Ahead of Nvidia (NVDA) earnings this week, the biggest debate is the company's growing role in financing the wider AI buildout.
Financial guarantees and the demand bet
Investors want proof that customers actually want to buy the product, that demand is real, and that Nvidia (NVDA) can build it while working with memory makers, TSMC (TSM), and possibly Intel (INTC) if Intel steps in later for future Rubin chips. They also want to see that Nvidia can afford these commitments, given the circular investing now happening - Nvidia putting money into the same firms that buy its chips.
This is a large bet on growth. If growth does not come, that hurts the company. But no one makes a bet this size without believing AI revenue will keep rising - people buying from OpenAI and Anthropic - and that this demand keeps growing strongly over the next three or four years.
Who controls supply
The real bottleneck for the whole market is TSMC (TSM) and the memory makers. There are three memory makers, and this constraint hits AMD (AMD), Nvidia (NVDA), and anyone trying to build custom chips. If you want to make a chip, TSMC manufactures it and the memory makers sell the memory. Whoever gets the allocation sells whatever they can make.
Nvidia has by far the largest allocation - the most high bandwidth memory, the most TSMC wafers, and the most CoWoS packaging capacity. That makes Nvidia the winner. AMD has an allocation too and sells it in its Helios platform. Someone else can step in if they secure wafer allocation, but everything comes down to getting that allocation, and Nvidia holds the most.
Current bottleneck: HBM4
The main constraint right now is high bandwidth memory 4 (HBM4), which is scaling relatively well. No problems are expected in the projections Nvidia gives - those numbers should be hit. Next year may bring updated guidance on Rubin Ultra. There have been reports of lower yield or different performance around the pooling stack, but that will depend on what Nvidia says. This is unlikely to be a real issue.
Stock reaction and the surprise problem
Nvidia's next-day earnings reaction has been down each of the last four quarters. More telling: one month after earnings, the stock has been down seven of the last eight quarters. That does not mean the numbers are bad. It means the market already understands the story.
Each of the past eight quarters was a blowout beat. If everyone assumes a beat going in, how much surprise is left? An even larger blowout is possible, and that would be surprising, but not enough to make me buy a pile of shares. The stock tends to hype into each earnings launch and fade after. That pattern fits what you would expect with a large retail presence, and retail investors have been more active in the market recently.
Vera Rubin production ramp
Vera Rubin questions will get partial answers, and new questions will follow. The Grock LPX rack is now in full production with VR Rubin. Late last week and over the weekend, VR Rubin production racks came online at Microsoft (MSFT) and OpenAI. Earlier in the year those sites ran pre-production racks to make sure hardware and software were ready; now they are installing the production racks. Manufacturing is working and customers are finally receiving deliveries, which should be a headline topic on the call. Expect updated production timelines that are very likely on time. Knowing the supply chain, there is no sign of major disruption to the current Rubin and VL72 planned to ship through next year.
The real surprise would be a miss
Asked which would be more surprising - Nvidia barely beating or missing, versus a huge beat despite known bottlenecks - the answer is that barely getting by or missing expectations would be the shock. There was a rumored 15% price hike; after that, revenue and growth clearly keep rising in the guidance, since the price increase alone drives it, even with supply chain bottlenecks.
Most of the reported trouble sits around next-gen - the four die, or quad die, Rubin Ultra. Nvidia has said nothing has changed on its side, even as supply chain reports suggest something might have. That is not necessarily bad. If Nvidia scales back the low yield, high performance chip, it can sell far more of a high yield chip that is still very high performance with a slightly lower peak. That delivers greater overall compute capacity on a more manageable timeline - easier for customers to buy and easier to supply the power for. It would be a massive surprise if Nvidia failed to hit expectations.
