GOOGL (Trans): AI mode lowered the search monetization; TPU external sales to ramp next yr

Dolphin Research summary of$Alphabet(GOOGL.US) FY26 Q3 earnings call transcript

I. Key takeaways

1) Shareholder returns: The BOD declared a quarterly cash dividend of $0.22 per share, payable in Dec. The call did not mention buybacks.

2) Outlook

a) Revenue: At spot FX, Q4 FX headwinds will be slightly worse than Q3, mainly in Search and YouTube ads. Google Services will also lap a tough base created by Search acceleration starting in Q3 2025.

b) Cloud: TPU systems sales will ramp further in Q4, with a meaningful revenue climb in 2027. Given supply constraints, third-party compute was activated at end-Q3 and will scale in Q4, pressuring near-term cloud margins; this capacity only came online at end-Q3, so impact this quarter was limited.

c) Capex: 2026 guidance of $195–205 bn unchanged. 2027 will step up materially, with details on the Q4 call. Depreciation and data center OpEx such as energy will continue to weigh on P&L.

3) Key financials this quarter

a) Aggregate: Revenue of $127.6 bn (+25% YoY, reported vs. FXN aligned), OP of $40.6 bn (+30%), OPM of 31.8%. Other income was -$8.2 bn, mainly unrealized losses from a SpaceX mark-down, driving declines in net income and EPS YoY.

b) Costs: Other cost of revenues +32% YoY, driven by TPU systems hardware and depreciation. R&D +36% YoY, driven by AI talent compensation and depreciation.

c) Cloud: OPM improved from 23.7% to 36.2%. Backlog reached $561 bn, mostly standard GCP contracts, with a little over 55% to be recognized over the next 24 months.

4) Cash flow and commitments

a) CFO was $42.3 bn (TTM $179.6 bn), weighed by TPU inventory build and the $5.2 bn EU fine payment from 2018.

b) Capex was $51.4 bn, approx. 65% for servers and networking (useful life 6 years) and 35% for data centers (7–40 years). FCF was -$9.1 bn for the quarter, TTM $19.7 bn.

c) Future purchase commitments of $924 bn and $152 bn of yet-to-commence leases and other obligations, both stated at maximum contractual values with variable pricing/volumes in some agreements.

d) Management reiterated ROIC over the full investment cycle will be well above the cost of capital. That said, infra build and TPU inventory will suppress FCF.

II. Earnings call details

2.1 Prepared remarks highlights

1) AI models

a) First-party model APIs processed nearly 32 bn tokens per minute, +40% QoQ and over 4x YoY. The Gemini app has reached 1 bn users, the fastest-growing product in Google’s history.

b) Gemini 4 Argon launched two weeks ago. Developer usage in its first week ramped at 3.5x the pace of Gemini 3, even under capacity constraints.Internal Chrome, Ads, and Search teams are using it for coding, with deliveries up 35% and errors down 30%.

c) Today Google announced Carbon: it ranks No. 1 on the Artificial Analysis index (score 59) and outperforms Argon across most domains. It is being rolled out first to select trusted cybersecurity defenders, with the U.S. Gov. testing; Google also launched the world model Genie 4.

d) AI is accelerating training and releases, helped by RSI (recursive self-improvement). Management expects further acceleration over the next 12 months and will ship more efficient versions.

2) Search and Ads

a) Query growth hit a five-year high YoY, including commercial queries. AI Mode queries are on avg. 3x longer than traditional search, and 1/5 use voice, images, or docs; visual search exceeds 50 bn monthly.

b) Ads: Product Listing Ads saw the biggest update in a decade, with satisfaction up 19%. Gemini helped advertisers generate over 1.5 bn creatives this quarter, more than 7x YoY.

c) AI Mode: Most queries are incremental and skew upper funnel. Click-through rates for sponsored vs. organic product listings are converging.Today Google expanded AI Mode ads to Intl markets and launched multi-turn conversational ads in the U.S.

d) Agentic: Agents will serve as a unifying layer across Search, the Gemini app, Chrome, Android, and Workspace. UCP now supports direct checkout for hundreds of thousands of brands.

3) YouTube

a) No. 1 in U.S. living-room watch time. In Q3, living-room viewing accounted for over one-third of YouTube ad revenue.Over 20 mn videos are uploaded daily.

b) YouTube Music and Premium subs (incl. trials) surpassed 150 mn, with subs growing faster than ads.

c) Amazon joined YouTube Shopping, allowing U.S. creators to tag Amazon products directly.

4) Google Cloud

a) 18 product lines now each generate over $1 bn in ARR. In the past 12 months, 2,600+ customers used over 10 bn tokens, and nearly 75 customers exceeded 10 tn tokens.

b) Nearly 90% of the Fortune 100 use Gemini Enterprise. For customers already on the platform by Q3 2025, spend is up more than 4x YoY.Cumulative knowledge-worker seats sold have surpassed 100 mn.

c) TPU: 8t training throughput improved up to 2.7x; 8i inference price-performance improved by 80%. Nearly 100 customers are using TPUs, including Anthropic (direct and via Broadcom) and Apple.

d) AI revenue in Cloud YTD has reached several tens of billions of dollars. BigQuery query speeds improved over 40% YoY, while processing costs fell over 50%.

5) Long-term bets

a) Waymo has logged over 300 mn fully autonomous miles across 15 U.S. cities, newly adding San Antonio, San Diego, and Tampa. It plans to enter Canada, Germany, Singapore, and Spain.

b) Quantum computing demonstrated magic state cultivation for the first time. Project Suncatcher is sending TPUs into space.

2.2 Q&A

Q: Will the next-gen models usher in a ‘golden era of productivity’ at Google, and how should we think about the opportunity?

A: We hit a productivity inflection with Argon internally, and Carbon marks a clear step-up again. We already saw step-changes when Flash models improved on long-horizon tasks and agent programming, but the real inflection came about four weeks ago with Argon’s internal rollout.It is one of the most meaningful internal productivity levers in software engineering we have seen, reflected in higher volumes and faster velocity of high-quality code submissions, with fewer rollbacks.

Employees are now tackling bigger problems, and we are starting to see innovation in products and consumer experiences. Carbon is being deployed internally and will be rolled out more broadly as compute becomes available; management expects the next year to be a ‘golden year’ for productivity.

Q: Query growth hit a five-year high, so why did Search revenue growth slow? Was it due to light ad load in AI Overviews and AI Mode?

A: FX was the largest factor, followed by vertical mix and lower monetization in AI Mode. Search grew 12% YoY to $63 bn in Q3.By vertical: in retail, some APAC advertisers were affected by the EU’s removal of de minimis for small parcels; in media & entertainment, Q2 was unusually strong due to the World Cup (more matches in Jun than Jul), with Q3 normalizing.

Strong growth in AI Mode queries with lower monetization also weighed on growth. Experiment results are encouraging, and we are expanding ads in AI Overviews and AI Mode to more users and queries, including Intl markets and U.S. multi-turn conversational ads.

Q: What are the technical blockers to launching agents to all users — model efficiency, compute, or maturity?

A: We need distilled models, robust harnesses, and secure runtimes, with consumer launch later this year. We are building distilled versions of Carbon to balance compute, latency, and quality at billion-user scale.A strong model is not enough: we also need a robust harness, a highly secure, 24/7 cloud runtime (where GCP comes in), plus skills and connector ecosystems. We launched enterprise first because requirements for security, governance, and controls are highest there, feedback is strong, and we are scaling; on the consumer side, we will start from the Gemini app’s 1 bn users, who are shifting from chat to task-based experiences.

Q: How big is the base effect in Search? What about Q4 and next Q1?

A: The Q3 base effect was about a 3ppt drag; Q4 should still see a few ppts of drag, with FX also a headwind. On a qoq basis and at spot FX, currency will be a headwind in Q4.On a yoy basis, this quarter begins lapping the Search acceleration in 2H25, with Q4 continuing that effect. No specific commentary on next Q1.

Q: With compute constrained, how do you prioritize allocation across businesses?

A: No granular ranking; we plan under a long-term ROIC framework rather than short-term allocation. Compute allocation requires forward planning and consumes significant management time.Each business submits long-term plans under a strict ROIC framework, with AI models underpinning Search, YouTube, Cloud, Waymo, and more. Demand signals are very strong across the board, and we are investing accordingly.

Q: How does Google Cloud balance TPU sales vs. meeting cloud customers’ compute needs?

A: No explicit split; we emphasize a diverse customer base and a three-prong approach to TPU across chips, software, and distribution. Cloud customers include digital natives, AI labs, traditional enterprises, and financial/capital markets, with 100+ major customers already trained on TPU on Google Cloud.Our approach: chips with 8t for training and 8i for inference, delivering big gains in performance and price-performance vs. prior gen; software by optimizing PyTorch for seamless migration to expand the addressable market; and distribution by deploying TPU systems into customer data centers and other providers, e.g., working with Oracle, and launching JV Crux AI with Blackstone.

Q: What portion of TPU systems sales is recognized as revenue, and how do margins compare with Cloud?

A: We do not disclose by product; no comment on revenue recognition mix or margins. Management emphasized that TPU systems are not just chips but integrated hardware, compute, storage, networking, and training/inference software that deliver significant value.Some systems have been recognized as revenue to date, with more expected in 2027.

Q: Does building first-party infra for labs like Anthropic reduce Alphabet’s capex intensity?

A: TPU systems sales run through inventory, not capex, creating a different cash cycle with upfront stocking and payment post-delivery. After contracting, orders go into backlog and we start stocking, lifting inventory balances (historically mostly hardware products).Cash inflow occurs after delivery and invoicing. During the ramp, inventory build will be visible in the cash flow statement; no direct comment on the impact to aggregate capex intensity.

Q: Will the Agentic shift be a headwind for Search, and can it enhance monetization long term?

A: We see Agentic as incremental and expansive, and we are confident we can design attractive ad models for it. We are still early and focused on trustworthy, safe, web-scale experiences, underpinned by UCP, Universal Cart, and the Agent Payments Protocol.Many tedious shopping steps suit AI, but users also value discovery, brand loyalty, and service beyond price. Agent flows will deliberately preserve choice, e.g., surfacing alternatives in insurance with higher premiums but lower deductibles and current discounts.There are few uniquely optimal decisions, which makes ads viable and opens paths to subs and other models. As AI dialogue broadened the scope of questions in Search, agents will expand the types of problems Search can solve, with assets like the Knowledge Graph enabling complex experiences.

Q: Is the TPU systems biz. attractive, and why commit capacity to it?

A: No product-level margin disclosure, but we expect margins to improve with scale. TPU systems offer price-performance advantages vs. peers amid accelerating and scarce compute demand.We only stock after signing, and all TPU inventory corresponds to contracted orders. Deploying in third-party facilities and customer data centers expands TAM (some sectors cannot move data to the cloud and need compute brought to the data) and saves us from pre-building our own sites for these systems.Management added: scale economies matter in semis; serving both internal and external demand grows TPU scale and directly improves economics.

Q: Any surprising early agent use cases on Gemini and AI Mode that changed your thinking?

A: We added a Tasks tab in the Gemini app; user adoption has a high threshold and scales only after crossing it. Repeat usage depends on handling longer tasks and delivering satisfying end results.Like coding, it may look inadequate before a threshold, and then it inflects once crossed. We are learning, closing gaps, and firming the foundation, and expect substantial progress across products later this year.

Q: Ex-TPU systems, how fast is Cloud growing? How binding are capacity constraints, and will Q4 accelerate with third-party compute online?

A: No ex-TPU growth disclosed; growth remains very strong, with no specific Q4 acceleration guide. GCP drove the vast majority of Q3 Cloud revenue.Growth comes from three areas: Gemini’s own revenue; agents built on Gemini driving compute, storage (agent memory), and databases (context for agents); and rising threats boosting security portfolio growth. New customer adoption is also accelerating.We could meet more demand with more capacity at any time, and we are prudently managing the parameters.

Q: Why collaborate with competitors in models and cloud, and why is that an opportunity vs. a risk?

A: We position TPU systems as a platform, and platforms serve all players, including competitors. We have always distinguished products from platforms: browsers, OS, and cloud are platforms open to all participants.On models, we collaborate with Apple, which competes in many areas. Subjecting TPUs to market tests and the most demanding workloads keeps teams on the frontier and enlarges the opportunity over time.

Q: Do consumers care about agent safety and trust, and how do you protect trust and privacy?

A: Users care deeply about safety and trust, which we view as a major differentiator. This is key to how users choose to engage with agents.We have earned trust over decades at multi-billion-user scale and invest end-to-end when building models and agents, with extensive pre-deployment testing and real-time monitoring at scale. We are hearing positive user feedback; no separate disclosure on privacy measures.

<End of text>

Risk disclosure and statements:Dolphin Research Disclaimer and General Disclosure