Nvidia & Marvell AI Outlook: The Circuit Breaks Down Earnings and Hot Chips

According to Ben Bajarin and Jay Goldberg on The Circuit, Nvidia’s latest earnings and guidance set the tone for AI infrastructure investing going into 2027–2028, while Hot Chips revealed a confusing but important shift toward bespoke silicon. The hosts frame the discussion around Nvidia’s claim that next year’s revenue can grow about 70%, even as the company remains supply constrained.

Goldberg argues that markets focused less on Nvidia’s quarterly beat and more on this aggressive growth outlook, which runs counter to recent expectations of sharp deceleration. At the same time, Bajarin says Hot Chips underscored how fragmented AI compute strategies have become, with many chip and system vendors pursuing divergent architectures and memory approaches.

The pair repeatedly emphasize that none of this is investment advice and that their goal is to interpret what Nvidia, Marvell, and others actually said. They also note that capacity, balance sheets, and custom silicon are emerging as competitive weapons, rather than just process nodes or raw performance.

Goldberg’s core takeaway from the Nvidia (NVDA) call is that the company guided to around 70% revenue growth for the next year and stressed that it could grow even faster if more manufacturing capacity were available. He notes Jensen Huang repeatedly said Nvidia is sold out for the current year and likely for next year, making TSMC capacity the main swing factor for upside.

Bajarin adds that supply-chain work he follows suggests that meaningful new capacity pockets may not really open until calendar 2028, with 2027 still looking tight across wafers, substrates, and related components. He argues Nvidia is therefore highly incentivized to help Intel ramp its foundry, potentially becoming a “customer zero” similar to how Apple was for TSMC.

On Marvell, Bajarin says management has now “firmly” chosen an XPU attach strategy rather than trying to build full monolithic compute tiles. He views this as making Marvell a companion to many ASIC programs instead of needing to win the whole socket. Goldberg notes Marvell guided to about 60% growth, which is strong in absolute terms but still implies Nvidia is gaining share in some key AI markets.

Goldberg points out that Nvidia’s quarter and near-term guide were essentially in line with expectations and still beat consensus, but investors “expect them to beat expectations,” so the bigger story was the explicit 70% revenue growth outlook for next year. He also highlights commentary that Nvidia could have grown more if it had additional supply, reinforcing the capacity-constrained narrative.

On profitability, Goldberg flags that Nvidia guided gross margins down from about 75% to roughly 72% in its Q4 (the January quarter) and expects margins to remain in the low 70s for three to four quarters. Nvidia attributed this mainly to memory and mix. Goldberg interprets this as evidence that memory suppliers now hold more leverage, and he speculates that hyperscalers may increasingly source memory directly, reducing Nvidia’s ability to mark up that content.

The hosts also discuss reports in The Wall Street Journal about a $500 billion SPV financing platform tied to Nvidia and note that Nvidia outlined roughly $579 billion in commitments and obligations, including large long-term agreements for lasers, substrates, and TSMC capacity. Goldberg says these off-balance-sheet-like obligations and new debt issues make him “nervous,” while Bajarin frames Nvidia’s balance sheet as a deliberate competitive advantage used to backstop Neocloud partners.

For Marvell, Bajarin focuses on the announced $120 billion potential revenue opportunity with Google, stressing that management described this as potential, not guaranteed, and declined to offer a conversion schedule, likely saving detail for an investor day in early October. He also mentions CXL as a modest but positive driver that could add “couples of billions” in revenue over time, though not a dominant piece.

Both hosts emphasize how difficult Nvidia is to model accurately. Bajarin notes that Nvidia can sell complete racks like an NVL72 system to Neoclouds, but also sells stand-alone CPUs, GPUs, and networking gear to hyperscalers, often without bundled memory or PCBs. He argues this mix makes it hard to infer content value or margin just from “racks shipped” estimates.

Goldberg adds that Nvidia’s new data center segmentation—ACIE (enterprise, cloud, sovereign) versus hyperscalers—implicitly signals different margin profiles, with hyperscalers likely paying lower margins than smaller Neoclouds buying full-stack solutions. He says his own Nvidia models are being tweaked to account for these margin differences but remain uncertain.

On financing, Goldberg raises the concern that Nvidia’s support for Neoclouds may resemble “circular financing,” though Nvidia explicitly denied that on the call. He concedes Nvidia’s argument that AI data centers require heavy upfront capex and that these could be viable businesses long term, but he worries that rapidly rising system prices make the economics tough for smaller clouds. Bajarin advises investors not to count on any upside from revenue-sharing structures until they’re clearer, even though the long-term potential could be large.

For Marvell, Goldberg notes the market’s negative reaction—he mentions the stock being down about 10%—and attributes it to Marvell announcing 60% growth just after Nvidia’s 70% outlook, making the relative positioning look weaker despite strong fundamentals.

Bajarin describes Hot Chips as one of the most technically rich conferences he has attended, with engineers from companies like Samsung Electronics presenting to other engineers and even competitors openly asking each other detailed questions. He stresses that recent Hot Chips sessions underscored what he calls an “age of ferment” in AI hardware, where many competing architectures coexist and it is hard to predict which will dominate.

A major theme he highlights is the push toward custom memory-on-base-die designs. Bajarin says “everybody walked away” convinced that custom base dies will be a major trend because they benefit GPU and XPU vendors. However, he argues that this reduces fungibility for memory suppliers like Micron and Samsung and does not help system-level memory costs. He worries it could even lead to scenarios where the same GPU performs differently depending on which vendor’s memory is attached, complicating performance expectations.

He also notes that Samsung talked about ZHBM and deeper 3D stacking, but he views those as longer-term, not near-term, commercial drivers. Some presentations on high-bandwidth flash sounded mixed to him—technically intriguing but also full of challenges that drew skeptical questions.

Goldberg adds that Waymo’s Hot Chips talk showed how application-specific AI can be. Waymo, he says, emphasized that autonomous taxis need higher precision than typical FP4 inference because they must interpret complex signs, time-of-day constraints, and safety-critical sensor data. He links this to AMD’s acquisition of Talis as evidence that semi-custom and model-specific chips could become more common, fragmenting AI silicon even further.

Looking beyond the immediate earnings reactions, Bajarin suggests investors watch how capacity evolves into calendar 2027–2028. His supply-chain conversations indicate that energized AI data center capex from orders made a year earlier will mostly manifest around 2028, when TSMC and substrate capacity may finally loosen somewhat, though he still expects Nvidia to be constrained even then.

Both hosts think Intel’s foundry roadmap is a key variable. Bajarin argues that Nvidia is “most incentivized” to help Intel succeed as a second major advanced manufacturing source and speculates Nvidia will act as a foundational customer, in the same way Apple helped TSMC scale early. He even jokes that Intel could theoretically move its own internal products fully to TSMC and sell all its internal capacity to Nvidia, though he admits this is unlikely.

On the customer side, Goldberg believes Neocloud viability and financing terms will be crucial. He notes that Nvidia has outlined roughly a trillion dollars in combined debt and obligations when broadly defined, tying its fate more deeply to AI data center operators. Bajarin says only a handful of companies can use their balance sheets as a competitive weapon this way and that Nvidia explicitly called this out as a strategy.

For Marvell, both expect more clarity at its investor day in early October, especially around the cadence of that $120 billion Google opportunity and how CXL and XPU attach scale in revenue terms. Goldberg casually advises that Marvell avoid reporting earnings the day after Nvidia in future quarters, given how difficult that comparison proved this time.