# Cerebras WSE-3

> Source: https://aiwiki.ai/wiki/cerebras_wse_3
> Updated: 2026-10-15
> Fact-checked: 2026-10-15
> Categories: AI Compute, AI Hardware, AI Infrastructure, Data Centers
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Cerebras WSE-3." aiwiki.ai, 15 Oct 2026. https://aiwiki.ai/wiki/cerebras_wse_3
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

The **Cerebras WSE-3** (Wafer-Scale Engine 3) is the third-generation wafer-scale [AI chip](https://aiwiki.ai/wiki/ai_chip) developed by [Cerebras Systems](https://aiwiki.ai/wiki/cerebras), announced on March 13, 2024, and described by Cerebras as the largest chip ever built.[^1] Fabricated on TSMC's 5nm process, it integrates 4 trillion transistors and 900,000 AI-optimized compute cores with 44GB of on-chip SRAM onto a single dinner-plate-sized die spanning 46,225 square millimeters, equivalent to nearly an entire 300mm silicon wafer.[^1][^2] Cerebras rates a single WSE-3 at 125 petaFLOPS of peak AI performance; that figure is 16-bit (FP16) compute that assumes heavy [sparsity](https://aiwiki.ai/wiki/sparsity), and The Register has put the dense equivalent at about 12.5 petaFLOPS.[^1][^19][^46] Tom's Hardware described the chip as theoretically equivalent to about 62 [NVIDIA H100](https://aiwiki.ai/wiki/nvidia_h100) GPUs, a comparison of sparse FP16 figures, and it has 54% more transistors than its predecessor, the WSE-2.[^1][^3][^19] CEO and co-founder Andrew Feldman called it "the fastest AI chip in the world, purpose-built for the latest cutting-edge AI work, from mixture of experts to 24 trillion parameter models."[^1] The WSE-3 powers the CS-3 system, Cerebras' third-generation AI supercomputer appliance, and underpins the Cerebras Inference cloud service launched in August 2024,[^8][^9] which by 2025 was a selectable fast-inference option in Meta's official [Llama API](https://aiwiki.ai/wiki/llama_api)[^21] and powered Mistral's [Le Chat](https://aiwiki.ai/wiki/le_chat) Flash Answers, Notion's enterprise search and Cognition's coding models.[^48][^49][^50] The WSE-3 was Cerebras' shipping chip when the company listed on Nasdaq under the ticker CBRS on May 14, 2026; the IPO raised about $5.55 billion at $185 a share (about $6.38 billion after the underwriters' option), and the stock closed its first day at $311.07, which CNBC said valued the company at about $95 billion.[^27][^29][^43] In August 2026 Cerebras introduced the CS-4, built from three higher-powered WSE-3 Turbo wafers, rather than a new WSE-4.[^44]

## When was the WSE-3 released, and how does it compare to earlier generations?

### WSE-1 (2019)

Cerebras Systems was founded in 2016 (it was incorporated in Delaware in April 2016) by Andrew Feldman, Gary Lauterbach, Michael James, Sean Lie, and Jean-Philippe Fricker with the explicit goal of redesigning the processor from first principles for AI workloads.[^15] In August 2019, at Hot Chips, the company unveiled the original Wafer-Scale Engine (WSE-1): the first and only trillion-transistor processor at that time. Manufactured on TSMC's 16nm process, it packed 1.2 trillion transistors, 400,000 cores, and 18GB of on-chip SRAM onto 46,225mm² of silicon.[^55]

The fundamental design premise departed from conventional chip design by treating the entire 300mm wafer as a single processor rather than dicing it into dozens of smaller chips. Conventional wafer yields benefit from smaller die sizes because any defect affects only one chip among many. Cerebras addressed yield by building redundant circuitry into the wafer, so defective cores are routed around rather than discarded.[^55]

### WSE-2 (2021)

In April 2021, Cerebras unveiled the second-generation WSE-2, produced on TSMC's 7nm process. The WSE-2 raised the transistor count to 2.6 trillion, added 850,000 AI-optimized cores, expanded on-chip SRAM to 40GB, and increased memory bandwidth to 20 petabytes per second. Total fabric bandwidth reached 220 petabits per second across the on-wafer interconnect mesh.

The CS-2 system based on the WSE-2 could support clusters of up to 192 nodes through Cerebras' weight streaming architecture, which separates parameter storage from the compute wafer.[^7] In this model, weights reside in an external MemoryX cluster and stream onto the wafer during computation, while the SwarmX fabric coordinates gradient reduction across nodes. Next Platform calculated that the first 32-node phase of Condor Galaxy 1, built from CS-2 systems, cost about $100 million, or about $3.13 million per node including Cerebras operating the cluster for G42.[^7]

By early 2024, the WSE-2 and CS-2 were deployed at research institutions, pharmaceutical companies, and large technology groups, and had trained notable models including the Jais-30B Arabic [language model](https://aiwiki.ai/wiki/language_model) and the Med42 medical language model via the Condor Galaxy supercomputer network.[^1]

### Generation summary

The three WSE generations moved from TSMC's 16nm to 7nm to 5nm processes between 2019 and 2024, raising the transistor count from 1.2 trillion to 2.6 trillion to 4 trillion:

| | WSE-1 (2019) | WSE-2 (2021) | WSE-3 (2024) |
|---|---|---|---|
| Process node | TSMC 16nm | TSMC 7nm | TSMC 5nm |
| Transistors | 1.2 trillion | 2.6 trillion | 4 trillion |
| AI cores | 400,000 | 850,000 | 900,000 |
| On-chip SRAM | 18 GB | 40 GB | 44 GB |
| Memory bandwidth | 9 PB/s | 20 PB/s | 21 PB/s |
| Die area | 46,225 mm² | 46,225 mm² | 46,225 mm² |
| System | CS-1 | CS-2 | CS-3 |
| Max cluster size | N/A | 192 nodes | 2,048 nodes |

The area has remained fixed across all three generations by design, as Cerebras sized its architecture to one full 300mm wafer and has not needed to change that baseline. The improvements come from shrinking the process node and redesigning the compute cores. In 2026 Cerebras added a higher-powered variant of the third-generation wafer, the WSE-3 Turbo, for its CS-4 system (see below).[^44]

## How is the WSE-3 built, and what is inside it?

### Die design and manufacturing

The WSE-3 is fabricated on TSMC's 5nm process (N5), a transition down from the 7nm used for the WSE-2.[^1][^2] The chip measures 21.5 centimeters on each side and covers 46,225mm², the same physical footprint as all prior WSE generations since the company designed its wafer processes around a fixed 300mm platform.[^1] Despite keeping the same area, the move to 5nm allowed Cerebras to raise transistor density enough to reach 4 trillion transistors, a 54% increase over the 2.6 trillion of the WSE-2.[^1]

The core count rose modestly from 850,000 to 900,000 AI-optimized cores, reflecting a design choice to widen each core's math pipeline rather than simply multiply core count.[^7][^6] The WSE-3 SIMD units are eight-wide for FP16 operations, doubled from the four-wide units in the WSE-1 and WSE-2, which accounts for the bulk of the 2x performance improvement at the same power envelope.[^7][^6] At peak, a single WSE-3 delivers 125 petaFLOPS of FP16 compute,[^1][^3] a figure that assumes sparse data: The Register described it as "highly sparse FP16," and Next Platform noted that Cerebras' sparsity gains range from 2x to 10x depending on the data.[^19][^7] Tom's Hardware summarized the chip as theoretically equivalent to about 62 NVIDIA H100 GPUs, though the comparison ignores precision and [clustering](https://aiwiki.ai/wiki/clustering) differences.[^3]

On-chip SRAM increased to 44GB from 40GB, a 10% gain.[^1] Memory bandwidth improved to 21 petabytes per second from 20 petabytes per second.[^1] The on-wafer network fabric bandwidth is 214 petabits per second, broadly similar to the WSE-2's 220 petabits per second.[^17][^38] These relatively modest gains in memory metrics reflect that bandwidth and capacity improvements at 5nm over 7nm are constrained by the same physical wafer area.

The WSE-3 does not include HBM or any form of stacked off-die memory integrated into the chip package. All 44GB of memory is SRAM distributed physically alongside the compute cores across the wafer surface. This co-location eliminates the off-chip memory bus bottleneck that limits GPU memory bandwidth at scale, but it also means models larger than 44GB must be split across several systems or have their weights streamed from external memory. Cerebras says the WSE-3's 21 petabytes per second of on-chip SRAM bandwidth is 7,000 times the HBM bandwidth of an NVIDIA H100, which it identifies as the underlying enabler of its inference throughput claims.[^8]

### On-wafer interconnect

Cerebras connects cores through a 2D mesh fabric [etched](https://aiwiki.ai/wiki/etched) onto the wafer itself. Because all wires are on-wafer, latencies are in the range of nanoseconds, and the aggregate bandwidth of 214 petabits per second across the mesh exceeds the interconnect bandwidth of much larger [GPU cluster](https://aiwiki.ai/wiki/gpu_cluster)s by orders of magnitude. Cerebras claims the CS-3's on-wafer fabric provides more than 200 times the interconnect bandwidth of an NVIDIA NVL72 rack containing 72 Blackwell GPUs.[^20]

One manufacturing challenge specific to wafer-scale design is that the silicon wafer is divided by narrow "scribe lines" used in conventional chip fabrication to separate individual dies before dicing. Cerebras worked with TSMC to run wires across the scribe lines, extending its 2D mesh from one die to the next without adding a separate packaging layer, an approach it has used since the WSE-1.[^18][^55]

### Weight streaming and MemoryX

Because 44GB of on-chip SRAM cannot hold the parameters of a [large language model](https://aiwiki.ai/wiki/large_language_model) (Llama 3 70B alone requires roughly 140GB), Cerebras employs a weight streaming execution model for training. In this approach, model weights are stored in an external MemoryX appliance rather than on the wafer. During a forward or backward pass, weights for each layer stream from MemoryX to the wafer compute cores layer by layer. After computation, updated gradients flow back.[^6]

At launch, Cerebras listed external memory options of 1.5TB, 12TB, or 1.2 petabytes for the CS-3.[^1] Its CS-3 blog described MemoryX SKUs of 24TB and 36TB for enterprise customers and 120TB and 1,200TB for hyperscalers, and said the 1,200TB configuration can store models with 24 trillion parameters; a single CS-3 can be paired with a single 1,200TB MemoryX unit.[^6]

The SwarmX fabric is a separate network designed specifically for the weight streaming architecture. It sits between MemoryX and the CS-3 compute nodes, broadcasting weights to all compute nodes and reducing gradients back to MemoryX after each layer, so that a cluster of up to 2,048 CS-3 systems can be programmed like a single device.[^6][^7]

### How does Cerebras keep a wafer-sized chip from failing? Defect tolerance

Producing a chip that spans an entire 300mm wafer creates a yield problem that does not exist in conventional chip manufacturing. In standard chip production, a single defect ruins only the small die containing it, leaving all other dies on the wafer usable. On a wafer-scale chip, one defect anywhere could theoretically disable the entire processor.

Cerebras addressed this with a dual approach. First, each AI core on the WSE-3 is extremely small, approximately 0.05mm² per core, or about 1% of the size of an NVIDIA H100 SM (streaming multiprocessor), which covers roughly 6mm². Because a random manufacturing defect disables whatever silicon area it occupies, a defect on the WSE-3 knocks out 0.05mm² of compute, while the same defect on an H100 would disable about 6mm² of compute. Cerebras argues this makes the wafer-scale design approximately 100x more fault tolerant per defect, at the silicon area level.[^36]

Second, Cerebras pairs a small reserve of spare cores with a routing architecture that reconfigures connections around defective cores using redundant communication paths. The WSE-3 has 970,000 physical cores, of which 900,000 are active in the shipping product, and Cerebras reports that it enables 93% of the wafer's silicon area.[^36]

### Sparsity acceleration

The WSE-3 includes hardware accelerators for both dynamic and unstructured sparsity. Most [AI accelerator](https://aiwiki.ai/wiki/ai_accelerator)s, including NVIDIA H100 and B200, support only structured sparsity (2:4 sparsity), which requires that exactly two of every four weights be zero. Unstructured sparsity, where any weight can be zero, is more flexible and can achieve higher compression ratios but is computationally harder to accelerate in hardware.

Cerebras claims its sparsity support speeds up training by up to 8x.[^1] At the WSE-3 launch it also announced a partnership with [Qualcomm](https://aiwiki.ai/wiki/qualcomm) to optimize models for Qualcomm's Cloud AI 100 Ultra inference accelerators using sparsity, speculative decoding, MX6 quantization and network architecture search; Cerebras claimed these techniques combined give a 10x improvement in performance per dollar.[^19]

## What is the CS-3 system?

The CS-3 is Cerebras' third-generation system appliance built around the WSE-3. Cerebras' CS-3 product page lists a 16 rack-unit system with 1.2 Tb/s of system I/O. Power consumption is 23 kilowatts per unit, the same as the CS-2 despite delivering twice the compute performance.[^59][^19][^1]

The CS-3 uses liquid cooling with redundant pumps to manage the heat load, and external connectivity is via fiber.[^4] Cerebras' memory options for the CS-3 are:

| MemoryX option | Notes |
|---|---|
| 1.5 TB, 12 TB | Listed in the March 2024 launch announcement[^1] |
| 24 TB, 36 TB | Enterprise SKUs[^6] |
| 120 TB | Hyperscale SKU[^6] |
| 1.2 PB (1,200 TB) | Hyperscale SKU; stores models of up to 24 trillion parameters[^1][^6] |

Clusters scale up to 2,048 CS-3 systems through the SwarmX fabric.[^1] A 2,048-node cluster delivers 256 exaflops of AI compute at FP16 precision (with sparsity) and contains a total of 1.84 billion compute cores.[^1][^7] By comparison, the CS-2 supported clusters of up to 192 systems, so the CS-3 increased the maximum cluster size by a factor of roughly 10.7.[^7]

Cerebras said the CS-3 costs the same as the CS-2.[^1] Next Platform guessed a price of about $2.5 million per node and estimated that a full 2,048-node cluster would cost more than $5 billion, perhaps $6 billion with full MemoryX capacity.[^7]

Software support includes native PyTorch 2.0 integration. Cerebras states that a standard implementation of a [GPT-3](https://aiwiki.ai/wiki/gpt-3)-sized model required 565 lines of code on Cerebras and that the CS-3 requires 97% less code than GPUs for large language models, because it abstracts away the distributed communication primitives that GPU users must manage manually. The system also supports pure data-parallel training for models from 1 billion to 24 trillion parameters without requiring [model parallelism](https://aiwiki.ai/wiki/model_parallelism) or [tensor parallelism](https://aiwiki.ai/wiki/tensor_parallelism) strategies.[^1]

## What are the Condor Galaxy supercomputers?

Condor Galaxy is a series of AI supercomputers built jointly by Cerebras and [G42](https://aiwiki.ai/wiki/g42), the Abu Dhabi-based technology holding company. The network was announced in 2023 as nine interconnected supercomputers totaling 36 exaflops of sparse FP16 AI compute.[^19]

| Supercomputer | Compute | Underlying hardware | Location | Status |
|---|---|---|---|---|
| Condor Galaxy 1 (CG-1) | 4 exaFLOPS | 64 CS-2 nodes | Santa Clara, California | Operational (2023) |
| Condor Galaxy 2 (CG-2) | 4 exaFLOPS | 64 CS-2 nodes | California | Operational (2023) |
| Condor Galaxy 3 (CG-3) | 8 exaFLOPS | 64 CS-3 nodes | Dallas, Texas | Announced as under construction in March 2024 |
| CG-4 through CG-9 | 20 exaFLOPS combined (2023 plan) | | Multiple locations | Planned in the 2023 announcement |

CG-1, launched in 2023, was among the largest AI supercomputers in the world at the time of its announcement, delivering 4 exaFLOPS and hosting 54 million compute cores.[^13] CG-1 and CG-2 used Cerebras CS-2 systems and were both deployed in California.[^1] CG-3, announced on March 13, 2024 (the same day as the WSE-3 announcement), is the first Condor Galaxy system built on CS-3 hardware.[^12] CG-3 uses 64 CS-3 nodes and delivers 8 exaFLOPS, double the capacity of each of its predecessors.[^12][^19] By May 2026, The Register reported that CS-3 systems powered the majority of the Condor Galaxy cluster Cerebras had built for G42.[^46]

The Condor Galaxy network has trained several publicly released models: Jais-30B (an Arabic foundation model), Med42 (a medical domain model), Crystal-Coder-7B (a code generation model), and BTLM-3B-8K (a general-purpose language model with 8,192 context length).[^1] G42 accounted for 83% of Cerebras' reported revenue for 2023, 87% of total revenue in the first half of 2024 and 97% of hardware sales in that half,[^15] reflecting how central the Condor Galaxy partnership was to Cerebras' commercial operations during this period.

## How fast is Cerebras Inference?

On August 27, 2024, Cerebras launched Cerebras Inference, a commercial cloud API offering LLM inference at throughputs it claimed were the fastest publicly available at the time.[^8][^9]

At launch, the service delivered:

| Model | Tokens per second | Comparison to GPU clouds |
|---|---|---|
| Llama 3.1 8B | 1,800 | ~20x faster than NVIDIA GPU-based hyperscale clouds |
| Llama 3.1 70B | 450 | ~20x faster than NVIDIA GPU-based hyperscale clouds |

Cerebras positioned the service against [Groq LPU](https://aiwiki.ai/wiki/groq_hardware)-based inference, claiming 2.4x higher throughput for the 8B model.[^8] Pricing started at $0.10 per million tokens for Llama 3.1 8B and $0.60 per million tokens for Llama 3.1 70B.[^9] Developers received one million free tokens per day at launch. Cerebras also advertised that the service uses native 16-bit weights rather than the 8-bit quantization common among inference providers, noting that 16-bit models score up to 5% higher than 8-bit counterparts on standard benchmarks.[^8] [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) independently measured launch throughput at above 1,800 output [tokens per second](https://aiwiki.ai/wiki/tokens_per_second) for Llama 3.1 8B and above 446 output tokens per second for Llama 3.1 70B.[^9]

The speed advantage comes from the WSE-3's architecture: Cerebras keeps the model's weights in on-chip SRAM, splitting larger models across a small number of CS-3 systems, so memory access per generated token avoids the movement between HBM and compute units that limits GPU systems.[^8]

When a model exceeds the memory of one wafer, Cerebras splits it at layer boundaries across several CS-3 systems; at launch, 20B models fit on a single CS-3 and 70B models on as few as four systems.[^8] Tokens flow through the pipeline sequentially, and the system's fast per-layer processing keeps latency low.

On October 24, 2024, Cerebras raised Llama 3.1 70B throughput to 2,100 tokens per second, which it described as a threefold boost over its prior release (and about 4.7 times the August launch figure), from rewritten kernels, asynchronous wafer I/O and speculative decoding.[^10][^23] On November 18, 2024, Cerebras reported running a customer workload on Llama 3.1 405B at 969 tokens per second, another claimed record for that model size.[^11]

### 2025 throughput records and Llama 4

Inference performance continued to scale through 2025. By April 2025, Cerebras cited a record 2,500 tokens per second on [Llama 3.3](https://aiwiki.ai/wiki/llama_3_3) 70B.[^22]

In April 2025, alongside Meta's release of [Llama 4](https://aiwiki.ai/wiki/llama_4) (the Scout and Maverick mixture-of-experts variants), Cerebras launched inference support and published Artificial Analysis-verified results of over 2,600 tokens per second on Llama 4 Scout, which it described as 19x faster than the fastest GPU solutions; Feldman said the leading GPU solution delivered 137 tokens per second.[^22] On the 400-billion-parameter Llama 4 Maverick model, Artificial Analysis benchmarked Cerebras at 2,522 tokens per second compared to 1,038 tokens per second on [NVIDIA Blackwell](https://aiwiki.ai/wiki/nvidia_blackwell), an explicit head-to-head result on a frontier model.[^30] In the same Maverick test, Artificial Analysis recorded SambaNova at 794 t/s, Groq at 549 t/s, Amazon at 290 t/s, [Google](https://aiwiki.ai/wiki/google) at 125 t/s, and [Microsoft Azure](https://aiwiki.ai/wiki/azure) at 54 t/s, placing Cerebras roughly 3x ahead of the next-fastest specialty silicon and 2.4x ahead of Blackwell.[^30]

These numbers come from Cerebras and from Artificial Analysis rather than from the MLPerf Inference suite. Without MLPerf results from Cerebras, apples-to-apples comparison with NVIDIA, [Intel](https://aiwiki.ai/wiki/intel), and AMD submissions is difficult.

### Meta, Mistral, and the Llama API

On April 29, 2025, Meta and Cerebras announced that Cerebras would provide fast inference in Meta's new **Llama API**, the company's hosted endpoint for its Llama family. Developers can choose Cerebras from the model options in the Llama API, and Cerebras said Llama 4 on Cerebras in the API would run up to 18 times faster than traditional GPU-based solutions.[^21]

Earlier, on February 6, 2025, the French foundation-model startup [Mistral AI](https://aiwiki.ai/wiki/mistral_ai) began using Cerebras Inference to power the Flash Answers feature of **Le Chat**, its consumer chat product; Cerebras said Le Chat reached over 1,100 tokens per second on Mistral's 123-billion-parameter [Mistral Large](https://aiwiki.ai/wiki/mistral_large) 2 model.[^48]

Together, the Meta and Mistral partnerships marked a shift in Cerebras' commercial profile. Where 2023 and the first half of 2024 had been dominated by G42 hardware sales, Cerebras' inference customers by 2025 included [Perplexity](https://aiwiki.ai/wiki/perplexity), [Hugging Face](https://aiwiki.ai/wiki/hugging_face), AlphaSense, Notion and [Cognition](https://aiwiki.ai/wiki/cognition_ai).[^37][^49][^50] Revenue, however, remained concentrated in the UAE: G42 and Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) together accounted for 86% of Cerebras' $510 million in 2025 revenue.[^38][^42]

### Six-data-center inference build-out

On March 11, 2025, Cerebras announced six new AI inference data centers to house CS-3 systems.[^37] Its site list showed Santa Clara, Stockton and Dallas already online, Minneapolis due in the second quarter of 2025, Oklahoma City and Montreal in the third quarter, and further sites in the Midwest or Eastern U.S. and in Europe in the fourth quarter. The build-out was projected to raise aggregate Cerebras inference capacity twenty-fold, to more than 40 million Llama 70B tokens per second, with 85% of capacity in the United States; the Oklahoma City and Montreal sites house hardware owned and operated by Cerebras, while the remaining sites are jointly operated with G42.[^37][^24] The build-out went ahead, though not on the announced schedule. The Oklahoma City site, a roughly 10 MW facility operated with Scale Datacenters that had been due online in June 2025, opened on September 22, 2025 with more than 44 exaflops of AI compute; in announcing it Cerebras described its Dallas and Minneapolis sites as high-capacity builds already in place.[^24][^34][^35]

The capacity expansion was funded in part by a $1.1 billion Series G round announced on September 30, 2025 at an $8.1 billion post-money valuation, led by Fidelity and Atreides; Cerebras said it would use the funds for processor, packaging and system design and to expand its U.S. manufacturing and data center capacity.[^25] In January 2026, Cerebras and [OpenAI](https://aiwiki.ai/wiki/openai) disclosed a deal for 750 megawatts of Cerebras inference capacity (described in the [OpenAI section](#openai) below).[^26]

### Competitive inference landscape

The rapid inference throughput numbers Cerebras published in 2024 came amid a broader competition among specialized inference hardware vendors. The main contenders for the fastest-LLM-inference title in 2024 were Cerebras (WSE-3), Groq LPU, and SambaNova.

Groq's Language Processing Unit (LPU) is built around a different philosophy than the WSE-3: it uses a large array of relatively small chips, each with a modest amount of SRAM, interconnected through a synchronous fabric, and it attracted early developer attention for low-latency LLM inference.

SambaNova uses a reconfigurable dataflow architecture and its [SN40L](https://aiwiki.ai/wiki/sambanova_sn40l) chip, released in late 2023, combined on-chip SRAM with HBM to allow larger models per system than either Groq or a single CS-3.

At its August 2024 launch, Cerebras claimed 2.4 times Groq's throughput on Llama 3.1 8B.[^8] By late October 2024, after software updates, Cerebras pushed its 70B throughput to 2,100 tokens per second.[^10] Through 2025, the competitive picture broadened to include NVIDIA Blackwell B200 GPU clouds on the high-throughput end, with Cerebras' Llama 4 Maverick result of 2,522 tokens per second per user roughly 2.4x ahead of Blackwell in published Artificial Analysis figures.[^30]

These comparisons are vendor-reported or come from a single independent benchmarker, and they have not been verified through a standardized benchmark suite such as MLPerf Inference. The methodologies differ: vendors vary [batch size](https://aiwiki.ai/wiki/batch_size)s, context lengths, and precision settings, making direct comparison difficult.

## How does the WSE-3 compare to the NVIDIA H100 and B200?

The WSE-3 and the NVIDIA H100 and [NVIDIA B200](https://aiwiki.ai/wiki/nvidia_b200) serve overlapping but distinct purposes. The H100 and B200 are general-purpose data center GPUs used across training, inference, and simulation workloads at a wide range of model sizes. The WSE-3 is optimized specifically for AI training and inference workloads where models fit within the weight streaming architecture.

| Specification | Cerebras CS-3 (WSE-3) | NVIDIA H100 SXM | NVIDIA B200 SXM |
|---|---|---|---|
| Process node | TSMC 5nm | TSMC 4N | TSMC 4NP |
| Transistors | 4 trillion | 80 billion | ~208 billion |
| Die area | 46,225 mm² | 814 mm² | Two reticle-sized dies |
| On-chip memory | 44 GB SRAM | 80 GB HBM3 | 192 GB HBM3e |
| Memory bandwidth | 21 PB/s (on-wafer) | 3.35 TB/s | 8.0 TB/s |
| Peak AI compute (FP16, with sparsity) | 125 petaFLOPS | 1.979 petaFLOPS | ~4.5 petaFLOPS |
| Interconnect | 214 Pb/s on-wafer | NVLink 900 GB/s | NVLink 1.8 TB/s |
| System power | 23 kW (CS-3) | 700 W (per GPU) | 1,000 W (per GPU) |
| Form factor | Rack-scale appliance (16 RU) | SXM module (server-integrated) | SXM module (server-integrated) |

The transistor and die area comparisons illustrate the wafer-scale approach: the WSE-3 integrates roughly 50 times more transistors than the H100 and covers approximately 57 times more silicon area.[^4][^19] Against the newer Blackwell generation, Cerebras' own comparison set the WSE-3's 4 trillion transistors and 125 petaFLOPS against the B200's 208 billion transistors and 4.4 petaFLOPS, or roughly 19 and 28 times more.[^47] However, the majority of the WSE-3's transistors go to compute logic and SRAM, not to HBM interface circuitry, since the WSE-3 does not use HBM at all.

For peak AI compute, a single CS-3 (125 petaFLOPS sparse FP16) outperforms a single H100 (1.979 petaFLOPS sparse FP16) by approximately 63x and exceeds a single B200 (approximately 4.5 petaFLOPS sparse FP16) by approximately 28x.[^3][^19] On dense FP16, the gap is much smaller, since Cerebras' figure falls by roughly a factor of ten without sparsity while NVIDIA's halves; The Register estimated in 2024 that a CS-3 would be slightly slower than a pair of DGX H100 servers on dense FP16.[^7][^19] NVIDIA systems are also deployed in racks: NVIDIA rates a GB200 NVL72 rack, with 72 Blackwell GPUs, at 360 petaFLOPS of FP16 with sparsity.[^56]

A March 2025 paper by researchers at the company Insaito, posted on [arXiv](https://aiwiki.ai/wiki/arxiv), compared peak specifications of rack systems at equal space and power and found that the CS-3 offers approximately 3.5x higher peak FP8 performance and 7x higher peak FP16 performance than an H100-based system, and about 1.1x (FP8) and 2.15x (FP16) more than a B200 system; on performance per watt per dollar, it found the B200 system about 1.5 to 3 times better.[^16]

The interconnect comparison is where the architectural gap is widest. The WSE-3's on-wafer fabric at 214 petabits per second vastly exceeds NVLink's 900 GB/s (7.2 terabits per second) per H100 or 1.8 TB/s (14.4 terabits per second) per B200. Cerebras states the CS-3 provides over 200 times the interconnect bandwidth of the full NVL72 rack.[^20] This matters most for training jobs where gradient synchronization across many units becomes the bottleneck.

The critical limitation is memory capacity. The WSE-3's 44GB of SRAM is much smaller than the 80GB per H100 or 192GB per B200, and smaller still than the 13.4TB of HBM3e across a GB200 NVL72 rack.[^56] Large models that do not fit neatly within Cerebras' weight streaming pipeline suffer performance degradation, whereas on GPU clusters they can often be tensor-parallelized or pipeline-parallelized with less architectural constraint. Models requiring more than roughly 24 trillion parameters require more than the largest available MemoryX configuration.[^1]

## Who uses the WSE-3? Customers and deployments

### G42

G42 is Cerebras' largest historical customer and a strategic investor. Under the Condor Galaxy partnership, G42 agreed in May 2024 to spend about $1.43 billion more with Cerebras by February 2025 to build out the rest of the Condor Galaxy infrastructure, and the partnership accounts for the dominant share of Cerebras' early hardware revenue.[^54][^15] G42 uses the Condor Galaxy network to train large Arabic and multilingual language models, including Jais-30B, a 30-billion parameter Arabic foundation model trained on Condor Galaxy 1.[^1] In 2025, G42 accounted for 24% of Cerebras' revenue and Mohamed bin Zayed University of Artificial Intelligence, another UAE institution, for 62%, so customer concentration remained high even after the broader inference cloud build-out.[^38][^42]

### Mayo Clinic

Mayo Clinic is a Cerebras customer for healthcare AI research. At the January 2024 J.P. Morgan Healthcare Conference, Mayo announced a multi-year collaboration with Cerebras to build AI models, including genomic models drawing on data from more than 100,000 patients.[^57][^60] On January 14, 2025, Cerebras and Mayo Clinic unveiled a genomic foundation model aimed first at predicting patient response to rheumatoid arthritis treatment; it combined public human reference genome data with exome data from 500 Mayo Clinic patients.[^14]

### Meta, Mistral, IBM, and the Llama API era

Beginning in early 2025, the customer profile shifted toward large-scale inference users. Meta added Cerebras as a fast-inference option in its Llama API in April 2025,[^21] and Mistral moved Le Chat's Flash Answers onto Cerebras in February 2025.[^48] Hugging Face and AlphaSense announced in March 2025 that they were adopting Cerebras inference, and Perplexity used it for instant search results.[^37] Notion adopted Cerebras in July 2025 to power enterprise search in Notion AI for Work.[^49] Cognition, the developer of the Devin [coding agent](https://aiwiki.ai/wiki/coding_agent), serves its SWE-1.5 and SWE-1.6 coding models on Cerebras at up to about 950 tokens per second in [Windsurf](https://aiwiki.ai/wiki/windsurf).[^50] IBM offers Cerebras inference through its watsonx Model Gateway.[^38] In April 2025, [DARPA](https://aiwiki.ai/wiki/darpa) awarded Cerebras a contract to build a real-time computing platform combining its wafer-scale processors with Ranovus co-packaged optics.[^51]

### GlaxoSmithKline, TotalEnergies, national laboratories and other early customers

GlaxoSmithKline is an early Cerebras customer.[^9] TotalEnergies Research & Technology USA selected a CS-2 in April 2022 for multi-energy research, and Cerebras reported more than 200 times the performance of an NVIDIA A100 GPU on a key stencil benchmark.[^53] Lawrence Livermore National Laboratory integrated a CS-1 into its Lassen supercomputer in 2020, and Argonne National Laboratory, which described itself as "a long-time partner of Cerebras," commented on the WSE-3 launch.[^52][^1] Cerebras' prospectus also refers to defense organizations and national laboratories among its users.[^38]

### OpenAI

[OpenAI](https://aiwiki.ai/wiki/openai) emerged as a major Cerebras customer in 2026. On January 14, 2026, Cerebras and OpenAI disclosed an inference deal, reported by CNBC as worth more than $10 billion, under which OpenAI would deploy 750 megawatts of Cerebras capacity to power inference workloads behind [ChatGPT](https://aiwiki.ai/wiki/chatgpt) and the [OpenAI API](https://aiwiki.ai/wiki/openai_api).[^26] Cerebras' April 2026 IPO filing valued the agreement, a master relationship agreement signed in December 2025, at more than $20 billion: Cerebras is to make 250 MW available each year from 2026 to 2028, and OpenAI can buy an additional 1.25 gigawatts of capacity through 2030.[^39][^42] As part of the arrangement, OpenAI lent Cerebras $1 billion as a working capital loan at 6% interest and received a warrant for 33,445,026 Cerebras shares at an exercise price of $0.00001 that vests as its purchases grow.[^38][^42] Cerebras said the agreement "represents a substantial portion of our projected revenues over the next several years."[^42]

## What happened in the Cerebras IPO?

Cerebras filed an S-1 registration statement with the Securities and Exchange Commission on September 30, 2024, disclosing its intention to raise funds through an initial public offering.[^15] The filing revealed $78.7 million in revenue for fiscal year 2023, $24.6 million for 2022, and $136.4 million for the first half of 2024.[^15]

The filing also disclosed that G42 accounted for 83% of Cerebras' 2023 revenue and 87% of its revenue (97% of hardware revenue) for the first half of 2024.[^15] The listing was held up by a national security review by the Committee on Foreign Investment in the United States ([CFIUS](https://aiwiki.ai/wiki/cfius)) of G42's investment in Cerebras; on March 31, 2025, Cerebras said it had obtained CFIUS clearance to sell shares to G42.[^41] Cerebras raised a $1.1 billion Series G at an $8.1 billion valuation on September 30, 2025, and days later, on October 3, 2025, withdrew its registration statement; CNBC reported that it withdrew the paperwork to add information on its financial performance and strategy.[^25][^40][^42] In February 2026 it raised a further $1 billion Series H at a valuation of about $23 billion.[^28][^42]

The company filed a new S-1 on April 17, 2026, disclosing $510 million in revenue for 2025, up 76% from $290 million in 2024.[^39][^42] Under GAAP it reported 2025 net income of $237.8 million, helped by a large non-operating gain, compared with a net loss of $481.6 million in 2024; net income attributable to common stockholders was about $88 million, the figure CNBC reported, and on a non-GAAP basis the company posted a net loss of $75.7 million.[^38][^39][^27] It also reported $24.6 billion in remaining performance obligations as of December 31, 2025.[^42]

On May 13, 2026, Cerebras priced its IPO at $185 per share, above the indicated $150 to $170 range, selling 30 million Class A common shares for gross proceeds of approximately $5.55 billion, with underwriters holding a 30-day option for an additional 4.5 million shares.[^29] Morgan Stanley, Citigroup, Barclays, and UBS were lead book-running managers.[^29][^43] Trading opened on the Nasdaq Global Select Market under ticker CBRS on May 14, 2026 at $350 per share and closed the first session at $311.07, a 68% gain that CNBC said valued the company at about $95 billion; CNBC called it the largest IPO for a U.S. technology company since Uber's in 2019.[^27] The offering closed on May 15, 2026 with the underwriters' option exercised in full, for 34.5 million shares and about $6.38 billion in gross proceeds.[^43]

Post-IPO, customer concentration remained the most visible risk identified in analyst commentary: G42 and MBZUAI represented 86% of fiscal 2025 revenue, and the OpenAI agreement is expected to account for a substantial share of revenue over the next several years.[^32][^38] The prospectus said the net proceeds, about $5.4 billion before the underwriters' option, would cover tax withholding for employee stock awards and general corporate purposes, including working capital, operating expenses and capital expenditures.[^38] In March 2026, Amazon Web Services signed a binding term sheet to become the first hyperscaler to deploy Cerebras systems in its own data centers, with a warrant for up to 2,696,678 Cerebras shares at $100 each.[^38][^42]

## What is next after the WSE-3? Successor and roadmap

As of mid-October 2026, Cerebras had not announced a fourth-generation Wafer-Scale Engine; the [WSE-4](https://aiwiki.ai/wiki/cerebras_wse_4) remained a label for the expected successor. On August 18, 2026, Cerebras instead introduced the CS-4, its fourth-generation system, built from three new WSE-3 Turbo processors in a redesigned rack-scale platform called Nexus. Cerebras says the CS-4 delivers twice as much power to the WSE-3 Turbo for higher operating frequencies, gives up to 2x the performance and up to 10x the throughput per watt of the CS-3, and runs inference up to 30 times faster than GPU systems; it said first CS-4 shipments would begin in the third quarter of 2026.[^44]

At Hot Chips on August 25, 2026, Cerebras said its next-generation WSE will debut in the CS-5, targeted for 2027 with goals of up to 10,000 output tokens per second per user on leading open models, and that a later CS-6 will stack DRAM on top of the wafer's compute and SRAM to expand memory capacity.[^45] Before these announcements, The Register had expected a WSE-4 to emphasize lower-precision data types such as FP8 and FP4 and possibly to add SRAM through TSMC's 3D chip stacking.[^46]

## What are the WSE-3's limitations?

The WSE-3 and the broader Cerebras architecture have several practical constraints.

Memory capacity is the most cited limitation. The 44GB of on-chip SRAM is small relative to HBM-equipped GPUs. Although weight streaming mitigates this for training workloads by keeping weights off-chip, inference workloads must also accommodate KV cache growth as context length increases. At high batch sizes and long contexts, the KV cache can consume a substantial portion of the 44GB, reducing effective [model capacity](https://aiwiki.ai/wiki/model_capacity). The Register estimated in May 2026 that a trillion-parameter model such as Kimi K2 would require between 12 and 48 WSE-3 chips.[^46]

Cerebras' published WSE-3 performance figures are 16-bit AI figures, and at the chip's launch the company said nothing about FP64 double-precision support, which many scientific applications require.[^58] General-purpose computing tasks that do not map well onto the 2D mesh topology also perform poorly.

The wafer-scale manufacturing approach involves tradeoffs in yield and cost. Although Cerebras' redundancy approach routes around defective cores, producing defect-free wafers at sufficient volume to meet enterprise demand requires tight manufacturing partnership with TSMC, and each wafer corresponds to a single processor rather than dozens, reducing the economies of scale from which conventional chipmakers benefit.

The software ecosystem is narrower than that surrounding NVIDIA GPUs. While the CS-3 supports PyTorch 2.0, it does not support the full [CUDA](https://aiwiki.ai/wiki/cuda) ecosystem, and workloads written for NVIDIA hardware typically require porting. Cerebras provides its own SDK and compiler, and the company claims that porting reduces code complexity, but the initial migration effort is a barrier for organizations with existing GPU-optimized codebases.

Independent benchmarking remains a contested topic. Cerebras' published inference results rely on its own measurements and on Artificial Analysis rather than on MLPerf Inference, the most widely cited industry benchmark suite, which makes direct comparisons with NVIDIA, Intel, and AMD inference submissions difficult.

Customer concentration is a financial rather than technical limitation but figures heavily in post-IPO analyst commentary. G42 and MBZUAI still represented 86% of fiscal 2025 revenue at the time of the May 2026 listing,[^32][^38] and the OpenAI agreement, which Cerebras says represents a substantial portion of its projected revenue, adds another large single customer over the 2026 to 2030 horizon.[^42] The prospectus also pointed to AWS's planned deployment as a possible future source of material revenue.[^38]

## See also

- [Cerebras Systems](https://aiwiki.ai/wiki/cerebras)
- [Cerebras WSE-4](https://aiwiki.ai/wiki/cerebras_wse_4)
- [AI chip](https://aiwiki.ai/wiki/ai_chip)
- [NVIDIA H100](https://aiwiki.ai/wiki/nvidia_h100)
- [NVIDIA B200](https://aiwiki.ai/wiki/nvidia_b200)
- [Groq LPU](https://aiwiki.ai/wiki/groq_hardware)
- [AI accelerator](https://aiwiki.ai/wiki/ai_accelerator)
- [AI inference](https://aiwiki.ai/wiki/inference)

## References

[^1]: Cerebras Systems. "Cerebras Announces Third Generation Wafer-Scale Engine." Press release, March 13, 2024. https://www.cerebras.ai/press-release/cerebras-announces-third-generation-wafer-scale-engine
[^2]: IEEE Spectrum. "Cerebras WSE-3: Third Generation Superchip for AI." March 2024. https://spectrum.ieee.org/cerebras-chip-cs3
[^3]: Tom's Hardware. "Cerebras launches 900,000-core 125 PetaFLOPS wafer-scale processor for AI." March 13, 2024. https://www.tomshardware.com/tech-industry/artificial-intelligence/cerebras-launches-900000-core-125-petaflops-wafer-scale-processor-for-ai-theoretically-equivalent-to-about-62-nvidia-h100-gpus
[^4]: ServeTheHome. "Cerebras WSE-3 AI Chip Launched 56x Larger than NVIDIA H100." March 2024. https://www.servethehome.com/cerebras-wse-3-ai-chip-launched-56x-larger-than-nvidia-h100-vertiv-supermicro-hpe-qualcomm/
[^5]: Data Center Dynamics. "Cerebras unveils four trillion-transistor giant chip, targets generative AI." March 2024. https://www.datacenterdynamics.com/en/news/cerebras-unveils-four-trillion-transistor-giant-chip-targets-generative-ai/
[^6]: Cerebras Systems. "Cerebras CS-3: the world's fastest and most scalable AI accelerator." Blog, March 2024. https://www.cerebras.ai/blog/cerebras-cs3
[^7]: Next Platform. "Cerebras Goes Hyperscale With Third Gen Waferscale Supercomputers." March 14, 2024. https://www.nextplatform.com/ai/2024/03/14/cerebras-goes-hyperscale-with-third-gen-waferscale-supercomputers/1642584
[^8]: Cerebras Systems. "Introducing Cerebras Inference: AI at Instant Speed." Blog, August 27, 2024. https://www.cerebras.ai/blog/introducing-cerebras-inference-ai-at-instant-speed
[^9]: Cerebras Systems. "Cerebras Launches the World's Fastest AI Inference." Press release, August 27, 2024. https://www.cerebras.ai/press-release/cerebras-launches-the-worlds-fastest-ai-inference
[^10]: Cerebras Systems. "Cerebras Inference now 3x faster: Llama3.1-70B breaks 2,100 tokens/s." Blog, October 24, 2024. https://www.cerebras.ai/blog/cerebras-inference-3x-faster
[^11]: Cerebras Systems. "Llama 3.1 405B now runs at 969 tokens/s on Cerebras Inference." Blog, November 18, 2024. https://www.cerebras.ai/blog/llama-405b-inference
[^12]: Cerebras Systems. "Cerebras and G42 Break Ground on Condor Galaxy 3, an 8 exaFLOPs AI Supercomputer." Press release, March 13, 2024. https://www.cerebras.ai/press-release/cerebras-g42-announce-condor-galaxy-3
[^13]: Cerebras Systems. "Introducing Condor Galaxy 1: a 4 exaFLOPS Supercomputer for Generative AI." Blog, 2023. https://www.cerebras.ai/blog/introducing-condor-galaxy-1-a-4-exaflop-supercomputer-for-generative-ai
[^14]: Cerebras Systems. "Cerebras Systems and Mayo Clinic Unveil Best in Class Genomic Foundation Model." Press release, January 14, 2025. https://www.cerebras.ai/press-release/cerebras-systems-and-mayo-clinic-unveil-best-in-class-genomic-foundation-model
[^15]: SEC. Cerebras Systems S-1 Registration Statement. September 30, 2024. https://www.sec.gov/Archives/edgar/data/2021728/000162828024041596/cerebras-sx1.htm
[^16]: Kundu et al. "A Comparison of the Cerebras Wafer-Scale Integration Technology with Nvidia GPU-based Systems for Artificial Intelligence." arXiv:2503.11698, March 11, 2025. https://arxiv.org/html/2503.11698v1
[^17]: Hot Chips 2024. "Cerebras Wafer-Scale AI." Presentation by Sean Lie. August 2024. https://hc2024.hotchips.org/assets/program/conference/day2/72_HC2024.Cerebras.Sean.v03.final.pdf
[^18]: Digitimes. "TSMC and Cerebras' chip cooperation across three generations embodies development of Moore's Law." March 2024. https://www.digitimes.com/news/a20240315PD203/tsmc-cerebras-ic-design-ai-supercomputing.html
[^19]: The Register. "Cerebras unveils its third-gen waferscale AI accelerator." March 13, 2024. https://www.theregister.com/2024/03/13/cerebras_claims_to_have_revived/
[^20]: Futurum Group. "Cerebras CS-3: Bring on the NVIDIA Blackwell Competition." 2024. https://futurumgroup.com/insights/cerebras-cs-3-bring-on-the-nvidia-blackwell-competition/
[^21]: Cerebras Systems. "Meta Collaborates with Cerebras to Drive Fast Inference for Developers in New Llama API." Press release, April 29, 2025. https://www.cerebras.ai/press-release/meta-collaborates-with-cerebras-to-drive-fast-inference-for-developers-in-new-llama-api
[^22]: Cerebras Systems. "Cerebras Launches World's Fastest Inference for Meta Llama 4." Press release, April 2025. https://www.cerebras.ai/press-release/llama4PR
[^23]: Cerebras Systems. "Cerebras Triples its Industry-Leading Inference Performance, Setting New All-Time Record." Press release, October 25, 2024. https://www.cerebras.ai/press-release/cerebras-triples-its-industry-leading-inference-performance-setting-new-all-time-record
[^24]: Data Center Frontier. "Cerebras Unveils Six Data Centers to Meet Accelerating Demand for AI Inference at Scale." March 18, 2025. https://www.datacenterfrontier.com/hyperscale/article/55273769/cerebras-unveils-six-data-centers-to-meet-accelerating-demand-for-ai-inference-at-scale
[^25]: Cerebras Systems. "Cerebras Raises $1.1 Billion at $8.1 Billion Valuation." Press release, September 30, 2025. https://www.cerebras.ai/press-release/series-g
[^26]: CNBC. "Cerebras scores OpenAI deal worth over $10 billion ahead of AI chipmaker's IPO." January 14, 2026. https://www.cnbc.com/2026/01/14/cerebras-scores-openai-deal-worth-over-10-billion.html
[^27]: CNBC. "Cerebras pops 68% in Nasdaq debut, pushing the AI chipmaker's market cap to $95 billion." May 14, 2026. https://www.cnbc.com/2026/05/14/cerebras-cbrs-stock-trade-nasdaq-ipo.html
[^28]: Next Platform. "With Its IPO Done, Cerebras Can Get Back To Pushing The AI Envelope." May 15, 2026. https://www.nextplatform.com/compute/2026/05/15/with-its-ipo-done-cerebras-can-get-back-to-pushing-the-ai-envelope/5241317
[^29]: Cerebras Systems. "Cerebras Systems Announces Pricing of Initial Public Offering." Press release, May 13, 2026. https://www.cerebras.ai/press-release/cerebras-systems-announces-pricing-of-initial-public-offering
[^30]: Cerebras Systems. "Cerebras Beats NVIDIA Blackwell: Llama 4 Maverick Inference." Press release, May 28, 2025. https://www.cerebras.ai/press-release/maverick
[^31]: Next Platform / The Register. WSE-4 and CS-4 roadmap analyst commentary, 2025-2026. See https://www.nextplatform.com/compute/2026/05/15/with-its-ipo-done-cerebras-can-get-back-to-pushing-the-ai-envelope/5241317 and https://www.theregister.com/ai-ml/2026/05/15/cerebras-wafer-scale-ai-bet-delivers-blockbuster-ipo/
[^32]: TechTimes. "Cerebras Raises $5.55 Billion in AI Chip IPO, But 86% Revenue Dependence on UAE Entities Unresolved." May 15, 2026. https://www.techtimes.com/articles/316698/20260515/cerebras-raises-555-billion-ai-chip-ipo-86-revenue-dependence-uae-entities-unresolved.htm
[^33]: Crypto Briefing. "OpenAI commits $20B for potential 11% stake in Cerebras." May 2026. https://cryptobriefing.com/openai-20b-cerebras-stake/
[^34]: Cerebras Systems. "The Fastest AI Datacenters will run on Cerebras: Meet OKC." Blog, September 22, 2025. https://www.cerebras.ai/blog/okc
[^35]: Data Center Dynamics. "Cerebras opens 10MW data center in Oklahoma City." September 23, 2025. https://www.datacenterdynamics.com/en/news/cerebras-opens-10mw-data-center-in-oklahoma-city/
[^36]: Cerebras Systems. "100x Defect Tolerance: How Cerebras Solved the Yield Problem." Blog, January 13, 2025. https://www.cerebras.ai/blog/100x-defect-tolerance-how-cerebras-solved-the-yield-problem
[^37]: Cerebras Systems. "Cerebras Announces Six New AI Datacenters Across North America and Europe to Deliver Industry's Largest Dedicated AI Inference Cloud." Press release, March 11, 2025. https://www.cerebras.ai/press-release/cerebras-announces-six-new-ai-datacenters-across-north-america-and-europe-to-deliver-industry-s
[^38]: SEC. Cerebras Systems Form 424B4 Prospectus. May 14, 2026. https://www.sec.gov/Archives/edgar/data/2021728/000162828026035214/cerebras-424b4.htm
[^39]: SEC. Cerebras Systems S-1 Registration Statement. April 17, 2026. https://www.sec.gov/Archives/edgar/data/2021728/000162828026025762/cerebras-sx1april2026.htm
[^40]: SEC. Cerebras Systems Form RW (request for withdrawal of registration statement). October 3, 2025. https://www.sec.gov/Archives/edgar/data/2021728/000119312525230382/d886849drw.htm
[^41]: CNBC (syndicated by NBC 5 Dallas-Fort Worth). "AI chipmaker Cerebras announces CFIUS clearance, a key step toward IPO." March 31, 2025. https://www.nbcdfw.com/news/business/money-report/ai-chipmaker-cerebras-announces-cfius-clearance-a-key-step-toward-ipo/3804980/
[^42]: CNBC. "AI chipmaker Cerebras files to go public after scrapping IPO plans last year." April 17, 2026. https://www.cnbc.com/2026/04/17/cerebras-new-ipo-ai-chips.html
[^43]: Cerebras Systems. "Cerebras Systems Announces Closing of Initial Public Offering." Press release, May 15, 2026. https://www.cerebras.ai/press-release/cerebras-systems-announces-closing-of-initial-public-offering
[^44]: Cerebras Systems. "Introducing Cerebras CS-4: The Fastest AI Just Got Faster." Blog, August 18, 2026. https://www.cerebras.ai/blog/introducing-cerebras-cs-4
[^45]: Cerebras Systems. "Ultrafast Frontier Inference: Cerebras Deep Dive at Hot Chips 2026." Blog, August 25, 2026. https://www.cerebras.ai/blog/ultrafast-frontier-inference-cerebras-deep-dive-at-hot-chips-2026
[^46]: The Register. "Cerebras risked it all on dinner plate-sized AI accelerators a decade ago. Today it's worth $66B." May 15, 2026. https://www.theregister.com/ai-ml/2026/05/15/cerebras-wafer-scale-ai-bet-delivers-blockbuster-ipo/5240821
[^47]: Cerebras Systems. "Cerebras CS-3 vs. Nvidia B200: 2024 AI Accelerators Compared." Blog, 2024. https://cerebras.ai/blog/cerebras-cs-3-vs-nvidia-b200-2024-ai-accelerators-compared
[^48]: Cerebras Systems. "Cerebras brings instant inference to Mistral Le Chat." Blog, February 6, 2025. https://www.cerebras.ai/blog/mistral-le-chat
[^49]: Cerebras Systems. "Cerebras Enables Notion to Deliver Real-Time Enterprise Search for 100 Million Workspace Users." Press release, July 8, 2025. https://www.cerebras.ai/press-release/cerebras-enables-notion-to-deliver-real-time-enterprise-search-for-100-million-workspace-users
[^50]: Cerebras Systems. "Case Study: Cognition x Cerebras." Blog, May 1, 2026. https://www.cerebras.ai/blog/case-study-cognition-x-cerebras
[^51]: Cerebras Systems. "DARPA Selects Cerebras to Deliver Next Generation, Real-Time Compute Platform for Advanced Military and Commercial Applications." April 1, 2025. https://www.cerebras.ai/news/darpa-selects-cerebras-to-deliver-next-generation-real-time-compute-platform-for-advanced
[^52]: HPCwire. "Cerebras-LLNL Deployment Teases Second Gen Wafer Scale AI Chip." August 19, 2020. https://www.hpcwire.com/2020/08/19/cerebras-llnl-deployment-teases-second-gen-wafer-scale-ai-chip/
[^53]: Cerebras Systems. "TotalEnergies Research & Technology USA Selects Cerebras Systems CS-2 to Accelerate Multi-Energy Research." Press release, April 22, 2022. https://www.cerebras.ai/press-release/totalenergies-research-technology-usa-selects-cerebras-systems-cs-2-to-accelerate-multi-energy-research
[^54]: Next Platform. "Cerebras Needs Wall Street To Expand Beyond One Core Customer." October 2, 2024. https://www.nextplatform.com/2024/10/02/cerebras-needs-wall-street-to-expand-beyond-one-core-customer
[^55]: TechCrunch. "The five technical challenges Cerebras overcame in building the first trillion-transistor chip." August 19, 2019. https://techcrunch.com/2019/08/19/the-five-technical-challenges-cerebras-overcame-in-building-the-first-trillion-transistor-chip/
[^56]: NVIDIA. "NVIDIA GB200 NVL72." Product page. https://www.nvidia.com/en-us/data-center/gb200-nvl72/
[^57]: Fierce Healthcare. "JPM24: Mayo Clinic deepens AI commitment with new Cerebras Systems deal." January 2024. https://www.fiercehealthcare.com/health-tech/jpm24-mayo-clinic-deepens-ai-commitment-new-cerebras-systems-deal
[^58]: XPU.pub. "Cerebras WSE-3 Doubles FP16 Perf to Target the Biggest AI Models." March 18, 2024. https://xpu.pub/2024/03/18/cerebras-wse-3/
[^59]: Cerebras Systems. "Cerebras CS-3." Product page. https://www.cerebras.ai/cs3
[^60]: Mayo Clinic News Network. "Mayo Clinic engages Cerebras to deliver potent computing power, scale AI transformation." January 2024. https://newsnetwork.mayoclinic.org/discussion/mayo-clinic-engages-cerebras-to-deliver-potent-computing-power-scale-ai-transformation/