Cerebras wafer-scale chip

Long-term owners

cerebras is 10 years down this rabbit hole and poised to dominate

Jun 2026 · comment

What hackers pick instead

• Google TPUGoogle's custom AI chips, rented through Google Cloud, are recommended for cheap, energy-efficient inference and large-scale clusters. Common caveat: hard to program and behind Nvidia at top performance.73recommend
• Taalas HC1Model-in-silicon chip praised for very fast, low-power inference; the main caveat is that the model cannot be updated without replacing the chip.59recommend
• Taalas hardwired LLM chipPeople recommend it for very fast LLM inference from a model etched into silicon. Critics note the chip is locked to one model and updates are limited.45recommend
• Cerebras CS-4People recommend it for very fast AI inference, above 1,000 tokens per second on large models. Critics say it is scarce, hard to access, and not affordable for individuals.12recommend
• Groq LPURecommended for fast, efficient AI inference, using SRAM instead of HBM. Critics point to small per-chip SRAM and hard deployment requirements.9recommend

Every quote

The 40 most detailed of 44 comments. Each links to the original.

1. The tokens per second performance numbers coming from Cerebrus/Talas are several orders of magnitude higher than models running on GPUs

   Recommends · Jul 2026 · in “Kimi K3, Qwen 3.8, and Anthropic's (Potential) Unravelling”

2. With the next generation of Cerebras chips expected to be 5–7× faster, peak throughput could reach the ~17,500 tokens-per-second range.

   Recommends · Feb 2026 · in “The path to ubiquitous AI (17k tokens/sec)”

3. Using a waffer sized chip doesn't sound great from a cost perspective when compared to using many smaller chips for inference.

   Mixed · Feb 2026 · in “GPT‑5.3‑Codex‑Spark”

4. Cerebras did the Really Hard Thing, something many other brilliant companies and people failed at for decades.

   Recommends · Feb 2026 · in “Byte magazine artist Robert Tinney, who illustrated the birth of PCs, dies at 78”

5. Cerebras is kind of insane in that they've built one massive wafer-scale chip with a comparative ocean of SRAM

   Mixed · Feb 2026 · in “Two different tricks for fast LLM inference”

6. Just doing what Cerebras did and making a wafer-sized chip which is mostly SRAM is a much better solution

   Recommends · Aug 2026 · in “AMD acquires Taalas to boost inference performance by etching models in silicon”

7. can give great result in the long run, but at the pace AI is growing it's not very practical yet

   Mixed · Sep 2026 · in “GPU World”

8. The initial cost of serving is very high, and while super performant not great for scaling up.

   Mixed · Apr 2026 · in “Cerebras S-1”

9. I am skeptical that it will be economically useful for models that are larger in size

   Mixed · Feb 2026 · in “GPT‑5.3‑Codex‑Spark”

10. the advantages of running >5x the token throughput per thread can't be overstated.

    Recommends · Jul 2026 · in “Apple Silicon Exec Explains Mac Mini AI Demand and On-Device Future”

11. for what does fit onto a Cerebras wafer or a Taalas chip, the speed is phenomenal

    Recommends · Sep 2026 · in “Kimi K3 (2.8T) at 1 token/s on a MacBook Pro, streamed from four SSDs”

12. Cerebras will be substantially better for agentic workflows in terms of speed.

    Recommends · Feb 2026 · in “GPT‑5.3‑Codex‑Spark”

13. wafer scale is great for very low latency, but loses to throughput per dollar

    Mixed · Apr 2026 · in “Cerebras S-1”

14. takes a lot of cool engineering and cooling to make it work, and is very cool

    Recommends · May 2026 · in “Accelerating Gemma 4: faster inference with multi-token prediction drafters”

15. just need way to many racks to be able to serve the full Deepseek model

    Mixed · Jan 2026 · in “VLLM Large Scale Serving: DeepSeek 2.2k Tok/S/H200 with Wide-EP”

16. Cerebras makes a giant chip that runs inference at unreal speeds.

    Recommends · Nov 2025 · in “Cerebras Code now supports GLM 4.6 at 1000 tokens/sec”

17. It's actually much faster than anything else for real workloads

    Recommends · Feb 2026 · in “GPT‑5.3‑Codex‑Spark”

18. One I like in particular is Cerebras.

    Recommends · Feb 2026 · in “Byte magazine artist Robert Tinney, who illustrated the birth of PCs, dies at 78”

19. much too small for the real 5.3-codex

    Mixed · Feb 2026 · in “Nvidia with unusually fast coding model on plate-sized chips”

20. that chip is really one of a kind

    Recommends · Feb 2026 · in “GPT‑5.3‑Codex‑Spark”

21. I'm a huge fan of their hardware

    Recommends · Jan 2026 · in “VLLM Large Scale Serving: DeepSeek 2.2k Tok/S/H200 with Wide-EP”

22. Yes, Cerebras is very cool.

    Recommends · Feb 2026 · in “Byte magazine artist Robert Tinney, who illustrated the birth of PCs, dies at 78”

23. very excited for cerebras

    Recommends · Feb 2026 · in “Nvidia with unusually fast coding model on plate-sized chips”

24. makes Cerebras rather cost intensive given that Nvidia can get dozens of GPUs out of a single wafer

    Criticizes · Jul 2026 · in “Kimi K3, Qwen 3.8, and Anthropic's (Potential) Unravelling”

25. They don't have it, nor do they have a way to get it since they are last in line at the fabs.

    Criticizes · Feb 2026 · in “GPT‑5.3‑Codex‑Spark”

26. a multi-million dollar machine with ~40GB of SRAM is not a good solution for big LLMs

    Criticizes · Aug 2026 · in “Processing in Memory: DRAM Is About to Do Math”

27. Cerebras is extremely expensive and difficult to scale, hence the limited release.

    Criticizes · Jun 2026 · in “Previewing GPT‑5.6 Sol: a next-generation model”

28. It is not real in the sense that almost nobody has access to them.

    Criticizes · Jan 2026 · in “The Lottery Ticket Hypothesis: finding sparse trainable NNs with 90% less params”

29. It doesn't scale, and won't ever be profitable.

    Criticizes · Sep 2026 · in “MongoDB CEO resigns to join Meta”

30. their approach of just copy/pasting a whole bunch of identical functional blocks, over-provisioning the chip by ~8%, and then just fusing off blocks with defects allows them to effectively have 100% yield

    Recommends · Aug 2026 · in “AMD acquires Taalas to boost inference performance by etching models in silicon”

31. GPUs are highly optimized but they still lose to different architectures that are better suited for inference

    Recommends · Feb 2026 · in “How Taalas "prints" LLM onto a chip?”

32. their packaging does offer inherent advantages in giving you maximal compute without external communication

    Recommends · Feb 2026 · in “Nvidia with unusually fast coding model on plate-sized chips”

33. it will be possible to scale these chips up to the point where a model like full Codex-5.3 can run on them

    Recommends · Feb 2026 · in “Two different tricks for fast LLM inference”

34. they have more options than non-wafer-scale solutions, the chips are going 3d

    Recommends · Sep 2026 · in “MongoDB CEO resigns to join Meta”

35. You should read about the Cerebras WSE configurable colour channel mesh.

    Recommends · Mar 2026 · in “Process-Based Concurrency: Why Beam and OTP Keep Being Right”

36. cerebras is 10 years down this rabbit hole and poised to dominate

    Recommends · Jun 2026 · in “Sophon PFG-1: a monolithic-3D AI ASIC with 330 GB of on-die DRAM and no HBM”

37. from a manufacturing perspective this approach seems cursed

    Mixed · Apr 2026 · in “Cerebras S-1”

38. take a look at Cerebras, who are doing wafer-scale compute

    Recommends · Sep 2026 · in “Kimi K3 (2.8T) at 1 token/s on a MacBook Pro, streamed from four SSDs”

39. Cerebras has already retired their technical risk.

    Recommends · Jun 2026 · in “Sophon PFG-1: a monolithic-3D AI ASIC with 330 GB of on-die DRAM and no HBM”

40. the architecture scales up to 1.6PB RAM

    Recommends · Jul 2026 · in “Apple Silicon Exec Explains Mac Mini AI Demand and On-Device Future”