# Marvell at $300: Stress-Testing Jensen's Trillion-Dollar Call
> Jensen Huang called Marvell the next trillion-dollar company. The stock rose 32% in a day to a record $302. We stress-tested the math behind the call.
- Author: Barebone Research, Barebone AI
- Published: 2026-06-04
- Canonical: https://barebone.ai/resources/marvell-trillion-dollar-math
- Publisher: Barebone AI (https://barebone.ai)
---
## Four Words, Sixty-Two Billion Dollars
On Monday, June 1, at Computex in Taipei, Jensen Huang stood on stage next to a CEO most retail investors couldn't pick out of a lineup and said four words that moved $62 billion:
> "The next trillion-dollar company, ladies and gentlemen."
The CEO was Matt Murphy. The company was Marvell Technology — ticker MRVL. The next U.S. trading session, Marvell rose **+32.5%**, its biggest one-day gain since 2000. The company walked into Computex week worth $192 billion and closed Thursday, June 4, at a record **$301.65** a share — roughly a **$260 billion** company that has more than tripled in 2026.
A trillion dollars is about 3.8 times that. From an all-time high. For a company that, at the June 4 close, trades at right around **100 times** its trailing adjusted earnings.
So we deployed Barebone to stress-test the call — Marvell's earnings, every revenue outlook management has given since September, and the full set of Wall Street forecasts. Two things came back. The receipts are real: revenue up 42% last fiscal year, profit growing twice as fast, guidance raised again and again. And the math gap is enormous: even Marvell's own dream-case targets for fiscal 2028 support a valuation *below* where the stock trades today.
Both things are true. That's what makes this interesting.
## The Wiring Problem
Start with why Jensen said it, because the logic is real even if the price is debatable.
An AI data center is not one giant computer. It's hundreds of thousands of chips wired together to behave like a single brain — and as the chips get faster, the wiring becomes the constraint. Jensen's own framing on the Computex stage: to "distribute it across the entire data center, what's necessary is connectivity."
The physics are unfriendly. Copper — computing's cable for half a century — is fast and cheap only over short distances, and the faster you push signals, the shorter its useful reach. At today's speeds that reach is roughly a meter: fine inside a rack, useless across a building. For data centers the size of football fields, the data has to travel as pulses of light through glass — photonics. Jensen again, same stage: "in the next 5-10 years we're going to use a ton of copper and we're going to use tons and tons of optics."
The laser makers — Lumentum, Coherent — already repriced violently on this thesis; we spent three articles on that supply chain this spring. But a laser carries zero information by itself; a chip has to encode data onto the light and decode it at the other end. That chip is the **optical DSP**, and at scale it is essentially a duopoly: Marvell — which bought its way in with the $10 billion Inphi acquisition in 2021 — and Broadcom.
Marvell's other engine is **custom silicon**: co-designing the in-house AI chips hyperscalers build to reduce NVIDIA dependence — Amazon's Trainium, Microsoft's Maia, Google's Axion CPU. Add the $3.25 billion December acquisition of Celestial AI, whose photonic fabric aims to deliver light directly to the compute die, and the portfolio is coherent: Marvell helps clouds build their own chips, then sells them the means of wiring those chips into one brain.
Murphy's pitch at Computex was that connectivity is the next AI bottleneck, after compute and memory. Data center is now 76% of Marvell's revenue. The market spent two years paying for compute. It has started paying for the connections between the compute.
## What NVIDIA Actually Bought
Here's the part of the story we got wrong — briefly.
In our GTC 2026 breakdown in March, we flagged NVIDIA's new integrated inference stack as a threat to the custom-chip vendors: if Vera Rubin plus the Groq LPU line is compelling enough, why would hyperscalers keep commissioning their own silicon from Broadcom and Marvell?
Twelve days after that keynote, on March 31, NVIDIA answered by wiring Marvell **$2 billion**.
The investment came with a strategic partnership: Marvell's custom XPUs (the catch-all term for non-NVIDIA AI accelerators) become compatible with **NVLink Fusion** — the rack-scale fabric that lets third-party chips plug into NVIDIA's GPUs, CPUs, and networking at up to 1.8 terabytes per second. The two companies also agreed to co-develop silicon photonics and AI-RAN — pushing AI compute into 5G and 6G networks. Announcing it, Huang said:
> "The inference inflection has arrived. Token generation demand is surging, and the world is racing to build AI factories."
Read the structure, not the press release. NVIDIA concluded hyperscalers are going to build custom chips no matter what — Counterpoint Research expects AI ASIC shipments to roughly triple between 2024 and 2027. If that's inevitable, NVIDIA would rather those chips live *inside* its racks, attached to its fabric, than inside someone else's open-standard alternative. The $2 billion isn't charity; it's ecosystem glue — the same playbook as NVIDIA's stakes in Nokia, Lumentum, and Coherent that we've covered this spring. Notably absent from the NVLink Fusion tent: Broadcom.
The market did the rest:
The entire move is ten weeks old, and every leg is an announcement: the NVIDIA investment in late March, a record quarter and raised outlook on May 27, then Jensen's four words. Which raises the only question that matters — what do the fundamentals actually support?
## The Receipts
Credit first: this is operationally the strongest stretch in Marvell's history, and management keeps raising its own bar in public.
On the March 5 call closing out fiscal 2026 (Marvell's fiscal year ends in late January), Murphy put it plainly:
> "We concluded fiscal 2026 on a strong note with revenue growing 42% year-over-year and non-GAAP EPS increasing 81%, roughly twice the rate of revenue growth, demonstrating the strong operating leverage in our business model."
Then the guidance staircase. September 2025, at a bank-hosted investor call: fiscal 2027 revenue of approximately **$9.5 billion**. On the December earnings call: approximately **$10 billion**. On May 27 this year: "approaching **$11 billion**" — in Murphy's words, raising the forecast "by almost another $1 billion." Same call, he reached two years out: fiscal 2028 revenue growing "close to 40% year-over-year, reaching approximately **$15 billion**," driving non-GAAP EPS "to well over **$5**."
The hottest line is custom AI silicon: from zero to **$1.5 billion** in fiscal 2026, and management continues to expect it "to at least double year-over-year." The latest quarter backed all of this up — record revenue of $2.42 billion, up 28%, data center up 27% to $1.83 billion, and a guide for roughly 35% growth next quarter. The fiscal year just ended and the two ahead of it:
| Period | Revenue | YoY growth | Non-GAAP EPS | Status |
|--------|---------|-----------|--------------|--------|
| FY2025 | $5.77B | +5% | $1.57 | Actual |
| FY2026 | $8.20B | **+42%** | $2.84 | Actual (reported Mar 5) |
| FY2027 | ~$11B | +34% | — | Outlook (raised May 27) |
| FY2028 | ~$15B | "close to 40%" | "well over $5" | Outlook (May 27) |
A 42% growth year, profit compounding at twice the rate of sales, three successively higher outlooks inside nine months, the custom business doubling. The problem is not the company.
The problem is the price.
## The Trillion-Dollar Math
At $301.65, Marvell trades at roughly **100 times** its trailing twelve months of adjusted earnings — about $3.02 per share, or roughly $2.65 billion of profit. Paying 100 times means paying $100 today for every $1 the business currently earns.
A trillion-dollar Marvell is 3.8x the June 4 market cap. The trap in that arithmetic: expensive stocks don't keep their multiples as they mature. Every company that has actually crossed the trillion-dollar line did it on tens of billions of profit at a much lower multiple — not on a 100x rating. So the profit has to do far more than 3.8x the work:
| If the multiple matures to... | Profit needed for $1T | vs. today's ~$2.65B |
|------------------------------|----------------------|---------------------|
| 100x (today's rating) | ~$10B | 3.8x |
| 40x (premium large-cap semi) | ~$25B | 9.4x |
| 30x (mature semiconductor) | ~$33B | 12.6x |
Now hold that against management's own best case. Suppose Marvell hits everything: $15 billion of fiscal 2028 revenue, EPS well over $5 — call it roughly $5 billion of annual profit on today's share count. Put a generous 40 times on that flawless execution and you get a **$200 billion company. That is less than Marvell is worth today.**
Read that chart carefully, because it's the whole article. The green bars are real. The purple bar is management's two-year dream case. The red bars are what a trillion dollars actually requires — and they sit five to seven times above the dream case. The June 4 price has already spent the fiscal 2028 report card and is borrowing against fiscal 2030's.
None of this makes Jensen wrong, exactly. He wasn't issuing a price target with a deadline; he was naming a bottleneck. A Marvell that compounds custom silicon and optics for a decade can grow into very large numbers. But "eventually, if everything keeps doubling" is doing an extraordinary amount of work in a stock that just repriced 3.6x in five months.
## The Uncomfortable Columns
The stress test surfaced four more things the victory lap leaves out.
**Customer concentration cuts both ways.** Data center is 76% of revenue, and the custom business is, practically speaking, a handful of contracts with Amazon, Microsoft, and Google — sockets that get re-bid every chip generation. The market has priced this risk before: in early 2025, on fears that Amazon would dual-source the next Trainium, Marvell fell from a January peak near $127 to below $69 by mid-March — nearly half its value, on a rumor about one customer. The concentration that produces 42% growth produces 46% drawdowns.
**Broadcom is the bigger rival, and it stayed outside the tent.** In custom AI silicon, Broadcom — partnered with Google's TPU, Meta's MTIA, and reportedly OpenAI — is projected by Counterpoint to hold roughly 60% of the custom ASIC market by 2027. It also conspicuously declined to join NVLink Fusion. If the open-standard ecosystem wins the next round of hyperscaler designs, NVIDIA's embrace of Marvell starts to look less like a gift and more like a constraint.
**Wall Street's own math doesn't reach the price.** Of the 40-plus analysts covering Marvell, not a single one rates it Sell — and yet the average price target sits in the low $230s, roughly a quarter *below* the June 4 close. Everyone says buy; nobody's model gets to $300. That's what a stock looks like when it's running on narrative ahead of numbers.
**The smart money bought the doubt, not the euphoria.** Sound Shore — a value shop — built an 860,000-share position in the fourth quarter of 2025 at an average cost near **$84**, and its Q1 letter called Marvell "a bright spot in the tech space." They're up roughly 3.6x. Meanwhile, every reported transaction by Marvell's own management this year has been a sale: Murphy sold 30,000 shares on March 26 — five days before the NVIDIA announcement — and another 7,500 on May 13 at prices near $177. Insider selling is a weak signal — executives sell on schedules for plenty of reasons; insider *buying* is the strong one. There is none. One footnote: on GAAP accounting — after stock compensation and acquisition charges from the Celestial AI deal — Marvell earned four cents a share last quarter, against the 80-cent adjusted figure everyone quotes.
## What This Means
Separate the two claims hiding inside Jensen's four words.
Claim one: connectivity is the next AI bottleneck, and Marvell sits on both sides of it — custom compute and the optical links between compute. We think that's right, and the receipts back it: 42% growth, custom silicon doubling, a guidance staircase that keeps climbing, and a $2 billion endorsement from the one buyer who sees everyone's roadmap.
Claim two: that this is worth paying 100 times trailing earnings for today, 29% above the average analyst target, after a 3.6x repricing in five months. That claim isn't supported by any number management has published — including its own fiscal 2028 targets.
The signals worth watching from here: whether custom revenue actually doubles this fiscal year; whether a second and third XPU customer ramps to scale; whether NVLink Fusion wins hyperscaler designs against Broadcom's open path; and what happens to a 100x multiple the first time the staircase pauses.
Jensen has been right about where the money flows often enough that we track his words for a living. But the trillion-dollar question isn't whether Marvell can get there eventually. It's how much of that journey the stock is asking you to prepay at the door.
---
*Data: Barebone | Sources: Marvell Q4 FY2026 results (March 5, 2026), Marvell Q1 FY2027 results (May 27, 2026), NVIDIA–Marvell strategic partnership announcement (March 31, 2026), Computex 2026 remarks, Sound Shore Q1 2026 investor letter, SEC Form 4 filings | Prices as of June 4, 2026 close*