Gemini 4 Is Already in Training — and Google Just Admitted It Needs to Catch Up

During Alphabet’s Q2 2026 earnings call on July 22, Sundar Pichai confirmed that Gemini 4 training is already underway — describing it as “our most ambitious pretraining run yet” — while simultaneously reporting that Google had just posted its first-ever quarterly negative free cash flow as a publicly traded company. The juxtaposition tells you everything about where Google thinks the AI race is headed and what it believes it will cost to stay in it.

The raw numbers are striking. Alphabet posted Q2 revenue of $119.8 billion, up 24% year over year, beating analyst estimates of $116.9 billion. Google Cloud grew 82% to $24.8 billion. The Gemini app crossed 950 million monthly active users, with daily active users tripling in a single year. AI Mode in Google Search crossed one billion monthly active users since its global launch in October. By any conventional income-statement measure, the quarter was one of Alphabet’s strongest on record.

The Cash Flow Statement Told a Different Story

Capital expenditures doubled year over year to $44.9 billion in a single quarter — overwhelmingly AI infrastructure — outrunning the $39.1 billion in operating cash flow the business produced. The result was free cash flow of roughly negative $5.9 billion, compared to approximately positive $24.5 billion just two quarters earlier. Alphabet CFO Anat Ashkenazi confirmed on the call that free cash flow will remain under pressure driven by continued large-scale infrastructure investments, and raised the company’s full-year 2026 capex guidance to between $195 billion and $205 billion — up from prior guidance of $180 billion to $190 billion. Analysts expect 2027 capex to reach at least $262 billion.

To finance the build-out, Alphabet is combining operating cash flow with significant debt accumulation and an $85 billion equity issuance. That is not the balance sheet posture of a company that believes AI infrastructure spending is nearing its peak. Shares fell more than 4% in after-hours trading on the guidance increase, suggesting investors are beginning to price in duration risk on the spending cycle even as the underlying business performs.

What Pichai Actually Said About Gemini 4 Training

Pichai’s comments on Gemini 4 were unusually direct for an earnings call, where frontier model roadmaps are typically kept vague. “We have started our most ambitious pretraining run yet for Gemini 4,” he said. “I am very excited by the progress I’m seeing internally. I’m confident that people will be pleased.” He framed the model’s target carefully: Google is building Gemini 4 to compete at the frontier level of where the frontier will be when Gemini 4 comes out — an acknowledgment that the competitive target is moving and that matching today’s top models is not the goal.

That framing also contains an implicit admission. Google is not currently at the frontier of agentic coding — Pichai said so explicitly, noting that a larger Gemini 4 base model is necessary to close that gap. He had made a similar admission in May on the Hard Fork podcast, where he said Google was “a bit behind” in agentic coding due to the lack of a developer-facing product that generates feedback data at scale. The Q2 call did not walk that back. It doubled down on the diagnosis while pointing to Gemini 4 as the solution.

The Gemini 3.5 Pro Problem

The context around Gemini 4 is made sharper by what is happening — or not happening — with Gemini 3.5 Pro. The model has been delayed. A late June update to its training data did not meet expectations, and coding performance issues have pushed the release to “as soon as it’s ready,” per Pichai. That is not a confidence-inspiring phrase from a CEO whose model was being awaited by partners. Google released Gemini 3.6 Flash one day before the earnings call, and separately released Gemini 3.5 Flash-Lite on July 21 targeting high-volume, lower-cost workloads — both efficiency plays rather than frontier pushes.

Gemini 3.6 Flash showed meaningful improvement: a 17% gain in token efficiency over its predecessor and a 12-percentage-point jump on the DeepSWE coding benchmark within six weeks. Pichai said Google now plans to ship updated models every month. But monthly efficiency updates and a delayed flagship are two different things, and the market has been watching the gap between Gemini’s current public offering and what Anthropic’s Claude Opus 5 and OpenAI’s GPT-5.6 family are delivering.

The Structural Bet Alphabet Is Making

Pichai’s remarks and the capex figures together describe a single strategic position: Google believes the AI race will be won or lost at the pretraining scale, not at the inference optimisation layer. That is why the company is willing to post negative free cash flow — something it has never done since going public — rather than slow the infrastructure build. The $514 billion cloud backlog and 82% Cloud revenue growth give Alphabet the credibility to make that argument to investors, even if the short-term share price reaction was negative.

The parallel with Google’s early cloud infrastructure decisions is not accidental. Google built data center capacity years before the revenue materialised, and that lead became a structural advantage. Pichai is explicitly betting that the same dynamic will play out with AI compute — and that arriving at Gemini 4 scale before rivals consolidate their positions is worth burning through cash reserves to achieve.

Whether that bet pays off depends on a variable Pichai cannot control: how quickly the frontier moves between now and Gemini 4’s release. The model is in pretraining today. It will not ship for months. In an environment where xAI is planning Grok 4.6 and 4.7 within four weeks, and where Anthropic has committed to an accelerated release cadence, the frontier Google is training toward is not standing still. The full earnings transcript is available via Google’s official blog.

Frequently Asked Questions

Is Gemini 4 confirmed?

Yes. Sundar Pichai confirmed during Alphabet’s Q2 2026 earnings call on July 22, 2026, that Google has begun pretraining for Gemini 4, describing it as “our most ambitious pretraining run yet.” No release date has been announced.

Why did Google post negative free cash flow in Q2 2026?

Google’s capital expenditures reached $44.9 billion in Q2 2026 — double the year-earlier figure — primarily for AI data centers, networking infrastructure, and custom chip deployments. This exceeded the $39.1 billion in operating cash flow produced, resulting in negative free cash flow of roughly $5.9 billion. It was the first negative free cash flow quarter in Alphabet’s history as a public company.

What is the status of Gemini 3.5 Pro?

Gemini 3.5 Pro remains delayed as of the Q2 2026 earnings call. A late June update to its training data did not meet expectations, and coding performance issues pushed the timeline back. Pichai said it will ship “as soon as it’s ready.”

How does Gemini 4 fit into Google’s competitive strategy?

Pichai framed Gemini 4 as Google’s answer to acknowledged gaps in agentic coding, where rivals like Anthropic and OpenAI have gained momentum. The model is being trained to match the frontier as it will exist at release — a recognition that the competitive target is moving faster than any single model can chase.