Google’s own memo says Kavukcuoglu, formerly DeepMind’s CTO and Google’s chief AI architect, will oversee Gemini model development, frontier-AI research, the Gemini app, and developer teams as senior vice president of Google DeepMind. Hassabis becomes DeepMind chair and Alphabet chief scientist, retains leadership of Isomorphic Labs, and says he will advise the model and research teams rather than run the unit’s daily operations.
Axios independently reported that the reorganization lands amid delayed model releases, senior-researcher departures and competitive pressure from OpenAI and Anthropic. But the company’s announcement is more specific about the operating model than SemiAnalysis’s “Gemini is cooked” framing: Google is moving the executive responsible for its technical architecture into direct control of the product, model, and research organization. That is a high-stakes reset, not evidence of a shutdown.
The more defensible part of SemiAnalysis’s argument is about Google Cloud Platform. Alphabet’s second-quarter results show Google Cloud revenue growing 82% year over year, to $24.8 billion, while the company began recognizing revenue from TPU system sales. Google Cloud is becoming a substantial commercial outlet for the same AI infrastructure that underpins Gemini, including external hardware deployments rather than only usage inside Google-operated regions.
For IT leaders, developers, and Windows administrators evaluating Gemini, Vertex AI, Gemini Enterprise, or Google Cloud’s growing accelerator portfolio, the practical implication is straightforward: Google is increasingly organized to sell AI capacity, tooling, security, and agent infrastructure to customers—even where the customer’s preferred model is not Gemini.
Koray Kavukcuoglu inherits the Gemini delivery problem
Google’s memo places a large and unusually broad remit under Kavukcuoglu. He is now responsible for the flagship-model pipeline, frontier research, consumer Gemini experiences, and developer-facing teams. That structure reduces the organizational distance between model research and shipping products, which has been a recurring weakness at large AI organizations trying to turn breakthroughs into dependable releases.
Hassabis is not leaving Alphabet, despite some initial social-media interpretations of the reshuffle. He remains DeepMind chair, becomes Alphabet chief scientist, continues to run Isomorphic Labs, and will advise Google DeepMind leadership. Still, the title change is meaningful: Kavukcuoglu reports to Pichai, not to Hassabis. The role of Google DeepMind CEO is effectively gone from the operating chart.
SemiAnalysis treats that as proof that DeepMind has ceased to be a serious frontier laboratory. The public record does not support that conclusion. Google says Gemini 4 has entered its “most ambitious” pre-training run and says Gemini 3.5 Pro remains in testing. It also explicitly says Kavukcuoglu will oversee frontier-AI research. Those statements do not guarantee that Google will regain a lead in model quality, but they directly contradict the claim that the company has stopped trying.
What has been independently confirmed is a more concrete problem: Gemini 3.5 Pro missed an earlier expected release window. TechCrunch reported in July that Google had been testing the model with partners after it failed to arrive alongside other Gemini releases. Ars Technica likewise reported that the model was still in testing when Google shipped Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and its specialized cybersecurity model.
That distinction is important. SemiAnalysis says Google “silently canceled” Gemini 3.5 Pro. As of August 8, Google’s public position is the opposite: the model is in testing. A delayed flagship is a material execution failure in a market where competitors are shipping rapidly; it is not proof of cancellation.
Jeff Dean’s departure matters, but the departure list needs scrutiny
Jeff Dean’s exit carries weight beyond a routine executive transition. He helped build Google’s early distributed systems, co-founded Google Brain, and was a central figure in the TPU program that has now become one of Google Cloud’s most valuable differentiators. Google confirmed that Dean and Senior Fellow Sanjay Ghemawat are launching an independent public-benefit corporation aimed at machine learning, science, and engineering.
Google also committed to work with the new company as a founding investor and Cloud partner. That is a revealing arrangement. Rather than simply losing two senior technical leaders to a rival hyperscaler or frontier lab, Google is attempting to preserve an economic and infrastructure relationship with them. The company can retain a customer, research collaborator, and potential source of future work while Dean and Ghemawat gain an independent vehicle.
SemiAnalysis identifies the company as Discovery Loop and says Quoc Le and Oriol Vinyals are joining its founders. Axios reported that Dean was leaving with three top Google AI colleagues, and other accounts have named Le and Vinyals. But Google’s official memo names only Dean and Ghemawat. The company has not publicly detailed Discovery Loop’s full founding team, its funding terms, ownership structure, access to Google IP, or whether its Cloud relationship includes preferential TPU capacity.
Those omissions matter because they determine whether this is primarily a talent loss, a Google-backed spinout, or both. A startup led by former Google Fellows that buys or receives large-scale Google Cloud capacity may still strengthen GCP revenue and utilization. It does not automatically strengthen Gemini’s internal research capability, particularly if former leaders take future breakthroughs outside Alphabet’s organizational boundary.
Google Cloud’s growth is real, but the 82% figure is not pure cloud consumption
The strongest factual element in SemiAnalysis’s report is Google Cloud’s acceleration. Alphabet reported 82% year-over-year Cloud revenue growth in the second quarter of 2026, and Pichai said the segment’s backlog reached $514 billion. Google cited demand for AI infrastructure and AI solutions, and it says nearly 90% of the Fortune 100 use Gemini Enterprise.
However, readers should not treat the 82% growth rate as a clean measure of recurring cloud-platform demand. Alphabet has confirmed that it began recognizing revenue from TPU system sales in the quarter: sales that package hardware, software, installation, support, and warranty services. That makes the reported Cloud number a mix of traditional consumption revenue, AI software and platform services, and increasingly large infrastructure deliveries.
The Information reported that Cloud revenue reached $24.8 billion in the quarter and that Alphabet raised its 2026 capital-expenditure forecast to between $195 billion and $205 billion as it expands capacity. Alphabet has said much of the revenue from TPU deliveries is expected in 2027. That supports the direction of SemiAnalysis’s thesis: Google Cloud has a significant hardware-and-capacity growth vector beyond conventional virtual machines, storage, and managed services.
But SemiAnalysis’s precise forecasts—including $35 billion per gigawatt, $150 billion of TPU system backlog, $200 billion of future external sales, and a three-dollar contribution to 2027 Alphabet earnings per share—are its own estimates. Alphabet has not published those figures. They should be read as a bullish analyst model, not as reported company guidance.
The existence of multi-gigawatt TPU commitments is independently documented. Broadcom disclosed in an April filing that Anthropic is expected to access approximately 3.5 gigawatts of next-generation TPU-based AI compute beginning in 2027, subject to Anthropic’s continued commercial success. The filing describes that capacity as part of a broader multi-gigawatt commitment and says the parties are working with operational and financial partners.
That is bigger than a normal cloud contract. It suggests Google’s TPUs are moving from an internal cost advantage and Cloud option into a product sold through external data-center arrangements. The resulting revenue may be less recurring and less comparable to ordinary Google Cloud consumption, but it is still revenue generated by Google’s accelerator design, supply chain, networking, and software stack.
The platform strategy creates a choice for enterprise buyers
Google’s leadership has openly embraced a broad infrastructure posture. In its earnings materials, the company highlighted support for its own TPUs and Nvidia platforms, along with native support for JAX, PyTorch, vLLM, and SGLang. It also promoted a network architecture intended to connect large accelerator fleets across multiple data centers.
For enterprise customers, that is more valuable than a narrow “use Gemini or leave” strategy. A Windows-heavy organization can keep existing Microsoft, Active Directory, endpoint-management, and developer workflows while using Google Cloud for selected AI workloads—particularly where TPU pricing, model serving, security tooling, or agent-development tools make economic sense. The company’s Agent Development Kit and Gemini Enterprise are designed to connect agents to enterprise systems and apply governance and cost controls, which is the work many IT teams care about more than a benchmark leaderboard.
The catch is vendor concentration. Google is asking customers to view it simultaneously as model developer, cloud provider, TPU supplier, security platform, agent platform, and partner to other AI labs. That can create flexibility at the model layer, but it makes procurement and exit planning more important. Organizations should ensure their prompts, retrieval pipelines, identity integration, logging, evaluation suites, and agent tools remain portable across model providers and accelerator back ends.
Google’s own support for PyTorch, vLLM, and SGLang is helpful here. Buyers should use those portability claims in contracts and technical architecture, rather than assuming that a workload launched on Vertex AI can be moved cleanly later. The decisive details are model availability, regional capacity, data-residency controls, committed-use terms, API compatibility, and whether a managed-agent implementation can be exported without rebuilding it.
SemiAnalysis is right to identify an internal tension: every TPU committed to external customers is capacity that cannot be instantly reassigned to a new Gemini training run. Yet Google’s current strategy is not an either-or choice between Gemini and Cloud. It is to make infrastructure sales fund the same hardware, networking, and data-center expansion that Gemini can use later.
The immediate test is less dramatic than whether DeepMind can ever lead the field again. Google has said Gemini 3.5 Pro is still testing and Gemini 4 is in pre-training; until either becomes a publicly available model with a model card, pricing, API access, and sustained developer adoption, the company’s frontier claims remain promises. Google Cloud’s revenue, TPU system deliveries, and enterprise AI tooling, by contrast, are already showing up in reported financial results.
References
- Primary source: SemiAnalysis
Published: August 7, 2026 at 2:32 AM UTC
Gemini is Cooked but GCP is Cooking
or why DeepMind's long term failure is GCP's short term gainnewsletter.semianalysis.com - Related coverage: theinformation.com
How Google’s New AI Architect Plans to Spread Gemini Everywhere — The Information
When Google’s billionaire co-founder, Sergey Brin, needs something done at the company’s artificial intelligence unit—AI chips allocated to a team, a researcher hired, an employee moved to a new department—there’s one person he messages: Koray Kavukcuoglu. Kavukcuoglu isn’t Google’s flashiest AI ...www.theinformation.com - Related coverage: techcrunch.com
Google releases three new Gemini models — but no 3.5 Pro | TechCrunch
Google released Gemini 3.6 Flash, 3.5 Flash-Lite, and Flash Cyber, but the continued absence of Gemini 3.5 Pro raises fresh questions about its AI strategy.techcrunch.com - Related coverage: swiftkvm.com
Has Gemini 3.5 Pro Been Delayed in 2026? Google Release Date, Postponement Reasons, and Latest Updates - Swiftkvm Blog
As of July 17, 2026, Gemini 3.5 Pro has missed Google's June window but is not cancelled. Flash is live; Pro remains in testing. Original timeline, delay reasons, and official signals to watch.
swiftkvm.com
- Related coverage: arstechnica.com
Google announces Gemini 3.6 Flash and cybersecurity AI, teases 3.5 Pro and Gemini 4 - Ars Technica
There are new 3.6 and 3.5 models today, but Google is already training Gemini 4.arstechnica.com - Related coverage: computerworld.com
Gemini 3.5 Pro is late but Gemini 4 will be great, says Google CEO – Computerworld
A more frequent model release cycle may bring faster improvements to capabilities and pricing, but CIOs will pay the price in testing, governance and validation, analysts say.
www.computerworld.com
- Related coverage: decrypt.co
Google Ships New Gemini Flash Models, But Pro Is Still Missing - Decrypt
As Gemini 3.5 Pro stalls in testing limbo, Google ships 3.6 Flash, 3.5 Flash-Lite, and a restricted cybersecurity model—and quietly teases Gemini 4.decrypt.co - Related coverage: tomshw.it
- Related coverage: magica.com
Gemini 3.5 Pro misses June rollout amid coding report | Magica
Gemini 3.5 Pro missed its expected June rollout. Google confirms partner testing but not the reported coding shortfall or a new release date.magica.com - Related coverage: packetnebula.com
Gemini 3.5 Pro missed July 17: launch vs leak | PacketNebula
Google's Gemini 3.5 Pro slipped its reported July 17 target, a third miss. The 2M context, Deep Think and pricing are all leaks. Here is what Google actually said.www.packetnebula.com - Related coverage: kie.ai
What Is Gemini 3.5 Pro? Google's 2M-Context Flagship
Gemini 3.5 Pro is Google DeepMind's unreleased flagship — reportedly 2M context, Deep Think reasoning, rebuilt from a new pretraining run.kie.ai - Related coverage: androidcentral.com
Google Gemini AI: Gemini 3.5 Flash, Nano Banana, Live, best features, linked apps, and more | Android Central
From Gemini 3.5 Flash to all the Google apps that use Gemini, this guide will break down every key AI feature and whether you need to pay for it.www.androidcentral.com - Related coverage: androidcentral.com
Google thinks Gemini 3.5 Flash can finally make AI agents more useful | Android Central
Google just supercharged AI speed with Gemini 3.5 Flash, but the Pro model is still weeks away.www.androidcentral.com - Related coverage: cincodias.elpais.com
Loading…
cincodias.elpais.com - Related coverage: tomsguide.com
Google just launched Gemini 3.5 Flash — here's all the upgrades | Tom's Guide
Google just launched Gemini 3.5 Flash alongside Gemini Spark, a new always-on AI agent designed to complete tasks, automate workflows and move beyond chatbot-style AI.www.tomsguide.com