Google scraps Gemini 3.5 Pro: does a six-month frontier gap erode Cloud's

Google spent most of 2026 unable to ship a flagship. Its Pro-tier Gemini 3.5 — promised for June, pushed to July, then internally scrapped when internal candidates failed to improve enough over the cheaper Flash line — left the company with no frontier model on the shelf. Even as of August it had not shipped and instead of launching it, Google was already hyping Gemini 4, which by early September sat in the post-training phase. That leaves a roughly six-month window, possibly longer, in which the enterprise contracts that pay top dollar for frontier intelligence run toward OpenAI and Anthropic. The stakes are Google Cloud, the segment growing fast enough to carry Alphabet's ~$4.1 trillion market cap and its roughly 33x forward multiple.

The scary read writes itself: no flagship, a Flash-first strategy built on high-volume, low-margin inference, and the model ranked 8th or 9th place by one widely-cited AI-infrastructure shop. If OpenAI and Anthropic pocket the high-margin enterprise commitments in this window, the argument goes, Cloud's premium revenue decays even while its volume grows — a slow bleed that eventually shows up in the growth rate and then in the multiple.

The trouble is that Cloud's actual numbers do not look like a business in decay. In the quarter ended June 30, Google Cloud grew 82% year-over-year to $24.8 billion, accelerating from 63% the quarter before, with operating income of $8.81 billion, up 212% and a cloud backlog of $514 billion. On those figures, management is telling a story of premium enterprise demand — not distress. "Wide adoption of Gemini Enterprise" with nearly 90% of the Fortune 100, CEO Sundar Pichai said on the call. If that is the whole picture, there is nothing to worry about, and the six-month gap is a rounding error against a backlog the size of a sovereign fund.

But Cloud and Gemini are not the same thing, and conflating them is where the real risk hides. Google is winning Cloud in 2026 partly at the layer underneath the model — by becoming the neutral shop that hosts rivals and sells them the glass. The Gemini Enterprise platform, the strong part of the premium story, reportedly draws much of its strength from including third-party models, such as Claude. And in a striking inversion, Google is now selling large chunks of its TPU compute directly to Anthropic — more than 20% of total TPU shipments from late 2026 through 2027, by one estimate — booking system sales at low-30s% margins. Google is monetizing the physical layer even as the model layer slips. That is a training-to-inference value migration in its purest form: the brand that can't win the frontier is selling the shovels to the ones that can.

The deceleration is already visible where it should be. First-party API token growth slowed from 60% in the first quarter to 38% in the second, and with it the growth of Gemini's first-party API revenue. The aggregate Cloud number holds because hardware sales and neutral hosting fill the hole the weakened model leaves. That is the tell: Cloud is growing, but increasingly its growth is bought, not earned at the frontier.

So the honest way to frame this is a falsifiable claim. The thesis — that the six-month gap erodes Cloud's premium mix and the multiple — is invalidated only if three things all hold: Gemini 4's post-training completes and its Pro tier ships on schedule; independent third-party benchmarks put the Flash line at or near frontier flagships; and Cloud's premium (first-party enterprise) growth holds once you strip out TPU system sales and third-party model hosting. Score the three today and the thesis is not beaten back yet.

On benchmarks, the strongest of Google's three defenses fails. The just-released Gemini 3.8 Flash — the third Flash release in six weeks, priced at $0.75/$3.75 per million tokens against the far pricier flagships from OpenAI and Anthropic — claims frontier-level coding and agentic results. But almost all the data so far comes from Google itself; the one clearly independent third-party test, on a patching benchmark, puts the Flash Cyber variant at 47.2% against 47.8% for the leading frontier model — behind, not ahead. Flash may be competitive on the workloads enterprise customers actually buy, but the evidence for it is self-reported at the moment.

On the premium mix, the second defense fails as well. The underlying first-party Gemini layer is already decelerating, exactly as the thesis predicts, even while the aggregate is propped up by compute sales and neutral hosting. The premium story is not holding on strength; it is holding on financialization.

That leaves Gemini 4 timing as the entire swing factor. Pre-training began in late July, and prediction markets put roughly an 89% chance of a launch before the end of 2026, with the Pro tier likely to follow into 2027. That is a plausible on-schedule outcome — and it is the one fact that would invalidate this whole argument. It is also the least certain one, because the same reporting that flagged the cancelled 3.5 Pro describes a DeepMind gutted by departures: co-founder Demis Hassabis out of day-to-day operations, Gemini co-lead Jeff Dean leaving for a new lab, and a deep-talent bench walking out the door. Shipping a frontier model on schedule through that churn is a harder promise than the prediction-market odds imply.

Here is where the risk actually lands on the multiple. Cloud exited the quarter at a run rate near $99 billion a year, growing ~82%. The sell-side consensus already assumes it decelerates to ~64% in 2027; the risk case is a move below that as premium leaks and the mix shifts toward lower-margin volume. Alphabet's forward multiple in the low 30s is being paid largely for Cloud's hypergrowth and the AI optionality inside it. Each sustained point of growth erosion, and each visible shift from premium contracts to hardware, trims the premium that multiple carries. A de-rating from ~33x forward to the high 20s — a 10-15% compression, several hundred billion in market value — is a reasonable sensitivity if the premium mix deteriorates even while search sits untouched.

I would not frame this as a Cloud collapse, because the evidence points the other way: Google is buying time and revenue by selling the layer under the model. But that is precisely the problem the thesis names. A Cloud whose growth rests on selling TPUs to Anthropic and hosting Claude is a lower-multiple, less durable business than one compounding a frontier model — even when the headline growth looks identical. The gap does not need to bankrupt Cloud to hurt the stock. It only needs to be confirmed by one more quarter of first-party API deceleration, one more self-reported Flash benchmark, and one missed Gemini 4 date. Which is to say: watch Gemini 4's post-training, not the price target.