Google's AI risk is execution

Google's AI platform story just shifted from capability to dependability.

Gemini 3.5 Pro was announced at Google I/O on May 19 with a promised 2 million-token context window and Deep Think reasoning, but its general availability has slipped from June to July; as of June 27, the model was still limited to a Vertex AI enterprise preview and waitlist. 1 The delay landed in the same week that Noam Shazeer left Google DeepMind for OpenAI, John Jumper left for Anthropic, and Bloomberg-reported departures of Jonas Adler and Alexander Pritzel pointed to more senior movement toward Anthropic. 2 3

For PMs, the product question is no longer "Can Gemini match GPT or Claude on benchmarks?" It is "Can Google ship the model your roadmap assumes, on the date your customer plan needs?"

Google DeepMind still has elite research assets. Gemini 2.5 Pro with Deep Think reportedly led MMLU-Pro at 89.8%, GPQA Diamond at 82.4%, and LiveCodeBench V6 in the same late-June window. 1 That matters, but it does not answer the PM problem. A model can be strong in evaluation and still create product risk if access, latency, pricing, or GA timing remains unstable.

The current signal is unusually concentrated. Shazeer was a Google engineering vice president, Gemini co-lead, Character.AI founder, and one of the eight co-authors of the 2017 Transformer paper; Google had brought him back in 2024 through a $2.7 billion Character.AI deal. 4 Jumper was the AlphaFold 2 scientist who shared the 2024 Nobel Prize in Chemistry with Demis Hassabis, and his move to Anthropic ended nearly nine years at DeepMind. 3 Adler was described as a Google AI coding lead, while Pritzel was described as a pretraining specialist and AlphaFold contributor. 2

Markets treated the departures as more than routine lab churn. Alphabet's GOOG shares fell about 5% on June 22, the largest one-day percentage drop since May 7, 2025, and Dow Jones Market Data put the one-day market-cap loss at about $225 billion, the largest single-day wipeout in Alphabet's history. 5 FAQ and Crypto Briefing reported a roughly $269 billion loss across the week, about 6% of Alphabet's estimated $4.4 trillion value. 1 6

A frontier-model vendor is not only selling intelligence. It is selling delivery confidence. When a CEO says a flagship model will be available "next month," product teams may staff integrations, write customer commitments, and delay alternative vendor work. 1 A two-month slip can turn a technical dependency into a launch dependency.

The talent story matters because it points to execution capacity, not only reputation. Fortune reported analyst Gil Luria's view that Google had briefly held the state-of-the-art model last year but had "fallen off since," and that the departures may mean Google is falling behind. 3 Build Fast with AI reported Demis Hassabis telling reporters at Cannes Lions on June 23 that talent mobility among leading labs is expected and that Google still has the largest and broadest research team in the industry. 7 Both statements can be true. Google can still have a huge bench, while PMs still face higher uncertainty around the next specific model they plan to use.

The most product-relevant detail is compute allocation. KuCoin/MetaEra and Build Fast with AI reported that Shazeer's team lost compute to DeepMind's London pretraining group shortly before his departure, leaving a Transformer-variant project stalled. 8 7 Treat that as reported context rather than confirmed internal record. The PM implication is still useful: in frontier AI, compute allocation is the research roadmap. If internal priority fights slow experiments, downstream product timelines can slip even when the vendor has enough money and GPUs in aggregate.

Google's reported AI infrastructure capex is about $190 billion per year. 1 That number should make PMs less forgiving, not more. Resource abundance does not remove execution risk if the scarce resource is research autonomy, launch discipline, or retained model leadership.

Start by adding a new row to your model-selection scorecard: vendor execution reliability. Put it beside quality, latency, cost, privacy, and security. The row should answer four questions:

For current Vertex AI or Gemini API plans, the next sprint action is a dependency audit. Identify every feature that assumes Gemini 3.5 Pro GA, long-context performance, or Deep Think reasoning. Mark each as launch-blocking, quality-enhancing, or optional. Launch-blocking features need a fallback provider or a degraded mode before July.

For roadmap commitments, stop promising model-specific capabilities to customers before the vendor has public GA, pricing, and latency documentation. Promise the user outcome instead. If the outcome is "analyze a 400-page contract," the implementation can prefer Gemini's long context when available and fall back to chunking plus retrieval on another model when it is not.

For architecture, separate the user-facing workflow from the model route. Use feature flags by model class, keep evaluation sets that compare Gemini, Claude, OpenAI, and open-weight alternatives on your own tasks, and log which model handled each request. The goal is not to abandon Google. The goal is to avoid making Google's July launch your own single point of failure.