$CBRS Culper's thesis hinges on wafers wearing out in ~2 years. The industry's toughest hardware qualifiers say otherwise with their own capex: AWS is racking CS-3s in its data centers on Nitro, Gimlet Labs signed ~100MW incl. CS-4, Oracle is bringing it to OCI.
2
206
1/ Culper's case against $CBRS rests on a single load-bearing assumption: that wafer-scale engines fail within ~1.5–2.5 years. Remove it, and the margin recalculation, the OpenAI table, and the "concealment" narrative fall with it. And that assumption is modeled, not measured.
1
1
181
4/ Hyperscalers don't put unqualified hardware on Nitro. And failures inside your own data center can't be hidden from you. "Unreliable hardware masked by accounting" is very hard to square with where this hardware is actually being installed.
1
15
The spotlight isn't on GPT-6 Astra. It's on GPT-6.1 Sol, where near-Astra performance at a fraction of the cost has developers hooked. With Sol Ultrafast reportedly days away, my cautious bet is that Cerebras silicon will be under the hood. The tell will be speed:
2
1
264
if it lands nearer 750 tok/s than 300, that's wafer-scale at work. $CBRS The added sentence is the speed tell from earlier: a number closer to 750 tok/s points to Cerebras, closer to 300 points to GPUs. It gives readers a way to check your prediction
1
234
low-batch GPU inference is not a real threat. Running GPUs at minimal batch sizes buys speed by giving up throughput and token economics, which is exactly where wafer-scale SRAM architecture is strongest. The 6x price premium on Astra Ultrafast tells you what that trade-off costs
1
39
NVIDIA's own moves make the point. It brought Groq's technology in-house and unveiled LPX to strengthen decode. Even the incumbent recognizes that disaggregated inference, with prefill and decode on purpose-built silicon, is the direction the industry is heading.
1
1
30
To set the record straight: the official NVIDIA confirmation covers GPT-6 Astra Ultrafast running on Blackwell. It says nothing about GPT-6.1 Sol Ultrafast. Those are two different products, and conflating them has driven much of this week's noise around $CBRS.
1
358
Why would low-batch GPU serving of fast models be a problem for Cerebras? The real constraint is WSE capacity, not demand. Running GPUs at low batch sizes is a deliberate trade: you give up throughput to buy speed, which is exactly the bet Cerebras was built on. If anything,
1
93
Beyond its existing partnerships with AWS Bedrock and AMD Helios, Cerebras is steadily expanding its foothold in disaggregated inference through recent collaborations with Gimlet and General Compute. And with GPT-6.1 SOL Ultrafast just around the corner, the next catalyst is
1
83