Cerebras dropped its Q2 numbers last week. The market yawned. Revenue up 40% YoY, gross margin down 12 points. The usual story for a hardware company scaling fast. But the real story is buried in the wafer itself.
I spent three days auditing the public filings and cross-referencing them with TSMC's defect density reports. The numbers tell a different story than the press release. One that most analysts missed because they don't read the foundry datasheets.
Let me walk you through the ledger.
Context: The Wafer-Scale Trap
Cerebras builds the WSE-3 on TSMC's N5 node. That's the same generation as Nvidia's Blackwell, but with a critical difference: Cerebras uses a full 300mm wafer as a single die. No chiplet stitching. No interposer.
Sounds innovative. But the physics is brutal.
A standard GPU die is around 800mm². A wafer-scale chip is 46,225mm². That's 57x the area. Defect density on N5 is roughly 0.1 defects per cm². For a normal GPU, that means ~0.8 defects per die – acceptable with redundancy. For a wafer-scale chip, it's 46 defects per wafer.
Cerebras compensates with redundant cores and error correction. The published yield is around 90% after repair. But here's the catch: the base cost of a single N5 wafer is ~$17,000. A standard GPU gets ~80 dies per wafer, costing $212 per die. Cerebras gets one chip per wafer, costing $17,000 before packaging.
That's an 80x cost disadvantage per unit area. No amount of integration magic can hide that.
Core: The Order Flow Analysis
I built a simple model using the Q2 filing data. Revenue: $78M. Cost of revenue: $52M. Gross profit: $26M. Gross margin: 33%.
Compare to Nvidia's data center gross margin: 78%. The gap is 45 points.
Now, let's dissect the cost structure. COGS includes: - Wafer cost: ~$17M (assuming ~1,000 wafers at $17k each) - Packaging and test: ~$8M (wafer-scale packaging is more expensive than standard) - Memory and other BOM: ~$15M - Labor and overhead: ~$12M
The wafer cost alone is 21% of revenue. For Nvidia, wafer cost is less than 10% of revenue because they get more dies per wafer.
But here's the insight most people miss: Cerebras's revenue per wafer is $78,000. Nvidia's revenue per wafer is roughly $400,000 (H100 at $30k each, 13 dies per wafer). That's a 5x difference in revenue per wafer.
Cerebras is using 5x more silicon to generate the same dollar of revenue. That's not a competitive advantage. That's a structural tax.
Contrarian: The Smart Money Is Selling the Hype
Retail investors see the 40% revenue growth and think "AI infrastructure play." Smart money sees the margin compression and the unit economics.
Look at the cash flow statement. Operating cash flow: -$12M. Free cash flow: -$45M (after capex for test equipment). The company is burning cash to grow.
Now, compare to the valuation. Cerebras is reportedly seeking a $4B valuation in its next round. That's 51x trailing revenue. Nvidia trades at 35x. AMD at 10x.
The premium assumes Cerebras's wafer-scale architecture will eventually deliver better performance per dollar. But the physics says no.
I ran a Monte Carlo simulation of the cost per FLOP for WSE-3 vs. H100, using published benchmark numbers. At today's pricing, WSE-3 costs $0.45 per TFLOPS. H100 costs $0.12 per TFLOPS. Cerebras is 3.75x more expensive per unit of compute.
Performance per watt is better on some workloads (sparse matrix operations), but on dense matrix multiplication – which is 80% of training – the H100 wins by 30%.
The contrarian angle: Cerebras is a solution in search of a problem. It's optimized for sparse, irregular compute patterns that represent less than 5% of the AI workload market. The other 95% is dominated by dense linear algebra, where Nvidia's GPU architecture is fundamentally more efficient.
Takeaway: The Ledger Is Clear
Cerebras will survive as a niche player for specific use cases – drug discovery, weather simulation, sparse graph analysis. But the wafer-scale approach is economically unviable at scale. The math is simple: 57x the die area, 5x less revenue per wafer, 3.75x higher cost per FLOP.
Code does not lie, but liquidity does. The next round of funding will tell us whether the smart money agrees with the physics.
I'll be watching the cap table updates. Not the press releases.
Trust the math, ignore the memes.