Alphabet (GOOGL) — Q3 FY2026 Analyst Q&A Breakdown
Fiscal period: Q3 FY2026 · Call date: 2026-10-28
Seven analysts questioned a panel of Sundar Pichai, Philipp Schindler, Thomas Kurian, and Anat Ashkenazi. Three themes dominated: whether agentic AI helps or erodes Search, what the TPU systems business actually earns, and how much of the Search slowdown is structural versus transitory. Every question unit is below, in order.
1. Douglas Anmuth, JPMorgan
(a) The question. On a recent podcast, Sundar called 2027 potentially the “golden age of productivity” for Google given the new class of models and their integration with other products — elaborate on the opportunity set. (b) For Philipp: query growth is the fastest in five years, yet Search growth decelerated — is the disconnect simply that fewer ads are shown in AI Overviews and AI Mode, which are becoming a bigger share of queries?
(a) Management’s response. Sundar said the step change began internally with 3.8 Flash as models improved at long-horizon tasks and agent decoding, then a “huge point of inflection” once deployed internally. It is “one of the most meaningful internal productivity levers we have seen on software engineering” — measured directly on high-quality code checked in, the rate of check-ins, and reduced rollbacks. He dated the deployment to only about four weeks before the call, said Carbon is now deploying internally as “a meaningful step up” from Argon, and that he is “even more excited since I said that in the podcast.”
(a) The “Analyst’s Take.” Direct and unusually concrete for a vision question: the code check-in/rollback metrics and the four-week timeline are new, and dating the inflection underscores how early the productivity curve is. The podcast framing was upgraded, not defended.
(b) Management’s response. Schindler recapped the print — Search up 12% YoY to $63 billion — then gave a three-driver bridge for the Q2-to-Q3 deceleration: FX was “the single largest driver”; retail was hit by the EU eliminating its de minimis exemption for some APAC advertisers; and media and entertainment normalized after an unseasonably strong Q2 driven by the FIFA World Cup, which put more matches — and ad budgets — in June than July. He then conceded the analyst’s hypothesis in substance: strong AI Mode query growth “had an impact because of the lower monetization we were seeing in AI Mode.” He closed by pointing to promising ad experiments and scaling ads in AI Overviews and AI Mode to more users and queries, including international markets and multi-turn ads in the US.
(b) The “Analyst’s Take.” The most direct growth-quality answer of the call: management named FX as the top driver and effectively confirmed the AI Mode monetization drag rather than disputing the premise. The bridge is falsifiable — each driver should wash out on a known schedule.
2. Brian Nowak, Morgan Stanley
(a) The question. On agentic: after rolling out Spark to some users during the year (analyst-provided), what are the key technological hurdles to shipping agentic products to the full user base — model efficiency, capacity, or technology readiness? (b) For Anat: explain the search performance comparison factor and how to think about the 4Q and 1Q compares.
(a) Management’s response. Sundar laid out a stack: the Gemini 4 series matters, and distilled Carbon versions are being built for billion-user scale across compute, latency, and quality. Beyond the model, “you need a robust harness… a highly secure runtime” — with GCP providing each user a secure runtime with 24/7 availability — plus an ecosystem of skills and connectors. Enterprise came first via the Gemini Agent for Enterprise because that buyer holds the highest security and governance bar; it has been “well received.” The same experience reaches consumers later this year, starting with the Gemini app’s user base (see the Highlights in the companion briefing), where users already “move beyond chat” to task- and agent-like queries; agents will extend as a unifying layer across Chrome, Android, and Workspace.
(a) The “Analyst’s Take.” Direct and structured — the harness/runtime/connectors framing is the new content, more systems engineering than model capability. The “Spark” rollout the analyst cited was never engaged by name; the “later this year” consumer commitment is the falsifiable line.
(b) Management’s response. Ashkenazi separated two dynamics: sequentially, FX is a headwind again in Q4 at current spot rates; on the year-over-year compare, Q3 began lapping the search growth acceleration from the back end of 2025 — lapping about 3 points of growth this quarter — with “a couple of points of lapping” continuing in Q4.
(b) The “Analyst’s Take.” The cleanest quantitative disclosure in the Q&A: the 3-point lap is new and quantifies how much of Q3’s deceleration is comp-aware rather than demand loss. It frames Q4 as continuation, not deterioration.
3. Eric Sheridan, Goldman Sachs
(a) The question. Given recent comments on compute constraints, what are the key priorities for allocating compute as it becomes available across the organization?
(a) Management’s response. Pichai framed it as long-range planning, “not a short-term allocation exercise”: each business area submits long-range plans under a disciplined ROIC framework. AI models underlie everything from Search and YouTube to Cloud and Waymo, making the investment “a highly leveraged approach,” with the full-stack approach driving efficiency and optionality. He cited “very, very robust demand signals” and said Alphabet is investing for that moment.
(a) The “Analyst’s Take.” Philosophical rather than granular — no explicit ranking of claimants on capacity. The new element is the leverage framing: one model base amortized across every business is the stated allocation logic, with ROIC as the referee.
(b) The question. For Sundar or Thomas: within Google Cloud, how is the view evolving on applying compute to cloud customer needs relative to external TPU demand?
(b) Management’s response. Kurian described a deliberately broad portfolio — digital natives, AI labs, traditional enterprises, financial and capital-markets customers — with over 100 significant TPU customers who start in Alphabet’s cloud and learn the systems there. A TPU system, he stressed, is hardware, system software, compilers, and optical networking assembled as a system. His three-part approach: build the right silicon (recapping the prepared 8T-for-training/8i-for-inference script), optimize software for seamless portability — TorchTPU/PyTorch support “opens up a much larger market” — and broaden distribution: TPU systems into customer data centers, an Oracle partnership, and a Crux AI joint venture launched along with Blackstone. Demand from AI labs, capital markets, and enterprises is “extraordinary,” and the 100-plus cloud customers are expected to grow in number and scale their usage.
(b) The “Analyst’s Take.” The cloud-to-data-center funnel is the strategic reveal: Google Cloud as the onboarding environment from which customers graduate to purchased systems. Direct and additive, though it stops short of how internal compute allocation trades against external sales.
4. Ross Sandler, Barclays
(a) The question. On TPU resale economics: the analyst offered that a gigawatt of TPU full system costs around $35 billion for compute and surrounding components (analyst-provided) — what percent of that is recognized as revenue, and what is the margin flow-through relative to where the Google Cloud segment sits? (b) As Alphabet helps companies like Anthropic bring on their own first-party infrastructure, what happens to Alphabet’s CapEx intensity as future CapEx hands off to AI labs?
(a) Management’s response. Ashkenazi declined product-level financials but walked through the mechanics: a signed agreement first appears in backlog, then inventory builds — cash going out ahead of the revenue stream — and cash comes in as systems are delivered and billed. “It’s not the traditional kind of CapEx… rather a different cycle of cash flow generation.” The inventory build is visible in the current cash flow statement. Some system sales have been recorded as revenue to date, with more expected through 2027. She framed the product as far more than a chip — hardware, compute, storage, networking, software, training and inference — carrying “significant added value.”
(a) The “Analyst’s Take.” Direct on mechanics, evasive on numbers: the $35 billion-per-gigawatt figure was neither confirmed nor disputed, and no margin view was offered. The backlog-to-inventory-to-billing sequence is genuinely new and gives analysts the template for modeling working capital, if not margin.
(b) Management’s response. Explicitly combined with (a): the CapEx-versus-cash distinction was the answer — TPU sales are contract-backed inventory cycles, not Alphabet CapEx — with the ramp of cash needs showing up now and revenue to follow through 2027.
(b) The “Analyst’s Take.” The intensity question went substantially unanswered: whether AI-lab self-build reduces Alphabet’s future CapEx burden was left as the reader’s inference from the cash-cycle mechanics, not a management statement.
5. Justin Post, Bank of America
(a) The question. Three parts: with the universal shopping cart launched, how will Alphabet manage Search’s transition to agentic use — could it be a headwind? (b) Could agentic actually drive more monetization of traffic long-term? (c) Again on TPU margins: is this a really good business, and why, if margins can’t be given?
(a)+(b) Management’s response. Schindler: it is “early days,” but agentic experiences are “clearly additive.” The build is anchored on agentic-commerce foundations — the Universal Commerce Protocol, Universal Cart, the Agent Payments Protocol. The design philosophy: automate the “grunt work” users dislike, but enhance — not automate away — the discovery experience users love, intentionally designing time and space for choice beyond price (brand loyalty, customer service). Given “there’s almost never only one perfect option,” he is confident Alphabet can provide “really interesting, compelling advertising models for agentic use cases,” extending to subscriptions and other payment methods. Pichai added the expansion framing: as AI chat did, agents expand the types of problems solved in Search — a “new frontier” — with people underestimating the accumulated components (Knowledge Graph, world understanding) behind the surface simplicity.
(a)+(b) The “Analyst’s Take.” Direct on thesis, silent on risk: no headwind scenario was engaged, and the answer’s substance — preserving discovery and choice — reads as a deliberate defense against cannibalization. The advertising-model confidence is new color; the mechanisms remain unshown.
(c) Management’s response. Kurian repeated that product margins are not broken out, then argued the business case on structure: the complete system (not just a chip), price-performance advantages versus competitors, growing demand for accelerated compute, and its scarcity — with the company “very focused on continuing to improve margins as we scale.” Two additions: component procurement happens only after contracts are signed, so all TPU inventory sits behind committed contracts; and placing systems in third-party facilities and customer data centers broadens the addressable market — bringing compute to data that cannot move to a cloud — while offsetting the need to pre-build facilities.
(c) The “Analyst’s Take.” Evasive on the number, new on the structure: the committed-contracts point reframes the inventory build as de-risked demand rather than speculative capacity, and the data-gravity argument explains the in-place distribution push.
6. Mark Shmulik, AllianceBernstein
(a) The question. On early agentic-style use cases in Gemini/AI Mode and third-party integrations: anything that surprised or changed the view — including in consumer productivity? (Then, for Thomas or Anat: the Cloud growth rate excluding TPU sales, and how much is capacity-constrained — and is it fair to assume acceleration as third-party capacity comes online in Q4?)
(a) Management’s response. Pichai pointed to the new task tab in the Gemini app where people type agentic queries. The bar: users want to be delighted, and repeat use comes from handling longer tasks to a satisfactory outcome. His coding analogy — “when you’re doing it at 70%, it looks like it doesn’t work until you cross a certain threshold and it really takes off” — framed the current phase as pre-threshold learning: identifying gaps and laying the foundations, with “a lot of progress across our surfaces… later this year in ‘27.” He tacked on a TPU aside: semiconductor economies of scale mean dual deployment — Alphabet’s data centers plus others’ — directly translates into better economics.
(a) The “Analyst’s Take.” Candid about immaturity: the 70%-threshold framing is an admission that current agentic experiences sit below the take-off point. No surprise was actually named — the task tab is observational, not a data point.
(b) Management’s response. Kurian delivered essentially the scripted segment recap — record Cloud growth, GCP the vast majority of revenue, ex-TPU growth again very strong — then added what was new: a three-part demand flywheel (Gemini revenue itself; Gemini driving attach of compute, storage for agent memories, and databases serving context for agent reasoning; and cybersecurity growth), plus accelerating new-customer adoption into GCP. On capacity: “we could obviously serve a lot more demand with capacity, but we are managing all of these parameters and growth carefully.”
(b) The “Analyst’s Take.” The recap was script; the flywheel was the addition — and it explains why Gemini pull-through matters beyond token revenue. The acceleration question was never answered: capacity constraint confirmed, no Q4 growth commitment offered.
7. Michael Nathanson, Moffett Nathanson
(a) The question. On working with companies that may be competitors in models and cloud: what is the philosophy, and why is it an opportunity rather than a risk longer term? (b) On agentic consumer behavior: what has Alphabet learned about safety and trust — do those variables matter to consumers — and how is consumer trust and privacy protected as agentic tools build out?
(a) Management’s response. Pichai drew the products-versus-platforms distinction: platforms must “work for everyone,” and Alphabet has a long history operating them — browser, operating system, cloud — striving to “provide it to all players and be a great customer.” The precedent: working with Apple on models while competing “in many, many areas.” TPU systems are treated the same way — a platform tested against market forces and the most demanding workloads, which keeps the teams “at the frontier.” Over time, he argued, this “always ends up in expanding opportunities.”
(a) The “Analyst’s Take.” Direct and doctrinally consistent: competitors-as-customers is the platform playbook, with Apple as the load-bearing precedent. The risk leg of the question — platform dependency arming rivals — was not engaged.
(b) Management’s response. Pichai confirmed the premise: people approaching agentic experiences are concerned about safety and trust — “a very, very high factor” in how they use these things. The protection case rests on decades of trust built across billion-user products, extensive testing as models and agents are built and deployed, and real-time monitoring at scale. He called trust a “huge differentiator” and said users “would vote with their feet as this frontier goes through.”
(b) The “Analyst’s Take.” Directional, not operational: no concrete privacy mechanics were offered beyond testing and monitoring. The notable move is elevating trust to a stated competitive moat — positioning Alphabet’s consumer track record as the differentiator as agentic features scale.
Synthesis
Clusters — one row per repeated concern; the repeated question is where the signal is.
| Concern | Analysts (Q#s) | Uniform management answer | Signal |
|---|---|---|---|
| Agentic: headwind or additive? | Anmuth (Q1b), Post (Q5a/b), Shmulik (Q6a) | “Clearly additive” — expansion opportunity; monetization scaled deliberately after UX proves out | Three independent approaches in one call mark agentic cannibalization as the street’s biggest medium-term worry — and management’s most rehearsed answer |
| TPU systems economics | Sandler (Q4a/b), Post (Q5c), Sheridan (Q3b) | No product-level disclosure; backlog→inventory→cash mechanics; margins to improve with scale | Management wants TPU valued as a systems platform but will not yet underwrite the margin math |
| Search growth composition | Anmuth (Q1b), Nowak (Q2b) | FX the largest sequential driver; ~3 points lapped in Q3 with a couple more in Q4; AI Mode monetization still lower | Management pre-emptively quantified the decel bridge instead of defending the headline — a confidence signal in the re-acceleration |
| Cloud capacity constraints | Sheridan (Q3b), Shmulik (Q6b) | Demand exceeds capacity; third-party bridging capacity ramps in Q4 at a margin cost | Supply, not demand, is the binding constraint on Cloud growth |
Deflections
| Topic (Q#) | What was asked | Deflection technique |
|---|---|---|
| TPU margin and revenue detail (Q4a, Q5c) | Percent of system cost recognized as revenue; margin flow-through; is it a good business | Decline product-level disclosure, then reframe to system value, contract-backed inventory, and scale economics |
| CapEx hand-off to AI labs (Q4b) | Does AI-lab self-build reduce Alphabet’s CapEx intensity? | Merged into the cash-cycle answer; the intensity question itself left unaddressed |
| Cloud acceleration on new capacity (Q6b) | Should Q4 Cloud accelerate as third-party capacity lands? | Scripted recap plus “managing all of these parameters” — constraint confirmed, no growth commitment |
| Agentic headwind scenario (Q5a) | Could agentic use deflate Search? | Reframe to “clearly additive” and new use cases; no headwind path engaged |
Contradictions — where the Q&A recontextualized or conflicted with the prepared remarks.
| Contradiction | Prepared-remarks claim | Q&A walk-back |
|---|---|---|
| Search momentum vs. growth quality | Sundar: query growth the highest in five years | Anmuth and Nowak draws concede a sequential decel bridged by FX, de minimis, the World Cup lap, and AI Mode’s lower monetization (Q1b, Q2b) |
| Frontier-model leadership vs. availability | Carbon announced as the strongest frontier model, ranked first on the Artificial Analysis Intelligence Index | Q&A reveals deliberately narrow access — a subset of trusted cyber defenders, US-government testing, distilled billion-user versions still in the works (Q1a) |
| CapEx framing | Prepared remarks discuss CapEx guidance and 2027 increases | Ashkenazi recontextualizes TPU sales as a contract-backed inventory cycle — “not the traditional kind of CapEx” — recasting the CapEx-intensity debate (Q4a/b) |
| Competitors vs. customers | Nathanson’s premise frames Oracle-type TPU buyers as competitors | Pichai’s platform doctrine recasts them as customers whose demanding workloads keep the platform at the frontier (Q7a) |
Unanswered
- Q4a (Sandler): what percent of a TPU system’s cost converts to revenue, and its margin flow-through — declined outright.
- Q4b (Sandler): whether AI-lab self-build lowers Alphabet’s future CapEx intensity — never directly addressed.
- Q6b (Shmulik): whether Cloud growth accelerates in Q4 as third-party capacity comes online — no commitment given.
- Q2a (Nowak): the “Spark” rollout premise (analyst-provided) was never confirmed or discussed by name.
- Q7b (Nathanson): specific privacy safeguards for consumer agentic tools — answered only with testing and monitoring generalities.