MRVL · Forward model · Data center · Bear case

Assumptions & reasoning

• Recalibrated to the 27 August 2026 outlook, not to consensus. The 22.6% first sequential step puts this line at $2,662.3M in fiscal 2027 Q3, against management's guided 'more than 20% sequentially and roughly 75% year over year' - the model lands +22.6% sequential and +75.4% year on year against fiscal 2026 Q3's $1,517.9M, so it sits inside both. With communications at its own guided decline the consolidated first projected quarter is $3,150.6M against the guided $3.150B mid-point, a residual of +0.02%.
• The annual fit: fiscal 2027 data center comes out at $9,795.6M, up 60.6% on the $6,100.3M of fiscal 2026 against a guided 'approximately 60%', and fiscal 2028 at $16,344.1M, up 66.9% against 'more than 60%'. Consolidated, that is $11,881.5M in fiscal 2027 against 'roughly $12 billion' (-1.0%) and $17,957.5M in fiscal 2028 against 'approximately $18 billion' (-0.2%).
• The decay had to move and that is the single biggest change in this revision. Holding the previous 0.11 while starting from 22.6% instead of 15.4% would have put fiscal 2028 data center at $18,971M, +92% year on year, and consolidated revenue at $20.6B against a guided $18B - $2.6B of revenue nobody guided. Raising the decay to 0.24 is what reconciles a faster basis quarter with an annual target management raised by less than the quarter beat by. The cost is at the other end of the horizon: the line grows 6.6% year on year in the twentieth quarter against 12.8% at the previous basis, and roughly $3.4B less annual revenue arrives in the terminal year. That is a mechanical consequence of the fit rather than a new view of the business, and it is by far the largest single lever on this page - holding the decay at 0.11 with everything else as published gives a base case of $267.54 against $132.79. The whole fall in fair value across this roll-forward, and more, is that one parameter.
• Note what the driver structurally CANNOT do, and the print made this worse rather than better: management says growth accelerates again in fiscal 2027 Q4, both sequentially and year over year, and a decaying sequential rate cannot accelerate. The fit is to the annual totals, and the shape inside fiscal 2027 is flatter than the shape management describes - the model's Q4 step is 13.4% against a Q3 step of 22.6%, where management guides the opposite ordering.
• Marvell publishes no XPU units, no ASPs, no wafer allocation, no capacity and no backlog for this line, and does not say what share of it is custom silicon versus optics, switching and storage. The call added colour without numbers - custom 'more than doubling' in fiscal 2028, scale-out switching 'on track to more than double this year', three connectivity businesses at or near a billion-dollar annualised run rate - but not one of those is a disclosed dollar figure for a quarter. Any unit- or capacity-driven model here would still be inventing the denominator, so this vertical runs on a growth rate and says so.
• Seasonality tested at the previous basis and NOT carried; the tenth quarter does not change that. Ratio to a centred four-quarter moving average gave 1.028 for fiscal Q1, 1.001 for Q2, 0.984 and 0.964 for Q3 and 1.073 for Q4, with only fiscal Q3 carrying two comparable windows and those disagreeing by 2.0 points. The apparent shape is the custom ramp, not a season, and the +18% sequential just printed in a fiscal Q2 whose factor was 1.001 is more evidence of that. Derived, method as stated; no factors are applied.
• Terminal growth of 1.5% a quarter, about 6% a year, is untouched and is now doing even more work after the exit multiple, because the faster decay reaches it sooner. It says the custom programmes that carry this line eventually behave like semiconductors again. Nothing Marvell has disclosed reaches past fiscal 2029, and the 6 October 2026 Investor Day is where the long-term numbers are promised - this model does not anticipate them.
• Margin start moves to 36.6%, the reported consolidated non-GAAP operating margin, from 35.0%. The terminal 37.0% was NOT moved and is now nearly reached at the basis: management guided non-GAAP operating margin into its 38-40% long-term target range in fiscal 2027 Q4 and to the upper end of it through fiscal 2028, so the print argues that this vertical's terminal margin is too low by two to three points. Left alone deliberately; a revision that moves it should move it for a reason of its own, not as a side effect of a roll-forward.
• Capex intensity moves to 4.6%, the $126.7M spent on property and equipment against $2,739.3M of revenue in the basis quarter, from 6.4% at the previous basis. Capex is lumpy at this issuer - 2.37%, then 6.44%, then 4.63% over three quarters - and the trailing four-quarter rate is 4.98%, so the basis-quarter figure is a fair but noisy anchor. The terminal 5.0% was not moved, so this line's capex now glides UP rather than down, which is a visible artefact of the roll-forward. Separately, the roughly $1B of capacity prepayments to suppliers Marvell expects to make in fiscal 2027 do not appear anywhere in this model: they sit in operating cash flow, not capex, and they are why operating cash flow fell sequentially on record revenue.
• Celestial AI (closed 2 February 2026, $3.5B) and XConn (closed 10 February 2026) sit inside this line from their acquisition dates. Neither revenue contribution is quantified, so the organic and acquired parts of the 45.7% year-on-year growth in the basis quarter cannot be separated here.
• Concentration is the risk the sliders do not express. One direct customer was 16% of consolidated revenue and one distributor 45% in the fiscal 2027 Q1 10-Q; the fiscal 2027 Q2 10-Q is not filed yet, so those percentages are one quarter stale and have not been restated. The 8-K of 18 August 2026 makes the point sharper rather than softer: the expanded agreement and warrant are with a single hyperscaler, and the same hyperscalers are both the customer and the potential competitor.