Alphabet’s $185bn AI Bet Recasts the Balance Sheet for the Gemini Era

AI infrastructure: from cost line to strategic moat

On Alphabet’s fourth-quarter call, the numbers told a story of a company recasting its balance sheet around AI infrastructure on a scale usually reserved for sovereigns, not software groups. Revenues for 2025 pushed past the $400bn mark for the first time, with the December quarter delivering $113.8bn, up 18 per cent year-on-year (17 per cent in constant currency). Yet it was the guidance, rather than the print, that framed the investment case: capital expenditure in 2026 is slated to soar to between $175bn and $185bn.

Sundar Pichai cast that escalation as the logical next step of a decade-long “AI-first” strategy. The bulk of this investment will again go into technical infrastructure: in 2025, roughly 60 per cent of the $91.4bn CapEx was in servers and accelerators, and 40 per cent in data centres and networking. Finance chief Anat Ashkenazi signalled a similar mix for 2026, but with a much bigger absolute base and a marked rise in depreciation charges to follow.

The rationale is twofold. First, Alphabet is betting that owning the “widest variety of compute options” – from Nvidia’s latest Vera Rubin GPUs to its own seventh-generation Ironwood TPUs – will become a durable competitive moat, both for Google Cloud and its own AI products. Second, management still finds itself in what Pichai described as a persistently “supply constrained” world for AI compute and power, even after years of stepped-up spending. The 2026 CapEx plan is, in effect, an attempt to get ahead of that bottleneck.

The company is also trying to improve the economics of this buildout. Pichai highlighted a 78 per cent reduction in Gemini serving unit costs through 2025, driven by model optimisation and better utilisation. Ashkenazi, in turn, walked investors through a sprawling efficiency programme that runs from data centre design to using AI agents inside Google’s own back office – including its treasury and payments teams – to free up capital and headcount for further AI investment.

Core businesses riding an “expansionary moment”

Beneath the capital intensity, Alphabet’s legacy engines are still accelerating rather than stalling in the face of AI disruption.

Search and Other advertising revenues rose 17 per cent to $63.1bn in the quarter, supported by “broad strength across all major verticals”, with retail the biggest contributor. Pichai described an “expansionary moment” for search, in which AI overviews and the new AI mode are lengthening sessions, making queries more complex and conversational, and pushing users beyond text into voice and image-based interactions. Nearly one in six AI mode queries now come via non-text inputs, and Circle to Search has reached more than 580m Android devices.

Crucially for shareholders, this AI shift is not yet cannibalising the text-based ad business. Philip Schindler, chief business officer, argued that Gemini models are improving query understanding across languages, unlocking monetisation of longer, more complex searches that were previously “difficult to monetise”, and reducing irrelevant ads. Tools such as AI Max and Performance Max helped advertisers generate some 70m creative assets in the quarter, with case studies like Aritzia and L’Oréal using AI-powered campaigns to find new high-value customers and lift direct-to-consumer revenue.

YouTube, meanwhile, remains a hybrid of media platform and subscription service. Advertising revenues rose 9 per cent to $11.4bn, held back by tough comparisons against 2024’s US election spending and some softness in brand-related verticals. But on an annual basis, YouTube’s combined ads and subscription revenues cleared $60bn. Subscription momentum – especially in YouTube Music and Premium, alongside a new lower-priced sports tier and a Premium Lite option – is now significant enough that shifts from ad-funded users into paying subs can show up as a drag on ad revenue line growth even as the overall business improves.

Engagement metrics in YouTube’s emerging formats remain striking. Shorts now average more than 200bn daily views, and in several markets, Shorts already earns more revenue per watch hour than traditional in-stream video. In the living room, YouTube has held the top streaming spot in the US for nearly three years, with long-form podcast viewing on TVs reaching over 700m hours in October, up 75 per cent in a year. AI appears to be deepening this ecosystem: over 1m channels a day used new AI creation tools in December, and more than 20m viewers used Gemini-powered tools to interrogate content.

Google Services as a whole generated $95.9bn in Q4 revenue, up 14 per cent, with operating income up 22 per cent to $40.1bn and margins rising to 41.9 per cent. Subscription, platforms and devices revenue climbed 17 per cent to $13.6bn, powered by YouTube subscriptions and AI-enhanced Google One plans. Across consumer services, Alphabet now counts more than 325m paid subscriptions.

Cloud and Waymo: from “Other” to core growth pillars

If the services business showed AI as an accelerant, Google Cloud demonstrated how deeply the group is now embedded in the AI spending cycle of enterprise and government clients.

Cloud revenues rose 48 per cent year-on-year to $17.7bn in the quarter, with operating income more than doubling to $5.3bn and operating margins leaping from 17.5 per cent to 30.1 per cent. Management stressed that Google Cloud Platform itself is growing even faster than that already-rapid segment rate, driven by “billions” in quarterly revenues from enterprise AI products.

The client list is increasingly diverse: frontier AI labs; high-frequency trading firms such as Citadel Securities; industrials like Mercedes-Benz; and Western governments using Google’s accelerators for high-performance computing. On top of that infrastructure demand, more than 120,000 enterprises now use Gemini models, including 95 per cent of the top 20 and over 80 per cent of the top 100 SaaS companies. These AI customers, Google says, consume 1.8 times as many products as non-AI clients.

Enterprise AI “agents” have moved beyond pilot status. Gemini Enterprise, the company’s AI workplace platform, has more than 8m paid seats across some 2,800 companies, including banks and cruise operators, and handled over 5bn customer interactions in Q4, up 65 per cent. Partners such as Wendy’s, Kroger and Woolworths have deployed these agents into customer service and operations. Workspace, too, delivered double-digit growth on the back of higher average revenue per seat and seat expansion.

The clearest datapoint of future demand, however, lies in the cloud backlog. It reached $240bn at year-end, more than doubling year-on-year and surging 55 per cent quarter-on-quarter. That increase, Ashkenazi underlined, was driven primarily by AI infrastructure and solutions commitments, with the number of deals over $1bn in 2025 exceeding the prior three years combined. A new collaboration with Apple – where Google will be the preferred cloud provider for the next generation of Apple foundation models based on Gemini – underlines how far the company has pushed into its rivals’ supply chains.

Waymo, long a loss-making “Other Bet”, is being repositioned as another AI-intensive asset with potentially global scale. The unit has just completed a $16bn investment round, a significant portion funded by Alphabet, and has now delivered more than 20m fully autonomous trips, at a current run rate of over 400,000 weekly rides. Its service footprint is expanding: a sixth market, Miami, went live recently, and operations in additional US cities as well as the UK and Japan are in the pipeline. That said, the business still carries heavy accounting costs; a $2.1bn stock-based compensation charge linked to the new funding round pushed Other Bets’ operating loss to $3.6bn in Q4.

Balance sheet strength versus depreciation drag

Alphabet can afford this spending spree – for now. The company closed the quarter with $120.8bn in cash and marketable securities against $46.5bn in long-term debt. Operating cash flow reached a record $52.4bn in Q4 and $160.5bn for 2025, translating into $24.6bn in Q4 free cash flow and $73.3bn for the year. Capital returns continued, albeit at a measured pace relative to cash generation: $5.5bn of share repurchases and $2.5bn in dividends during the quarter.

Still, the P&L is beginning to feel the weight of past and present AI bets. Depreciation rose 38 per cent in 2025, from $15.3bn to $21.1bn, and Ashkenazi warned of a “meaningful” increase in 2026 as the new CapEx cohort flows through. Total cost of revenues climbed 13 per cent in Q4, reflecting higher depreciation on infrastructure and content acquisition costs, particularly for YouTube. Total operating expenses rose 29 per cent, driven by AI talent, the Waymo stock-based compensation charge, and increased marketing for the Gemini app and AI search products.

Despite this, operating income still rose 16 per cent to $35.9bn and net income 30 per cent to $34.5bn. The operating margin, at 31.6 per cent, absorbed the Waymo charge; absent that, it would have been higher.

Management is trying to reassure investors that this will remain an investment cycle, not a margin free-for-all. Ashkenazi described a “highly rigorous” internal capital allocation framework that weighs near and long-term returns on AI projects, sets a top-down CapEx envelope, and then allocates across units – with over half of machine learning compute earmarked for Cloud in 2026. At the same time, she emphasised a continuous hunt for efficiencies across the organisation, from code generation (half of Google’s code is now written by internal coding agents before review) to small AI deployments inside finance.

Gemini as product, platform and distribution

Although the call was nominally an earnings discussion, the subtext was Gemini’s ambition to be both Alphabet’s internal AI spine and the industry’s engine. Pichai presented a three-layered AI stack: infrastructure (chips and data centres), models and tools (Gemini 3 Pro, multimodal systems like Vio, Imagine and Lyria, and the nascent Anti Gravity development platform), and products (Search, Chrome, Gmail, Android, Pixel and Cloud).

On the consumer side, the Gemini app has reached 750m monthly active users just a year after launch, with usage intensity – query length, session depth, retention – rising sharply since the December release of Gemini 3. The models now handle more than 10bn tokens per minute via direct API calls from customers, up from 7bn the previous quarter. The company is pushing Gemini into every surface: personal intelligence in the Gemini app and AI mode in Search; “agentic” features in Chrome such as Autobrowse; new Gmail and Vio capabilities; and Android and Pixel hardware updates, including a forthcoming Pixel 10a.

In parallel, Google is trying to shape the commercial protocols of the AI economy itself. At the National Retail Federation show, it unveiled the Universal Commerce Protocol, an attempt to standardise “agentic commerce” – how AI agents discover products, compare offers and complete checkouts across retailers. Built with large retail partners, the protocol is designed to reconcile user convenience (frictionless action from discovery to purchase) with merchant control over offers, promotions and catalogue representation. Early experiments include direct offers in AI mode and a forthcoming ability to complete purchases directly within AI mode and the Gemini app from selected merchants.

For now, monetisation of the Gemini app is focused on a free tier and subscriptions, rather than ads. But Schindler framed advertising as an eventual, and likely inevitable, layer: if done well, he argued, ads can be “helpful commercial information” even in conversational interfaces. The more pressing question for investors, then, is less whether Alphabet can monetise Gemini than how fast it can scale premium AI features without diluting the economics of its highly profitable search franchise.

Sector context and the investor trade-off

Alphabet’s Q4 numbers drop into a sector where AI exuberance and infrastructure anxiety are running in tandem. Cloud hyperscalers are scrambling to lock up chips, power and land, even as investors worry about whether AI workloads will ultimately prove as margin-rich as the old cloud compute and search advertising businesses they are replacing.

On this call, management tried to position Alphabet on the optimistic side of that ledger. Gemini 3 Pro, they insisted, is seeing the fastest adoption of any model in the company’s history, with token throughput tripling that of its predecessor. AI features, rather than eroding the ad base, are expanding query volumes and making high-value searches more legible and monetisable. Google Cloud, once a laggard, now has growth, margin and backlog metrics more typical of a high-growth software company than a struggling No. 3 cloud provider.

Yet the trade-off is not trivial. A $175–185bn CapEx year will test the patience of investors accustomed to Alphabet’s prodigious free cash flow. Depreciation, data centre operating costs and the risk of overbuild in a volatile macro and regulatory environment remain live issues. And outside the headline segments, some familiar drags persist: network advertising revenues fell 2 per cent in the quarter, and Other Bets still post multi-billion dollar operating losses despite Waymo’s apparent commercial traction.

For now, though, the numbers suggest that Alphabet is turning its AI thesis into revenue and profit fast enough to keep its balance sheet – and shareholders – on side. The coming years will reveal whether the Gemini era’s infrastructure binge proves to be a moat, or merely an expensive hedge in a market reshaped as much by regulation and open models as by proprietary silicon.