# Cerebras CS-4 Is Fast. Its 30x Uses a Mid-Pack GPU Baseline

> Cerebras announced the CS-4 on August 18, 2026: one rack with three WSE-3 Turbo wafers, the same 4-trillion-transistor, 44 GB-of-SRAM silicon as the WSE-3 run at about twice the clock, sold as up to 30x faster than GPUs at inference. Measured in pixels, the GPU bar behind its 4,400 tokens per second on gpt-oss-120b is about 180, while Artificial Analysis measured Microsoft Azure at 336 and Baseten and Nebius near 290 in the 72 hours to October 6, which makes the lead 13 to 15x. The CS-4 page promises 1,000 tokens per second on models over 50 trillion parameters, but the datasheet says 10 trillion, and Cerebras's own latency chart, extended, uses the whole 1 ms budget at about 46 trillion.

URL: https://blog.bokvi.com/blog/cerebras-cs-4-30x-claim/
Published: 2026-10-06
Tags: hardware, infrastructure, benchmarks

**Cerebras [unveiled the CS-4 on August 18](https://www.cerebras.ai/press-release/cerebras-unveils-cs-4-up-to-30-times-faster-than-gpu-based-solutions-3347356) as "up to 30 times faster than GPU-based solutions": one rack, three wafer-scale chips, and more than 4,400 tokens a second for a single user on OpenAI's gpt-oss-120b. The release doesn't say which GPU it beat. Cerebras's own launch chart shows it, though, and when I measured its bars in pixels, the GPU behind the 30x does about 180 tokens a second, slower than five of the general-purpose clouds serving the same model today.** The CS-4's lead is still one nobody else is near. Against the quickest of those clouds, though, it's 13 to 15 times faster, not 30.

I read the [product page](https://www.cerebras.ai/cs4), the [datasheet](https://cdn.sanity.io/files/e4qjo92p/production/bb028ae2422089dfbcc9fd2455c021f3625c7a75.pdf), the launch blog and the Hot Chips write-up against each other, then checked the charts against Artificial Analysis's provider measurements. One more number didn't survive: the page's 1,000 tokens a second on 50-trillion-parameter models.

![A black Cerebras rack with three power drawers sticking out of its front and three slim modules mounted on its back, labelled power delivery in the front of the rack and backpack contains wafer scale engine](https://blog.bokvi.com/_astro/cerebras-cs-4-rack.CTXF3ZxC_Z19VJ5k.webp)

*The CS-4 rack: power shelves at the front, three Wafer-Scale Backpacks at the back, one wafer in each. Image: Cerebras.*

## The WSE-3 Turbo is the WSE-3, run twice as hard

The chip isn't new. According to the datasheet, each WSE-3 Turbo has the same 4 trillion transistors, 900,000 cores, 44 GB of on-wafer SRAM and 46,225 mm² of TSMC 5 nm silicon as the WSE-3 from 2024. Everything that did change doubled together: 250 PFLOPS of sparse FP16, 43.2 PB/s of memory bandwidth, 2.4 Tbit/s of I/O. That's what a clock bump looks like, and [The Register](https://www.theregister.com/systems/2026/08/19/cerebras-cs-4-rack-systems-juice-chips-for-every-last-drop-of-ai-performance/5289286) reports Cerebras told it as much, estimating the clock went from about 1.4 GHz to 2.8 GHz.

The engineering went into feeding it. Cerebras moved power conversion to 0.5 mm from the wafer, against about 50 mm on a GPU board, and says that delivers twice the power to the chip. It hasn't published what a rack draws. The Register estimates about 46 kW per backpack and 120 to 140 kW per rack, roughly half the 240 to 250 kW racks due from AMD and Nvidia later this year.

![An exploded render of one Wafer-Scale Backpack: a finned cold plate on top, an orange wafer below it, a power module sub assembly directly beneath the wafer, and an I/O board on each side](https://blog.bokvi.com/_astro/wafer-scale-backpack.DhKJ6nWW_1yq2GM.webp)

*One backpack taken apart. The power module sits right under the wafer, which is the whole trick behind the doubled power. Image: Cerebras.*

What didn't grow is the SRAM: 44 GB per wafer, 132 GB per rack. That shapes what the CS-4 is for. Cerebras also pitches it as the decode half of a split system, with a "purpose-built prefill engine" such as AMD Helios or AWS Trainium reading the prompt and the wafers writing the answer. I think that's the honest position for a chip whose strength is streaming weights out of SRAM fast. In that setup, the CS-4 does nothing for the time it takes to read a long prompt.

## The 30x is measured against a GPU doing about 180 tokens a second

Here's the claim in full: "In a head-to-head comparison on GPT-OSS-120B, when given identical prompts, the CS-4 delivers more than 4,400 tokens second per user (TPS/user), up to 30 times faster than GPU solutions." There's no GPU model, batch size or context length, the footnote says throughput "varies by model architecture, context length, precision, and serving configuration," and the CS-4 page's fine print says comparisons are "based on third-party benchmarking or internal testing." The same page carries the chart behind it.

![A dark bar chart titled CS-4 is transformative, up to 30x faster, comparing tokens per second per user for GPU and CS-4 across Gemma 4 31B, Llama 3.3 70B, GPT OSS 120B, GLM 4.7 355B, Kimi K2.7 1T, GPT 5.4 and GPT 5.6 Sol, with tall orange CS-4 bars and barely visible gray GPU bars](https://blog.bokvi.com/_astro/cerebras-speed-chart.Cp9Kcr1a_Zm1RXW.webp)

*Cerebras's launch chart. The gray GPU bars are 11 to 20 pixels tall on a 1,080-pixel image, so I measured them. Image: Cerebras.*

The gridlines sit 110.5 pixels per 1,000 tokens a second, so the bars can be read off. The CS-4 bar for gpt-oss-120b comes to about 4,490, which matches "more than 4,400." Its GPU bar is 20 pixels, about 180 tokens a second. The other GPU bars land between about 100 and 145, and with bars that short each reading can be off by about 10%, so treat the ratios below as approximate.

Read that way, "up to 30x" isn't even the top of Cerebras's own chart. Llama 3.3 70B comes out near 35x, Gemma 4 31B and gpt-oss-120b near 25x. The ratio falls on the big models, which are the ones the CS-4 is sold for: GLM 4.7 and Kimi K2.7 at about 2,000 tokens a second, GPT 5.4 at about 1,500, GPT 5.6 Sol at about 1,350, roughly 11 to 18 times their gray bars.

## The fastest clouds already serve gpt-oss-120b at nearly twice that GPU

180 tokens a second is a plausible GPU number. It just isn't a fast one. [Artificial Analysis](https://artificialanalysis.ai/models/gpt-oss-120b/providers) benchmarks 18 public providers of gpt-oss-120b, and over the 72 hours to October 6 its medians put Microsoft Azure at 336 tokens a second, Nebius at 292, Baseten at 287 and Crusoe at 265.

![A horizontal bar chart of gpt-oss-120b output speed by provider: Cerebras CS-4 claim 4,400, Cerebras today 1,763, SambaNova 700, Groq 471, Microsoft Azure 336, Nebius 292, Baseten 287, Crusoe 265, DeepInfra Turbo 210, the GPU in Cerebras's chart about 180, then Novita, Together AI, Parasail and Scaleway between 135 and 160 and four providers under 80](https://blog.bokvi.com/_astro/gpt-oss-120b-speed-by-provider.C7oz6DX6_ZTgbFa.webp)

*Cerebras's GPU sits mid-pack, between DeepInfra Turbo and Novita. Artificial Analysis tests public endpoints with its own prompts, so this is what a customer gets, not a rerun of Cerebras's head-to-head. Data: Artificial Analysis and Cerebras.*

Against Azure, 4,400 is 13 times faster. Against Baseten and Nebius it's 15. Two other custom chips sit in between: SambaNova at 700 and Groq at 471. 13x is still an enormous lead, and nothing else on that chart is close. My problem is that 30x is the number that ends up in procurement decks.

The chart also shows Cerebras's own service. Artificial Analysis measures it at 1,763 tokens a second on today's CS-3 hardware, while Cerebras [said in August](https://www.cerebras.ai/blog/how-cerebras-serves-gpt-5-6-sol-at-up-to-750-tokens-per-second) that its public endpoint serves gpt-oss-120b at 3,000. The CS-4's "2x the speed of CS-3" lands between the two, depending on which CS-3 figure you start from.

## 1,000 tokens a second at 50 trillion parameters doesn't fit Cerebras's own chart

The CS-4 page says: "By reducing wafer-to-wafer interconnect latency to 2 microseconds, CS-4 delivers more than 1,000 tokens per second on models exceeding 50 trillion parameters." The datasheet and the [launch blog](https://www.cerebras.ai/blog/introducing-cerebras-cs-4), both from August 18, make the same claim with 10 trillion. The press release uses 50 trillion only for what CS-4 clusters can hold, not how fast they run it, and Cerebras's [Hot Chips write-up](https://www.cerebras.ai/blog/ultrafast-frontier-inference-cerebras-deep-dive-at-hot-chips-2026) gives interactive speed on 50-trillion-parameter models to the CS-5, targeted for 2027.

Why 10 and not 50 matters: a big model doesn't fit on one wafer, so Cerebras splits it into a pipeline. In its description of how it serves GPT-5.6 Sol, "For every token, the activation moves from one wafer to the next until the final stage emits the result." Every hop costs latency, and 1,000 tokens a second leaves 1 ms per token for all of it, the math in every wafer and every hop between them.

The same CS-4 page has a chart of the hop part: interconnect latency across all hops against model size, a straight line from 0.026 ms at 1 trillion parameters to 0.22 ms at 10 trillion. I extended it.

![A line chart of interconnect latency per token against model size from 1 to 50 trillion parameters: Cerebras's CS-4 line rises from near zero to 0.22 ms at 10T, a dashed straight-line extension crosses a white 1 ms budget line at about 46T and reaches 1.08 ms at 50T, and Cerebras's GPU line climbs faster up to 0.58 ms at 10T](https://blog.bokvi.com/_astro/cs-4-latency-extended.DQLBRGaM_EPCcc.webp)

*The extension is mine. Cerebras's chart stops at 10 trillion parameters. Data: Cerebras.*

The line crosses 1 ms at about 46 trillion parameters and reaches 1.08 ms at 50, before any wafer has done any math. The SRAM tells the same story. At four bits per weight, 50 trillion parameters is 25 TB, about 570 wafers of 44 GB each, and at one hop per wafer, 570 hops at 2 microseconds is 1.1 ms. The ways out I can see are speculative decoding, which can emit several tokens per trip through the pipeline, or fewer than four bits per weight, and Cerebras mentions neither for the CS-4. I read the page as two claims merged into one sentence; the datasheet's 10 trillion is the version that adds up. Even that is big iron: 10 trillion parameters at four bits is 5 TB of weights, or 38 CS-4 racks before any room for the KV cache.

## For anyone calling these models, it speeds up the writing, not the reading

You won't buy a CS-4; you'll rent tokens from one. Today that means Cerebras's own inference cloud, or OpenAI's Ultrafast mode for GPT-5.6 Sol, in [limited preview since August 13](https://thenextweb.com/news/openais-new-ultrafast-mode-runs-gpt-5-6-sol-14-times-faster-on-cerebras-chips) at about 750 tokens a second, up to 14 times OpenAI's standard tier. That runs on CS-3; the CS-4 chart puts the same model at about 1,350. On September 29 the neocloud General Compute [signed a multi-year deal](https://www.cerebras.ai/press-release/general-compute-selects-cerebras-to-bring-ultra-fast-inference-to-agentic-coding) to offer Cerebras inference for coding agents from Q1 2027, without naming the system.

For agents, output speed compounds. A step that writes 2,000 tokens of GPT-5.6 Sol takes about 1.5 seconds at the chart's 1,350 tokens a second and 20 seconds at its GPU bar of 100. But a coding agent also re-reads a large context on every step, and in the split design that reading happens on the prefill partner's hardware, not on the wafers. If your loop is dominated by long prompts and short answers, the CS-4 changes less than its headline suggests.

Two things are still missing. Cerebras hasn't published a price, and "first CS-4 shipments begin this quarter," written on August 18, is still on the product page six days after that quarter ended. I couldn't find an announcement that one has shipped. Until there's a customer number and an independent benchmark on a trillion-parameter model, I'd read the CS-4 as 13 to 15 times the clouds you'd otherwise use, and the 50-trillion line as 10.