The semiconductor sector is holding its breath this week as two of the most important AI chip designers report earnings within 48 hours of each other. Nvidia reports on Wednesday. Marvell follows on Thursday. Both companies are fabless, meaning they have no chip fabrication of their own. Both are completely dependent on TSMC for their existence. But the market is not asking the same question of each. The market wants to know if Nvidia can maintain its 90% grip on the AI training market. The market wants to know if Marvell can finally convert its custom ASIC design wins into actual revenue growth. The answer to both questions is locked inside the same physical constraint: TSMC's CoWoS packaging capacity. Ledgers bleed, but code remembers the truth. So let's read the code.
Context: The CoWoS Chokehold
TSMC's CoWoS (Chip-on-Wafer-on-Substrate) advanced packaging is not a technical footnote. It is the single most important constraint on AI chip supply in 2025. Nvidia's Blackwell B200 uses a dual-die design that demands the most complex CoWoS-L packaging TSMC has ever produced. Marvell's custom ASICs for Amazon and Google are equally dependent on the same advanced packaging lines. Here is the math that matters: TSMC's CoWoS monthly capacity was approximately 32,000 wafers by the end of 2024. Nvidia alone takes more than half of that output. TSMC has announced plans to expand this to 80,000 wafers per month by the end of 2025, but equipment delivery cycles run 12 to 18 months. The bottleneck is not going away.
Liquidity is just trust, quantified in gas. In the case of CoWoS, it is quantified in silicon.
Core: What the Numbers Actually Say
Let me walk you through the technical and financial landscape of both companies. This is not a recommendation. This is a forensic review of the data available.
Technical and Process Position
Nvidia is a fabless design house. Its current Hopper H100 and H200 chips use TSMC's 4N process, which is a customized version of the 5nm class technology. Blackwell B200 uses the 4NP variant, also based on the 5nm class. The Rubin platform, scheduled for 2026, is expected to move to TSMC's N3 or N2 process. Marvell's custom ASIC, including Amazon's Trainium 2 and Google's Axion, uses 5nm and 3nm class processes from TSMC. Neither company uses Gate-All-Around (GAA) transistor architecture yet. TSMC's N2, which will be the first GAA node, is not expected to enter production until 2025. The technological gap is approximately one process node. Nvidia compensates with advanced packaging and architectural innovation.
Yield and Packaging
Fabless companies do not directly bear yield risk. The yield risk sits with TSMC. For the 4nm/5nm class nodes, TSMC's yield is already mature, above 90%. But the packaging is the critical bottleneck. The Blackwell B200's dual-die design doubles the complexity of CoWoS packaging. TSMC's capacity expansion is the real story. Every quarter of delay in the CoWoS expansion is a quarter of constrained supply for Nvidia. Based on my audit experience, the CoWoS bottleneck is the single most important factor to watch in both earnings calls.
IP and Ecosystem
Nvidia's CUDA ecosystem is its moat. The hardware is replaceable. The software stack is not. Developers are locked into CUDA through years of accumulated code and training. This is why Nvidia maintains its dominant position. The custom ASIC from Marvell does not have such an ecosystem. It depends on the buyer's willingness to design around its silicon. This is a fundamental difference in business models.
The Market Demand: An Inflated Bubble or a Real Foundation?
AI Training and Inference
AI training chip demand is in a state of excess demand. Nvidia's Blackwell B200 and GB200 orders are booked through the end of 2025. The market believes that AI inference demand will surpass training demand by 2025-2026. This is a critical transition point. Training is a concentrated workload. Inference is a distributed one. If inference demand is truly exploding, it changes the demand profile from a few hyperscalers to a broader market.
The Hyperscaler Spending Spree
The hyperscalers — Microsoft, Meta, Google, Amazon — are projected to spend a combined $300 billion on capital expenditures in 2025. Most of this is directed at AI infrastructure. This is a massive bet. If AI demand growth slows, these companies will face a severe hit to their earnings.
Pricing Power
Nvidia's pricing power is extreme. H200 and B200 GPUs are priced between $30,000 and $40,000 per unit. There is no pressure to lower prices because demand is higher than supply. Marvell's custom ASIC pricing is project-based and yields a lower gross margin of about 40-50%, compared to Nvidia's 70%+.
Inventory Cycle
AI chip inventory is in a restocking phase. Nvidia's inventory turnover is around 60-70 days, which is healthy given the supply shortage. Channel inventory for AI chips is extremely low. This is a sign of a healthy supply-demand imbalance, not a bubble. The imbalance is expected to persist until 2026.
Financial and Valuation Analysis
Nvidia's gross margin is around 75%, up from 65% three years ago. This is a reflection of its dominant pricing power. Marvell's gross margin is around 45-50%, and it is under pressure from the custom ASIC business, which has lower margins.
Nvidia's return on equity is over 100%. Its return on invested capital is over 80%. This is the strongest value creation in the industry. Marvell's ROIC is below its weighted average cost of capital, meaning it is not creating value. This is a red flag.
Nvidia's stock trades at around 50 times trailing earnings. That is high, but the growth justifies it. Marvell trades at 80 times trailing earnings. That's rich, and the growth needs to deliver.
Geopolitical and Export Controls
Nvidia faces export controls on its high-end chips to China. The H100, H200, and B200 are banned from China. Only the reduced-capability H20 and B20 are allowed. China is about 15-20% of Nvidia's revenue. Export controls are reducing this, but growth elsewhere is offsetting the loss. The compliance costs are real. The uncertainty is real. If the controls are further tightened, the impact will be more significant.
Contrarian Angle: What the Market Is Getting Wrong
The market is focused on Nvidia's revenue guidance. If the guidance is above $50 billion for Q1 of FY2026, the market will see it as a confirmation of AI demand. But there is a deeper problem. The market is not paying enough attention to the concentration risk. Nvidia's top five customers, which include Microsoft, Meta, Amazon, Google, and Tesla, account for 40-50% of its revenue. This is a concentrated customer base. If one of these customers reduces their capex, Nvidia will feel it.
The bigger risk is the supply chain. TSMC is a single point of failure. If there is a geopolitical event, or if TSMC's CoWoS expansion is delayed, Nvidia's shipments will be constrained. The market is not pricing this risk.
Marvell's risk is different. It has a high concentration of customers, including AWS and Google, which could account for more than 60% of its revenue. If these customers decide to design their own chips in-house, Marvell will face a sharp drop in revenue. The trend of CSPs designing their own custom chips (Trainium, TPU, Maia) is a long-term threat to both Nvidia and Marvell.
Security is a myth until the bridge breaks. The bridge here is TSMC's CoWoS.
Takeaway: What to Watch in the Earnings Calls
Every exploit is a lesson paid for in ETH. For these earnings, the lesson is in silicon.
Watch Nvidia's revenue guidance. If it exceeds $500 billion for Q1 FY2026, it will confirm that AI demand is not slowing. Watch the gross margin. If it stays above 75%, Nvidia has pricing power. Watch the prepayments. If prepayments increase, it means Nvidia is confident about future demand.
For Marvell, watch the AI revenue share. If AI-related revenue exceeds 30% of total revenue, it will confirm the custom ASIC trend. But also watch the debt. Marvell's net debt-to-EBITDA is 3-4 times, and high interest rates will eat into profits.
The deeper question is: How long can the AI capex cycle last? Hyperscalers are spending $300 billion a year. At some point, they will need to see a return on that investment. If AI revenue does not materialize, the capex will be cut, and the entire AI chip supply chain will be hit.
Yields vanish when the herd arrives at the gate. The herd is already here. The question is whether the gate is open or closed.
Logic cuts through the noise of the bull run. The logic of these earnings is not about the AI story. It's about the physical constraints: CoWoS capacity, HBM supply, and customer concentration. Those are the metrics that will tell us whether the AI story is real or just a dream.