After processing the offload on Cerebras this is what I concluded:
What you’re highlighting here is the real Cerebras story — not the tech, but the scalability physics.
The TSMC lock‑in, the non‑portable reticle stitching, and the fact that the OpenAI deal is effectively a cloud‑service commitment (not a hardware sale) all point to the same thing:
Cerebras isn’t fighting Nvidia or Groq — it’s fighting its own ability to scale a custom architecture.
Interesting take on Semtech. I glanced at their DesignCon notes for 200G LPO and it appears they are doing some clever tricks to co-optimize the host Tx FIR and module CTLE. I'm not completely sure how the overall power math plays out, but it seems like some of the processing/power burden is pushed to the host DSP. Its still promising though.
Loved the irrational interview and this is such a useful follow up Vik. Would love the deep dive on Semtech. Esp if you think it will be bottleneck or have a big expansion in demand from here
Hi Vik, Great blog from you as always. From what i understand, the active copper cable (ACC) semtech product will have higher BER (bit-error rate), lower reach compared to AEC since they don't do any DSP/active equalization in TX/RX compared to Credo, Marvell etc. Might be best to compare both the products on these aspects as well instead of power comparison alone I feel.
Expanding further on the systems‑level tension in Cerebras’ model.
Nvidia and Groq scale because their ecosystems compound — modular hardware, portable software, interchangeable supply chains. Cerebras is the opposite: every layer is bespoke. Silicon, packaging, interconnect, power delivery, even the operational model.
That creates performance upside, but ecosystem drag.
A custom stack doesn’t just have to work — it has to scale across fabs, customers, and operational realities.
The real question isn’t whether wafer‑scale works.
It’s whether a custom architecture can ever become a default in a world where modular ecosystems always win long‑term.
The GaN vs SiC battle is one of the few bottlenecks that AI progress can't software-patch away.
No amount of model improvement reduces the physics constraints on power conversion efficiency or thermal management. These are materials science problems with decade-long qualification cycles. The hard part is fab capacity, reliability testing and supply chain buildout. That IS the value. There is no shortcut.
This is the purest counterexample to the "AI solves everything" narrative running through public markets. Some bottlenecks stay physical.
I’m not 100% convinced how they can do 200G at 10% of power budget. It’s not like Marvell’s analog guys are sleeping. There is some tradeoff I am not seeing.
After processing the offload on Cerebras this is what I concluded:
What you’re highlighting here is the real Cerebras story — not the tech, but the scalability physics.
The TSMC lock‑in, the non‑portable reticle stitching, and the fact that the OpenAI deal is effectively a cloud‑service commitment (not a hardware sale) all point to the same thing:
Cerebras isn’t fighting Nvidia or Groq — it’s fighting its own ability to scale a custom architecture.
The tech is brilliant.
The business model is the risk.
100%
Interesting take on Semtech. I glanced at their DesignCon notes for 200G LPO and it appears they are doing some clever tricks to co-optimize the host Tx FIR and module CTLE. I'm not completely sure how the overall power math plays out, but it seems like some of the processing/power burden is pushed to the host DSP. Its still promising though.
Mind sharing the designcon note? Want to see more of the actual methods used
Loved the irrational interview and this is such a useful follow up Vik. Would love the deep dive on Semtech. Esp if you think it will be bottleneck or have a big expansion in demand from here
Vik, I would be very interested in reading your organized collection of your thoughts after you complete it. Thank you for sharing.
Thanks. Thats the plan.
Hi Vik, Great blog from you as always. From what i understand, the active copper cable (ACC) semtech product will have higher BER (bit-error rate), lower reach compared to AEC since they don't do any DSP/active equalization in TX/RX compared to Credo, Marvell etc. Might be best to compare both the products on these aspects as well instead of power comparison alone I feel.
Agree. Need to compare BER, reach, cable gauge, materials. All this matters for a 1:1 comparison.
Expanding further on the systems‑level tension in Cerebras’ model.
Nvidia and Groq scale because their ecosystems compound — modular hardware, portable software, interchangeable supply chains. Cerebras is the opposite: every layer is bespoke. Silicon, packaging, interconnect, power delivery, even the operational model.
That creates performance upside, but ecosystem drag.
A custom stack doesn’t just have to work — it has to scale across fabs, customers, and operational realities.
The real question isn’t whether wafer‑scale works.
It’s whether a custom architecture can ever become a default in a world where modular ecosystems always win long‑term.
The GaN vs SiC battle is one of the few bottlenecks that AI progress can't software-patch away.
No amount of model improvement reduces the physics constraints on power conversion efficiency or thermal management. These are materials science problems with decade-long qualification cycles. The hard part is fab capacity, reliability testing and supply chain buildout. That IS the value. There is no shortcut.
This is the purest counterexample to the "AI solves everything" narrative running through public markets. Some bottlenecks stay physical.
Semtech is a sleeper :) Nice to see it start to get rewarded by the market but yes their strength is still lowkey
I’m not 100% convinced how they can do 200G at 10% of power budget. It’s not like Marvell’s analog guys are sleeping. There is some tradeoff I am not seeing.
Looking forward to seeing what you find Vik!