Cerebras CS-4 Review: Specs, Performance

On August 18, 2026, Cerebras Systems (NASDAQ: CBRS) pulled back the curtain on the CS-4, a rack-scale AI accelerator it calls “the fastest AI accelerator in the industry.” The headline number is 30x faster inference than GPU systems, delivered by wiring three of its new Wafer Scale Engine 3 Turbo processors into a single rack. But the announcement landed in a complicated moment for the company: its stock had been drifting below its IPO pop, hardware sales had just fallen 23% year over year in the June quarter, and the growth story had quietly shifted from selling chips to renting compute. The CS-4 is the test of whether Cerebras can convert its engineering moat into a durable business.

Key Takeaways:

The CS-4 combines three WSE-3 Turbo wafers into one rack delivering 750 PFLOPS, 129.6 PB/s of memory bandwidth, and up to 30x faster inference than GPU systems, per Cerebras’ own benchmarks.

Wafer-to-wafer latency drops to 2 microseconds, letting a single CS-4 cluster serve models above 10 trillion parameters at more than 1,000 tokens per second, a vendor claim not yet independently reproduced.

Disaggregated inference with AMD Helios and AWS Trainium splits prefill from decode, letting operators pair efficient GPUs or ASICs with Cerebras’ low-latency decode.

The commercial story is shifting from hardware to cloud: Q2 2026 core revenue hit $210 million, but physical hardware sales fell 23% while cloud revenue rose 281%.

All performance figures come from Cerebras internal benchmarking or its partner OpenAI; independent third-party verification is still pending.

What the CS-4 Actually Is: Three Wafers, One Rack

Cerebras’ core bet has never changed since its founding in 2015: instead of stitching hundreds of small GPU dies together and paying the latency and power cost of moving data between them, it builds one processor that spans an entire silicon wafer. The WSE-3, its current flagship, measures 46,225 square millimeters and packs 4 trillion transistors and 900,000 AI-optimized cores, with 44 GB of SRAM integrated directly on the wafer, according to the company’s chip product page. That SRAM is the key architectural difference: it is faster to access than the off-chip high-bandwidth memory GPUs rely on, which is why wafer-scale systems can stream tokens so quickly.

The Nexus Rack Platform and the Backpack Design

The CS-4 takes that philosophy one step further by connecting three of these wafers into a single rack-scale system, each using the new WSE-3 Turbo variant. The WSE-3 Turbo doubles per-wafer AI compute to 250 PFLOPS and doubles memory bandwidth to 43.2 PB/s compared with the original WSE-3. In a CS-4 rack, those three wafers sum to 750 PFLOPS of AI compute, 129.6 PB/s of memory bandwidth, 160.5 PB/s of total compute fabric bandwidth, and 7.2 terabits per second of system I/O, according to Cerebras’ launch release.

Sean Lie, Cerebras’ CTO and co-founder, framed the design goal at the launch event: the next leap in AI infrastructure cannot come from improving one component in isolation. Compute, power, cooling, and I/O have to move forward together. That is why the CS-4 is not just a faster chip but a redesigned rack, a new power-delivery scheme, and a new I/O subsystem all at once. The result is a system built to generate fast tokens for interactive experiences while also delivering the total token capacity that large-scale operators need.

Close-up of a computer processor circuit board representing the Cerebras wafer-scale engine hardware — The wafer-scale approach keeps model weights on-chip in SRAM, removing the memory-bandwidth bottleneck that dominates GPU inference.

Why Wafer-Scale: The Memory Bandwidth Argument

To understand why the CS-4 matters, you have to understand the bottleneck it attacks. Fast frontier inference is fundamentally a data-movement problem. On a GPU, running a large model means repeatedly transferring model weights between on-chip memory and off-chip storage to generate each successive token. That shuttling is slow, and it is why large models generate tokens slowly even on expensive GPU clusters.

Cerebras’ contrarian answer, explained in its Ultrafast blog post, is to pack 44 GB of SRAM on each wafer-sized chip so that weights stay on-chip and tokens flow uninterrupted through model layers pipelined across wafers. The company argues this scales smoothly with model size, which is what gives it a continued speed advantage as frontier models get bigger. The WSE-3’s internal memory bandwidth of 27 PB/s was already more than 200 times what Nvidia’s NVLink interconnect offered, according to SiliconANGLE. The WSE-3 Turbo doubles that to 43.2 PB/s per wafer.

This is a genuine architectural difference, not a marketing gloss. A conventional GPU cluster shards a model’s weights across hundreds of dies and pays a synchronization penalty every time they need to communicate. A wafer-scale system keeps the entire model on one physical substrate, so the data never has to leave the chip. The trade-off is that the chip is enormous, expensive, and hard to manufacture at yield, which is why Cerebras had to invent fail-in-place routing to work around inevitable defects rather than discard whole wafers. The company’s chip page describes this as “designing to withstand defects, not avoid them,” with redundant compute cores, redundant routing, and a fail-in-place architecture that shuts flawed regions down and routes around them.

The 30x Claim: Where It Comes From and What It Means

The “up to 30x faster” figure is the centerpiece of the launch, but it needs careful reading. In a head-to-head comparison on the open-weight GPT-OSS-120B model, Cerebras reports the CS-4 delivering more than 4,400 tokens per second per user, up to 30 times faster than GPU solutions on the same prompts, according to the launch release. The company’s blog post cites “Artificial Analysis and internal benchmarking (August 2026)” as the source for the broader 30x claim, and the site’s own footer carries the caveat that “observed inference speed improvements versus GPU-based systems may vary depending on workload, cfg, date and models being tested.”

These are vendor figures. Independent evaluators have not yet reproduced the full CS-4 benchmark set, and Cerebras discloses that “prf comparisons are based on third-party benchmarking or internal testing.” That matters because the company’s track record shows vendor claims running ahead of verified results. Teams evaluating the CS-4 should treat the 4,400 tokens-per-second figure as an upper bound to test on their own workloads, not a guaranteed service level.

The speed story is also about scale, not just raw speed. By cutting wafer-to-wafer interconnect latency to as low as 2 microseconds, Cerebras says a CS-4 cluster can serve models exceeding 10 trillion parameters at more than 1,000 tokens per second, a claim the company marks as “extrapolation from internal benchmarking (August 2026).” The low-latency fabric is what lets the system preserve interactive decode performance as model size grows, rather than collapsing like a multi-GPU setup where weights must be sharded across thousands of dies. Sean Lie put the practical value this way: being 30 times faster does not just make a response feel fast, it gives an agentic system room for more than an order of magnitude as much reasoning, verification, or tool use in the same wall-clock time.

Dylan Patel, founder of SemiAnalysis, offered a rare independent endorsement in the launch release, saying the CS-4 “makes dramatic improvements in system deployability, reliability, and networking” and will “scale ultrafast tokens for larger models and significant user volumes.” That is notable because SemiAnalysis has been one of the more skeptical voices on AI hardware economics, so its willingness to go on record about the CS-4’s time-to-market advantages carries some weight, even if it does not independently verify the raw speed figures.

CS-3 vs. CS-4: The Specification Jump

The most concrete way to understand the CS-4 is to compare it directly with the CS-3 it replaces. The table below pulls the key specifications from Cerebras’ launch release and the Digitimes report of the announcement.

Metric	CS-3 (one wafer)	CS-4 (three wafers)
AI compute	125 PFLOPS	750 PFLOPS
Memory bandwidth	21.6 PB/s	129.6 PB/s
On-chip fabric bandwidth	26.7 PB/s	160.5 PB/s
System I/O bandwidth	1.2 Tb/s	7.2 Tb/s
I/O latency	5 microseconds	2 microseconds

Source: Cerebras CS-4 launch release and Digitimes.

Cerebras positions the CS-4 as up to twice as fast as the CS-3 on token generation, with up to 10x more throughput per watt in certain solutions and three times more density per rack, per the launch release. The throughput-per-watt improvement is the economic argument: because fast tokens are worth more than slow tokens, the company argues that a data center gets both higher-value tokens and more total tokens within a fixed power budget. That framing is aimed squarely at hyperscalers who pay for power, not just for chips.

The per-wafer improvements are worth spelling out because they explain the 2x gain. The WSE-3 Turbo doubles AI compute to 250 PFLOPS per wafer and doubles memory bandwidth to 43.2 PB/s. On-chip fabric bandwidth and off-chip I/O both double, to 53.5 PB/s and 2.4 Tb/s respectively, and I/O latency shrinks from 5 microseconds to as low as 2 microseconds. Since memory bandwidth is the determining factor in driving speed and throughput, Cerebras argues this translates to a step-change improvement in both. The 600-megawatt deployment footprint the company has online or under contract by the end of next year, per MarketBeat, is the scale at which these per-watt economics start to compound.

Disaggregated Inference: The AMD and AWS Strategy

The most strategically interesting part of the CS-4 is not the silicon but how Cerebras wants it used. Disaggregated inference splits the two phases of model serving across different hardware. A purpose-built prefill engine processes the incoming prompt and prepares the model state; that state is then handed to the Cerebras system, which performs the ultra-low-latency decode and generates the response. The company frames this as pairing its decode strength with complementary prefill platforms including AMD Helios and AWS Trainium.

The AWS relationship is the deepest. In March 2026, AWS announced it would bring the WSE-3 to its cloud via Bedrock and develop a disaggregated architecture combining WSE-3 with Trainium, connected through AWS’ Elastic Fabric Adapter, according to SiliconANGLE. The arrangement assigns Trainium to prefill and WSE-3 to decode, targeting a fivefold increase in output speed. AWS’ David Brown described the logic plainly: by splitting inference across Trainium and CS-3, “each system does what it’s best at,” producing inference an order of magnitude faster with higher performance than what is available today.

Cerebras’ own director of product marketing, James Wang, added a nuance that is easy to miss in the partnership headlines: disaggregated inference is ideal when you have large, stable workloads, but most customers run a mix of workloads with different prefill-to-decode ratios, where the traditional aggregated approach is still better. He said Cerebras expects most customers to want access to both modes. That is an honest admission that the disaggregated architecture is not a universal win, just a powerful option for the right workload shape.

The AMD partnership, announced alongside the CS-4, follows the same blueprint. Mark Papermaster, AMD’s CTO, said the combination of GPU prefill and Cerebras decode could provide inference 10 times faster than GPUs alone and five times more throughput than Cerebras alone, per MarketBeat’s coverage of the launch event. This is a notable departure from the usual adversarial chip-vendor dynamic: AMD and Cerebras are collaborating on inference even as Cerebras positions its hardware against GPU systems, because both see the disaggregated approach as a way to win workloads neither could take alone. The deal builds on a broader partnership with Arista Networks, whose CEO Jayshree Ullal spoke at the launch event about the networking requirements of scale-up, scale-out, and geographically distributed “scale-across” AI deployments.

The Nexus Rack Platform and the Backpack Design

Underneath the CS-4 sits the new Nexus Platform Architecture, a redesign of how the rack itself is built. Cerebras breaks the rack into three foundational elements: compute, power, and I/O. Each is a self-contained modular assembly, which the company says reduces components by 50% and shortens data-center deployment from days to hours, according to the launch blog post. The modularity is a manufacturing and logistics bet as much as an engineering one: instead of treating the rack as a tightly coupled collection of parts, Nexus turns it into a platform of purpose-built modules that can be manufactured, installed, expanded, and upgraded independently.

The centerpiece is the rear-mounted “Wafer-Scale Backpack,” a self-contained assembly that folds power conversion, direct liquid cooling, high-speed I/O, and control electronics into a compact three-dimensional package built directly around the wafer. Cerebras says this is what reduces deployment time from days to hours, a meaningful shift for hyperscalers who need to stand up capacity quickly.

The power delivery is notably aggressive. By moving power conversion from roughly 50 millimeters away from the processor on a conventional GPU board to about 0.5 millimeters, Cerebras says it nearly eliminates board-level power loss and can deliver twice as much power to the WSE-3 Turbo, enabling higher operating frequencies and faster token generation. This is the kind of detail that matters more than the headline speed number to a data-center operator, because power delivery efficiency directly determines how much compute you can extract from a given electrical feed.

The I/O subsystem is also new. The programmable Wafer I/O Module supports standards-based RoCE v2 RDMA over Ethernet for integration with existing infrastructure, plus a new “Direct Wafer Links” mode for switch-free connections within and across racks. That direct mode is what enables the 2-microsecond wafer-to-wafer latency and, in turn, the ability to scale massive CS-4 clusters to serve frontier models at more than 1,000 tokens per second. Cerebras says the Nexus architecture is built not just for the CS-4 but for future CS-5 and CS-6 systems, with a roadmap that calls for doubling speed annually and delivering up to 20 times higher throughput by 2027.

The OpenAI Anchor: Ultrafast and the HLE Benchmark

The OpenAPI partnership is the commercial anchor that gives the CS-4 a production reference beyond a laboratory benchmark. OpenAI’s GPT-5.6 Sol runs in a new “Ultrafast” mode on Cerebras hardware, delivering up to 750 output tokens per second, according to Cerebras’ blog. OpenAI’s head of compute infrastructure, Sachin Katti, framed it as adding “a dedicated low-latency inference solution to our platform” that produces “faster responses, more natural interactions, and a stronger foundation to scale real-time AI to many more people.”

The benchmark that got the most attention is Humanity’s Last Exam (HLE), a 2,500-question test designed to be answerable only by people with PhDs in fields like chemistry, economics, and literature. Cerebras reports that GPT-5.6 Sol on Ultrafast mode answered all 2,500 questions in 11 hours and 11 minutes, while Claude Fable 5 needed 78 hours and 27 minutes, more than three days of continuous compute, to arrive at the same conclusions. That is roughly seven times faster at comparable accuracy. Cerebras disclosed the benchmarking methodology: it ran GPT-5.6 Sol Ultrafast with Codex on “xhigh” reasoning on July 10, and Claude Fable 5 with Claude Code on xhigh reasoning on July 13-15.

We covered the Ultrafast tier in detail in our analysis of how to speed up GPT-5.6 Sol for production, noting that the figures come from OpenAI and Cerebras and have not been independently reproduced. The same caveat applies here: the HLE head-to-head is a Cerebras-run comparison against a competitor’s model, not a neutral third-party test. It is directionally useful evidence that wafer-scale inference is fast, but it should not be treated as an independent confirmation of the 30x claim.

OpenAI’s Thibault Sottiaux added an important strategic detail at the launch event: OpenAI would like Ultrafast to become the default experience over time, though the company is still early in deploying it, and it reserves some Ultrafast capacity for incidents, security matters, and major internal projects. That suggests the CS-4’s most important customer is still in a proof-and-expand phase rather than a fully committed at-scale deployment. The OpenAI relationship is also reflected in the ownership structure: OpenAI holds warrants to purchase 33.4 million Cerebras shares, subject to its purchases of compute capacity, according to Wikipedia’s overview.

The Financial Reality Behind the Hardware

The CS-4 is a technology story, but it lands inside a financial one that investors are scrutinizing hard. Cerebras went public in May 2026 at $185 per share, raising $5.55 billion in the largest U.S. tech IPO of the year, as we covered in our analysis of the Cerebras IPO.

The June-quarter results, reported August 12, show why the stock has been volatile. Cerebras beat expectations with $210 million in core revenue, more than double the year-ago quarter and ahead of the $191 million analysts expected, but physical hardware revenue fell 23% year over year to $54.1 million, while cloud and other services revenue climbed 281% to $126 million, according to Cryptopolitan’s coverage. The company swung to a GAAP net loss of $450.5 million, driven largely by $386.6 million in stock-based compensation tied to the IPO. Management raised full-year 2026 core revenue guidance to $880 million to $890 million, above the $867.6 million analysts expected.

The trend is unmistakable: Cerebras increasingly rents its compute rather than sells its boxes. CEO Andrew Feldman characterized the hardware dip as “lumpy” delivery timing, noting customers sometimes lack the data-center space to install systems the size of Cerebras’. CFO Bob Komin said the company intends to more than triple revenue in 2027 as production scales and component costs come down.

The balance sheet is strong after the IPO: Cerebras ended June with $8.6 billion in cash, cash equivalents, restricted cash, and short-term investments, plus $25.4 billion in remaining performance obligations. The OpenAI deal, worth over $10 billion for 750 megawatts of compute through 2028, is the demand anchor, and the company has data centers across North America and Europe, including Santa Clara, Toronto, Dallas, Minneapolis, Montreal, Oklahoma City, Alabama, Lyon, Norway, and Mikkeli, Finland, per MarketBeat.

A Practical Latency and Throughput Example

For a technical team weighing the CS-4 against GPU alternatives, the useful exercise is to model what the speed difference actually buys in a real workload. The code below compares the generation time and cost of serving a coding-agent task across a GPU baseline and the CS-4’s advertised peak, using realistic output sizes rather than a toy prompt. The point is not to reproduce Cerebras’ benchmark but to give a template for plugging in your own measured numbers.

Note: The following code is an illustrative example and has not been verified against official documentation. Please refer to the official docs for production-ready code.

from dataclasses import dataclass

@dataclass
class InferencePlatform:
 name: str
 output_tokens_per_second: float
 cost_per_million_output_tokens_usd: float

def estimate_task(platform, output_tokens):
 generation_seconds = output_tokens / platform.output_tokens_per_second
 cost_usd = platform.cost_per_million_output_tokens_usd * (output_tokens / 1_000_000)
 return {
 "platform": platform.name,
 "output_tokens": output_tokens,
 "generation_seconds": round(generation_seconds, 2),
 "est_cost_usd": round(cost_usd, 4),
 }

# Replace these with measurements from your own serving stack.
# The CS-4 figure is Cerebras' vendor-reported upper bound (4,400 t/s/user
# on GPT-OSS-120B). The GPU baseline is a representative production cluster.
platforms = [
 InferencePlatform("GPU cluster baseline", 150.0, 30.0),
 InferencePlatform("CS-4 (vendor peak)", 4400.0, 30.0),
]

# Realistic outputs for a multi-turn coding agent, not a one-line reply.
workloads = {
 "single_function_patch": 400,
 "code_review_feedback": 1500,
 "multi_file_refactor": 6000,
}

for workload, tokens in workloads.items():
 print(f"\n{workload}")
 for p in platforms:
 print(estimate_task(p, tokens))

# Note: production evaluation must also measure prompt processing (prefill),
# time to first token, queueing delay, tool-call round trips, retries,
# and concurrency. This example models only the decode phase, which is
# where the CS-4's on-chip SRAM advantage is concentrated.

The asymmetry this exercise exposes is the core of Cerebras’ pitch. A 6,000-token refactor that takes 40 seconds on a 150-token-per-second GPU cluster finishes in under 1.4 seconds at the CS-4’s advertised peak. For an interactive coding agent, that is the difference between a developer staying in flow and context-switching to something else while the model churns. The same logic applies to the agentic workloads Cerebras is targeting: Cognition, which builds coding agents, reported faster end-to-end task completion on Cerebras hardware, and Scott Wu, Cognition’s CEO, said running an agent at roughly 1,000 tokens per second lets you optimize all parts of the agent together, including context retrieval, UI, and model behavior.

The caveat is that decode speed is only one part of the latency budget. If a coding agent spends most of its time waiting on a slow repository search or a serialized tool call, paying for faster decoding will not fix the bottleneck. The engineering response is to combine faster serving with parallel retrieval, caching, and removal of unnecessary sequential calls, as we discussed in the Ultrafast analysis. The 4,400 tokens-per-second figure also describes a single stream; aggregate throughput across concurrent users is a separate number that Cerebras has not published for the CS-4 in a form independent evaluators can verify.

Trade-offs and Limitations

For all the speed claims, the CS-4 carries real trade-offs that independent practitioners should weigh before committing. The most obvious is cost: wafer-scale systems are expensive, with earlier-generation nodes priced as much as $3 million and power draw around 25 kW per node, according to Wikipedia’s overview of Cerebras Systems. A three-wafer rack multiplies that economics, and Cerebras has not published CS-4 pricing, so a total cost of ownership comparison against a GPU cluster is not yet possible on public numbers.

Software compatibility is the second constraint. Cerebras uses its own software platform rather than the full CUDA ecosystem, so pipelines that rely on custom kernels or libraries that are not ported will need rework. The company’s own materials concede that observed gains “may vary depending on workload, cfg, date and models being tested.” A team running a standard open-weight model may see far less than 30x if its serving stack is not tuned for the wafer-scale fabric, and the effort to port a CUDA-heavy pipeline can be substantial.

Customer concentration is the third risk. If any of those relationships shift, the revenue base is exposed. The company’s own SEC risk factors flag its “dependence on a limited number of significant customers” including OpenAI, G42, Mohamed bin Zayed University, and AWS, and its “historical reliance on sales of hardware systems” in an early-stage cloud market.

There is also a strategic tension in the disaggregated strategy. Cerebras depends on partners like AMD and AWS that also compete with its GPU ecosystems, and the company simultaneously markets its hardware against GPU systems while asking GPU vendors to route prefill to it. That cooperation is working today because the workloads are complementary, but it is a fragile arrangement if any partner decides to build its own low-latency decode capability. Finally, the manufacturing constraint is real: Cerebras’ chips are fabricated by TSMC, which Wikipedia notes is currently the only company with the ability to manufacture them, concentrating supply risk in a single foundry.

What to Watch Through 2026

Four signals will tell whether the CS-4 is a real inflection or a benchmark exercise. First, independent measurement: a standardized Artificial Analysis or MLPerf test of sustained CS-4 throughput would confirm whether the 4,400 tokens-per-second and 1,000-tokens-on-10T-model claims hold under load and concurrency. Second, CS-4 pricing and general availability, which the company says begins shipping this quarter, with full specifications in the CS-4 datasheet.

Third, capacity conversion: whether the 600 megawatts of power under contract turns into paying cloud revenue rather than idle racks, and whether the OpenAI Ultrafast tier expands beyond limited preview to become a general enterprise offering. Fourth, the stock. In our prior analysis, we projected Cerebras shares would trade above $300 by December 31, 2026, on Ultrafast capacity expansion. The CS-4 launch and its 2-microsecond fabric are the concrete evidence that capacity is coming; whether it converts into durable revenue is the open question.

I expect the CS-4 to firm up Cerebras’ position as the low-latency decode specialist in heterogeneous AI infrastructure, even as its hardware-only sales remain lumpy. The disaggregated partnerships with AMD and AWS give it a path into workloads it could not win alone, and the OpenAI Ultrafast tier gives it a marquee production reference. The risk is that the 30x marketing frame outruns what independent benchmarks show, repeating the pattern where vendor claims lead verified results by 10 to 20%. For technical buyers, the CS-4 is worth a controlled pilot on a real workload, with the 30x figure treated as an upper bound to test, not a guarantee. The hardware is genuinely novel; the question is whether the economics and the ecosystem can keep pace with the marketing.

More in-depth coverage from this blog on closely related topics:

Sources and References

Sources cited while researching and writing this article: