NVIDIA Triton Wins CNCF’s Radar as Kubeflow Stays at Trial

NVIDIA Triton, DeepSpeed, TensorFlow Serving and BentoML took the adopt ring in a Cloud Native Computing Foundation survey of 302 developers. SlashData ran the Q3 2025 Technology Radar and CNCF released it on November 11, 2025, at KubeCon in Atlanta.

The same survey left two CNCF projects, Kubeflow and Argo Workflows, in trial. Production inference and long-running workflow tools won the vote; the foundation’s own ML platform stack did not.

NVIDIA Triton Took Adopt as Kubeflow Took Trial

SlashData asked professional developers who already use cloud native tools to score AI inference engines, ML orchestrators and agent platforms on familiarity, usefulness, maturity and whether they would recommend them. Composite scores, not CNCF’s sandbox-to-graduated ladder, placed each tool in adopt, trial, assess or hold. Personal views made up 75% of that score; usage filled the rest.

CNCF CTO Chris Aniszczyk used the release to push teams toward the adopt list.

Choosing technologies rated ‘adopt’ helps reduce risk and increase productivity.

Chris Aniszczyk, CTO, Cloud Native Computing Foundation

The full Q3 2025 radar findings make the split plain. NVIDIA Triton led inference. Airflow and Metaflow led orchestration. Model Context Protocol and Llama Stack led agent platforms. Kubeflow, a CNCF incubating project at the time, sat in trial for both inference and orchestration. Argo Workflows, already a graduated CNCF project, sat in trial for orchestration.

The radar note in the PDF is blunt about that gap: these rings do not have to match sandbox, incubating or graduated status. Graduation is a governance badge. Adopt is a user verdict from this sample.

What the Inference Scores Show

Two hundred two developers said they knew the inference category. Maturity scores came from 192 people who currently used a tool or had used it before. NVIDIA Triton Inference Server is a cloud product that exposes an HTTP or gRPC inference service so clients can call models the server is managing, with a design aimed at NVIDIA GPUs. That ops-shaped product is what this crowd rewarded.

AI INFERENCE RADAR POSITIONS

Triton’s combined 4- and 5-star maturity share is 80%. DeepSpeed and TensorFlow Serving trail on that cut and still clear adopt. Adlik, a Linux Foundation AI and Data project, did not. It won loyalty instead of volume.

ollama split the room. It drew 34% 5-star maturity scores and the highest pile of 1- and 2-star marks, at 23%. kgateway, a CNCF incubating project, matched DeepSpeed at 71% combined 4- and 5-star maturity and still missed the four-tool adopt set, which means usage or recommendation dragged it down. Envoy AI Gateway is a graduated CNCF project and posted the second-highest negative maturity share, at 16%.

CNCF PROJECTS ON THE INFERENCE CHART

• Kubeflow: Trial, and incubating when the survey ran.
• kserve: On the chart as an incubating project, outside the adopt four.
• Kubernetes Kueue: Graduated, still outside adopt.
• Envoy AI Gateway: Graduated, with a 16% 1- and 2-star maturity share.
• kgateway: Incubating, 71% combined 4- and 5-star maturity, not adopt.
• KAITO: Sandbox, listed with the rest of the field.

The foundation can graduate a project and still watch developers park it off the adopt ring. That is the scoreboard this survey actually printed.

A 134-Person Sample Behind Triton

The charts do not print a sample size under every bar. Lawrence Hecht, an enterprise IT analyst, called that out on the day of the release, saying the report is hard to trust because it does not say how many people are using or even aware of the specific technologies in the charts.

Liam Bollmann-Dodd, senior market research consultant at SlashData and the report’s author, answered for Triton the next day.

Good to see you, Lawrence!
"134 provided opinion on Triton, 77% are active users, the others are previous users" – Liam Bollmann-Dodd, who led the analysis and authored the report

— SlashData (@SlashDataHQ) November 12, 2025

THE TRITON USER POOL

• Opinions collected: 134 current or former users rated Triton.
• Still active: 77% of that group were active users; the rest were previous users.
• Usefulness: 41% gave 5 stars and 38% gave 4 stars, a 79% combined top-two share.
• Survey base: 302 developers overall, 202 of them familiar with the inference category.

Those usefulness stars are a Triton-user share, not a share of all AI developers. Adopt, in this design, is a verdict from people who already touched the tool. A 50% 5-star maturity score on Triton is a strong score inside that pool. It is not a census of the AI market.

HOW THE RADAR WAS BUILT

1. Q3 2025: SlashData fields the survey with 302 professional developers who use cloud native tools, scoring only the products they know.
2. October 2025: The PDF carries an October cover date and Q3 2025 running headers.
3. November 11, 2025: CNCF publishes the results at KubeCon + CloudNativeCon North America in Atlanta.

Usefulness meant how well a tool meets project needs. Maturity meant stability and reliability. Recommendation answers were turned into a net promoter score and folded into the same composite that produced the rings.

Airflow’s Usefulness Scores Had No Low Marks

Only two orchestrators reached adopt: Airflow and Metaflow. Metaflow led maturity, with 84% of its raters giving 4 or 5 stars. Airflow led usefulness and recommendation, and it received no 1- or 2-star usefulness marks at all. Metaflow, Airflow and Feast each drew 43% 5-star usefulness scores, so the 5-star peak was tied; Airflow’s edge was the empty basement.

ML ORCHESTRATION RADAR POSITIONS

BentoML is the split decision in the file. It made adopt as a serving tool and dropped to trial as an orchestrator. The PDF treats that as a warning about multi-role products: they can work in more than one job and still fail to lead all of them.

Flyte and Seldon Core landed in the middle of the pack, with Flyte drawing mostly 3-star marks. Bollmann-Dodd used those names as the room-to-run examples.

These findings show just how diverse the AI/ML toolchain has become. Metaflow and Airflow are excellent examples of how developer trust is earned through stability and fit-for-purpose design. But even newer projects like Flyte and Seldon Core are showing traction that signals opportunity for differentiation.

Liam Bollmann-Dodd, Senior Market Research Consultant, SlashData

Argo Workflows is the awkward row. It is graduated at CNCF and it still lost the adopt ring to an Apache scheduler and a Python-first ML framework. Teams that already run DAGs on Airflow had no reason, in this sample, to move.

MCP Took Adopt While A2A Took the Recommendations

Model Context Protocol and Llama Stack were the only agent platforms placed in adopt. MCP is an JSON-RPC protocol for tools and data, with hosts, clients and servers sharing context and exposing capabilities. That boring plumbing is what this sample treated as ready.

AGENT PLATFORM HIGHLIGHTS

• MCP: Adopt, with 80% 4- and 5-star usefulness and the broadest base among the top tools; 33% 5-star maturity and 73% combined 4- and 5-star maturity.
• Llama Stack: Adopt, with a 35% 5-star maturity share.
• agentgateway: Highest 5-star maturity share in the group, at 38%, without an adopt label in the announcement.
• Agent2Agent: 94% of users would recommend it, the highest recommendation rate in the report, with lower maturity than the adopt pair.

Agent2Agent, also called A2A, is an open standard for agent communication. Google created it and the Linux Foundation launched the project on June 23, 2025, a few months before this survey. The PDF’s own gloss is that excitement and perceived path can run ahead of what the code can do today. Adlik’s 92% recommendation rate in inference is the same shape of result: a smaller, happier user base, not a volume win.

CNCF’s write-up tied MCP’s usefulness scores to structured, agent-based design in jobs such as AI-powered customer support. The protocol won because developers could plug it in. A2A won because the people who had tried it wanted others to try it too.

Most AI Developers Still Are Not Cloud Native

The radar is easy to read as a production stamp. The companion SlashData study CNCF issued the same day undercuts that reading. Despite heavy infrastructure work, 41% of professional AI developers are cloud native. That 41% is a share of AI developers, not a share of Triton users, and it is the figure that belongs on the market, not on the inference table.

Thirty percent of AI developers use Machine Learning as a Service platforms that hide the machines. The people in this radar are the other path: they already live with containers, schedulers and service meshes, and they were asked to grade AI tools against that bar. The same companion release puts the wider craft at 15.6 million cloud native developers, with 77% of backend developers using at least one cloud native technology. AI is the lagging slice of that map, not the leading one.

Bollmann-Dodd said the workloads still lean on cloud native patterns even when the job title does not.

While only 41 percent of professional AI developers currently identify as cloud native, their workloads largely depend on it on a day-to-day basis.

Liam Bollmann-Dodd, Senior Market Research Consultant, SlashData

So the adopt list is a shortlist for the minority that already runs this way. NVIDIA Triton, Airflow, Metaflow and MCP are what those operators kept. Kubeflow and Argo Workflows, sitting on CNCF’s own project list, still have to win that same room. Aniszczyk told teams they cannot pick AI tools the way they did five years ago. This sample’s answer is narrower than a boom: use the serving stack and the workflow engine that already survived production, and treat the foundation’s ML platform projects as trial until those same 302-developer scores move.

Frequently Asked Questions

How Is the CNCF Technology Radar Different From CNCF Graduation?

Sandbox projects are early experiments. Incubating projects have a set technical vision and a growing contributor base but are still maturing in adoption, stability and governance. Graduated projects are widely adopted, with mature policy and governance. The radar’s adopt, trial, assess and hold rings are a separate user-perception score from this survey and, the PDF says, do not have to line up with those three stages.

How Did SlashData Define Usefulness and Maturity?

Usefulness was defined as how well a technology meets project requirements. Maturity was defined as stability and reliability. Both used a 1-to-5 star scale. Likelihood to recommend was converted into a net promoter score and then combined with usage to place each tool on the radar.

Who Selected the Tools on the 2025 AI Radar?

CNCF and CNCF’s End User Community picked the products for relevance and importance. Developers then scored only the ones they already knew. Respondents came from around the world, covered a wide range of specialties, and were recruited from third-party panels rather than from a single vendor list.

What Do Adopt, Trial, Assess and Hold Mean on This Radar?

Adopt tools are treated as reliable choices for most use cases. Trial tools are worth exploring to see if they fit a specific need. Assess tools need careful evaluation before a team commits. Hold tools are viewed as less mature or less useful in their current form.