Quick Answer
Alphabet raised its 2026 AI spending forecast to between $195 billion and $205 billion, and posted its first negative free cash flow quarter as a result. For users, the practical consequences are a faster model release cadence, three new cheaper Gemini models, and confirmation that Gemini 3.5 Pro is still only in testing.
Key Takeaways
- 2026 capital spending guidance raised to $195–205 billion, up $15 billion, with 2027 expected higher again
- Record quarterly capex of $44.9 billion pushed free cash flow negative for the first time, at -$5.9 billion
- The Gemini app now has 950 million monthly users; AI Mode in Search has passed 1 billion
- Gemini 3.5 Pro remains in testing, while pretraining has begun on Gemini 4
- Google says it is still supply constrained — demand for AI compute is outrunning what it can build
What did Google actually announce?
Alphabet’s second-quarter results, reported on July 22, 2026, were strong by almost every operating measure: revenue up 24% year over year to $119.8 billion, Google Cloud up 82% to $24.8 billion, and operating income up 30%. Cloud’s order backlog reached roughly $514 billion.
Then the company told analysts it would spend more than planned. Capital expenditure guidance for 2026 rose to $195–205 billion, up from $180–190 billion a quarter earlier — an increase of about $15 billion, and, as CNBC reported, well above the roughly $188 billion analysts had modelled. Chief financial officer Anat Ashkenazi added that 2027 spending would increase significantly again.
The market’s reaction was telling. Despite the revenue beat, shares fell around 3% in extended trading. The reason sits in one number: quarterly capex hit a record $44.9 billion, which pushed Alphabet’s free cash flow to negative $5.9 billion — the first negative free cash flow quarter in the company’s history. A business that has printed cash for two decades briefly stopped, because it is pouring money into servers and data centres faster than the money comes in.
Why is Google spending this much?
Because it cannot build fast enough to meet demand. Ashkenazi told analysts the company remains supply constrained and has been for several quarters running, and that the guidance increase reflects an acceleration in delivering capacity rather than a change of strategy. Roughly 60% of the spending goes to servers, with the remaining 40% on data centres and networking equipment.
The usage figures behind that are genuinely large. Google says its model APIs now process about 22 billion tokens per minute, up from 16 billion a quarter earlier, with more than nine million developers building on its models each month. On the consumer side, the Gemini app has reached 950 million monthly active users, and AI Mode in Search has passed a billion. Nearly 90% of the Fortune 100 use Gemini Enterprise.
Ashkenazi framed the commitment simply: Google will keep investing as long as the return looks attractive. That is a real position, but it is also an unfalsifiable one until the returns actually arrive — which is the crux of the wider debate we cover in the 2026 AI spending reckoning.
What does “supply constrained” mean for you?
It is the most consequential phrase in the entire earnings call for ordinary users, and it rarely gets translated out of finance language.
Supply constrained means there is not enough computing capacity to serve everyone who wants it. When that happens, companies ration — and consumers are the cheapest place to ration. In practice that shows up as rate limits on free tiers, slower responses at peak times, features that launch for paying subscribers first, and capable models quietly reserved for higher-priced plans while free users get the lighter version.
It also explains a pattern you may have noticed across the industry: the aggressive push toward smaller, cheaper models. Those are not just a budget option for developers. They are how a provider serves a billion people without needing a billion people’s worth of frontier compute. If your free AI assistant feels faster but occasionally less capable than it did six months ago, this is the underlying reason.
What is happening with Gemini 3.5 Pro and Gemini 4?
Chief executive Sundar Pichai gave the clearest public update yet on both. In his earnings remarks, he confirmed that Gemini 3.5 Pro is “currently in testing” — which, in plain terms, means the flagship that was expected in June and then rumoured for July is still not generally available. We track that timeline in detail in our guide to the Gemini 3.5 Pro release date.
More striking is what comes next. Pichai said Google has begun its “most ambitious pre-training run yet” for Gemini 4, the next generation of frontier models. So the company is simultaneously finishing one flagship and training its successor — a reminder of how compressed these cycles have become.
He also said Google now plans to release new models almost every month. For users, that means less waiting for a single annual leap and more incremental improvement arriving continuously, which is a genuinely different rhythm from the one the industry had two years ago.
What are the new Gemini models?
Google announced three the day before earnings, all in the cheaper, faster tier rather than at the frontier:
| Model | Positioning |
| Gemini 3.6 Flash | The new workhorse — cost-effective and efficient for everyday tasks |
| Gemini 3.5 Flash-Lite | Lighter and cheaper again, for high-volume, simple work |
| Gemini 3.5 Flash Cyber | Security-focused; paired with Google’s CodeMender agent to find and fix vulnerabilities |
Pichai said the Flash series is seeing enormous demand because it hits the balance of performance and cost. The Cyber variant is the interesting one: Google claims it delivers frontier-comparable performance on security work at a fraction of the size of far larger models — a direct answer to the concern that AI is getting dangerously good at finding software flaws faster than defenders can patch them.
There is competitive pressure here too. Cheap, capable open-weight models from Chinese labs have put real downward pressure on token prices, and businesses watching their AI bills are increasingly willing to switch. A cheaper Flash tier is partly a response to that.
Is this level of spending sustainable?
Honestly, nobody knows, and anyone claiming certainty in either direction is guessing. What can be said is what the numbers show and what the risks are.
The case for it: demand genuinely exceeds supply, Cloud is growing 82% a year with a $514 billion backlog, and roughly 90% of the Fortune 100 are already using the enterprise product. That is not speculative demand — it is contracted.
The case against: a negative free cash flow quarter is a meaningful signal, depreciation on this much hardware will weigh on margins for years, and the company has told investors 2027 will be more expensive still. Data centres also need power, and power ordered now does not arrive for years, which is a constraint money alone cannot solve quickly.
The consumer-facing risk is straightforward: this spending eventually has to be recouped. That usually arrives as higher subscription prices, more aggressive monetisation of AI features, or advertising inside AI products. If you have wondered why free AI tiers feel generous right now, the answer is that the market is still buying growth. That phase does not last forever. The same dynamic is reshaping hiring, which we cover in our analysis of tech layoffs and AI spending, and it is why chipmakers keep posting figures like Nvidia’s record AI earnings.
What does it mean if you use Google products?
Four practical takeaways. Expect more frequent, smaller model updates rather than occasional big launches — the near-monthly cadence Pichai described is now the plan. Expect the cheap tiers to keep improving, because that is where Google is competing hardest and where most users actually live.
Do not plan around Gemini 3.5 Pro until Google publishes an official model card; “in testing” is not a release date, and this flagship has already slipped twice. And treat today’s free tiers as a promotional phase rather than a permanent entitlement — reasonable to enjoy, unwise to build a business process on without a fallback.
If you are new to Google’s AI stack, our explainer on what Google Gemini is covers the basics, and our roundup of everything announced at Google I/O 2026 sets out where this year’s roadmap started.
FAQ
How much is Google spending on AI in 2026?
Alphabet raised its full-year 2026 capital expenditure guidance to between $195 billion and $205 billion, up from a previous forecast of $180–190 billion. About 60% goes toward servers and 40% toward data centres and networking. The company has said 2027 spending will increase significantly again.
Why did Alphabet’s stock fall despite beating estimates?
Revenue rose 24% and Cloud grew 82%, but investors focused on the $15 billion increase in capital spending. Record quarterly capex of $44.9 billion pushed free cash flow to negative $5.9 billion, the first negative quarter in the company’s history, and shares fell about 3% in extended trading.
Has Gemini 3.5 Pro been released?
No. On the Q2 2026 earnings call, Sundar Pichai said Gemini 3.5 Pro is “currently in testing.” It was originally expected in June 2026 and has slipped past several rumoured dates. Google has not published an official release date or model card, so treat any specific date as unconfirmed.
What is Gemini 4?
Gemini 4 is Google’s next generation of frontier models. Pichai confirmed on the July 2026 earnings call that the company has begun what he called its most ambitious pre-training run yet for the model. No release timeline, specifications, or pricing have been announced, and pretraining is an early stage of development.
Will AI stay free if companies are spending this much?
Not necessarily in its current form. The spending has to be recouped eventually, which typically arrives as higher subscription prices, tighter free-tier limits, or advertising inside AI products. Generous free tiers today reflect a market competing for growth. It is reasonable to use them, but unwise to build critical workflows around them without an alternative.