# Cerebras Unveils CS-4 Chip It Claims Is 30 Times Faster Than Nvidia GPUs > Source: > Published: 2026-08-20 09:44:47+00:00 *Cerebras has made a blunt promise with CS-4: much faster AI answers without waiting for Nvidia's next GPU rack to arrive.* Cerebras unveiled the CS-4 at its Supernova 2026 event in San Francisco on August 18, and the company is selling it on one number you can understand without a chip diagram: up to 30 times faster inference than GPU systems on frontier models. The pitch is simple. You ask an AI system a question, and Cerebras wants the reply to arrive quickly enough that the delay stops shaping how you use the tool. Investor's Business Daily reported that the CS-4 uses three wafer-scale engines in one system and delivers 750 petaflops of AI compute, 7.2 terabits per second of I/O bandwidth and 129.6 petabytes per second of memory bandwidth. Cerebras also says the system gives up to twice the speed of CS-3 and up to 10 times more throughput per watt. Those are company claims, so don't treat them as settled law. Treat them as the bar Cerebras has now set for itself. That part matters. Inference isn't only about how much math a chip can do in a brochure. Large language models have to keep reading weights from memory as they produce each token, and that memory movement is where a lot of the wait comes from. GPUs solve the problem with more chips, more networking and more high-bandwidth memory. Cerebras is still making the opposite bet: keep compute and memory close together on a wafer-scale processor and avoid as much of that back-and-forth as possible. No hops, no tax. [Groq raises $650 million for its neocloud second act after selling its soul to Nvidia for $20 billion](https://startupfortune.com/groq-raises-650-million-for-its-neocloud-second-act-after-selling-its-soul-to-nvidia-for-20-billion/) Groq is raising $650 million to fund its pivot from AI chip maker to inference neocloud provider, following a $20 billion LPU licensing deal with Nvidia that paid out shareholders and handed the founding team to its former competitor. Existing investors Disruptive and Infinitum are backstopping the full round. The raise signals durable investor... - [Groq raises 650 million for AI inference neocloud](https://startupfortune.com/groq-raises-650-million-for-its-neocloud-second-act-after-selling-its-soul-to-nvidia-for-20-billion/) - [AI chip startup licensing technology to Nvidia deal](https://startupfortune.com/groq-raises-650-million-for-its-neocloud-second-act-after-selling-its-soul-to-nvidia-for-20-billion/) The company has been working toward this moment for years. In an August 12 earnings release, Cerebras said its cloud revenue nearly quadrupled from a year earlier, that it had 600 megawatts of data center capacity under contract, and that manufacturing capacity would rise more than 10 times in 2026. It also said it had secured more wafer supply from TSMC and didn't rely on HBM, CoWoS packaging or 3-nanometer fabrication, three areas where AI hardware buyers have spent the past year running into shortages. Good timing, in other words. ## The Power Bill Is The Real Audience The 30 times faster headline gets attention, but the power claim is the one data center buyers will read twice. Cerebras is comparing CS-4 with the 240 to 250 kilowatt class of next-generation GPU racks from Nvidia and AMD. That's the number that counts. If the company can deliver materially better tokens per watt, it changes the conversation from chip bragging to site planning, because power availability is now one of the main limits on AI buildouts. Power is the bill. That doesn't mean Nvidia is suddenly weak. Nvidia remains the default choice for training, the software stack is familiar, and buyers know they can get a broad menu of systems, cloud instances and support around its GPUs. You don't displace that with a faster demo. You displace it only when real customers decide the speed gain is worth changing infrastructure, code paths and procurement habits. ## Groq Shows The Other Path Groq chose another route. Bloomberg Law reported this week that Groq raised $350 million at a $3.5 billion valuation in a round led by Disruptive, with Nvidia also set to invest. That came after Groq announced a $650 million raise in June to expand its AI inference cloud, and after its December 2025 licensing agreement with Nvidia. Groq now describes itself as a neocloud for fast inference. The company says it operates 13 data centers across North America, Europe, the Middle East and Asia-Pacific, serves more than five million developers and plans to scale toward 200 megawatts by the end of 2027. That's a different business. Groq is turning speed into cloud capacity. Cerebras is trying to prove its own wafer-scale systems should sit at the center of that same fast-inference market. You can see why Cerebras wants the timing. Mistral has already used Cerebras hardware for Le Chat, where the companies promoted very fast response speeds for users. G42 and Core42 have also been part of Cerebras's buildout in the Middle East. The CS-4 gives the company a cleaner story to take to model makers: if your product depends on quick back-and-forth with users, faster tokens are not cosmetic. [The AI Boom Just Pushed Nvidia's RTX 5090 Past $4,900 on Nvidia's Own Store](https://startupfortune.com/the-ai-boom-just-pushed-nvidias-rtx-5090-past-4900-on-nvidias-own-store/) The RTX 5090 now costs $4,929.99 on Nvidia's own store, more than double its $1,999 launch price, as Samsung, SK Hynix, and Micron divert memory chip capacity to AI data centers. MSI warns gaming hardware prices could rise another 15 to 30% before the year is out. - [why is nvidia RTX 5090 so expensive now](https://startupfortune.com/the-ai-boom-just-pushed-nvidias-rtx-5090-past-4900-on-nvidias-own-store/) - [nvidia RTX 5090 price increase AI demand 2026](https://startupfortune.com/the-ai-boom-just-pushed-nvidias-rtx-5090-past-4900-on-nvidias-own-store/) Nvidia isn't waiting either. Its next racks will arrive with the advantage of scale, developer familiarity and a supply chain built around customers that already buy by the data center. The test is close. Cerebras says CS-4 systems start reaching customers in the third quarter of 2026, which means the 30 times claim should move from stage demo to production traffic soon enough. **Also read:** [The AI Boom Just Pushed Nvidia's RTX 5090 Past $4,900 on Nvidia's Own Store](https://startupfortune.com/the-ai-boom-just-pushed-nvidias-rtx-5090-past-4900-on-nvidias-own-store/) • [CISA Gives Agencies Three Days to Patch a Critical Flaw in Ray](https://startupfortune.com/cisa-gives-agencies-three-days-to-patch-a-critical-flaw-in-ray/) • [How To Evaluate AI Agents Before Production With a Real Eval Harness](https://startupfortune.com/how-to-evaluate-ai-agents-before-production-with-a-real-eval-harness/)