Convequity
AI Value-Chain & Enterprise Technology Research
Better power delivery appears to be helping Cerebras work through wafer-scale power and thermal limits more than expected. CS-4 gets higher clocks and improved I/O from WSE-3T, despite the same core count and SRAM capacity as WSE-3. Oracle, OpenAI and AWS choosing to work with Cerebras carries weight. But higher clocks do not answer Culper Research's allegation that its wafers fail after about two years; how representative those failures are of current systems remains unclear. x.com
Customer warrants make Cerebras's reported gross margin a poor measure of its compute delivery costs relative to Nvidia's. Cerebras reported 14% in Q2 2026 versus Nvidia's roughly 75%, but 41% after adding back warrants and other items. Cerebras could challenge on selected workloads without matching Nvidia's whole platform. A longer wait for Feynman at scale would give Cerebras more time to win customers, while the 3D-stacked SRAM we expect in Feynman could narrow its fast-inference advantage. x.com
Cerebras looks attractive because the short report's two-year wafer-life thesis rests on unvalidated reliability assumptions, not representative failure data. Energy per useful token, lower warranty utilization and ordinary repair accounting also undercut its financial conclusions. OpenAI has integrated Cerebras into production; AMD and AWS have announced integrations, while Oracle plans 2027 availability for ultra-fast inference, adding commercial validation despite the report. x.com x.com
Vicor's efficient delivery of key power modules is part of the case for Cerebras meeting and exceeding its 2028-29 numbers. As models improve, more security breaches and deployments should make Cerebras look increasingly like a cybersecurity play. Its revenue ramp looks constrained by capacity, not yet by demand. Bullish on Cerebras but cautious on the broader market, with no short-term earnings prediction. x.com x.com x.com x.com
Cerebras's yield problem is not solved, because the fault tolerance it claims comes at a price. Wafer-scale chips need testing, retesting of repairable defects and repair, and long bypass routes around defects can fail higher-speed specifications. @damnang2 lays out the details in a yield model for the WSE-3 built from Cerebras's public disclosures, though by his own account the figures rest on his assumptions, not Cerebras's actual yield. x.com
Oracle's year-long evaluation of Cerebras' costs, performance, customer uses and equipment life strengthens the case for its inference technology. Oracle chose to bring the equipment into its data centers; that real-world buyer decision conflicts with the short seller report's criticism. x.com
Cerebras should dominate the enormous fast-inference market because agentic workflows often run steps sequentially, making each step's speed critical. Its newly announced Oracle partnership brings its ultra-fast inference to Oracle Cloud Infrastructure. Cerebras delivers inference speeds up to 13x faster than Nvidia GPUs, and its chips should eventually reach every major AI cloud. x.com
Cerebras faces brutal competition from SpaceX, and competing in AI models or data centers requires cash at scale. SpaceX is already trashing Cerebras and CoreWeave. x.com
Culper Research's wafer-failure complaints do not justify shorting Cerebras at $100. The customer complaints sound illogical, although quality issues would not be surprising while the chips are in the early-adopter phase. x.com x.com
Cerebras may be the only company willing to stitch together a true high-speed solution. If so, it retains upside optionality from high-speed pricing and gains a first-mover advantage. x.com x.com
Cerebras is not worth owning because its proprietary stack faces limited broad adoption. The attached analysis says NVIDIA-aligned cloud operators are unlikely to deploy Cerebras hardware, while Groq LPUs fit NVIDIA's software ecosystem more easily. It also flags Cerebras' scaling and deployment friction, despite solid product-market fit with marquee customers. x.com
Cerebras is the purest bet on decode as disaggregated inference becomes the next shift in AI model serving. Prefill runs on AMD Helios, Amazon Trainium or Nvidia GPUs, while decode needs extreme on-chip bandwidth, where Cerebras' Wafer-Scale Engine claims 1,000-2,000 times the effective bandwidth of GPU HBM. Splitting the work yields more tokens per watt, lower latency and better economics at scale. It is already shipping in Cerebras' pairings with AMD and Amazon. x.com
