AppLovin (Trans): Model upgrade lagged; Q3 re-accelerated

Dolphin Research's Trans of AppLovin FY26Q2 earnings call

I. Key takeaways

1) Shareholder returns: In Q2, the company repurchased and withheld ~1.14 mn shares for $551 mn; shares outstanding were 335 mn at quarter-end, with ~$1.8 bn remaining under the authorization. The buyback pace slowed vs. roughly $1.0 bn in Q1 as FCF was lighter this quarter. This does not signal a change in confidence or how the authorization will be used.

2) Q3 guide: revenue of $2.05–2.085 bn (+46%–48% YoY, +7%–8% QoQ); Adj. EBITDA of $1.71–1.74 bn (+48%–50% YoY) with ~83% margin. The guide embeds already-deployed model improvements, ongoing scale-up in consumer verticals, normal seasonality, and higher training/inference costs. It does not assume any yet-to-be-deployed new model releases.

3) Q2 headline metrics: both revenue and profit came in below the company’s own bar. Revenue: $1.92 bn (+53% YoY, +4% QoQ), slightly below the midpoint of guidance. Adj. EBITDA: $1.61 bn (+58% YoY), margin expanded by ~300 bps YoY, landing slightly below the low end of guidance.

Costs: sequential cost increase was mainly compute for training existing models and developing new ones. About 70% of the QoQ cost delta flowed through to Adj. EBITDA.

4) Cash flow and balance sheet: FCF was $863 mn, with conversion below normal due to timing of international cash taxes and interest payments. This was a timing effect, not a change in underlying profitability. Management expects conversion to improve in Q3 and to be ~75% of Adj. EBITDA for the full year. Quarter-end cash was $3.05 bn, total debt $3.7 bn, and net leverage was ~0.1x last-12-month Adj. EBITDA, well below the long-term operating range of ~1x.

5) SEC inquiry closed: it was a voluntary information request, which the company never deemed material. The SEC recently informed the company that the inquiry is closed and it is not recommending any action.

II. Call details

2.1 Management highlights

1) Gaming ads: model performance remains the single biggest growth driver. When models improve, advertisers can deploy more budget at target ROAS and spend naturally steps up. The Q2 issue was purely timing — the cadence of in-quarter, substantive model gains was softer than usual, and the next model step-up landed just after the quarter-end.

No signs point to weaker demand or a changed competitive landscape: Max publisher revenue grew double digits QoQ, and the company’s share in publishers’ waterfalls stayed stable. With improvements live and seasonality strengthening, the business is re-accelerating, and Q3 is off to a strong start.

2) Consumer (e-com) verticals: standout performance in Q2. Advertiser spend hit a new record, 28% above Q4 2025, which is typically the seasonal peak for these advertisers. Outperforming the peak season during an off-season underscores the steepness of the curve.

That said, consumer remains too small to fully smooth quarterly volatility. As it scales each quarter, this will change.

3) Tech and compute investment: incremental spend is going exactly where it should. The company has been re-architecting to support more complex models that benefit more from additional training compute; this includes extra compute supporting the model improvements already live for Q3. Every dollar is spent with ROI discipline.

When extra compute yields materially higher revenue via better model performance, the company will do that trade every time. Higher training and inference costs are reflected in next quarter’s guide.

4) Platform opening and customer acquisition paths. The platform was opened to the public under its original name, AppLovin Ads Manager, with no expectation of an overnight business shift. The ramp is deliberate: start with mid-market advertisers where performance is strongest, then unlock the long tail as data accumulates, mirroring the playbook used in gaming.

Execution is partner-led, and the company will keep investing in this motion.

5) Four execution priorities and long-term growth framework. First, improve the core models — the near-term growth engine. Second, advance architecture so the company can harvest more gains from scaling compute, which should unlock meaningfully larger benefits over time.

Third, enhance creative tools and ad formats to deliver better results to advertisers. Fourth, bring more high-quality advertisers onto the platform via strategic partnerships. Running multiple advertiser categories in one auction extends the opportunity window with each new category; with gaming improving and consumer expanding, management believes the business can compound at ~30% annually.

2.2 Q&A

Q: This seems like the first time you’re leaning on partners to bring in more customers. Can you expand on partnership opportunities? (Citi)

A: We have already closed several third-party partnerships, including with a large e-com analytics provider (Triple Whale). Going directly to where target advertisers are served on the other side is a more precise way to bring the right advertiser types onto the platform. Versus buying ads upfront to pull in the long tail — which is harder at the current model maturity — these partnerships yield highly qualified customers.

Q: In gaming ads, what exactly was the 'model breakthrough that didn’t land,' and what happened timing-wise? (Jefferies)

A: Over the last 12 quarters, growth was strong each quarter, with Q2 the only single-digit quarter; since Axon 2, nearly every other quarter was double-digit. Q2 is seasonally the weakest. Every quarter with outsized growth coincided with model improvements — those improvements are ongoing even if not broken out. This time, Q2 did not realize a gain of the usual magnitude, and a meaningful step-up hit right after quarter-end, hence the strong Q3 start and guide.

Q: You said consumer categories set a record in Apr. How did Q2 progress, and how much came from existing vs. new advertisers? (Jefferies)

A: New advertisers helped, but the base is already solid, and new logos are not the primary growth driver. Up 28% vs. Q4 2025 is a significant step-up for this category. Typically e-com drops sharply in Q1 vs. Q4, recovers in Q2, then Q3–Q4 carry the year, with half of spend in Q4. Delivering this growth in Q2 shows existing clients on the platform are seeing strong success.

Q: E-com advertisers say they are far from hitting efficiency ceilings on your platform. Is that a rosy sample, or a broader trend? (UBS)

A: We are a new platform. These companies manage budgets slowly; our product has been live ~18 months, and such companies plan budgets 1–4 quarters ahead, mostly to social and search, with us in a test bucket that takes time to graduate. For current spenders, given results, we are confident scale will be larger in 12 and 24 months. The ramp should be faster than on mature platforms because our current wallet share is lower, and as data accumulates, familiarity grows, budgets rise, confidence in long-term outcomes improves, and incremental studies are run, the climb should accelerate.

Q: You prioritized mid-market before the long tail and were cautious on very large advertisers. Has that view changed over the past 3–6 months, given some gaming advertisers are quite large? (UBS)

A: On gaming, models are mature, data is rich, and the architecture is more refined. A new game can onboard and hit return targets quickly with virtually no learning budget; the models are precise enough to run any genre. On e-com/consumer, we are still early: data penetration is lower and models are less granular. Small shops are less likely to hit targets and scale with minimal spend — some can, many cannot — while mid-market brands understand learning costs, spend more, hit targets, and then scale, making mid-market the sweet spot now.

Q: Why didn’t Q2 deliver the expected model uplift, and can this be diagnosed after the fact? (William Blair)

A: This is R&D — there is no promise of a step every three months. The team is always improving, usually via multiple smaller gains stacking rather than a single leap. In Q2, the impact was modest, followed by a substantial gain early in Q3. Experiments are hypothesis-driven A/B tests across the system; some periods yield little, others deliver large gains that drive +12%, +13%, +15% QoQ quarters.

Q: How fast have improvements come post-quarter, and what does the guide include? (William Blair)

A: The framework is consistent with prior guidance philosophy, and confidence is high. The difference this time is the guide only includes already-released model improvements and known increases in compute costs.

Q: With compute costs rising, how should we think about the margin structure? (William Blair)

A: Margins may show some short-term volatility over time, consistent with our past messaging. Investors should view this as a positive signal: we spend cautiously and only when incremental revenue is visible behind the spend. When you see near-term fluctuations, recognize we are improving off those costs; long term, we remain highly confident in maintaining EBITDA margin in the low-80% range.

Q: Six weeks after GA of the platform, how are customer adds tracking vs. expectations, and what is the advertiser count trajectory into 2027? Will consumer skew to many small advertisers or a few large ones, and how long will the 'precision' go-to-market last before adoption inflects like Meta/Google? (Bank of America)

A: Today it skews toward the latter — fewer advertisers contributing more — and progress is in line with expectations. We did not plan broad marketing for GA, and our marketing and spend target mid-market brands. We must first win higher-ticket customers: not the very largest for now, and not the very smallest — SMB/long tail adds count but are harder to run today. The focus is mid-market, which already constitutes most active spenders in consumer; once we nail this, we will expand outward.

That yields broader data coverage while buying time to sharpen models and build reputation into the larger market. Building such systems takes years: Google took decades, Facebook well over a decade; we ourselves took 14 years to fully penetrate mobile gaming to where not spending with us is unwise. This will take time, but growth is fast because clients are succeeding and are still far from their performance-supported ceilings; templates and models will keep improving, each new mid-market client improves the data, and over time we can expand — timing remains uncertain.

Q: Technically, what is ahead of plan and what lags? GenAI creatives are in focus — what milestones should we track? (Bank of America)

A: We are early in a sizable, fast-growing business, so it is less about being ahead or behind. The mandate is clear: acquire more customers, gather more data, build more complex models, improve ad templates, and deliver gains for existing and new customers so results compound over time. As more success stories and external narratives emerge and more agencies recognize us as a strong third anchor, customers will come organically — that is the formula.

Q: At the start of the year, 57% of qualified self-serve leads converted, with creative gaps the main leak. Has this changed with creative tools, and what common feature asks are you hearing post-GA (better targeting, measurement, more automation)? (Loop Capital)

A: The answers are related. Creative remains the largest bottleneck, and that has not materially changed. The common unit is a long-form video with an interactive end card; we can now auto-generate the interactive end card efficiently. But we still cannot consistently hand advertisers a turnkey 30–60s high-quality video — this is in progress with mixed results. Once we can, or we deliver alternative templates that do not require video, we can offer one-click campaign setup and fix funnel conversion.

By contrast, on open web, social, and search, many ads are dynamic product catalogs or static promos, rarely 30–60s videos. Unless an advertiser is large and also runs TV, they typically lack such assets. Thus mid-market and above are more likely to have creative matching our formats, while direct-signup SMBs typically do not; to penetrate the long tail, we must solve this first.

Q: How healthy is the mobile gaming ecosystem? Devs cite rising CPI and weaker ROAS, and third-party data show downloads down ~10%–15% YoY. (Wells Fargo)

A: The category must be healthy for us to perform, which is why we cited Max market growth at double digits QoQ — a very large market growing strongly given our ad-monetization penetration. The ad-supported market is expanding fast. IAP data can mislead because measurement firms cannot track off-platform purchases; as more IAP games move transactions off-platform, those revenues have not vanished and often increase on a net-basis — they just are not measured.

So neither our reported services revenue nor third-party data capture that incremental. On installs, the multi-year trend has shifted from high-install, low-quality hyper-casual toward casual and deeper IAP titles. Deeper funnels mean higher CPI and higher LTV; rising cost does not equal falling value — everything is judged against ROAS.

Finally, we ourselves catalyze this market — to our knowledge, we are the largest at the UA discovery layer globally. A quarter without a model lift is not great for the category; conversely, after the early Q3 release, the same advertisers report higher installs, lower CPI, and better performance. At our operating scale, we are a key catalyst for UA in gaming, and we take that responsibility seriously to keep devs achieving the success our platform has long delivered.

Q: You rebranded to Axon for web and reverted at GA — does limited brand awareness for AppLovin or Axon hurt SMB acquisition? Do you have the right brand to onboard the next 100k long-tail e-com and web advertisers? (Wells Fargo)

A: The name would not stick — after switching to Axon, people still used the old name, so we switched back. If you ask 100 e-com prospects not using us today, awareness is indeed an issue, and many may still think of us as a gaming platform. A decade ago, many gaming clients did not know us either; they spent far more with Facebook and Google. Today, every gaming client spends on our platform.

Brand matters, but loyalty is earned by performance. As we compound improvements in tech, templates, and case studies, clients will discover the platform over time. Google took decades to become the de facto search standard; Facebook took well over a decade. Those 100 brands will not all know us next quarter, but if we execute in coming quarters, recognition will compound to standard status.

Analytics partners also help: they talk to their customers and, if they see we are the second- or third-best channel for clients already live, it is in their interest to recommend us. They also benefit economically via partnerships we sign with them.

Q: Advertisers say Android spend has been higher than expected this year. Can you further widen the gap on Android and grow wallet share long term? (Evercore)

A: We are competitive on both platforms. The market, however, has one very large company that owns Google Play and provides substantial installs and spend for Play clients. We see opportunities as comparable across platforms: as models improve, our ability to scale rises on both, with Android more competitive given a very large, capable rival there.

Q: Did the World Cup create a headwind for IP advertisers, and were there notable cyclical factors in Q2? (Evercore)

A: There are spikes from World Cup-related advertisers during matches, but at a quarterly scale, it is not material. It does not meaningfully affect whether IAP or ad-monetized gaming clients can spend on our platform. We avoid over-concentration in any advertiser type and do not do brand advertising; our focus remains gaming advertisers.

Q: By region, U.S. revenue accelerated QoQ while Intl was flat vs. Q1. Any color on that divergence? (Piper Sandler)

A: Nothing fundamental. Intl had strengthened for several quarters and then moderated slightly without a specific driver. This view is by user location and reflects population mix, and note that web consumer spend is currently more concentrated in Western markets than Intl.

Q: Outside gaming but excluding e-com, markets discuss short-form dramas and prediction markets adding to gaming. Thoughts? (Piper Sandler)

A: Prediction markets, mentioned earlier with the World Cup question, are one such category, but these are not a current priority. If you split the world into consumer, web e-com, and non-gaming apps, the first is much larger and our initial focus. The second will come later and is not a current modeling priority.

Q: Your org is lean. On A/B testing, what do more iterations unlock, and is there diminishing return on R&D? (RBC)

A: You need smart A/B tests — a mass of undifferentiated, low-quality tests does not help. We want very high talent density, which we have, and we still need more research scientists and more A/B tests; more importantly, we use AI to accelerate testing. We have done a lot here; with LLMs getting much better at coding, our internal testing velocity has indeed increased over the past two years. The focus remains: hire very smart people, leverage LLMs for heavy coding, and let them run more A/B tests.

Q: Partnerships can help near-term logo growth. For the second layer — data accumulation — what else can you do, like credits or more managed service to help scale? (BTIG)

A: Each new client’s spend does not determine the data we get — they must share their data with us to plug into deep learning and get strong results. The goal is to acquire more mid-market customers so the models see more transactional behavior. As we do this over time — by bringing more scaled customers onto the platform — we will capture more data, build more complex models, and deliver better outputs. It compounds: more customers drive better performance; we aim to reach a tipping point where customer acquisition accelerates and the system enters a natural fast lane.

Q: How is LeadGen progressing — scaling or directional trends? (BTIG)

A: It is ongoing. We remain in customer tests and have nothing new to disclose.

Q: Beyond Q3, how should we think about incremental tech and compute spend? Will elevated investment persist, and how much visibility do you have into future compute costs? (Macquarie)

A: We do not expect to deviate from our high-level framework that roughly $0.10 of compute spend accompanies each incremental $1 of revenue. The guide is consistent with that, as is the data center disclosure in the 10-Q, and we remain on that track. We do not anticipate changes, though volatility is possible; if we add compute, we will explain why. Engineers are doing R&D — if someone finds a way to train and run a much larger, more complex model that drives material revenue, we will pursue it, even mid-quarter.

This re-acceleration in the guide came from exactly that — a model release with more complexity that re-accelerated revenue. That is always good for our business.

Q: Is the prior 'first-year $70k spend per new customer' still the right baseline, and are model breakthroughs the main driver to raise it? (Macquarie)

A: As we add partnerships and target mid-market, that number likely goes up. It is an indicator of mix: how many customers come from marketing-driven self-serve signups — which skews to SMBs and low-GMV shops — versus strategic partnerships that bring fewer but higher-GMV customers. Given the focus on the latter, you should expect the number to rise.

Q: On integrating Unity Vector runtime data — you already capture most ad signals well, while runtime offers engine-level signals you do not natively access. Do the channels become complementary, and will these new signals expand the total pie? (Unidentified)

A: Everyone’s data and models differ. When we IPO’d, a bear case said this was zero-sum; we believe the last five years proved it is not. As marketing platforms improve, UA scale and growth rise, end customers’ P&Ls improve, and they can reinvest more in marketing. We are clearly the largest and want the market to improve. Max market’s double-digit QoQ growth is a very large incremental for an already large market given our monetization penetration.

This is not zero-sum because models trained on differentiated data raise the ceiling for gaming customers. That expands the overall spend potential.

Q: With consumer GA, will you re-focus on Wurl and CTV? How portable is consumer-side experience to CTV? (SSR)

A: On consumer, we are still budget-constrained and early. If we expand supply, the natural path is non-gaming apps and other open web inventory. Consumer verticals create the conditions to expand supply — gaming is more niche, with little gaming ad presence on open web and social, while e-com ads are ubiquitous.

Thus, the first supply expansion will be device-side, then CTV. CTV is attractive because that full-screen lipstick ad on a phone should translate well to a TV; TV clearly drives shopping at scale. The opportunity is there, but we are not yet ready to execute.

Q: CTV is migrating toward performance, which is your strength. Is the challenge less about porting to CTV and more about educating interfaces, audiences, and advertisers? (SSR)

A: Think of it this way: consumer advertisers are performing well but are far from their ceilings on mobile, and we do not have surplus budget to shift. Simply launching CTV and reallocating money would weaken our mobile position and not truly expand revenue. To enter new supply, we must believe incremental budgets are available. Step one is non-gaming apps, including running non-gaming ads inside gaming apps that do not yet carry them; step two is open web; step three is CTV.

We will do all of these once there is budget to chase because they are obvious growth levers. The sequence ensures expansion is additive, not cannibalistic.

<End of content>

Risk disclosure and statement:Dolphin Research Disclaimer and General Disclosure