Marvell Finally Put a Number on the Denominator

Current FY2031 view (October 7, 2026): My Base assumes roughly $55B of revenue and $20 EPS at 26x, implying about $520/share. The Investor Day scenarios below reflect my October 6 snapshot; the Marvell holding record contains the current assumptions and live returns.

I went into Marvell’s Investor Day with one question.

Not “how big is the AI TAM?”

Please. We have enough giant AI TAM slides to wallpaper a house at this point.

The question was much more boring: after all the acquisitions, stock comp, preferred stock and warrants, how much of this growth is actually going to belong to each share I own?

That was the part of the MRVL thesis I had gotten increasingly uncomfortable with.

Today, management finally answered it.

Marvell is now targeting $70B–$90B of revenue in FY2031 and, much more importantly to me, more than $30 of non-GAAP EPS. It also raised FY2028 revenue from about $18B to ~$20B and lifted the FY2029 custom-silicon target from more than $10B to more than $12B. (Marvell Investor Day; Reuters; MarketWatch)

That does not make the target true.

It does make my old model obsolete.

The $80B number is wild. The $30 EPS number matters more.

The midpoint of management’s FY2031 revenue framework is about $80B.

For context, Marvell did $8.2B in FY2026.

So yes, the headline is basically “we think this company can become ten times larger in five fiscal years.”

Absurd enough to get attention.

I care more that management paired it with a per-share target. Marvell expects more than $30 of non-GAAP EPS in FY2031, alongside a 56%–59% non-GAAP gross-margin framework. (Barron’s)

The reason is pretty simple.

Revenue belongs to the company. EPS belongs to the share.

My biggest issue with Marvell was that the operating thesis could work beautifully while the shareholder math got uglier. Celestial AI cost about $3.5B and included roughly 24.5M shares. NVIDIA invested $2B through convertible preferred. The Google agreement includes a warrant for up to roughly 59M shares. That sounds awful in isolation, so the structure matters: about 1.36M shares vest with time, while the rest vest in 240 tranches tied to qualifying product revenue, one tranche per $500M. Full performance vesting therefore requires roughly $120B of qualifying purchases. First-half SBC was already more than $530M. (Q2 10-Q; Google agreement)

A huge revenue number does not fix that.

More than $30 per share would.

Now management has to actually get there.

Interconnect quietly became the bigger story

The other thing that changed for me today is how I think about what Marvell actually is by 2031.

At the $80B midpoint, management sees roughly:

FY2031 midpoint mix	Revenue
Interconnect	~$37.5B
Custom silicon	~$30B
Switching + storage	~$10B
Communications + other	~$2.5B

The mix matters because the market story around MRVL has often been “the other custom-silicon company next to Broadcom.”

Investor Day makes that framing look too narrow.

The interconnect business is actually larger than custom in the midpoint framework. (Barron’s)

That gives me a better reason to care about Celestial AI and XConn. Marvell is trying to own more of the plumbing around accelerated compute: optical DSPs, scale-up connectivity, switching, PCIe/CXL, retimers, and eventually photonic fabric.

If hyperscalers keep fragmenting the architecture into more custom compute and more specialized networking, Marvell does not need to win every accelerator socket to participate in the build.

It can get paid to move the data around the accelerator too.

The result is a stronger thesis than the one I had a week ago.

The denominator still gets a seat at the table

There is an easy way to get way too excited here.

Take management’s >$30 EPS target, slap 35x on it, get a four-digit stock price and go celebrate.

No thanks.

Four years is a long time in semiconductors. Hyperscaler capex can change. Programs slip. Customers dual-source. Custom silicon carries lower gross margin. AI architectures will look different in 2031 than they do now.

And the >$30 number is non-GAAP EPS.

The share count is in the denominator, which is good. Stock-based compensation is still excluded from the expense line, which is less good.

Investor Day did not make the dilution issue disappear. It gave me a measurable hurdle for it.

Now there is something concrete to underwrite.

My old model is dead

Before today I was using a roughly three-year model with $5 / $8 / $11 of normalized EPS and $140 / $304 / $462 of implied value.

The base case was giving me a pretty mediocre expected return from the then-current price. That was the whole reason I had started thinking about whether I should simply let the taxable lot turn long-term in December and rotate the capital somewhere else.

Investor Day changed the inputs enough that keeping that model would be fake precision.

I moved the checkpoint to the end of FY2031 and rebuilt it from scratch.

These are my assumptions, not company guidance:

Case	FY2031 revenue	FY2031 EPS	Exit P/E	Implied value	Total return from ~$297	Annualized
Bear	~$45B	$12	22x	~$264	-11%	-2.7%
Base	~$65B	$23	28x	~$644	+117%	+19.6%
Bull	~$85B	$34	32x	~$1,088	+266%	+35.0%

The reference price is roughly $297 intraday on Oct. 6, when I ran the model.

Notice what I did not do.

My base case does not assume management’s $80B midpoint. It uses about $65B.

It does not assume >$30 of EPS. It uses $23.

Deliberately. Management’s Investor Day framework should be something Marvell can beat my model with, not the thing I need to copy for the investment to work.

The bear case is still ugly. Marvell can become a much bigger company and produce a lousy shareholder return if the long-range targets get cut, margins disappoint and the denominator keeps growing.

The base case now clears the hurdle by a lot more than it did before.

This changes the December decision

This is probably the most important portfolio consequence.

I had been leaning toward a pretty clean decision: once the original taxable lot became long-term in December, reassess MRVL and probably rotate if the expected return still looked weak.

That logic made sense with a base case around mid-single-digit annualized returns.

It does not make sense to carry it forward automatically after a material underwriting change.

If my current base case is even roughly right, selling solely because the tax clock hit one year would be backwards. The expected value changed. The tax treatment did not become the thesis.

Would I take the entire current position in cash and buy all of it back at ~$297 today?

I still do not think that is an obvious yes.

That question matters because the stock has already moved a lot, the Investor Day targets reach out to 2031, and I already own a lot of the same hyperscaler-capex ecosystem elsewhere.

But I also would not be eager to own zero anymore.

For me, that changes the decision.

The latest $1,000 still went to FXAIX

This is where portfolio construction saves me from turning a better thesis into a reflex buy.

Invested assets are around $325K now, and I just put another $1,000 into FXAIX.

I did not redirect it into Marvell after Investor Day.

At the ~$297 model reference price, the existing MRVL position is roughly 3.1% of the portfolio against a 0%–5% framework. The stake is already large enough for the new model to matter if it works.

Could I justify adding based on a ~20% base-case CAGR? Sure. The number is good.

But it is still a four-year model built around hyperscaler spending, custom-program execution, interconnect adoption, margin leverage and a share count that has been anything but boring.

The next dollar does not have to chase every improved thesis.

So the stance stays hold / no add.

What I care about now

The next checkpoints are much cleaner than they were before Investor Day.

FY2028 needs to actually land around the new $20B target. FY2029 custom revenue needs to clear the new >$12B bar. The interconnect ramp has to start looking capable of supporting that frankly ridiculous $37.5B FY2031 midpoint framework.

Then I want the shareholder math.

Does diluted share count flatten? Do buybacks start offsetting SBC and the warrant/preferred overhang? Does operating leverage keep improving even as custom mix pressures gross margin?

If those answers keep coming back yes, $23 of FY2031 EPS may end up being too conservative.

If revenue keeps going vertical while per-share earnings lag, then the old concern comes straight back.

Where I landed

Investor Day did something I genuinely wanted Marvell to do.

It stopped asking me to believe that a giant revenue opportunity would eventually become shareholder value and put an actual per-share target next to the story.

My expected value moved materially higher.

It is not enough to make me forget the financing, the customer concentration, the long horizon or the fact that I already have plenty of AI-infrastructure exposure.

So the tax lot hitting one year is no longer a sale signal. A $90B slide is not a buy signal either.

The position stays where it is while the business gets time to prove the denominator can finally stop being the annoying part of the thesis.