Cerebras CS-4 Debuts With 30x Faster AI Inference Using Three Wafer-Scale Chips

Cerebras CS-4 rack systems juice chips for every last drop of AI performance

If high-speed AI inference is what you're after, memory bandwidth is the bottleneck to beat. At a mind-numbing 21.6 petabytes per second (PB/s) of memory bandwidth, Cerebras' dinner-plate-sized AI accelerators were already 1,000x faster than Nvidia's or AMD's best GPUs. The chip newcomer unveiled its next-gen Wafer Scale Engine (WSE) and Nexus rack systems on Tuesday. Cerebras aims to extend that lead by boosting throughput per watt tenfold over the previous generation. Putting the 'T' in Turbo Cerebras accomplishes this in a couple of ways. But, from what we can tell, the primary lever comes from squeezing its chips for every hertz they've got. The newly announced WSE-3T -- the "T" here stands for "Turbo" -- promises twice the compute, memory fabric, and I/O bandwidth of the now two-year-old WSE-3. Yet, if you look at the chart below, you'll notice it accomplishes this using the same process tech, wafer area size, transistor count, core count, and SRAM capacity. That's because the WSE-3T isn't new silicon. Instead, Cerebras tells us it's just pushing its existing wafer scale engine harder. The main innovation this time around seems to be related to power delivery, which is apparently so efficient that they're able to push twice the power through the chip, which "enables higher operating frequencies and faster token generation." How much higher does it clock? By our estimate, Cerebras is now running the silicon at 2.8 GHz, up from 1.4 GHz last gen, which would be quite the accomplishment. In any case, each WSE-3T boasts 250 petaFLOPS of AI compute, 44 GB of SRAM (that's not a typo, there really is that much SRAM on there), good for 43.2 PB/s of memory bandwidth, and 2.4 Tbps of off-die connectivity. On paper that sounds more impressive than it really is. AMD and Nvidia's latest GPUs offer 4 to 5 petaFLOPS of dense FP16 compute or 35 to 50 petaFLOPS at FP4. Cerebras' headline performance figure relies heavily on sparsity, which as a general rule doesn't benefit LLM inference. Assuming the same 10x sparsity we saw with the WSE-3, the WSE-3T's dense FP16 performance should be closer to 25 petaFLOPS, which is still impressive, just not as impressive as the chipmaker would have you believe. We also suspect the WSE-3T's peak memory bandwidth is purely theoretical. During LLM inference, the WSE-3 lacked the compute necessary to saturate its SRAM on its own, and we have no reason to believe the Turbo variant will be any different. However, this time around Cerebras isn't trying to run the entire inference stack on its own accelerators. Instead, it has partnered with Amazon Web Services (AWS) and AMD to offload the compute-intensive prompt processing bits of the inference pipeline onto their respective Trainium XPUs and Instinct GPUs. At least for inference, Cerebras' chips now function primarily as decode accelerators, similar to how Nvidia is using Groq -- not to be confused with Elon Musk's Grok family of models -- LPUs in its LPX rack systems. The major benefit for Cerebras is its chips have a whack ton of SRAM on board. So, instead of needing 2,000 LPUs to run a trillion-parameter model, Cerebras can get away with using a few dozen, depending on the precision at which the weights are stored. Curiously, Cerebras opted to double performance this generation rather than boost SRAM capacity, which hasn't increased meaningfully since the WSE-2 launched five years ago. In a disaggregated inference environment where prefill is handled by GPUs, we'd have expected to see Cerebras prioritize SRAM capacity over compute. However, given that these disaggregated compute architectures are a relatively new phenomenon, it's possible Cerebras was already too far along in production to pivot. This likely explains the Turbo naming convention. If Cerebras plans to continue down this path, we expect the WSE-4, which is presumably still coming, to offer only modest performance gains at FP16 while roughly doubling SRAM capacity. Our sibling site The Next Platform has drawn up some predictions of what the WSE-4 might look like if you're interested. Cerebras goes rackscale Cerebras' latest generation of wafer scale accelerators also sees the company get serious about rack-scale compute architectures. Much like Nvidia's NVL72 and AMD's Helios racks, Cerebras' CS-4 makes the leap from a monolithic system to a modular architecture that breaks out compute, power delivery, and cabling for easier deployment, maintenance, and upgrades. The company's chips are now housed in what it calls a "backpack" form factor, a sort of self-contained system with all the control electronics on board. Each CS-4 can be equipped with up to three of these backpacks, which, as their name suggests, plug into the back of the rack, while the front of the rack is dedicated to the power shelves used to feed the beast. Cabling presumably runs down the center of the system. We're guessing it's probably not an OCP-compliant design, but when your chip is the size of a charcuterie board, it never was going to be. With 3x the accelerators per system, power consumption has unsurprisingly increased. Cerebras hasn't said how much power the racks will suck back, but it has said that its more efficient power delivery means it can push twice as many watts through the chip. The WSE-3 was already a hot chip at 15 kW at the wafer level and around 23 kW at the system level. This means we're probably looking at around 46 kW for each CS-4 backpack and a total system power of between 120 kW and 140 kW. A few years ago, that'd have been a monstrous amount of power, even for a liquid-cooled machine. Today, the machine looks positively conservative next to the 240 to 250 kW rack systems coming from AMD and Nvidia later this year. Killing latency by ditching the switch Even with 132 GB of SRAM memory per rack, you're still going to need a lot of racks to run any reasonably sized model, which means I/O is a major consideration. It just so happens that the new racks and silicon that power them have gotten some beefy upgrades in this respect. Each chip is equipped with 2.4 Tbps of chip-to-chip bandwidth, up from 1.2 Tbps. But arguably the more important improvement is to latency, which thanks to some clever tricks has been cut from five microseconds down to just two. That's because unlike GPUs, which rely heavily on bandwidth-intensive tensor and expert parallelism to multiply their compute and effective memory bandwidth, Cerebras' chips are already so fast they can get away with using pipeline parallelism. Pipeline parallelism is about as simple as multi-accelerator inference gets. Model weights are distributed across each accelerator and work is performed sequentially, one chip after another, hence the name pipeline parallelism. While interconnect bandwidth isn't as much of an issue, pipeline parallelism does benefit from very low latency connections. Cerebras is able to lower its interconnect latency in a pretty simple way: Get rid of all the extra switches and just have the chips talk to one another. This is a pretty big departure from what a lot of AI chip designers have been doing with their own rack-scale designs. AWS for example ditched its chip-to-chip mesh in favor of a switched fabric in its Trainium3 accelerators, which we looked at in detail late last year. Cerebras tells us it's using a 2D torus, which you can think of as a grid where the ends wrap around to the other side. The topology, the company says, can support models up to 50 trillion parameters in size, though no such model currently exists to our knowledge. And for models that do exist, it appears the parts will be quite performant, achieving speeds of up to 4,400 tok/s per user in gpt-oss-120b on a single CS-4 system, compared to around 350 tok/s on the fastest GPU-based inference service today, according to the benchmarking gurus at Artificial Analysis. Cerebras' new mesh topology isn't mandatory. If you wanted to connect the chips using a switched fabric, there's nothing stopping you. Like basically every modern chip, it supports RDMA over converged Ethernet (RoCE), too. The compromise is your latencies may be a bit higher. The chip upstart expects the first CS-4-based systems to come online later this quarter. ®

Cerebras launches new server chip and system designed to speed AI chatbots

SAN FRANCISCO, Aug 18 (Reuters) - Cerebras Systems (CBRS.O), opens new tab announced on Tuesday a new version of its server hardware that includes its dinner-plate-sized chips that it says will speed AI chatbot queries. Cerebras makes AI hardware and chips that compete with Nvidia (NVDA.O), opens new tab and targets the portion of AI called inference, the computing process of generating an answer in a chatbot such as Anthropic's Claude. The new system, called the CS-4, is a server rack powered by three of the large chips the company designs, which Cerebras said translates into better performance. The ⁠hardware is based around the company's Nexus server architecture, which includes pluggable modules that house the chips. In part, Cerebras' chips gain a speed advantage because they are large enough to avoid the energy and slowdown of moving data from one chip to another. The new machine is available in the third quarter and the chips are fabricated with the TSMC 5-nanometer manufacturing process, the company said. Cerebras also designed the new system to be easier to set up with 50% fewer components, which Chief Technology Officer ⁠Sean Lie said at a media briefing in San Francisco would speed data center construction. The server rack Cerebras designed includes a chip called WSE-3 Turbo and new networking components that Lie said would speed data movement between the chips. Cerebras plans another generation of ⁠the chip and server in 2027. The company expects to deliver 600 megawatts' worth of computing power by the end of 2027, CEO Andrew Feldman said. The company's engineering plans are focused on speeding the ⁠amount of data its future chips and systems can crunch. "We're going to get four times as fast between now and the end of the year, end ⁠of 2027, and we're going to get 20 times more throughput," Feldman said at the briefing. Last week, Cerebras reported an adjusted loss of $6.9 million on sales of $180.1 million. Reporting by Max A. Cherney in San Francisco; Editing by Lisa Shumaker Our Standards: The Thomson Reuters Trust Principles., opens new tab * Suggested Topics: * Technology Max A. Cherney Thomson Reuters Max A. Cherney is a correspondent for Reuters based in San Francisco, where he reports on the semiconductor industry and artificial intelligence. He joined Reuters in 2023 and has previously worked for Barron's magazine and its sister publication, MarketWatch. Cherney graduated from Trent University with a degree in history.

Cerebras launches the CS-4, its first multi-wafer system, though the chip inside is not new

Cerebras has put three of its dinner-plate-sized processors into a single rack for the first time. The CS-4, unveiled on Tuesday at the company's Supernova event and shipping this quarter, is pitched as an inference machine for frontier models, and Cerebras says it runs them up to 30 times faster than GPU-based systems. The launch lands five days after OpenAI's Ultrafast mode, which runs GPT-5.6 Sol roughly 14 times faster on Cerebras silicon, went live. It is also the first hardware the company has shipped since its $5.55bn Nasdaq debut in May, the largest US tech listing since Snowflake. On paper the system is formidable. Each CS-4 carries three WSE-3 Turbo wafers for a combined 750 petaflops of sparse FP16 compute, 129.6 petabytes per second of memory bandwidth, and support for models above 50 trillion parameters. Wafer-to-wafer latency falls to two microseconds from five, and the rack uses half as many components as its predecessor. Cerebras has moved power conversion, in its own phrasing, a hundred times closer to the processors, mounting it in a removable backpack at the rear of the chassis. What is less clear is whether the chip inside is actually new. The WSE-3 Turbo carries the same four trillion transistors, the same 900,000 cores, the same 44GB of on-chip SRAM, and the same TSMC 5nm node as the WSE-3 it replaces. The Register concluded it is not new silicon so much as the existing die pushed from roughly 1.4GHz to 2.8GHz. Per-wafer compute and bandwidth both exactly double, which is the signature of a clock bump rather than a redesign, and Cerebras has a genuinely new generation scheduled for 2027 anyway. The 30-times claim measures tokens per second per user on a single model, gpt-oss-120b, against unnamed GPU systems, and the 250 petaflops per wafer is a sparse FP16 number against 25 petaflops dense. Where the design does look genuinely differentiated is power. The Register puts the CS-4 at an estimated 120 to 140 kilowatts per rack, roughly half what comparable AMD and Nvidia rack systems draw, which gives the claimed tenfold gain in throughput per watt somewhere plausible to start from. Andrew Feldman, the chief executive, framed the launch around latency rather than raw throughput. "In AI, speed is productivity," he said in the announcement, and he told Reuters the company expects to get "four times as fast between now and the end of 2027, and 20 times more throughput". Sean Lie, the chief technology officer, tied the speed argument to agents rather than chatbots. Being 30 times faster, he said, "gives an agentic system room for more than an order of magnitude as much reasoning, verification, or tool use", which is the part of the pitch aimed at enterprise buyers rather than benchmark tables. Cerebras named OpenAI, G42, MBZUAI, and AWS alongside the launch, though it disclosed no CS-4 customer agreements and no pricing. AMD's Helios rack, unveiled in July, sits in the partner list, which is consistent with a company that has said it will work with everyone in AI hardware except Nvidia. The financial backdrop is mixed. Second-quarter revenue came in at $180.1m, up 74% year on year with cloud revenue nearly quadrupling, but it fell sequentially from $193.4m in the first quarter, and the margin pressure Cerebras flagged in June has not lifted. The quarter produced a GAAP net loss of $450.5m, an adjusted loss of $6.9m, and $25.4bn in remaining performance obligations. Cerebras guides to $880m to $890m of full-year revenue and 600 megawatts of data centre capacity live and under contract by the end of 2027. Concentration remains the structural question. G42 and the Mohamed bin Zayed University of Artificial Intelligence together accounted for around 86% of 2025 revenue, and the OpenAI contract signed in January, worth more than $10bn at signature, is the company's answer to that, alongside second-quarter additions including Cognition, Lovable, CrowdStrike, Block, and Figma. First CS-4 shipments are due before the end of the quarter. The harder test comes in 2027, when the next generation will have to arrive on new silicon rather than a faster clock.

Cerebras CS-4 server system claims 30x faster AI inference

Cerebras Systems introduced the CS-4 on Tuesday, a new rack-scale server system the company says delivers up to 30 times faster AI inference than GPU-based systems and twice the speed of its predecessor, the CS-3. The CS-4 is built around three of Cerebras' newly announced Wafer Scale Engine 3 Turbo processors and is the first product on the company's Nexus platform architecture. The system delivers 750 petaflops of AI compute, 129.6 petabytes per second of memory bandwidth, and 7.2 terabits per second of I/O bandwidth, the company said. In a benchmark on a 120-billion-parameter model, the system produced more than 4,400 tokens per second per user, the company said.

Cerebras launches CS-4 AI accelerator, boasting 30x speed over GPUs

Cerebras Systems unveiled the CS-4, a next-generation rack-scale AI accelerator, at its Supernova event in San Francisco on August 19, 2026. The company claims the CS-4 delivers up to 30 times faster inference than GPU-based solutions, positioning it as a significant competitor to Nvidia in the AI inference market. The CS-4 incorporates three WSE-3 Turbo processors, each offering 250 petaFLOPs of AI compute and 43.2 petabytes per second of memory bandwidth. This configuration allows the CS-4 to achieve a total of 750 petaFLOPs of compute and 129.6 petabytes per second of memory bandwidth. The system reportedly processes over 4,400 tokens per second per user on GPT-OSS-120B, in stark contrast to approximately 350 tokens per second on the fastest GPU-based inference service. The CS-4 introduces a modular "backpack" architecture, which reduces the number of components by 50 percent compared to its predecessor. This design aims to simplify deployment and accelerate data center construction. Additionally, the system features switchless chip-to-chip connections with a 2D torus topology, lowering interconnect latency from five microseconds to two. First shipments of the CS-4 are scheduled to begin later this quarter. Cerebras CEO Andrew Feldman emphasized the importance of speed in AI, stating, "In AI, speed is productivity." The company also aims to deliver 600 megawatts of computing power by the end of 2027. Prior to the CS-4 launch, Cerebras reported an adjusted loss of $6.9 million on revenue of $180.1 million.

Cerebras Systems: Cerebras launches new server chip and system designed to speed AI chatbots

Cerebras makes AI hardware and chips that compete with Nvidia and targets the portion of AI called inference, the computing process of generating an answer in a chatbot ⁠such ⁠as Anthropic's Claude. Cerebras Systems announced on Tuesday a new version of its server hardware that includes its dinner-plate-sized chips that it says will speed AI chatbot queries. Cerebras makes AI hardware and chips that compete with Nvidia and targets the portion of AI called inference, the computing process of generating an answer in a chatbot ⁠such ⁠as Anthropic's Claude. The new system, called the CS-4, is a server rack powered by three of the large chips the company designs, which Cerebras said translates into better performance. The hardware is based around the company's Nexus server architecture, which includes pluggable modules that house the chips. In part, Cerebras' chips gain a speed advantage because they are large enough to avoid ⁠the energy and slowdown of moving data from one chip to another. The new machine is available in the third quarter and the chips ⁠are fabricated with the TSMC 5-nanometer manufacturing process, the company said. Cerebras also designed the new system to be easier to set up with 50% fewer components, which Chief Technology Officer Sean Lie said at a media briefing in San Francisco would speed data center construction. The server rack Cerebras designed includes a chip called WSE-3 Turbo and new networking components that Lie said would speed data movement between the chips. Cerebras plans another generation of the chip and server in 2027. The company expects to deliver 600 megawatts' worth of computing power by ⁠the end of 2027, CEO Andrew Feldman said. The company's engineering plans are focused on speeding the amount of data its future chips and systems can crunch. "We're going to get four times as fast between now and the end of the year, end of 2027, and we're going to get 20 times more throughput," Feldman said at the briefing. Last week, Cerebras reported an adjusted loss of $6.9 million on sales of $180.1 million.

Cerebras CS-4 Generates In 1 Second What A GPU Rack Needs 30 Seconds For, Powered By 4-Trillion-Transistor WSE-3 Turbo

Cerebras, the creators of the wafer-scale engine chip, have released their latest CS-4 solution that packs its brand-new WSE-3 Turbo chip. Cerebras CS-4 Is Powered By Wafer-Scale Chips, Packing Up To 250 PFLOPS Per Wafer & Double The Bandwidth of 43.2 PB/s Using 44 GB SRAM Back in 2024, Cerebras unveiled its 3rd Gen Wafer Scale Engine, or WSE-3. This wafer acts as a single chip and offers lots of compute capabilities. Cerebras has been doing Wafer-Scale chips since their advent, and today, the company unveiled the next chapter in its wafer-scale journey. "In AI, speed is productivity," said Andrew Feldman, CEO and co-founder of Cerebras. "Historically, fast inference meant using smaller and less capable models. Cerebras CS-4 delivers industry-leading speeds on the largest frontier models, fundamentally changing the paradigm. Every aspect of the design has been optimized to deliver the highest speeds with massive throughput. With the CS-4, AI is so fast that it fundamentally reshapes product experiences." Starting with this new chip, the WSE-3 Turbo, or WSE-3T, is a follow-up to the WSE-3. It retains the "largest AI processor" positioning and also packs the same number of transistors and cores, at 4 trillion and 900,000, respectively. The entire chip measures 46,225mm2 and also comes with 44 GB of SRAM, which is directly integrated on the same wafer. In terms of some high-level performance aspects, the WSE-3T is said to offer double the compute in AI with 125 PFLOPs (250 PFLOPS with sparsity) per wafer, and bandwidth increases to 43.2 PB/s. The on-chip fabric offers 53.5 PB/s of bandwidth while the off-chip I/O offers 2.4 Tb/s bandwidth. Updates to the wafer-scale engine also help slash latency from 5ms to 2ms. That's all for the chip that is being disclosed for now, but today's AI is different. Chips are a vital component, but AI customers don't go after chips; they go after systems and a one-stop solution. As such, Cerebras is outfitting its latest CS-4 rack-scale solution with the WSE-3T chips. According to Cerebras, CS-4 is built on its new Nexus Platform Architecture, which centers around a modular concept that includes three layers: Compute, Power, and I/O. Each rack features a pluggable backpack design so that the compute subsystem goes into the rear as a "backpack" that is attached vertically to the power array. Each of these self-contained wafer-scale "backpacks" folds the power conversion, direct liquid cooling, high-speed I/O, and control systems into a compact package. The new design also helps eliminate power losses almost entirely by moving power conversion closer to the wafers. Since losses are reduced, more power can be directed to the WSE chips, & that results in 2x power for the engine to feed on, enabling higher frequencies & faster compute. The company is also claiming some big compute throughput capabilities in models such as GPT-OSS 120B. A single CS-4 rack offers over 4,400 tokens per second in this model. A GPU-based AI solution was also compared. What took CS-4 one second to generate, the GPU rack took 30 secs. Overall, Cerebras is claiming a hefty 10x increase in throughput per watt for CS-4 versus CS-3. Each CS-4 rack will offer three WSE-3T (Turbo) chips, pumping out 750 PFLOPs of AI compute, 7.2 Tb/s of I/O bandwidth, and 129.6 PB/s of SRAM bandwidth. With the new interconnect solution, AI factories can scale CS-4 to large-scale clusters to support models with over 50 trillion parameters. The first CS-4 shipments with WSE-3T are expected to begin this quarter, and NVIDIA is one of the main rivals these are meant to tackle. Cerebras has also partnered with AMD to utilize its wafer-scale rack solutions alongside the Helios AI rack to drive higher AI throughput, similar to how NVIDIA utilizes Groq 3 LPX racks to offload the KV cache to the large SRAM housed on these chips. Follow Wccftech on Google to get more of our news coverage in your feeds.

Cerebras Extends the Lead In Fast Inference Space with New Launch: Analyst - Cerebras Systems (NASDAQ:CBR

Cerebras Systems Inc. (NASDAQ:CBRS) on Tuesday unveiled the CS-4, a rack-scale AI system powered by three of its Wafer Scale Engine-3 Turbo chips and designed to speed up response generation from AI models. CEO Andrew Feldman said Cerebras expects to deliver 600 megawatts of computing capacity by the end of 2027. He also expects performance to improve fourfold and throughput to increase 20 times by then. Analyst View Following the news, Needham analyst Quinn Bolton reiterated a Buy rating and maintained a price forecast of $300. The analyst noted that CS-4 also delivers a 10x improvement in throughput per watt compared with CS-3 and is powered by three WSE-3 Turbo chips, which provide twice the PFLOP performance of the existing WSE-3 chip. Latest Private Market Opportunities Join 400,000+ Investors CS-4 is expected to begin shipping this quarter, while CS-5 is planned for 2027, added the analyst. Bolton wrote that the new "backpack" subsystem is designed to enable faster and more cost-efficient deployment at scale by separating the power supply from the compute system and allowing the power supply to attach vertically to the power array. The analyst said that Cerebras redesigned its compute subsystem with a rear-mounted "backpack" that vertically connects to the power array. The self-contained Wafer Scale Backpack integrated power conversion, direct liquid cooling, high-speed I/O and control electronics around the wafer. By separating compute from the power supply, the design simplified manufacturing and, according to Cerebras, cut deployment time from days to hours. It also used 50% fewer components and 60% more automated manufacturing than the prior-generation system, supporting lower-cost and more scalable production, added the analyst. Recent Earnings Snapshot Cerebras reported a second-quarter adjusted loss of 4.5 cents per share, narrower than the analyst estimate for a loss of 17 cents per share. Meanwhile, sales of $180.11 million fell short of the $194.20 million consensus estimate. Core revenue reached a record $209.9 million, up 103% year over year and above the company's guidance. Core gross and operating margins also exceeded expectations. For the third quarter of 2026, Cerebras expects core revenue of $214 million to $216 million, a gross margin of 38% to 40% and an operating margin of negative 25% to negative 23%. Cerebras raised its fiscal 2026 outlook to revenue of $880 million to $890 million, a gross margin of 41% to 43% and an operating margin of negative 19% to negative 17%. CBRS Stock Price Activity: Cerebras Systems shares were down 4.25% at $210.66 at the time of publication Wednesday. Tech Cathie Wood Says Elon Musk's Tesla Playbook Explains Why She's Avoiding SK Hynix and Micron Cathie Wood explains why Elon Musk's Tesla strategy shapes her cautious view on SK Hynix, Micron and the AI memory boom. 3 min read Read this article Photo via Shutterstock Market News and Data brought to you by Benzinga APIs To add Benzinga News as your preferred source on Google, click here.

Cerebras CS-4: Up To 30X Faster Than Nvidia GPUs - Cerebras Systems (NASDAQ:CBRS)

Cerebras Systems Inc. (NASDAQ:CBRS) unveiled a new AI system Tuesday that it says can generate answers up to 30 times faster than GPU-based alternatives, taking another shot at Nvidia Corp. (NASDAQ:NVDA) in the fast-growing market for AI inference. The new CS-4 hit speeds of more than 4,400 tokens per second per user on GPT-OSS-120B in a company benchmark, roughly doubling the per-user speed of its predecessor. The first systems are expected to come online this quarter. Same Silicon, More Speed The CS-4 still uses the same underlying 5-nanometer WSE-3 silicon as its predecessor. SemiAnalysis says Cerebras roughly doubled clock speeds by increasing power delivery and improving cooling. Cerebras also takes a different approach from Nvidia. Instead of connecting thousands of GPUs, it builds a processor roughly the size of an entire silicon wafer, with 900,000 cores and fast on-chip memory. Latest Private Market Opportunities Join 400,000+ Investors That reduces the need to move data between separate chips, one reason Cerebras can deliver unusually fast inference. Is 30X Believable? The claim mainly applies to the kind of ultra-fast inference Cerebras is built for. SemiAnalysis estimates CS-4 could reach about 4,000 tokens per second per user on frontier models, compared with roughly 100 to 200 for Nvidia's Blackwell chips. The performance jump comes partly from running the existing silicon much harder. SemiAnalysis estimates a three-wafer CS-4 rack at roughly 125 to 135 kilowatts and says performance per watt improves only modestly over the previous generation. Even so, The Register notes that is well below the 240 to 250 kilowatt racks Nvidia and AMD are preparing to ship later this year. OpenAI Is Already Testing the Thesis OpenAI is already putting that speed to work. The company last week previewed an Ultrafast tier powered by Cerebras that runs its flagship GPT-5.6 Sol model at up to 750 output tokens per second, up to 14 times faster than standard processing. Cerebras is betting that faster inference can win customers from traditional GPU-based systems. Prediction markets suggest demand for Nvidia compute could remain firm. Kalshi traders put a 64% chance on H200 rental prices ending the year above $6.69 an hour and a 61% chance on H100 prices staying above $3.38. The launch comes during a rough stretch for AI stocks. Cerebras shares fell 12.7% Tuesday, erasing Monday's 15% rally, as surging bond yields pressured high-growth technology stocks. Image: Shutterstock Kalshi and Benzinga have an existing data collaboration agreement. Markets 'Bond King' Jeff Gundlach Compares Nvidia's $500 Billion AI Financing Push to Bonds Backed by Bananas: 'Will Not Age Well' Jeff Gundlach warns Nvidia's $500 billion AI financing push may not age well as Wall Street takes on more debt to fund the AI boom. 2 min read Read this article Market News and Data brought to you by Benzinga APIs To add Benzinga News as your preferred source on Google, click here.

Cerebras's New Server Chip and Rack System Aims to Deliver Max Speed to AI Chatbots

They aim to deliver 600 MW worth of computing power by end-2027 with plans focused on speeding up the amount of data that its future chips and systems can crunch Chip newcomer Cerebras Systems has announced a new version of its server hardware that includes dinner-place sized chips aimed to speed up AI chatbot queries. The next-gen Wafer Scale Engine (WSE) and Nexus rack systems also seeks to extend their lead bandwidth provision by boosting throughput per Watt by tenfold over the previous generation. The company says that it accomplished these feats in a couple of ways with the primary solution coming from squeezing its chips for every hertz they've. The company's new WSE-3T has a turbo mode that promises twice the compute, memory fabric and I/O bandwidth of its earlier model that released two years ago christened WSE-3, a company statement says. Having posted lower than expected losses over sales worth $180.1 million last quarter, the company's move is likely to bolster investor sentiment, given that Cerebras makes AI hardware and chips that compete with Nvidia and targets the inference part of the AI ecosystem - the compute process working on generating answers when users query the chatbot. The company revealed that a major step forward was achieved because their chips are now large enough to avoid the energy loss and slowdown involved in moving data from one chip to another. The company said the new machines would become available in the third quarter. The chips are being fabricated with the TSMC 5-nm process. Another innovation appears related to power delivery whereby Cerebras is capable of pushing twice the power through the chip that enables higher operating frequencies and faster token generation. Experts believe that they could be running the chip at 2.8 GHz as against half that number in the previous generation. In addition, Cerebras has also designed the new system in a way that it makes it easier to set up with 50% fewer components, Chief Technology Officer Sean Lie told the media in San Francisco after the announcement of the new releases last night. As for the server rack, its designs include the WSE-3T and new networking components that would speed data movement between the chips, Lie said, noting that the company already has plans for an upgrade of the chip and server in 2027. CS-4 is the first iteration of the new Cerebras Nexus Platform Architecture. It is built around a modular concept with three foundational elements: Compute, Power and I/O. Cerebras brings significant innovation to each element. Modularity enables each element to scale independently so innovations get to market faster. The modular architecture also supports extremely rapid deployment and upgrades, the statement notes. CEO Andrew Feldman noted that the company expects to deliver 600 MW worth of computing power by the end of 2027 with its engineering plans focused on speeding up the amount of data that its future chips and systems can crunch. "We're going to get four times as fast between now and the end of the year, end ⁠of 2027, and we're going to get 20 times more throughput," he said.

Cerebras powers OpenAI's GPT-5.6 Sol ultrafast mode at 750 tokens/sec By Investing.com

SUNNYVALE, Calif. - Cerebras Systems (NASDAQ:CBRS) announced today that it is powering Ultrafast mode, a new service tier in the OpenAI API for GPT-5.6 Sol, according to a press release statement. The service, available initially in limited preview to OpenAI customers, runs GPT-5.6 Sol at up to 750 output tokens per second and up to 14 times faster than Standard processing. The company states that GPT-5.6 Sol Ultrafast maintains the same intelligence as GPT-5.6 Sol Standard. "GPT-5.6 Sol on Ultrafast is proof that speed and intelligence are no longer mutually exclusive," said Andrew Feldman, CEO and co-founder of Cerebras. Sachin Katti, VP Compute Strategy & GPT-Infra at OpenAI, said the companies are "exploring what becomes possible when customers can get the intelligence of our most capable models with significantly lower latency." OpenAI plans to start with a small group of customers to learn where the speed creates value before expanding the service. Based on output speeds for Anthropic models reported by Artificial Analysis, Ultrafast is 5 times faster than Claude Opus 4.8 in Fast mode and 11 times faster than Claude Fable 5. On Humanity's Last Exam, a 2,500-question benchmark spanning graduate-level subjects, GPT-5.6 Sol Ultrafast answered the full question set in just over 11 hours, compared to more than three days for Claude Fable 5. On GDP-Val, a benchmark of knowledge-work tasks, Ultrafast delivered a 5.6 times end-to-end speedup. The speed comes from Cerebras' Wafer-Scale Engine architecture, which keeps model weights on-chip with 44 GB of SRAM on each wafer-sized chip, eliminating the memory-bandwidth bottleneck that constrains frontier-model inference speed on conventional hardware. This article was generated with the support of AI and reviewed by an editor. For more information see our T&C.

Cerebras' CS-4 Launch Reinforces Thesis About Differentiation Company Architecture Offers for Fast Inference, UBS Says

Cerebras Systems Inc. is an artificial intelligence (AI) infrastructure company that designs and manufactures an AI compute platform comprised of proprietary systems and software. The Company's products include inference Cloud, Training Cloud, CS-3 system, AI supercomputer, Wafer Scale Engine and model development. The Company's pioneering Wafer-Scale Engine (WSE), a chip encompassing an entire silicon wafer, was specifically designed to enable higher performance and speeds than GPUs for the computational demands of inference, Generative AI (GenAI), and other AI applications. It offers deployment services to assist customers with data preparation, model architecture design, training management, inference optimization, and, in select cases, ongoing system operations and management. It also offers a subscription service providing access to an ongoing stream of software updates and upgrades for purchasers of its hardware.

Cerebras Systems Unveils CS-4 AI Accelerator

Cerebras Systems introduced the Cerebras CS-4, the fastest AI accelerator in the industry. The CS-4 is a rack-scale solution built from three new Wafer Scale Engines and revolutionary rack and system designs. The CS-4 is the first member of the next-generation Cerebras Nexus rack-scale platform architecture. It is up to twice as fast as the CS-3, bringing the CS-4s advantage in tokens-per-second-per-user over GPUs to up to 30x more. The CS-4 solution also delivers up to 10x more throughput per watt than the CS-3, vastly improving data center economics. The CS-4 delivers both higher-value tokens and more total tokens within a given power budget?enabling datacenters to be vastly more profitable. The rack scale CS-4 is built from three of the newly released Wafer Scale Engine 3 Turbo (WSE-3T). The CS-4 delivers 750 PFLOPs of AI compute, 7.2 terabits per second of I/O, and 129.6 petabytes per second of memory bandwidth. Total compute fabric bandwidth jumps to 160.5 petabytes per second, and wafer to wafer latency drops as low as two microseconds, enabling the creation of very large clusters and the support of models with over 50 trillion parameters. The CS-4 sets a new high watermark for inference speed. 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. CS-4 is powered by the newly announced WSE-3 Turbo (WSE-3T). Like the WSE-3, the WSE-3T is the largest AI processor ever built, containing four trillion transistors and 900,000 AI-optimized cores across 46,225 square millimeters of silicon, with 44GB of SRAM integrated directly on the wafer. The WSE-3T doubles AI compute to 250 PFLOPS per wafer and doubles memory bandwidth to 43.2 petabytes per second. The on-chip fabric bandwidth and off-chip I/O both double to 53.5 petabytes per second and 2.4 terabits per second respectively. I/O latency shrinks from five microseconds to as low as two microseconds. CS-4 is the first iteration of the new Cerebras Nexus Platform Architecture. It is built around a modular concept with three foundational elements: Compute, Power and I/O. Modularity enables each element to scale independently so innovations get to market faster. The modular architecture also supports extremely rapid deployment and upgrades. Cerebras has fundamentally re-imagined the compute subsystem into a rear mounted ?backpack? that attaches vertically to the power array. Each Wafer-Scale Backpack is 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. By decoupling compute from the power supplies, the Wafer-Scale Backpack simplifies manufacturing and reduces deployment time from days to hours. Compared with the prior-generation system, the Wafer-Scale Backpack has 50% fewer components and uses 60% more automated manufacturing. The Nexus Platform Architecture drives significant power delivery improvements. By moving power conversion 100x closer to the processors ? from roughly 50 millimeters away from the processor as on conventional GPU boards to approximately 0.5 millimeters ? CS-4 nearly eliminates board-level power loss. This delivers twice as much power to the WSE-3T, enabling higher operating frequencies and faster token generation. CS-4 introduces a new programmable I/O subsystem that supports two connectivity modes while doubling I/O bandwidth and reducing latency. The fully programmable Wafer I/O Module supports standards-based RoCE v2 RDMA over Ethernet for seamless integration into existing infrastructure and with an ecosystem of heterogeneous systems. Aggregate off-wafer bandwidth doubles to 2.4 terabits per second per wafer and 7.2 terabits per CS-4 rack solution. The Wafer I/O Module also supports a new communication mode, called Direct Wafer Links, which enables wafers to be linked within and across racks without a switch. Direct Wafer Links enables wafer-to-wafer latency as low as two microseconds. This low latency communication will allow the creation of massive CS-4 clusters and the ability to support models with more than 50 trillion parameters. Programmable low latency I/O is particularly beneficial for heterogeneous disaggregated inference, in which a purpose-built prefill engine processes an incoming prompt and then hands it to Cerebras for ultra-low latency decode. The combination of programmability, standards-based interfaces, and very low latency will allow the rapid creation of disaggregated solutions from different Cerebras ecosystem partners, like AMD Helios and AWS Trainium. Metric comparison: CS-3 (one wafer) vs. CS-4 (3 wafers): AI compute: 125 PFLOPS vs. 750 PFLOPS; Memory bandwidth: 21.6 PByte/s vs. 129.6 PByte/s; On-chip fabric bandwidth: 26.7 PByte/s vs. 160.5 PByte/s; System I/O bandwidth: 1.2 Tbit/s vs. 7.2 Tbit/s; I/O latency: 5 microseconds vs. 2 microseconds. First CS-4 shipments begin this quarter. Full system specifications are available in the CS-4 datasheet.

Cerebras Systems Launches New CS-4 AI Accelerator

Cerebras Systems Inc. is an artificial intelligence (AI) infrastructure company that designs and manufactures an AI compute platform comprised of proprietary systems and software. The Company's products include inference Cloud, Training Cloud, CS-3 system, AI supercomputer, Wafer Scale Engine and model development. The Company's pioneering Wafer-Scale Engine (WSE), a chip encompassing an entire silicon wafer, was specifically designed to enable higher performance and speeds than GPUs for the computational demands of inference, Generative AI (GenAI), and other AI applications. It offers deployment services to assist customers with data preparation, model architecture design, training management, inference optimization, and, in select cases, ongoing system operations and management. It also offers a subscription service providing access to an ongoing stream of software updates and upgrades for purchasers of its hardware.

Cerebras launches new server chip and system designed to speed AI chatbots

Cerebras Systems Inc. is an artificial intelligence (AI) infrastructure company that designs and manufactures an AI compute platform comprised of proprietary systems and software. The Company's products include inference Cloud, Training Cloud, CS-3 system, AI supercomputer, Wafer Scale Engine and model development. The Company's pioneering Wafer-Scale Engine (WSE), a chip encompassing an entire silicon wafer, was specifically designed to enable higher performance and speeds than GPUs for the computational demands of inference, Generative AI (GenAI), and other AI applications. It offers deployment services to assist customers with data preparation, model architecture design, training management, inference optimization, and, in select cases, ongoing system operations and management. It also offers a subscription service providing access to an ongoing stream of software updates and upgrades for purchasers of its hardware.

Cerebras Systems Powers OpenAI's GPT-5.6 Sol Ultrafast Mode

Cerebras Systems Inc. is an artificial intelligence (AI) infrastructure company that designs and manufactures an AI compute platform comprised of proprietary systems and software. The Company's products include inference Cloud, Training Cloud, CS-3 system, AI supercomputer, Wafer Scale Engine and model development. The Company's pioneering Wafer-Scale Engine (WSE), a chip encompassing an entire silicon wafer, was specifically designed to enable higher performance and speeds than GPUs for the computational demands of inference, Generative AI (GenAI), and other AI applications. It offers deployment services to assist customers with data preparation, model architecture design, training management, inference optimization, and, in select cases, ongoing system operations and management. It also offers a subscription service providing access to an ongoing stream of software updates and upgrades for purchasers of its hardware.