A K2 stream carrying 1 TB a day to three consumers costs about $165 a day at anticipated prices, and nearly three quarters of that is the read side. The beta is a sandbox that competes in 2027, and Cloudflare will not be the cheapest home for fan-out.
At its beta limit of 30 MB/s, a K2 stream fills its 10 GB storage allowance in about five and a half minutes (10,000 MB divided by 30). Cloudflare announced the event streaming service on October 1. It is a sandbox for Workers Paid accounts, is free during the beta, and cannot yet be accessed by a Kafka client.
The anticipated pricing is more settled than the product. Cloudflare expects to charge $0.04 per GB produced, $0.04 per GB consumed and $0.02 per GB-month retained. Consumption is billed per consumer. K2 targets Kafka-shaped workloads that can tolerate higher latency, and its read fees grow with every consumer in a fan-out.
Take a stream carrying 1 TB a day, kept for the 7 days (604,800 seconds) that the post uses as its default retention, with three independent consumers. Producing costs $40 a day. Consuming costs $120, because each consumer reads the full terabyte. Retaining roughly 7 TB costs about $5 a day. The total is near $165, nearly three quarters of it reads.
A fourth consumer adds $40 a day, as much as the entire production cost. Storage, often the line buyers focus on, barely registers. A team sizing a stream by write volume alone may underestimate what subscribers cost; the bill rises with every team that wants the data.
This matters because the pitch invites exactly that pattern. Consumers lease batches for five minutes and ack, nack or extend them. Retries are per batch, not per message. Fan-out comes either from separate subscriptions or from shared consumer pools. Every extra team that wants its own view of the data is another consumer on the meter.
The arithmetic makes it unlikely that Cloudflare will undercut every competitor. Cheap object storage lowers the storage cost, but per-consumer read fees make K2 expensive for heavy fan-out. A shop with ten downstream readers of one firehose would pay $400 a day in read fees on that terabyte alone. K2 is more competitive for streams with one or two consumers and longer retention, a narrower market than its headline pitch suggests.
K2 is a partitioned durable log built on R2, Cloudflare's object storage. It batches segments into files and uses R2 atomic operations for ordering and offsets, so teams do not have to run, patch or pay for a separate coordinator. This keeps the system simple and inherits R2's durability, with an initial produce latency of about one second at p99.
K2 began as the ingestion buffer for Cloudflare Pipelines because Cloudflare's edge, which spans more than 335 cities, often cannot run Kafka-style systems. In 2025, Cloudflare bought Arroyo, the stream processing company founded by Micah Wylde; Arroyo now powers Pipelines. K2 is an internal component that Cloudflare is turning into a product, bringing both the advantages and rough edges of that route.
On the Hacker News thread, a K2 engineer wrote that "every data system that doesn't need <100ms latency is moving to object storage." The engineer added that the K2 team sits beside the R2 team and can develop the two products together. An S2 commenter said its serverless stream gets about 50 ms p99 acknowledgments in the same region by flushing multi-tenant objects, suggesting that one-second latency is a design choice. Other commenters focused on object-store egress fees, a cost K2's pricing will have to address.
The main competitors have been around longer. Confluent is now part of IBM, which completed its roughly $11 billion all-cash purchase ($31 a share) on March 17, 2026. Confluent also owns WarpStream, a Kafka-compatible service on object storage that it acquired on September 9, 2024. S2 is the smaller, faster alternative. Both already support existing clients or offer lower latency, while K2 does neither yet. Its roadmap lists multi-GB/s streams, message keys with key based ordering, push based Worker consumers, an Express tier with lower latency and "drop-in support for Apache Kafka clients." None of these features is in the beta.
K2 may suit teams building on Workers who want a durable buffer beside their code, analytics ingestion where a second of delay is acceptable, or streams with a couple of consumers. Cloudflare distinguishes it from its other products: Queues tracks individual work items with retries and dead letters, while Pipelines is for data headed to R2 or Iceberg. K2 handles high scale movement, retention and fan-out when custom processing is needed.
Teams with Kafka estates that depend on client compatibility, or workloads requiring responses under 100 ms, are not ready to use K2. The beta is a curiosity for those buyers; its 10 GB and 30 MB/s caps also make it unsuitable for a serious load test. A platform team can try it on a side project this quarter, but cannot yet plan around it.
Kafka client support and keyed ordering will determine whether K2 belongs in migration plans, since Kafka estates rely on both. We expect Cloudflare to finalize pricing when the beta ends, perhaps with fan-out discounts to attract the multi consumer workloads its current price sheet penalizes. Buyers should model their consumer count first. The first 2027 roadmap update should show whether Kafka client support and keys have release dates.