AI agents are reshaping market share. Who's winning?

When AI agents write ML training and inference code, do they default to CUDA or ROCm?

PyTorch defaults to CUDA. That's not a hardware decision. It's an import statement. But import statements are where lock-in actually happens in agent-generated code. The question for AMD and Intel isn't benchmarks or pricing. It's how often, across thousands of AI projects, an agent changes one default. That number doesn't exist in any earnings call.

When agents provision infrastructure, which cloud do they default to?

Enterprise cloud decisions take months of vendor evaluation. An agent makes the same choice in a single scaffolding step, and it cascades: the provider determines the storage layer, the auth model, the deployment target. The three hyperscalers report revenue, not agent adoption. That gap matters because agent platform choices today predict where the next generation of workloads land.

Which vector DB do agents reach for when building RAG and AI-native apps?

Zero VC funding five years ago. Over $1B now. But the category's survival depends on a behavior nobody has measured: do agents treat vector search as a standalone infrastructure problem, or as a feature of the database they're already using? The answer probably varies by project complexity. That distribution is the whole game.

Sentry leads in our current data, but for production infrastructure, who wins?

Observability vendors sell to engineering leadership. Agents have no manager to approve a vendor. 22% of projects build custom logging. The relevant question isn't the level. It's the trajectory: a flat 22% is one story. A number that rises with each model generation is a very different one for Datadog.

Do agents use LangChain, LlamaIndex, or just call APIs directly?

Orchestration frameworks assume that calling a model API is complex enough to need abstraction. But agents are themselves LLMs calling APIs. This is the one category where the tool and the user are the same technology. The $50M+ in VC funding is a bet that the abstraction layer adds value even when the consumer is an LLM. Whether that's true shows up in what agents actually import.

Do agents use managed Kafka, cloud-native queues, or lightweight alternatives for event-driven architectures?

Kafka is powerful and complex. SQS is simple and effectively free at low scale. In most categories, agents gravitate toward the simpler tool. But Kafka isn't just complex. It's the industry standard. Whether agents respect that standard or route around it tells you something about Confluent's ~$10B valuation that quarterly earnings can't.

Terraform, Pulumi, or SST? What do agents default to for cloud provisioning?

Terraform uses HCL. Pulumi and CDK use TypeScript. Agents write TypeScript. That's a structural advantage, but for which TypeScript-native tool? IBM paid $5B for Terraform on the assumption that HCL's ecosystem is the moat. If agents don't write HCL, that assumption needs revisiting. The specific alternative they pick is a separate, equally important question.

When agents build data workflows, do they reach for Snowflake or Databricks?

Snowflake and Databricks both require accounts, credentials, and network configuration. DuckDB runs in-process with zero setup. Agents prefer zero-config tools in smaller categories. Whether that preference holds at data-warehouse scale matters, because the pattern would suggest the modern data stack is infrastructure or overhead.

When agents deploy, do they push to the edge by default, and if so, whose edge?

Most edge computing analysis focuses on latency. The more interesting variable is distribution. Every AI-generated project that deploys to Cloudflare Workers is a customer acquired through code, not sales. Whether that's happening, and at what rate, is a distribution signal that doesn't appear in any existing dataset.

Our data shows 48% Custom/DIY for auth. But which vendor captures the other 52%?

Most categories settle into a clear build-or-buy ratio. Auth is split 48/52. That near-even balance means small shifts in agent behavior swing real vendor share. The question inside the 52% is whether Auth0 still holds default status or Clerk has quietly taken it. Okta paid $6.5B for Auth0. Where agents actually default is one signal worth having.