Tooling and framework companies (SDKs, orchestration, observability, agent harnesses) sell to developers directly. Community, documentation and integrations are closest to the core product here.
• Model labs compete on capability, but also on distribution: SDKs, open standards, evals, partner programs and the habits of the developers who build on their APIs. For a lab, DevRel is an ecosystem and lock-in function more than a support forum.

A fourth actor sits underneath all three: cloud providers. They host the models, often hold equity in the labs, and earn revenue when developers build on them. Section III explains why that gives them a direct stake in the retention problem, and why developer community may increasingly be run from the cloud layer as well as the model layer.

Where a claim applies to one group and not the others, the text says so.

Executive Summary

The problem, in one example. Take an AI company with $10M in ARR and 40% gross revenue retention (GRR), the median ChartMogul reports for AI-native companies. It loses about $6M of existing ARR each year. Replacing the 600 lost accounts at a $6,000 blended CAC costs about $3.6M a year before any growth. Roughly 90% of its acquisition spend refills the base instead of growing it. This is an illustrative company built from market medians, not a case study. [ADD: one anonymized real company with actual cohort numbers.]

The evidence. ChartMogul's December 2025 data (~3,500 companies, ~200 AI-native above $250k ARR) puts AI-native median GRR at 40% and net revenue retention (NRR) at 48%. NRR is roughly on par with consumer apps (49%) and far below B2B SaaS (82%).

The thesis, in three parts.

1. Retention is a first-order driver of enterprise value in AI software. Revenue that is contractually optional, experimental or usage-based is priced below durable ARR.
2. Model advantages decay quickly; relational assets decay more slowly. Open-weight models trail the closed frontier by about four months on a capability index (Epoch AI), and applications multi-home across providers. Documentation, integrations, peer networks and champions survive model releases. Capability parity is not production parity, so this is a claim about direction and speed.
3. Community and DevRel are a credible, testable response, not a proven one. No published causal study links community engagement to NRR. The paper proposes a measurement plan and a decision gate.

The case for acting. Every AI release reopens the customer's evaluation. The assets that survive a release are documentation, integrations, peer networks, trained teams and champions, and they accumulate only if someone owns building them. Today most AI companies treat that work as a cost center scattered across marketing and support. The argument here is that it should be run as a retention function with a single accountable owner, its own budget line, and the same financial instrumentation as any other revenue lever.

The test. The paper does not ask for a leap of faith. It proposes a 90-day pilot sized to what finance is comfortable with, a scorecard tied to financial statements, and a pre-registered stop rule: if the leading indicators do not extrapolate to the break-even for the pilot's size, do not scale. The downside is capped by design, and the upside is retained revenue and cheaper acquisition that compound every year. Appendix A gives the break-even math for finance teams that want it.

Recommendation. Treat DevRel as a retention function with its own budget line. Start with a small, instrumented pilot. Decide at the gate.

---

Section I: The Loyalty Problem in AI

The market's instinct is to treat AI churn as a growing pain: a young category, heavy experimentation, retention that will mature. Part of that is right. But the structure of the problem argues against waiting.

In ChartMogul's December 2025 study, AI-native companies above $250k ARR posted median GRR of 40% and NRR of 48%. B2B SaaS posted median NRR of 82% and B2C 49%. The like-for-like comparison is NRR to NRR. (GRR is reported for the AI-native segment only, and GRR can never exceed NRR for the same cohort, so it should not be set against other segments' NRR.)

Table 1. Median retention by segment (ChartMogul, Dec 2025)

Reading the trend honestly. AI median GRR rose from 27% in January 2025 to 40% in September 2025. That supports the maturation view: experimental "AI tourist" revenue is washing out. The median will plausibly keep rising. The argument here is different. The forces that make AI revenue easy to leave (interchangeable models, multi-homing, contractual opt-outs) do not mature away, and as the median improves, the gap between durable and non-durable products matters more.

What the median hides. Three limits on the baseline, each of which a stronger version of this paper should address with data:

• Sample. ~200 AI-native companies above $250k ARR skew early-stage, prosumer and wrapper-like. Enterprise and infrastructure-layer companies may retain very differently.
• Distribution. A median says little about a bimodal market. The interesting group is the top quartile: if community and integration depth matter, they should show up in what the best retainers share. [ADD: top-quartile and NRR above 100% cohort data, enterprise-only cuts, and developer-tool versus non-developer-tool splits, from ChartMogul, ICONIQ, Bessemer or Ramp.]
• Pricing artifacts. Usage-based revenue fluctuates without a customer leaving. [ADD: Orb, Metronome or Paddle data separating billing volatility from logo churn.]

Revenue quality, not only churn. ChartMogul reports that AI contracts commonly include three-month opt-outs, which it says 70 to 80% of buyers exercise, and that experimental spend is counted as ARR. Some measured churn is a pricing artifact, but the same structures make revenue contractually optional, and acquirers and public markets pay lower multiples for optional ARR.

Structural drivers of leaving.

• Interchangeability. Fradkin et al. (arXiv:2504.15440, OpenRouter data from early 2025) find substantial multi-homing and rapid adoption shifts after each release. The sample skews toward developers and hobbyists, so the result is directional for enterprise.
• Developer behavior in the wild. The clearest live example is coding assistants, where developers move between tools within weeks of a release. [ADD: switching or share data for Cursor, Claude Code, Codex and Copilot; note that the conversation about these tools happens largely on Reddit and X rather than in vendor-run forums, which matters for how a community strategy is designed.]
• Price and contract length do not buy loyalty. RevenueCat's data on 75,000 consumer subscription apps shows high-priced monthly plans retaining 6.7% of subscribers at twelve months versus 36% for low-priced annual plans. This is consumer evidence, used only to illustrate the principle.
• Fit drives retention. The a16z/OpenRouter "glass slipper" analysis argues that durable retention attaches to users whose workload fits the product, not to whoever has the best model. The inference that education, integration and peer proof help users reach fit is this paper's, not a16z's.

The narrow conclusion: retention is a major determinant of enterprise value in AI software, and several of the mechanisms behind it are non-technical.

---

Section II: Commoditization and Open or Local Alternatives

Two developments shorten the shelf life of a model-quality advantage: the narrowing gap between open-weight and closed frontier models, and the move of local inference toward production.

The capability gap is measured in months. Epoch AI (May 2026) finds that since January 2026 the best open-weight models have trailed closed frontier models by about four months on the Epoch Capabilities Index (ECI), an average of about 8 ECI points.

Table 2. Selected ECI scores cited by Epoch AI (May 2026)

Two cautions about this table. First, Gemini 3.5 Flash is a speed-tier model, so it understates the distance to the top closed model. [ADD: ECI for the top closed model, which is the correct comparator for the "frontier" claim.] Second, benchmark parity is not production parity. Long-horizon agentic reliability, tool-use robustness, latency, cost and safety behavior are where frontier models still earn premium share. The claim here concerns the direction and speed of decay, not parity today.

Capability is not adoption. If open-weight capability is catching up but enterprise usage is not shifting, the commoditization story needs qualifying. The paper should show both curves. [ADD: share of enterprise tokens and spend on open-weight models (for example Menlo Ventures' enterprise reports, which put it in the low teens in 2025; verify the latest), OpenRouter open-weight token share, and Hugging Face download trends.]

Local inference has reached production for some workloads. Ollama reports growth from ~1.5 million active users (May 2024) to ~5 million (May 2026), per Presenc AI's compilation (third-party, directional). User counts are a weak proxy for production workload share, so treat this as a signal of direction only.

Pricing power cuts the other way. Top-tier closed models have largely held premium pricing despite open-weight catch-up. [ADD: evidence on price per token at the frontier, and on frontier-model share of enterprise API spend.] The thesis does not need frontier pricing power to collapse. It needs the lead on any single model to be short-lived, so that retention cannot rest on it.

Combined with multi-homing, every release is a re-evaluation event. What depreciates more slowly is what accumulates around the product: documented know-how, workflow integration, team habits and trust.

Caveat for each audience. Most application companies do not train frontier models, so their exposure is a differentiation layer built on someone else's model. For labs, the equivalent risk is that developers treat the API as a swappable endpoint.

---

Section III: The Capital Structure Context

This section is context, not part of the argument's proof. The AI sector's financing is tightly coupled. The IMF's October 2026 Global Financial Stability Report names an AI-driven correction as one of three salient near-term risks, and the BIS (Bulletin 137, October 2026) finds that 55.2% of incoming investment into AI firms from 2021 to 2025 came from other AI firms. Both institutions note that such structures can be rational and need not be destabilizing. The link to this paper is narrow. Application-layer retention is part of the revenue base behind those commitments, so a company whose renewals rest on demonstrated utilization and integration depth has more to show in a hard renewal meeting than one whose differentiation is a model lead. This paper does not forecast a correction, and it does not claim community-based companies are immune.

A direct implication: cloud providers have a financial reason to run developer community. Much of the circular financing runs through cloud providers. A provider takes an equity stake in, or signs a large compute agreement with, an AI company, and that company spends a large share of the proceeds on the provider's compute. The provider's return then depends on sustained utilization: the AI company's usage staying high, and developers building on the models through the provider's own managed services and marketplace. That is the retention and utilization problem this paper describes, seen from the other side of the ledger. A provider with equity exposure, committed-spend backlog and a model-hosting business has reason to fund documentation, reference architectures, credits, workshops, certifications and partner programs. This is an inference from incentives, not documented intent. [ADD: sourced examples, such as the share of provider cloud backlog attributable to AI labs, named investment-and-compute agreements, and provider developer programs with disclosed spend.]

Three consequences follow for the argument:

• The community function may sit at the cloud layer. Because platforms host several vendors' models, provider communities tend to lock in the platform (identity, data, committed spend, tooling) more than any single model. That fits the thesis that the model advantage is short-lived and the surrounding ecosystem is not.
• Labs depend on partners for part of the developer relationship. Co-marketing, joint reference architectures and marketplace presence matter, and the lab's own DevRel has to coordinate with a partner that also hosts its competitors.
• Credit-funded usage is not retention. Credits and committed-spend arrangements can make utilization look healthy while the revenue is subsidized, which also feeds the circularity risk above. A cloud-funded program should be measured on credit-to-paid conversion and retention after credits end, not on consumption alone.

---

Section IV: DevRel and Community as a Response

When differentiation is transient, the remaining assets are relational and structural. DevRel and community are one way to build them deliberately.

Six mechanisms.

• Education and activation bring users to first value before a trial lapses.
• Canonical docs and cookbooks become the default reference a developer returns to.
• Communities generate peer support and social proof at low marginal cost.
• Feedback loops route practitioner reality into the roadmap.
• Advocates carry the product into new teams.
• Deep integrations raise the literal cost of leaving.

The claim that activation is the highest-leverage retention intervention is widely repeated, but the author has not located a primary source. It is a hypothesis the 90-day plan can test.

Where each mechanism matters most. In a model lab or API company, the community lever that matters is often not a forum. It is SDK quality, open standards, reference implementations, partner programs and the presence of the company's tools in the workflows developers already use. For a framework company, it is the installed base of code written against the framework. For an application company selling to non-developers, community is the weakest lever, and workflow depth, proprietary data and customer success are likelier to be the durable assets.

What the field reports. Draft.dev's 2026 Developer Marketing Survey finds 51% of dev-marketing and DevRel leaders ranking community and events as their highest-ROI channel (a selection effect: respondents rate their own function). The CMX 2026 Community Industry Report finds budgets split between increases and cuts while ROI confidence fell. The 11th State of Developer Relations Report (September 2024) finds 61% of practitioners struggle to prove influence and only 18% link revenue influence to their metrics. The field has an evidence problem.

A structural example: framework-led ecosystems. LangChain's documentation positions the framework around model interoperability, which productizes model commoditization: the switching cost sits in customer code, team fluency and integrations, which a new model release does not invalidate. Vercel's AI SDK follows a similar pattern. These are structural illustrations only. GitHub stars measure reach, not retention, and the author holds no retention data for either company.

The contrast case: the thin wrapper (hypothetical). Consider an AI support company at ~$8M ARR whose differentiation is a prompt pipeline and clean interface over a single model provider. Its switching cost is one endpoint. If the vendor ships a comparable native feature and an open alternative reaches parity within months, every customer's evaluation reopens at once, and opt-out clauses become the exit.

What actually retains developers on an API. Community competes with levers that probably matter more for a lab or API company, and an honest version of this argument has to rank them:

• Reliability and uptime
• Rate limits and capacity
• Pricing stability
• Deprecation and migration policy
• Compliance and data residency
• Eval and prompt portability, meaning how costly it is to re-tune after switching models
• Support quality
• Customer success and onboarding

A counterpoint belongs here too: AI-assisted migration is lowering the cost of rewriting code against a new provider, which weakens the "code and integrations as switching cost" mechanism over time. The durable parts are likelier to be team habits, trust, eval suites and tacit knowledge than the integration code itself.

The evidence gap, stated honestly. No published causal study shows that community engagement raises NRR. The available evidence is correlational: survey rankings, cohort differentials and activation-retention linkages. Engaged users may simply be better-fit users. [ADD: whatever quasi-causal evidence can be assembled: Manchanda, Packard and Pattabhiramaiah ("Social Dollars," Marketing Science, 2015; verify citation and limits); public-company net expansion against developer-ecosystem strength (Datadog, MongoDB, Snowflake, Twilio, GitLab, Elastic); companies that cut or closed community or DevRel and what happened to activation or NRR; vendor-reported ticket-deflection and retention data (Atlassian, Salesforce Trailblazer, Microsoft MVP), labeled as vendor-reported; Common Room and Orbit benchmarks.] Until that is assembled, the recommendation is operational: each company should generate its own evidence through the matched-cohort method in Section VI.

---

Section V: Counterarguments and Rebuttals

---

Section VI: Practical Implications: A Staged Investment Framework

A. The Financial Case

The question is not "how much should we spend?" It is what must be true for this function to pay for itself, and how quickly can we find out? Three levers carry the answer.

• Retention. At the AI-native median of 40% GRR, an AI company spends most of its acquisition budget replacing customers it already won. LTV:CAC sits near 1.9:1, well below the 3:1 benchmark investors look for. Moving GRR from 40% to 60% lifts that ratio to about 2.9:1. Each point of retention keeps revenue the company would otherwise have to buy again.
• Acquisition. Community-sourced pipeline and referrals can lower blended CAC. This is the least-evidenced and most influential lever, so the pilot measures it directly.
• Support cost and activation. Peer support deflects tickets, and faster time-to-first-API-call shortens payback. These effects show up within a quarter and give finance early proof.

The function does not need to win on all three. A CAC benefit can substitute for part of the retention gain, and the gate is designed to tell which lever is working.

Funding without a headline. The pilot should be funded from existing budget, with total spend held flat. The cut should come from work whose payoff is a short-lived capability lead, not from work that drives conversion or retention. Appendix A shows why protecting conversion matters.

What the model does not claim. The lift is an input, not a finding. No study shows DevRel spending delivers any particular lift. Appendix A is a way to ask what must be true, and every company should replace its assumptions with its own data.

B. The CFO-Ready DevRel Scorecard

Activity metrics (members, stars, attendance) describe reach. The CFO-ready equivalents map to financial statements.

For cloud-funded or credit-backed programs, add two rows to the scorecard: credit-to-paid conversion rate, and GRR in the twelve months after credits expire. Both separate durable utilization from subsidized usage.

C. Closing the Measurement Gap: A 90-Day Plan

• Month 0: Set the rules before the data. Interview three to five CFOs or heads of finance and record what each would accept as evidence for a retention program. [ADD: findings.] Agree with finance on a deliberately small pilot size. Write down, before looking at results, the break-even lift for that size (Appendix A) and the minimum participation and engaged-account lift that would reach it.
• Month 1: Baseline. Join community IDs, product accounts and CRM records into one warehouse view. Baseline ticket deflection and cost per ticket, median time-to-first-API-call, community-influenced pipeline, and GRR and NRR by engagement decile. Output: a one-page baseline the CFO signs off on.
• Month 2: Cohort tracking. Define the engaged threshold (for example, three or more meaningful interactions per quarter). Track matched cohorts, controlled for plan, segment and acquisition channel. Where feasible, stagger the rollout by region or segment to create a comparison group, because matching alone cannot separate effect from fit. Log touchpoints as CRM campaign members so influence is attributable.
• Month 3: Analysis and gate. Review ρ, Δe, deflection savings, influenced pipeline and CAC mix. Compare the community lift against other levers in the same data (onboarding, customer success, reliability incidents). Note that retention effects take 12 to 24 months to show fully, so the 90-day gate relies on leading indicators (activation speed, early cohort retention, pipeline) and a forecast with stated uncertainty. Scale only if the leading indicators extrapolate to at least the break-even L written down in Month 0. Otherwise hold the pilot flat or stop.
• Ongoing. Report the scorecard quarterly alongside CAC and NDR. If a metric cannot be tied to a financial outcome after two cycles, retire it.

D. Priorities for the Board

1. Fund DevRel as a retention function with its own budget line, reported alongside CAC and NRR.
2. Start with a pilot, sized small enough that finance can absorb a null result.
3. Make activation velocity a tracked KPI, and test whether it predicts retention in your own data.
4. Build switching costs deliberately: workflow depth, eval suites, champion programs and team-level fluency, not only integration code.
5. Choose what to cut with care. Protect work that drives conversion and retention.
6. Fix the reliability, pricing and deprecation basics first. No community program offsets an unstable API.
7. Hold the decision to the data, using the rule written in Month 0.

E. What Would Change My Mind

The thesis is wrong, or at least not worth funding, if any of the following hold in a company's own data:

• Engaged and non-engaged cohorts converge once plan, segment and acquisition channel are controlled for (the gap is selection).
• A staggered rollout shows no retention or activation difference between exposed and unexposed groups.
• Participation among revenue-weighted accounts is too low for any plausible Δe to reach the break-even L.
• Community-sourced pipeline does not lower blended CAC.
• Reliability, pricing or support incidents explain churn better than any engagement variable.
• Enterprise cohorts already retain far above the AI-native median, leaving little deficit to close.

Across the market, the thesis is weakened if top-quartile retainers show no common ecosystem or community traits, or if open-weight share of enterprise spend stays small while frontier pricing power holds.

---

Section VII: Conclusion

The argument has four steps. AI-native products retain at 40% GRR and 48% NRR, close to consumer apps and far below B2B software, and part of their ARR is contractually optional (Section I). Model advantages are shortening to a clock measured in months, though capability parity is not production parity (Section II). The sector's financing is interlinked, which makes application-layer retention a shared interest (Section III). Where models cannot hold a customer, community, documentation and integration depth may, which is the hypothesis this paper puts forward (Sections IV to VI).

The evidence for the last step is correlational, and the paper has said so. That is why the prescription is staged. Start small, write the break-even down in advance, measure with the CFO-ready scorecard, and decide at the gate. If the data supports it, the function pays back through retention, lower acquisition cost and lower support cost. If it does not, the company has lost a small amount and learned something its competitors have not.

---

Appendix A: Sizing Model (for finance review)

This appendix shows break-even lift at several illustrative program sizes. It is provided so finance can test the logic, and it is not a recommended budget.

The model asks a narrow question: if a company moves an amount N per year from R&D into community and DevRel, what must be true for the move to pay for itself? It is illustrative, on stated assumptions, and is neither a forecast nor a valuation opinion.

Variables.

• N: annual program cost in $M, funded from R&D (total budget unchanged). Equivalent forms: N as a percentage of ARR, or of R&D.
• ρ (participation): the share of ARR, revenue-weighted, held by accounts that actively engage with the program.
• Δe: the GRR lift among engaged accounts, in points.
• L = ρ × Δe: the blended, company-wide GRR lift. This is what reaches the P&L.
• Ramp: how quickly L builds. Two cases are shown: immediate (full effect from Year 1) and ramped (25%, 50%, 75% and 100% of L in Years 1 to 4).
• c: change in blended CAC from organic acquisition. d: loss of new-logo conversion caused by the R&D cut.

Fixed assumptions. 1,000 accounts at $10K ACV ($10M ARR), 70% gross margin after inference, $4M go-to-market spend at a $6,000 blended CAC (about 667 new accounts, or $6.67M new ARR a year), baseline GRR of 40% (ChartMogul AI-native median) with expansion adding 8 points (NDR 48%). ARR next year = NDR × ARR this year + new ARR. Revenue in a year is the average of opening and closing ARR, gross profit is 70% of revenue, and there is no discounting. Net value is cumulative gross profit versus the status quo, minus N per year.

Unit economics. LTV = ACV × gross margin ÷ (1 − GRR). At 40% GRR, LTV is $11,667 and LTV:CAC is 1.9:1, below the common 3:1 benchmark. A 20-point lift (60% GRR) raises LTV to $17,500 and LTV:CAC to 2.9:1. A further 20% CAC reduction takes it to 3.6:1 and CAC payback from 10.3 to 8.2 months. CAC payback improves only if CAC falls, because retention does not change the first-year margin on a new account.

Result 1: break-even scales with N. Holding CAC unchanged, the blended GRR lift needed for the program to recover its cumulative cost within five years is:

Table 4. Break-even blended GRR lift L (points), five-year horizon, CAC unchanged

Result 2: the lift among engaged accounts must be much larger than the blended lift. Community reaches a minority of accounts, so the blended lift is diluted by participation.

Table 5. Lift among engaged accounts (Δe, points) needed to reach the ramped break-even, by participation

GRR cannot rise by more than 60 points from a 40% base, and most of any observed engaged-versus-non-engaged gap is likely selection rather than effect.

Table 6. Five-year cumulative net value versus status quo, $M (CAC unchanged)

Result 3: CAC and the R&D cut change the picture.

• CAC. Organic acquisition is the least-evidenced and most influential input. A 20% CAC reduction alone (no retention lift) more than pays for a small program. At N = $1.0M with a 10% CAC reduction, the ramped break-even falls from 12.8 to 3.5 points. The pilot should measure community-sourced pipeline and blended CAC directly.
• The R&D cut. A 10% loss of new-logo conversion costs about $3.2M of gross profit over five years on its own, and raises the break-even lift by about 5 points at every N shown. The cut should target work whose payoff is a short-lived capability lead, not work that drives conversion or retention. If the program can be funded from a source other than R&D, this penalty disappears.
• Expansion. The model holds expansion at +8 points. Community plausibly affects expansion (more teams, more use cases) as much as GRR, and a lift there adds to the result. The pilot should track expansion separately from GRR.

Avoided replacement cost, not double-counted. Each point of blended GRR lift on 1,000 accounts retains about 10 accounts a year, avoiding about $60K of replacement acquisition spend at a $6,000 CAC. Do not add the retained accounts' future margin on top, because the replacement accounts would have produced the same ARR. The case rests on compounding: retained revenue lets the same GTM budget grow the base instead of refilling it.

What the model does not show. The lift is an input, not a finding. No study shows DevRel spending delivers any particular lift, and the model assumes no valuation multiple. It is a way to ask what must be true, and Tables 4 to 6 give the answer for any N.

Modeling note. This appendix is an illustrative model on stated assumptions, not sourced data or a forecast. The ramp shape, participation rates, engaged-account lift, CAC change and new-logo penalty are scenario inputs. Break-even figures were computed from the roll-forward described above and reproduce the earlier draft's N = $2.2M results.

---

Sources

Primary sources

1. ChartMogul, "The SaaS Retention Report: The AI Churn Wave," Kyle Poyar, Dec 10, 2025. AI-native median GRR 40% and NRR 48%; B2B SaaS NRR 82%; B2C NRR 49%; GRR trend 27% (Jan 2025) to 40% (Sep 2025); three-month opt-outs; thin-wrapper prediction. Sample ~3,500 companies, ~200 AI-native above $250k ARR. chartmogul.com/reports/saas-retention-the-ai-churn-wave/
2. RevenueCat, State of Subscription Apps 2025. 12-month retention of 6.7% (high-priced monthly) versus 36% (low-priced annual). Consumer, illustration only.
3. IMF, Global Financial Stability Report, October 2026, Chapter 1. AI-driven correction as a salient risk; circular transactions.
4. BIS, Bulletin 137, "Circular relationships among AI firms," Oct 1, 2026. 55.2% of incoming investment into AI firms from other AI firms (2021 to 2025).
5. Epoch AI, "Open models lag state-of-the-art closed models by 4 months," May 29, 2026. Four-month average lag; ~8-point ECI gap; Table 2 scores.
6. Fradkin, Cunningham, Eckles and Rock, "Demand for LLMs," arXiv:2504.15440. Multi-homing and rapid adoption shifts; OpenRouter data from early 2025.
7. a16z/OpenRouter, "The Cinderella 'Glass Slipper' Effect," Dec 8, 2025. Qualitative finding on workload fit only.
8. Draft.dev, 2026 Developer Marketing Survey. 51% rank community and events highest ROI. Self-rated.
9. CMX, 2026 Community Industry Report. Split budgets and falling ROI confidence.
10. DevRel.Agency, 11th Annual State of Developer Relations Report, Sep 2024. 61% struggle to prove influence; 18% link revenue influence to metrics.
11. GitHub repository metrics, October 2026. Reach indicators only.
12. Presenc AI, "Ollama Ecosystem State 2026," May 2026. Third-party compilation, directional.

To add before publication (not yet sourced or verified)

• Open-weight share of enterprise spend and tokens: Menlo Ventures enterprise reports, OpenRouter, Hugging Face.
• AI retention and revenue quality: Ramp AI Index, ICONIQ and Bessemer State of AI, a16z enterprise CIO surveys, Orb, Metronome or Paddle usage-billing data, Sensor Tower or Appfigures for consumer AI apps.
• Developer sentiment and behavior: Stack Overflow Developer Survey, GitHub Octoverse, Postman State of the API, Anthropic Economic Index and OpenAI usage studies, coding-assistant switching data.
• Community-to-retention evidence: Manchanda, Packard and Pattabhiramaiah (2015), public-company net expansion cross-section, community shutdown case studies, vendor-reported programs, Common Room and Orbit benchmarks.
• Primary research: three to five CFO interviews; one anonymized real company case; independent model review.