Marvell (MRVL) at $310: Its Israeli CTO Names the Bottleneck the Market Already Paid to Solve
Noam Mizrahi has been saying the same thing for two years, and the market has only just decided to believe him. Marvell’s corporate CTO, based at the company’s Israeli site and a Technion graduate, has argued since the early innings of the AI build-out that the constraint on the next leap was never going to be the processor. It was going to be the wire between the processors — and then the optics, when copper ran out of reach. The industry called this a backwater. Marvell bet the company on it. The bet has now compounded into one of the most violent re-ratings the semiconductor tape has produced this cycle.
The framing Mizrahi uses is clean. A frontier model is no longer trained on a chip; it is trained on a data center wired to behave as a single machine. Disaggregate the computation, spread it across hundreds of thousands of accelerators, and the binding question stops being how fast each one computes and becomes how efficiently they talk. At that scale every non-optimized nanosecond is a tax levied across the whole cluster, and every wasted milliwatt, multiplied across a million parts, becomes a megawatt the operator has to buy. This is the seam Marvell sells into: optical interconnect inside the rack, long-reach optics between buildings, switching silicon, and custom ASICs designed one customer at a time.
The thesis acquired a face at Computex. Matt Murphy walked the audience through the bottleneck migration that has defined the entire AI capital cycle — compute first, which made Nvidia the first five-trillion-dollar company; then memory, which minted a fresh cohort of trillion-dollar names; and now connectivity, which Murphy claims as Marvell’s territory. He noted that the data center, once under a tenth of revenue, was three-quarters of it last quarter and still accelerating. Then Jensen Huang stood beside him and called Marvell the next trillion-dollar company out loud, to a room, with Nvidia’s own two-billion-dollar March investment sitting behind the sentence. The stock printed its best single session in its history.
So the narrative is intact and the endorsement is the strongest a chip company can receive. The problem is what it costs to own.
The valuation has already priced the speech
MRVL trades near $310 after touching an all-time high of $329.88 on June 18. It has returned north of three hundred percent in twelve months off a fifty-two-week low in the low sixties. Market capitalization sits around $272 billion. Forward earnings clear seventy-five times, trailing earnings north of a hundred. Management has guided fiscal 2028 revenue up to $16.5 billion, a double-digit raise over a number set only three months earlier. Every input to the bull case is moving the right direction.
And yet the consensus twelve-month price target sits below the spot price. The Street’s average objective is roughly $239 against a stock at $310 — a configuration that says, in the plainest terms analysts ever speak, that the shares have outrun the models even as the models keep rising. The price-target upgrades arriving in clusters — $385, $345, new street highs — are chasing the tape, not leading it. That is a tell. When targets are being revised up toward the price rather than above it, the premium is no longer a forecast. It is a position.
What the bull case has to defend
The connectivity story is real, but it is not uncontested, and Marvell is not always early to it. The Teralynx T100 switch unveiled at Computex — 102.4 terabits per second, lower power, lower latency — arrived after Broadcom’s Tomahawk 6 was already shipping and after Cisco’s Silicon One G300 was announced. Marvell leads in optical DSPs and custom silicon; in merchant switching it is the challenger, not the incumbent. The same Nvidia relationship that anointed the stock is also a dependency: a customer, an investor, and a partner whose own roadmap could insource the very interconnect Marvell sells. Alignment with Nvidia is the bull case and the concentration risk in one entity.
The structural argument for owning Marvell is that it is the smallest pure way to play the part of the AI stack that compounds last and longest. Compute and memory got their re-ratings; connectivity is the layer that determines whether a billion-dollar cluster runs at sixty percent efficiency or ninety-five. If Mizrahi’s “Optics Everywhere” thesis plays out, the revenue line keeps bending up and the multiple is retroactively justified. The bear case does not require a miss. It requires only that the cohort derates — that the market decides AI infrastructure names should trade at forty times forward instead of seventy-five — and a stock priced for perfection gives back a third of its value without a single bad quarter.
Mizrahi diagnosed the bottleneck correctly and early, and Marvell built the products to stand in it. The company has earned its place at the center of the AI infrastructure trade. What it has not done is make the stock cheap. The CTO named the constraint years ago; the buyer paying $310 is no longer being paid to be early to that insight. He is being asked to underwrite it at the price where everyone already agrees.