Thirty-eight analysts cover Marvell Technology. None of them rate it a sell. The average twelve-month target sits at $256.91 against a $217.07 close on August 12, which reads like modest, orderly upside until you look at the distribution behind the average. The low target is $126. The high is $400. That is a 3.2x spread on a company carrying a $191 billion market capitalisation, and it is not a disagreement about the quarter. It is a disagreement about the terminal number, and the event that speaks to the terminal number is not the August 27 earnings print. It is the Investor Day on October 6.
The distinction matters because Marvell has already spent its near-term guidance. Matt Murphy raised the revenue outlook for both fiscal 2027 and fiscal 2028 on the May call, citing exceptional AI-related bookings, and guided fiscal second quarter revenue to a $2.565 billion to $2.835 billion band against a $2.01 billion year-ago base. Consensus has settled near the $2.70 billion midpoint, roughly 34 to 35 percent year-over-year. A print anywhere inside that band confirms a model the market has already bought. A print outside it breaks one. Neither outcome tells you whether the company is worth $126 or $400 per share, because that question is about fiscal 2029 and beyond, and the fiscal 2029 framing has not been formally delivered to investors since before the company became a different business.
That gap is the underappreciated part of the calendar. Marvell postponed its June 2025 Investor Day citing macroeconomic uncertainty and did not reschedule until now. In the interval it closed Celestial AI for at least $3.25 billion in cash and stock with an earnout reaching $5.5 billion, closed XConn, took a $2 billion convertible preferred investment from Nvidia representing roughly 2.5 percent of the company, licensed itself into the NVLink Fusion ecosystem, issued nearly $1 billion of 5.300 percent senior notes due 2036, and grew data centre to about 76 percent of revenue. The October 6 session in New York is therefore not a routine update to a long-term model. It is the first coherent statement of what the assembled company is, delivered by Murphy and the full senior team, and the first opportunity for the sell side to reconcile a $126 target and a $400 target against the same set of disclosed assumptions.
The specific numbers under contest are already in circulation and will either be defended or revised on that morning: eighteen XPU and XPU-attach sockets ramping, $75 billion of lifetime revenue potential in the pipeline, a $94 billion data centre total addressable market by calendar 2028, and a stated intention to move from roughly 13 percent share to 20 percent. Multiply the last two and the argument for the upper half of the target range assembles itself. Which is precisely why the composition of the eighteen sockets is the disclosure that matters more than any of the headline figures.
Marvell’s competitive position is routinely described as custom silicon, and custom silicon is the weakest part of it. Designing a hyperscaler’s accelerator is a business with one customer per programme, a renegotiation at every node transition, and a permanent structural threat from the customer’s own engineering organisation. Amazon, Microsoft, Meta and Google all have the balance sheet and the headcount to take more of that work in house over time, and each incremental generation makes it easier. The durable franchise sits one layer out. Marvell’s SerDes and optical DSP position attaches to the socket regardless of who designs the compute die, and the 800G and 1.6T scale-out optics roadmap plus the 51.2T Ethernet switching line, extended by the 102.4 Tbps switch announced in June, sells into racks that Marvell did not win the accelerator in. Attach revenue survives an in-sourcing decision. The XPU itself may not. Celestial’s Photonic Fabric extends the same logic upward, targeting package, system and rack-level optical interconnect, a layer hyperscalers have shown the least appetite to internalise because it is a physics and packaging problem rather than an architecture problem.
The Nvidia stake should be read in the same frame rather than as a validation headline. A 2.5 percent economic interest means Nvidia now participates in Marvell design wins that contain no Nvidia silicon at all. That is the actual content of the $2 billion, and it is a more durable arrangement than the NVLink Fusion integration it was announced alongside.
The sequencing into the print is unusually hostile to clean price discovery. Nvidia reports August 26, Marvell August 27, Broadcom September 2. Marvell prints in the middle of the sandwich, inheriting whatever tone Nvidia sets the previous evening, and then gets repriced within three sessions by Broadcom’s commentary on custom ASIC share and pricing. With a beta of 2.20 and 3.71 percent of the float short, the window in which the stock trades on its own disclosed fundamentals is roughly four sessions wide. Hyperscaler third-quarter capex guides arrive in late October, after the Investor Day, which means the October 6 model will be tested by customer spending disclosure almost immediately.
Base case has Marvell printing inside the guide, reiterating the fiscal 2028 raise, and delivering a fiscal 2029 framework on October 6 consistent with the $94 billion and 20 percent share language already public. That supports the $220 to $270 zone and validates the $256.91 consensus without requiring anyone to change their model. Bull case requires a third-quarter guide above $3 billion or a named new hyperscaler socket, with Google’s SRAM-based Frozen v2 inference programme the most-flagged candidate for 2027, which would put the June record of $329.88 back in play and make the $400 high target a 2027 conversation rather than a fantasy. Bear case is not a Marvell miss. It is a cohort derating: a Broadcom comment on ASIC pricing, or a single hyperscaler trimming capex, compresses the entire custom-silicon group at once, and 38.5x forward earnings on 34 percent growth has a long way to fall toward the mid-twenties. That path runs to $140 to $160 and makes the $126 low a live number rather than a rhetorical one. The stock already sits about 34 percent below its June high after a 144 percent year-to-date advance, so the cohort has demonstrated it can do this without a company-specific cause.
Acquisition accounting will keep obscuring the GAAP picture through this, with a $331.8 million contingent consideration charge in the last quarter as the reminder. The Celestial earnout pays its first third on cumulative revenue of at least $500 million by the end of fiscal 2029 and pays in full above $2.0 billion, against management’s stated expectation of initial contribution in the second half of fiscal 2028 ramping to a $500 million annualised run rate in the fourth quarter. Those thresholds are the closest thing to a disclosed management forecast the company has published for the asset.
The decision-relevant disclosure on October 6 is not the pipeline figure or the share target. It is whether Marvell breaks the eighteen sockets into compute and attach. Investors currently price a $75 billion pipeline without knowing what fraction of it survives a customer deciding to design its own accelerator. That single split converts the in-sourcing risk from a narrative into a number, and the target spread cannot close until someone publishes it.