There is AWS, Microsoft and Google Cloud; but let’s not forget Alibaba Cloud.
We were reminded of this at the Google Cloud Summit Paris 2026. We looked at how customers and partners perceived the building blocks of agentic stacks among the “three big” American cloud players, with a particular angle: convergence or divergence?
Alibaba was mentioned for its promptness in implementing OpenClaw / Hermes, supported by a dedicated sandbox. NVIDIA was often cited in parallel, having also positioned itself on the topic, under the banner OpenShell.
Google Cloud, ahead on first-party models
From TPUs to prepackaged agents, Google’s command of its stack did not escape the attendees (around 3,600). And the Gemini Enterprise Agent Platform, a rebrand of Vertex AI, seems to have made its impact.
This mastery – or at least its demonstration – was particularly visible at the level of LLMs. In the keynote opening the Summit, there was little focus on anything other than Gemini. A family that is increasingly steering toward the “quad-modal” approach (text, image, audio, video). Gemini Omni and Gemma 4 are its latest exemplars. We were reminded, in this regard, that Gartner had classified Google as a “leader” in multimodality.
AWS also has in-house models (Nova). But the market remains far more cautious about them. So much so that it still treats Amazon’s subsidiary as an LLM broker. As for Microsoft, following its notably close partnership with OpenAI, it diversified only late. Its catalog includes first-party models, but they have not yet scaled.
On data, a more siloed Microsoft, a more agnostic AWS
Divergences also exist on the data side. In these fields, the Microsoft Foundry foundation is seen as more fragmented than the stacks of Google Cloud and AWS. The latter takes a more agnostic stance (“come as you are”), we are told, even if the other two are gradually opening up.
On the infrastructure side, the opinion largely converges. The word “convenience” comes up often. We are pointed to chunking SaaS at a “negligible cost” and the standardization of tokenizers, which favors the portability of RAG.
MCP, common denominator but not universal
There is, more globally, a certain parity on the elements necessary to industrialize (security, authorization, evaluations…), we’re told. The arrival of industry standards helps. MCP is at the forefront, even if, we’re told, it is not necessarily the best way to expose APIs. Google Cloud has indeed weighed in: if it has “MCP-ified” its services, it also offers other interaction channels, such as Workspace CLI and Antigravity CLI.
If the infrastructure isn’t a major differentiator among clouders, it gives them an edge in AI labs. Claude today offers a better cost/performance ratio on GCP than on Anthropic’s servers, according to a Google Cloud partner.
A convergence in the push for agentic AI, less in the targets
Clients and partners generally agree that the “three giants” are converging in how they push agentic AI forward. Their histories, however, still matter. AWS remains B2B-oriented. It is, for example, the only one able to claim real use cases in mainframe modernization. Microsoft, by contrast, is B2C-oriented. Google is largely the same. It is also credited with a more open mindset… at the risk of abandoning products more readily. We’re reminded of Veolia’s disappointment when Google Cloud Print was discontinued, even though its fleet had standardized on Chromebooks.
When directly asked to Google Cloud teams, they primarily emphasize the group’s history in fundamental research. In the showcase, DeepMind, responsible for the landmark paper “Attention Is All You Need”… and, among others, AlphaFold, which a partner spontaneously cited to us (this open model predicts protein structure from amino acid sequences).
The separation of responsibilities challenged
Convergence or divergence, the agentic paradigm nonetheless challenges certain best practices. Notably, the separation of responsibilities in systems, according to Nacim Rahal, VP data & AI at Doctolib, who spoke during the keynote. “We can have an agent capable of doing many different things. Things that affect the systems themselves, but also the design of the organization that will enable them to be produced. We will have product managers who will directly modify the prompts of the systems. Data engineers, ML engineers and software engineers who end up working together on the same kind of task.”
At RATP Dev, out of 25,000 employees, “more than 1,000” have switched to Gemini Enterprise. “We were not early adopters,” admits Hiba Farès, chairwoman of the executive board. The company began experimenting with solutions in 2023, but only truly pushed beyond test & learn by the end of 2025. “I was starting to perceive maturity,” explains Farès. “People talking about real use cases that could bring a meaningful ROI.” “Our data started circulating a bit too much outside the company,” she also admits, citing “50 shadow AI requests per employee on professional devices.”
“Seeing people who truly have business authority learn this unlocked something,” the executive notes about the management committee, which went to train at Google. RATP Dev also has 50 volunteer “champions”… who do not come from IT or digital. “They aren’t domain experts. They’re the people you want to talk to around the coffee machine.” These “field sensors” raise daily needs and friction points.
