Daily Silicon: Cerebras Hits $180M: Scaling AI Chips Without HBM
A practicing semiconductor engineer's daily read on global silicon news - Cerebras Hits $180M: Scaling AI Chips Without HBM
🆓 무료 글 · 누구나 무료로 읽을 수 있는 글입니다
A practicing semiconductor engineer's daily read on global silicon news:
Cerebras reached $180.1M in Q2 revenue, but the real tests are converting 600MW of inference capacity into revenue and qualifying custom HBM4E at scale.
Chase's Take - I care less about 281% growth than where an architecture without HBM and CoWoS moves the system bottleneck. Here, TSMC wafer supply, data-center power, manufacturing test, and customer deployment determine the ramp instead of packaging slots.
The next checks are Q3 core gross margin of 38-40% and the actual commissioning schedule for the contracted 600MW.
1. Cerebras Growth Is Gated by 600MW Deployment, Not HBM
TL;DR - Q2 GAAP revenue rose 74% to $180.1M, but 14% gross margin and the activation rate of 600MW in contracted capacity will determine the quality of that growth.
Source: Cerebras SEC 8-K - Cerebras Growth Is Gated by 600MW Deployment, Not HBM
2. Valens at 61.5% Gross Margin: The Long Road After an A-PHY Design Win
TL;DR - Q2 revenue of $18.105M and 61.5% gross margin show the strength of the product mix before automotive SerDes design wins convert into production revenue.
Source: Valens Semiconductor SEC 6-K - Valens at 61.5% Gross Margin: The Long Road After an A-PHY Design Win
3. Micron Custom HBM4E: Dual-or-Single Sourcing and the 4:1 Trade-Off

TL;DR - Custom HBM4E and a nearly 4:1 capacity trade-off mean customer-specific logic-die and stack qualification will set available supply.
Source: TrendForce News - Micron Custom HBM4E: Dual-or-Single Sourcing and the 4:1 Trade-Off
4. Japan Manufacturing Sentiment Improves for a Third Month on Chip Demand
TL;DR - The Reuters Tankan improvement shows AI demand spreading into equipment and materials orders, but backlog and utilization remain the operating checks.
Source: Reuters - Japan Manufacturing Sentiment Improves for a Third Month on Chip Demand
What to Watch Next
- Cerebras Q3 core gross margin of 38-40%
This will show whether the contracted 600MW is converting into utilization and margin. - HBM4E customer qualification cycle
It will determine dual-source feasibility and actual 2027 supply. - Valens quarterly automotive revenue
This will show how quickly MIPI A-PHY design wins move into OEM production. - Japan equipment and materials backlog and utilization
These metrics will verify whether stronger business sentiment is translating into actual wafer starts.
Signal Radar — More Headlines
- Transient Bit-Flip Sensitivity in Homomorphic Encryption Computation (arXiv)
- Where Faults Become Critical in CKKS Multiplication (arXiv)
- Switched-Capacitor Analog CIM Architecture (arXiv)
- Dynamic-Shape Matrix Multiplication on Ascend NPUs (arXiv)
- Synthesizing Saturating Counters with Differential Privacy Guarantees (arXiv)
- CNN Inference Cost Prediction and Deployment Screening (arXiv)
- AMD Hires University of Rhode Island Faculty Startup Team (University of Rhode Island)
- Cerebras Files Q2 10-Q (SEC)
- Synopsys Discloses General Counsel Transition Plan (SEC)
- Intel Underwriters Fully Exercise 31.57M-Share Overallotment (SEC)
Further Reading on VLSI Korea
- The Next RTL Simulation Bottleneck Is the CPU Front End: How ARMOR Node Compression Delivered 2.7×
Useful technical background on how system bottlenecks move when the primary constraint is compressed. - When the Wafer Becomes the Computer ⑤: The I/O and Memory Wall
Deeper context on the I/O, memory, and deployment constraints behind wafer-scale computing.
Frequently Asked Questions
What did Cerebras report for Q2 revenue and margin?
Cerebras Q2 GAAP revenue rose 74% to $180.1M, with a 14% gross margin. The quality of that growth depends on the activation rate of 600MW in contracted capacity; the watch point is Q3 core gross margin of 38–40%.
What did Valens report for Q2?
Valens reported Q2 revenue of $18.105M and a 61.5% gross margin. The results show product-mix strength before automotive SerDes design wins convert into production revenue.
What trade-off does custom HBM4E carry?
Custom HBM4E carries a nearly 4:1 capacity trade-off, so customer-specific logic-die and stack qualification will set available supply. Dual- or single-sourcing decisions follow that qualification cycle.
What did Japan's manufacturing sentiment show?
The Reuters Tankan improvement, a third consecutive monthly gain, shows AI demand spreading into equipment and materials orders. Backlog and utilization remain the operating checks.
Related reading
- Daily Silicon: Cerebras Explodes, Trump-Xi Summit Ends Empty-Handed
- Daily Silicon: Intel-Apple Chip Deal Surges 10%, Samsung-SK Cap Gap…
- Daily Silicon: Intel 18A-P Risk Production, SK Hynix Nasdaq ADR, TSMC…
- Daily Silicon: Big Tech Funds EUV Scanners — The Memory Supercycle…
- Daily Silicon: Analyst Consensus Surge vs TSMC April Slowdown,…