Cerebras' trillion-parameter pitch is impressive, but ask for the bill

Is the AI inference market still Nvidia's by default? Cerebras would like you to think not, and its latest engineering post is the strongest argument yet. It's also a reminder that a good argument is not the same as a purchase order.

What Cerebras actually showed

In a post on its own blog, Cerebras explains how it splits a multi-trillion-parameter model across several WSE-3T wafers. Each wafer has 44 GB of on-chip memory. The example is the open source Kimi K3, at roughly 3 trillion parameters. The company says it can serve such models at 1,000–2,000+ tokens per second. That is Cerebras' own figure, not an independent measurement.

The trick is a pipeline. Layers stay put on each wafer, and only activations hop between wafers over Ethernet, at roughly 2 microseconds per hop according to Cerebras. Many requests flow through at once, so no wafer sits idle.

The architecture argument is stronger than the marketing

We think the interesting part isn't the headline speed. It's that the design sidesteps the usual bottleneck. Cerebras argues that generation is limited by how fast weights move from memory to compute, and that keeping them on-chip avoids that traffic. A rival like Groq made the same bet on SRAM, as Spheron's comparison notes, just with far smaller chips. When two different companies bet against HBM, the thesis deserves respect.

It also matters that this is an open model. Cerebras says Kimi K3 scores 88.3% on Terminal-Bench 2.1 against 88.8% for GPT-5.6 Sol, and even beats leading proprietary models on some benchmarks. If frontier-level open weights keep arriving, the hardware that serves them fast becomes the differentiator. That's where Nvidia's software moat may matter less and where specialized silicon can win.

But "20x faster" is a claim, not a quote from your invoice

Look at how the numbers are built. Cerebras' CS-3 versus B200 page claims 21x faster and 32% lower cost. That headline figure rests on a SemiAnalysis benchmark of one model (Llama 3 70B) in one scenario. The page does cite other benchmarks for gpt-oss-120B and Llama 4 Maverick, but they are still a handful of models, and the company's own footnotes admit that speed gains versus GPUs "may vary depending on workload, configuration, date and models being tested."

The Groq comparison shows the same pattern. Spheron found Cerebras' own figure for GPT-OSS-120B at roughly 3,000 tokens per second, while Artificial Analysis measured about 1,700 independently. That's still fast, but it's a very different number. Spheron also notes that Groq is cheaper per million tokens on the two models it tracked, and that neither chip offers self-serve fine-tuning or runs some popular model families. Those are the practical limits of the speed story.

And notice what we didn't find in the new post. It explains the memory math, but we saw no price per million tokens for Kimi K3 and no latency under load. It says throughput per megawatt at equal output speed is the meaningful metric, and we agree. We'd like to see that figure published.

What buyers should do

Treat this as a reason to run a bake-off, not to switch vendors. Ask for time-to-first-token and end-to-end latency at your real prompt lengths and concurrency. Ask for price per million tokens on the exact model you'll deploy. Ask what happens to speed when the context grows toward the 256k requests the post uses in its memory example. Compare against an independent benchmark, not a vendor chart.

Our verdict: Cerebras has earned a place on the shortlist, and Nvidia looks like no longer the only serious answer for inference. But the proof of a challenger is a published cost per token under production load, and we haven't seen it yet. Until it shows up, keep your wallet closed and your benchmarks open.