Nvidia reports Wednesday. Marvell follows Thursday. Two fabless designs, one supply chain choke point. The market is watching one number: revenue guidance. I'm watching three others — prepayments, CoWoS allocation, and the quiet rise of custom silicon.
Here's the framework I'm using to decode this week's earnings.
First, the technical baseline. Nvidia's Blackwell B200 is a dual-die design on TSMC's 4NP process. That's a 5nm-class node, not the 3nm the marketing implies. Marvell's custom ASICs — Amazon's Trainium2, Google's Axion — sit on the same 5nm/3nm family. Both companies are fabless, which means neither carries yield risk. They offload that to TSMC. The real bottleneck is upstream: CoWoS advanced packaging.
Blackwell's dual-die architecture requires CoWoS-L, the most complex variant. TSMC's CoWoS monthly capacity was roughly 32,000 wafers at end-2024, with Nvidia taking over half. Expansion to 80,000+ wafers is underway but the equipment lead time runs 12-18 months. That's the constraint. Not EUV. Not yield. Packaging.
This creates an unusual dynamic for a semiconductor company. Nvidia's gross margins sit at 75%, yet its actual supply ceiling is controlled by a packaging line in Taiwan. In my own audit work on staking derivatives and on-chain protocols, I've learned to follow the physical bottleneck. For AI chips, that's CoWoS. The capex intensity of a fabless model is remarkably light — 5-8% of revenue, versus 35-45% for TSMC. But Nvidia's prepayments to TSMC, SK Hynix, and Samsung tell the real story. When prepayments rise sharply, management is signaling demand conviction.
This is the hidden variable I'll be tracking. A large prepayment increase combined with conservative guidance would be a signal that capacity, not demand, is the constraint. That's a bullish setup.
The demand side is where the numbers get interesting. Data center revenue is now 80%+ of Nvidia's total, growing at triple-digit rates. Blackwell orders are backlogged into 2026. This is textbook supply-constrained growth — the kind that builds pricing power.
But here's where I diverge from the consensus narrative. The market is laser-focused on Nvidia's hyperscaler demand. Microsoft, Meta, Google, Amazon. Everyone tracks those capex numbers. The under-watched variable is the 800G and 1.6T interconnect market that Marvell serves. Data center interconnect (DCI) is the lagging indicator for AI infrastructure spend. GPUs get deployed first, networking follows. If Marvell's DCI segment is accelerating, that's confirmation that AI clusters are moving from pilot to production scale.
This is a signal for the broader ecosystem. The GPU numbers are the tide. The interconnect numbers show whether the tide is rising broadly.
The contrarian angle this earnings season centers on the realism of custom ASICs. The narrative is that CSPs will continue to shift to in-house silicon — Amazon Trainium, Google TPU. This is a real threat to Nvidia, but the timeline is consistently overestimated.
CSP custom ASICs require 2-3 years from design to production. The software stack needs time to mature. CUDA has a 10-year head start with over 4 million developers. That's not a moat — that's a fortress.
In my experience with algorithmic trading, when you're the market leader, the biggest risk isn't the challenger — it's your own execution. Nvidia's supply constraints are actually protecting its margins. The pricing power stays in their hands. The question is whether they can ramp CoWoS capacity fast enough to maintain the growth trajectory.
Marvell has a different problem. Its top five customers represent over 60% of revenue. AWS and Google are both partners and potential competitors. If those customers shift more to in-house silicon — and they will — Marvell's margin structure comes under pressure. This is a structural risk, not a cyclical one.
Geopolitics adds another layer of complexity. Export controls on China have reduced Nvidia's Chinese revenue to 15-20% of the total. The company navigates this with compliant chips like H20, but I don't expect a full export license to China in the near term. The AI chip race has become a national security issue. The Middle East is the emerging wildcard — Saudi Arabia and the UAE are building out AI infrastructure, and they're not subject to the same export restrictions.
A geopolitical supply shock is a real tail risk. The Taiwan Strait scenario. TSMC's fabs produce all of Nvidia's advanced chips, all of its CoWoS packaging. There is no short-term alternative. Samsung lacks the yield, Intel lacks the momentum. This is a concentration risk that the market has priced in, but it's not fully hedged.
Now, the financials. Nvidia's ROE is over 100%. ROIC is over 80%, versus a WACC of about 10%. This is extreme value creation. Every dollar of capital generates eight dollars of return. That's what AI monopoly looks like on a P&L.
Marvell, by contrast, has an ROIC of around 8%, below its WACC. The stock trades at 80x trailing earnings — for a company with a lower return on invested capital than the cost of capital. The market is pricing in a future that doesn't yet exist in the fundamentals. I'm not saying Marvell's narrative is wrong — the custom ASIC story is real. But the financials are pricing in flawless execution.
So, what am I looking for in the numbers?
First, Nvidia's revenue guidance. The market is looking for over $50 billion for the next quarter. I'm looking at the rate of growth, not the absolute number. If guidance implies deceleration, that's a signal. If it's accelerating, the stock has more room to run.
Second, Nvidia's prepayments. A significant increase in prepayments to TSMC and SK Hynix is a signal of demand conviction. If they're willing to commit capital, they see demand ahead.
Third, Marvell's DCI growth. This tells us whether AI infrastructure spending is broadening beyond GPUs. If the data center interconnect segment is accelerating, the whole ecosystem is healthy.
And fourth, the CoWoS commentary. If Nvidia mentions supply constraints again, we know the bottleneck persists — and pricing power stays strong. If they say capacity is catching up, margins may be near a peak.
The market is pricing in a flawless execution. Any sign of hesitation, and the multiple compresses quickly. Keep the bias, watch the signals, and respect the risk.
Math doesn't lie. Sentiment does.
