Executive summary
For buyers, two things follow. The benchmark is single-sourced: no competing measurement of Cerebras against Groq, SambaNova, or current-generation Nvidia inference stacks appears anywhere in the public record cited here. What Cerebras' enterprise deployment terms specify, and whether they match the published public-tier speed, is not disclosed in any source reviewed here.
Alibaba is both Cerebras' model supplier and a full-stack AI competitor, having reported cloud revenue up 45% year-over-year in its June 2026 quarter alongside net income down 75% on a 75% rise in capex tied partly to "higher prices across a broad range of chip components," per Alibaba's CNBC-reported results. That cost pressure makes third-party inference hardware more attractive industry-wide, though no source ties that dynamic directly to Cerebras' contract economics.
What Cerebras' dedicated, non-public tier actually costs and caps at scale is undisclosed, and no customer or third party has benchmarked it independent of Cerebras' own marketing.
For practitioners
The cached-token billing detail compounds this rather than sitting beside it. Most inference providers discount or exclude cached context from rate accounting, because the point of prompt caching is to let an application resend a large system prompt or retrieved-document context on every call without paying the full compute cost again. explainx.ai's analysis of the Qwen 3.8 27B launch reports that Cerebras' public tier counts cached tokens against the 150,000-per-minute cap at full price. For an agentic loop, a tool-calling pattern where the model resends a growing conversation history or retrieved context on every step, this means the token-per-minute budget is consumed by re-sent context long before it is consumed by new generation.
This is the mechanism behind the "poor fit for long-running agentic workflows" framing, and it is a mechanism, not a caveat: the 128,000-token context ceiling and the cache-billing rule interact multiplicatively. A model with a long context window is attractive precisely because it lets an application carry more state, retrieved documents, multi-turn history, tool outputs, across a session. Cerebras' architecture supports the full 131K context on Qwen3-235B, per its own Qwen3-235B press release, but the rate-limit structure penalizes exactly the pattern that long context is meant to enable. The workloads best suited to the 1,500 tokens/second figure are therefore short, single-shot, low-concurrency prompts: code completion, one-off summarization, single-turn chat, where the cache-billing rule never triggers because there is nothing to re-send.
None of this appears in either of Cerebras' own posts. It surfaces only in the explainx.ai account, which is also the only source in the public record that describes the tier's constraints rather than its throughput.
For a team evaluating Cerebras against a GPU-based provider on a workload with meaningful context reuse, RAG pipelines, multi-step agents, anything that resends a system prompt, the relevant comparison is not tokens-per-second on a clean run. It is total tokens billed per completed task, a number neither Cerebras' marketing nor the explainx.ai piece calculates directly, and one no source in the record supplies for the enterprise tier at all.
Deep dive
The evidentiary record for the 1,500 and 2,400 tokens-per-second figures reduces to two documents: Cerebras' own Qwen3-32B launch post and its Qwen3-235B press release. Both attribute the underlying measurement to "Artificial Analysis." Neither post links to a benchmark report, a methodology page, or a set of raw run logs. Artificial Analysis does not otherwise appear in the sourced record. The claim that Qwen3-32B returns a first token in 1.2 seconds and sustains 2,400 tokens/second rests entirely on Cerebras characterizing what a third party found, with the third party's own account absent from the public record this analysis draws on.
That absence matters more than it would for a single, static number, because both figures are rates, and rates are sensitive to the conditions under which they were measured. A tokens-per-second figure depends on batch size, sequence length, whether the run was single-stream or concurrent, and which GPU generation and serving stack formed the "best GPU result" baseline that Cerebras says its 40x figure beats. None of that is specified in either post. The Qwen3-32B launch post states flatly that "every GPU based inference API tops out at 150 tokens per second." This comparison is presented as a settled ceiling rather than a snapshot of a specific set of providers tested at a specific point in 2025.
Our read is that a ceiling asserted without a tested-provider list or a date range cannot be checked against subsequent changes in GPU inference optimization — the research does not identify what was tested or when the comparison would go stale. The 150 tokens/second figure may have been accurate against whatever Cerebras tested in May 2025. Whether it describes the field eighteen months later is not something this record can answer, because the record does not name what was tested.
The pricing claims sit on similarly narrow footing. Cerebras' Qwen3-235B press release states $0.60 per million input tokens and $1.20 per million output tokens, "less than one-tenth the cost of comparable closed-source models." That comparison requires knowing which closed-source models, at which providers, at which point in time, formed the denominator — none of which the release specifies. A tenth of the cost of GPT-4-class pricing in July 2025 is a different claim than a tenth of the cost of whatever closed models are priced today. The ratio is not wrong on its face; it is simply not reproducible from what is published, because the comparison set is not named.
What the record does support, independent of Cerebras' framing, is the architecture. explainx.ai's analysis of the Qwen 3.8 27B launch confirms that the Wafer Scale Engine holds model weights in on-chip SRAM rather than external memory. It treats this as the plausible mechanical explanation for the throughput advantage, a claim about how the chip is built, which is checkable in principle against Cerebras' published hardware specifications, rather than a claim about how fast the chip ran on a particular day. That distinction is the one this record can actually support: the architecture is real and documented independently of Cerebras' marketing copy; the specific speed multiple over GPU inference is not independently documented at all.
The Alibaba side of the record adds a variable that cuts against both the bull and bear case without resolving either. The macro condition that would make Cerebras' pitch land is documented. Whether it has landed with anyone is not.
What would change this: a benchmark of Cerebras' dedicated tier run by a party with no commercial relationship to Cerebras, publishing batch size, concurrency, and the specific competing endpoints tested, alongside a named enterprise customer's total-cost-per-completed-task figure under real rate limits. None of the four sourced accounts of Cerebras' Qwen deployment — the two company posts, the explainx.ai piece, and Alibaba's own results — contains any of those three elements.