Marvell’s Make-or-Break AI Moment: Can Tonight’s Earnings Finally Launch MRVL Into Nvidia’s League?

Marvell reports after the bell with the AI crowd in a mood that can only be described as “prove it.” MDB and CRDO just delivered monster beats and guidance, and now investors want to see if Marvell can turn its AI data center story into hard numbers and, ideally, a stock that doesn’t fall apart on contact with an earnings release. Expectations are elevated, the rumor mill is humming with a potential Celestial AI acquisition and earlier SoftBank takeover interest, and the stock has already bounced off its lows – which means the bar for “good enough” is higher than usual.

Analysts are looking for roughly $2.07 billion in revenue, up about 36–37% year-on-year, and adjusted EPS of $0.74. That’s broadly consistent with management's own Q3 guide for $2.06 billion in revenue and non-GAAP EPS in a $0.69–$0.79 range. The mix matters: management flagged data center revenue as flat sequentially, with double-digit growth in electro-optics offset by a digestion phase in custom AI ASIC shipments. The Street will be looking for any sign that custom is already reaccelerating into Q4 and that the “one-quarter pause” doesn’t quietly morph into something more structural.

In the AI data center stack, Marvell is a leverage play on interconnect and custom compute rather than the headline GPU story. Data center already accounts for about 74% of revenue, driven by custom AI silicon and the electro-optics franchise inherited from Inphi. On the optics side, Marvell supplies DSPs and high-speed connectivity used to stitch together AI clusters at 800G and eventually 1.6T, a business that has quietly become an estimated ~$3 billion ARR engine with strong double-digit growth. On the compute side, Marvell is a key custom ASIC partner for hyperscalers like AWS, Microsoft, and Google, building “cloud-optimized” silicon such as Trainium- and Inferentia-linked parts and future XPU-attach devices. In other words, Marvell is selling the roads and some of the engines in an AI world that’s obsessed with the cars.

Last quarter set up tonight’s report in two ways. First, the numbers were undeniably strong: Q2 revenue came in at $2.006 billion, up 58% year-on-year and 6% sequentially, with data center revenue up 69% year-on-year and contributing nearly three-quarters of the total. Non-GAAP EPS was $0.67, up 123% year-on-year, and non-GAAP operating margin was a healthy 34.8%. Cash generation was solid too, with $462 million in operating cash flow, reduced inventory, and $200 million of buybacks as the company digested its $2.5 billion all-cash divestiture of the automotive Ethernet business to refocus fully on AI and cloud.

Second, the guidance was deliberately conservative, and the stock paid for it. Management guided Q3 revenue to $2.06 billion, implying only 3% sequential growth and flat data center sales quarter-on-quarter, as a double-digit jump in optics is offset by lower custom ASIC shipments. They were clear this was a one-quarter digestion phase, promising a “strong fourth quarter for custom” and reiterating that custom revenue should be up in the second half versus the first. Still, the cautious near-term framing triggered about a 7% pullback post-print and reinforced Marvell’s reputation as a name that tends to disappoint traders on earnings day even when the long-term AI narrative remains intact.

Heading into tonight, the AI setup around Marvell is more nuanced than the straight-up euphoria you see in names like Nvidia. Bulls emphasize the long-term roadmap: management’s custom silicon event in June laid out an expanded data center TAM of $94 billion by 2028 and a path to grow share from roughly 13% in 2024 to 20% by 2028, driven by custom XPUs, XPU-attach devices, and high-speed optics. Projects like AWS’s Project Rainier – with plans to deploy more than 500,000 Trainium2 chips initially and scale toward 1 million accelerators – support the idea of sustained demand for Marvell’s custom and connectivity content. On top of that, Marvell has won more than 20 ASIC projects (including six XPU-class designs) and sees a second wave of growth later in the decade as Microsoft’s Maia ramp and additional XPU-attach wins convert into revenue.

Skeptics, however, are laser-focused on competitive dynamics and expectations. Broadcom is widely viewed as the stronger hand in AI ASIC, with dominant share across several hyperscalers and deeper visibility into its roadmap. Some analysts argue Marvell’s share at AWS on Trainium2/2.5 is limited and could fall further on Trainium3 as competitors like Alchip and Amazon’s internal Annapurna team take on more of the design. Others worry that even if the long-term TAM is real, the Street is assigning too much value too early. A forward P/E around 30–32x and EV/sales near 10x puts Marvell at a premium to the broader semi group, even if it still screens cheaper than Broadcom. In a tape that has become more selective after the November tech correction, that leaves very little room for disappointment.

Against that backdrop, there are several key items to watch on tonight’s call. First, the Q3 print itself: any upside to the $2.07 billion / $0.74 consensus, and especially any hint that data center is growing sequentially instead of flat, will be well received. Second, the Q4 outlook: investors will want concrete confirmation that custom is indeed set to bounce back strongly, with color on AWS Trainium shipments, Meta and Google ramps, and the timing of future Microsoft Maia contributions. Third, mix and margins: another quarter of ~60% non-GAAP gross margin and mid-30s operating margin would support the idea that Marvell can fund heavy AI investment while still expanding profitability.

The strategic news flow adds more complexity. Reports that Marvell is in advanced talks to acquire Celestial AI in a cash-and-stock deal that could exceed $5 billion, including earnouts, suggests a push deeper into optical interconnect and advanced photonics – exactly where hyperscaler AI roadmaps are headed. At the same time, earlier stories that SoftBank explored a takeover of Marvell underscores the strategic scarcity value of its AI franchise, even if nothing came of those discussions. Tonight’s commentary may not address either topic directly, but any hints around capital allocation, M&A appetite following the auto divestiture, and how management prioritizes buybacks versus strategic deals will be parsed through the lens of those headlines.

Finally, price action and positioning matter. The stock is still down meaningfully year-to-date and roughly 30% below its all-time high, even as data center AI has grown into the overwhelming majority of revenue. That underperformance, combined with a series of “beat but guide conservatively” quarters, has created a strange cocktail of skepticism and FOMO around the name. After MDB and CRDO both exploded higher on clean AI beats and bullish guidance, Marvell is now under pressure to show that its AI opportunity isn’t just a slide deck story but a sustained earnings and cash flow engine. If management can deliver a solid Q3, a convincing Q4 recovery path for custom, and a steady hand on margins and capital allocation, tonight’s print has a chance to finally break the pattern of post-earnings hangovers. If they wobble on any of those fronts, the premium valuation probably does the talking.