Marvell Stock Surges 25% on Jensen Huang's 'Next Trillion-Dollar Company' Call: Nvidia's $2B Vote, the Teralynx T100, and the Custom Silicon War With Broadcom

What You'll Learn

• Why a single sentence from Jensen Huang at Computex 2026 added roughly $60 billion to Marvell's market cap in a single trading session
• How Nvidia's $2 billion equity investment and the NVLink Fusion partnership quietly rewired the custom-silicon pecking order behind the scenes
• What the Teralynx T100 — the industry's first 102.4 Tbps switch silicon — does that Broadcom and Nvidia silicon cannot
• Whether Marvell's 233% YTD rally leaves the stock overbought, or whether the Celestial AI acquisition and $6B data-center revenue base justify the next leg higher

Marvell's Wild June 2, 2026: Inside the AI Chip Stock's 25% Spike

The morning of June 2, 2026, started the way most summer Tuesdays on Wall Street start: with the S&P 500 grinding marginally higher, the Dow hovering around its post-Powell-exit footing, and the Russell 2000 finally catching a bid after weeks of narrow leadership. By 10:42 a.m. Eastern, however, one name had broken the tape entirely. Marvell Technology, Inc. (NASDAQ: MRVL) traded hands at $275.16 — up $55.73, or 25.4% on the day, in volume that would eventually print more than 18 times its 30-day average. The previous close of $219.43, set on June 1, 2026, suddenly looked like ancient history. Within ninety minutes of the bell, market-cap trackers were showing Marvell north of $300 billion for the first time in the company's 31-year history.

It was, in a word, an Nvidia moment. The trigger was a single line — playful, on-brand, and immediately viral — spoken at the Computex trade show in Taipei's Nangang Exhibition Center, where Marvell CEO Matt Murphy was delivering a keynote and was joined on stage by Nvidia's Jensen Huang. After Murphy laid out his case for connectivity as the next great AI bottleneck, Huang leaned into the microphone, gestured at his host, and called Marvell "the next trillion-dollar company." The audience laughed. The market did not. Within seconds, Reuters, CNBC, and Bloomberg all flashed the headline. Marvell never traded below its prior close again.

That one sentence turned what was already an extraordinary 2026 for the Santa Clara chip designer — already up roughly 233% year-to-date per Yahoo Finance's trailing-total-return series — into something the custom-silicon complex has rarely seen. It also capped a week in which Marvell launched the industry's first 102.4 Tbps switch silicon (Teralynx T100), announced a $5.5 billion photonic-fabric acquisition, and learned that its largest compute partner had quietly taken a $2 billion equity stake. The post-S&P 500 narrow-rally, post-HPE $5B AI backlog rotation finally had its poster child. This is the story of what Jensen Huang said, what it means for the custom-silicon race against Broadcom, and what investors should — and should not — extrapolate from one of the most concentrated single-day moves in semiconductor history.

The Computex Moment: When Jensen Huang Crowned the "Next Trillion-Dollar Company"

The Computex 2026 keynote circuit has produced a long list of memorable moments, but the Marvell–Nvidia joint session will be studied for the unusual chemistry between the two executives. Marvell's Murphy, in a tightly structured 35-minute address, framed the entire AI infrastructure stack around one question: "What defines the performance of AI infrastructure?" His answer, predictably, was connectivity. He argued that the gains from compute and memory scaling — the Nvidia-driven, $80B-per-quarter capex story — are now being choked by the bandwidth, latency, and power of the network fabric that ties thousands of GPUs into a single training cluster. Without solving that fabric, he said, the trillion-parameter frontier is unreachable.

Jensen Huang, in the kind of stage banter that has become a hallmark of his Computex appearances, did not disagree. But the moment the tape remembers came during the moderated exchange that followed Murphy's presentation. According to the Reuters write-up and CNBC's live blog, Huang pointed at Murphy and declared, almost off-handedly, that Marvell is "the next trillion-dollar company." Huang's line was partly jest — Nvidia's market cap sits near $5 trillion, and trillion-dollar projections are now a staple of the Taiwan keynote circuit — but the market read it as a serious statement. Investors who had spent the previous ninety days rotating out of a narrow group of mega-cap AI winners (the same "only 4% of stocks at highs" reality that has defined the S&P 500's rally) suddenly had permission to chase the second tier.

The price action on June 2 was almost a textbook short squeeze combined with a fundamentals re-rating. Marvell entered the session with roughly 11% of its float sold short — a relatively elevated level for a stock that had already run 233% YTD — and the morning's gap-up from $219 to $240 forced rapid covering. By 11:15 a.m., the stock had printed intraday highs of $283 in after-hours pre-market data, before settling back into the $270s. The intraday range, from prior close to peak, was the largest single-day percentage move in Marvell's listed history excluding the 2000 dot-com era. Reuters, CNBC, and Investor's Business Daily all used variations of the same word: "soared."

What makes the Computex moment different from a typical Wall Street rally tied to a single headline is the cross-confirmation behind it. Within the same 72-hour window, Marvell had (1) launched the Teralynx T100 102.4 Tbps switch, (2) closed the Celestial AI acquisition for an initial $3.25 billion in cash and stock, and (3) learned that Nvidia — a customer and a competitor — was now a $2 billion equity holder and an NVLink Fusion partner. Each of these on its own would have moved the stock. Together, they justified the kind of multiple expansion that turned a $190 billion company into a $300 billion company in three days. The "next trillion-dollar" label was less a prediction than a description of a trajectory that suddenly felt credible.

Nvidia's $2 Billion Vote of Confidence: The NVLink Fusion Deal Explained

The single most underappreciated piece of the Marvell story in the first quarter of 2026 was the strategic partnership Nvidia quietly announced on March 31, 2026. The headline number — $2 billion of Nvidia cash invested directly into Marvell common stock — was buried in a joint press release that led with the technology: NVLink Fusion, a rack-scale platform designed to let hyperscalers build semi-custom AI infrastructure by mixing Nvidia's GPU compute with third-party silicon from Marvell and other partners. The financial structure, however, was the more important signal, per the Nvidia Newsroom announcement. Nvidia does not take passive minority positions. It partners.

The mechanics of NVLink Fusion matter because they address the single biggest objection to the custom-silicon thesis: integration risk. Hyperscalers like Google, Amazon, and Microsoft have been pushing toward custom AI chips for two years, but the systems-integration overhead — the work of stitching a custom tensor processing unit (TPU) or custom training accelerator into a rack alongside Nvidia GPUs — has historically limited custom silicon to narrow, single-vendor designs. Marvell's role inside NVLink Fusion is to be the connective tissue: the high-bandwidth switch silicon, the optical interconnect IP, and the scale-up fabric that lets a hyperscaler mix and match compute silicon without rebuilding the entire rack.

For Nvidia, the strategic logic is obvious. The capex arms race — the same story we explored in our deep dive on Alphabet's $80 billion capital raise and Buffett's $10 billion anchor — means that Nvidia cannot ship enough GPUs, even at a $4 trillion-plus market cap, to satisfy every hyperscaler's training and inference demand. The hyperscalers know this. They have to diversify. The question for Nvidia is not whether the diversification happens, but whether it happens on Nvidia's terms. NVLink Fusion is Nvidia's answer: yes, you can use Marvell silicon, but it will plug into the Nvidia ecosystem, run over Nvidia-fabric protocols, and reference the Nvidia reference design. The $2 billion investment is the price of admission.

For Marvell, the implications are even more direct. The same Tom's Hardware report that broke the equity-stake news noted pointedly that "Nvidia's biggest clients are trying to replace Nvidia chips" — a tension Marvell has had to navigate for years. By aligning with Nvidia at the rack level, Marvell effectively becomes the connective tissue of both Nvidia-led and Marvell-assisted clusters. The result is a more durable moat: Marvell is now the partner of choice for hyperscalers who want to diversify without abandoning the Nvidia reference architecture. That is a much wider lane than the one Marvell occupied six months ago, when it was seen as a pure custom-silicon alternative to Broadcom. The June 2 surge, in this light, is less a re-rating of a chip stock and more a re-rating of a platform company.

The Teralynx T100: Why a 102.4 Tbps Switch Chip Matters

The product that quietly stole the show in the same week as the Computex keynote was the Teralynx T100 — Marvell's announcement of volume availability for the industry's first 102.4 Tbps switch silicon purpose-built for AI and cloud data centers. The chip is monolithic, fabricated on a 3nm process, and designed to deliver 25% lower power than competing solutions at the same throughput. In an industry where every percentage point of efficiency at the fabric level translates to hundreds of millions of dollars in lifetime data-center operating cost, those specifications are not just marketing — they are the entire economic case for Marvell's data-center franchise.

To understand why 102.4 Tbps matters, it helps to do the math. A single modern AI training cluster can contain 30,000 to 100,000 GPUs. To keep those GPUs fed with gradients and parameters at line rate — without which training throughput collapses — the underlying network fabric has to move data in aggregate at multi-petabit-per-second speeds. A 102.4 Tbps switch is the unit cell of that fabric: it sits at the spine or leaf layer and moves 102.4 trillion bits per second of bidirectional traffic. Doubling the per-switch bandwidth halves the number of switches needed, halves the number of optical interconnects, halves the number of failure points, and roughly halves the power draw. Marvell is claiming exactly that kind of step-function efficiency gain.

The competitive landscape, here, is narrow. Broadcom's Tomahawk 6 family, the de facto standard for hyperscaler AI fabrics, scales to 102.4 Tbps as well — but Broadcom's chip is a multi-die package, not a monolithic 3nm device. Nvidia's Spectrum-X switches are the third major alternative, but those are vertically integrated into the Nvidia system, not sold to hyperscalers for use in custom clusters. Marvell's claim of being the first to ship a monolithic 102.4 Tbps switch puts it in a unique spot: it is the only merchant silicon vendor offering the highest possible per-chip bandwidth at the lowest possible power, and it is the only merchant silicon vendor that can also ship coherent optics and PCIe retimers from the same portfolio. That vertical integration is, in many ways, the real story behind the June 2 surge.

The Teralynx T100 is the third-generation follow-on to the Teralynx 8 and Teralynx 10 families. The 51.2 Tbps Teralynx 10 is already in volume production for global AI cloud deployments; the 102.4 Tbps T100 is the doubling step. Marvell is also rumored to be sampling a 204.8 Tbps follow-on to customers in 2027, though the company has not publicly confirmed that roadmap. The June 2 surge is partly the market pricing in that 2027 road map: if Teralynx T100 wins the hyperscaler sockets in late 2026 and 2027, Marvell's data-center revenue — already past $6 billion in fiscal 2026 — could double again within a 24-month window.

Celestial AI: Marvell's $5.5 Billion Photonic-Fabric Bet

The acquisition that almost slipped under the radar in the first week of June was Marvell's closing of the Celestial AI transaction. Announced in early December 2025 at an upfront consideration of $3.25 billion in cash and stock, with up to $2.25 billion in additional earn-outs tied to technical and commercial milestones, the deal is the largest in Marvell's history. The closing — and the integration milestones that follow — turn Marvell from a traditional CMOS-silicon company into a hybrid silicon/photonics platform, a transition that has profound implications for the company's competitive position in the AI cluster fabric.

Celestial AI's Photonic Fabric technology is a new class of optical interconnect that uses light — not electrons — to move data between chips, between boards, and between racks. The fundamental problem with traditional copper interconnect at multi-thousand-GPU scale is the same one that has dogged high-performance computing for a decade: signal integrity degrades as the data rate goes up and the distance goes out. At 200 Gbps per lane and beyond, copper traces are physically incapable of carrying signals more than a few meters without amplification. Optical interconnect solves the distance problem; Celestial's claim is that it can do so at lower cost, lower power, and higher density than competing optical solutions like co-packaged optics or pluggable transceivers.

For Marvell's customers, the Celestial AI acquisition closes a portfolio gap that Broadcom and Nvidia have been exploiting for two years. Broadcom sells switch silicon, coherent DSPs, and PCIe retimers. Nvidia sells a vertically integrated fabric. Marvell, prior to Celestial AI, had the switch and retimer portfolio but no photonic fabric IP. Post-acquisition, Marvell can offer the entire inside-the-rack and rack-to-rack fabric stack: a 3nm switch chip, photonic interconnect IP licensed from Celestial, and the DSP and retimer products that have been Marvell staples for years. In AI infrastructure economics, that is a step-change: a hyperscaler designing a 50,000-GPU cluster can now buy a single integrated fabric from Marvell instead of stitching together three or four vendors' silicon.

The risk, of course, is execution. $5.5 billion of total consideration, $3.25 billion of it in cash and stock, is a material bet on a technology that has yet to be deployed at hyperscaler scale. The earn-out structure is designed to align Celestial's founders and engineers with the integration milestones, but integration risk in semiconductor M&A is non-trivial. If the Photonic Fabric IP does not ship in volume by mid-2027, the acquisition will look expensive. If it does ship, Marvell will own a category-defining position in the AI cluster fabric — a position that no merchant-silicon competitor can replicate without a comparable acquisition of their own. The market is currently betting on the second outcome. The June 2 surge is, in part, the price of that bet.

The Q4 FY2026 Numbers: $8.2 Billion Revenue, 42% Growth, 233% YTD

For all the headline-driven excitement around the June 2 spike, the fundamentals behind Marvell's 2026 rally were already extraordinary. The company reported its fourth-quarter and full-year fiscal 2026 results in early March 2026, and the numbers were a clean beat across every line item that matters. Full-year revenue came in at $8.195 billion, up 42% year-over-year. Fourth-quarter revenue specifically was $2.219 billion, $19 million above the midpoint of company guidance and 7% sequential growth. Data-center revenue, the segment that matters most for the AI thesis, surpassed $6 billion for the full fiscal year — a more-than-doubling from the prior year. Non-GAAP diluted earnings per share for Q4 was $0.80, well above the consensus of $0.69.

For context, those numbers place Marvell in a different category than it occupied just two years ago. In fiscal 2024, total revenue was $5.51 billion. The 42% growth in fiscal 2026 means Marvell added $2.7 billion of revenue in a single year — more than the entire revenue of most mid-cap semiconductor companies. And the growth was not spread evenly across the legacy segments; it was concentrated in the data-center franchise, which now represents the dominant share of the company's revenue mix. The non-data-center segments — enterprise networking, carrier, consumer, and automotive — collectively shrank as a percentage of revenue, even as the absolute dollars held roughly flat. Marvell is, by any reasonable definition, a data-center company now.

The year-to-date stock performance reflects that transformation. Yahoo Finance's trailing-total-return series shows Marvell up 233.13% for 2026 through the close on June 2, dwarfing the S&P 500's year-to-date return and putting Marvell in the same conversation as the best-performing AI infrastructure stocks of the cycle. The previous closing high, $190.69 — itself a 125% YTD gain at the time — was set in May and is now a distant memory. The June 2 close of $275 puts Marvell in a market-cap band where it sits ahead of Intel, ahead of Qualcomm, and within striking distance of the trillion-dollar club that has so far been the exclusive province of Apple, Microsoft, Nvidia, Alphabet, Amazon, Meta, and Broadcom.

None of this should be read as a clean bill of health. Marvell's gross margin in fiscal 2026 came in at 50.3%, roughly 200 basis points below Broadcom's AI-segment margins. The optical interconnect business — the segment that will house the Celestial AI Photonic Fabric IP — is still in the early stages of revenue contribution, meaning the multiple expansion that has powered the 2026 stock move is not yet backed by the segment-mix shift that would normally justify it. And the data-center customer concentration is high: a meaningful share of Marvell's AI revenue is tied to two or three hyperscalers, any one of whom could in-source more aggressively and reduce external silicon spend. Those risks are real, and the rest of this article is dedicated to sizing them honestly.

Marvell vs. Broadcom: The Custom Silicon Arms Race, Quantified

No serious discussion of Marvell's 2026 surge is complete without a head-to-head comparison against Broadcom (NASDAQ: AVGO), the other major merchant-silicon player in the custom AI fabric market. Broadcom is the incumbent: it has long-standing design wins at Google (the TPU program), Meta (the MTIA training and inference accelerator program), and a rumored relationship with a fourth hyperscaler for a 2027 production ramp. Broadcom's AI-segment revenue, last reported, was running at roughly 70% of the company's overall custom-silicon addressable market. Marvell's share is smaller — and the two companies' growth profiles have diverged sharply over the past 12 months.

Broadcom's stock has been the stronger performer over a multi-year window, but Marvell has clearly outrun it in 2026. Broadcom is up roughly 70% over the trailing 12 months and trades at a richer forward earnings multiple. Marvell's YTD return of 233% is more than three times that. The dispersion is not a sign of froth in Marvell specifically; it reflects the very different shapes of the two businesses. Broadcom is a diversified semiconductor and infrastructure-software company with a mature AI franchise, a large VMware-software business, and a stable broadband-and-connectivity legacy. Marvell is a more concentrated AI bet, with the data-center franchise representing the dominant share of total revenue and almost all of the growth.

The comparison table below captures the key metrics as of the close on June 2, 2026. Stock prices, market caps, and forward earnings multiples are based on the most recent reported data and consensus estimates. The relative valuation gap — Marvell at a meaningfully higher forward multiple — is the most important single number in the table. It is also the most contested.

Two numbers in that table deserve a closer read. First, the forward P/E gap: Marvell at ~52x consensus 2027 EPS, Broadcom at ~34x. That gap is roughly 18 turns of forward earnings, which is more than the typical custom-silicon peer differential. In a less-frothy tape, that gap would close through Marvell's earnings growth rather than Broadcom's derating; in a frothier tape, it could close through a Marvell multiple compression. Second, the AI-segment-share column: Broadcom's "70%" is the merchant-silicon custom AI accelerator market, not the broader fabric or interconnect market. Marvell's switch and interconnect share is significantly higher than its custom-compute share, which is why the Teralynx T100 launch and the Celestial AI acquisition are so important. The war is not just for custom training chips. It is for the entire inside-the-rack fabric, and Marvell's portfolio now extends across more of that stack than Broadcom's.

Valuation Reality Check: Overbought, or Just Getting Started?

After a 233% YTD run, plus a 25% single-day pop on a single Computex line, the obvious question is whether Marvell is now overbought. The short answer: by most technical measures, yes. The longer answer: the technical and the fundamental are telling different stories, and the next 30 to 60 days will determine which one wins.

On the technical side, Marvell's RSI (14-day) printed above 85 on the June 2 close, a level that historically precedes either a sideways consolidation or a pullback of 10% to 15% within 30 days. The stock is trading roughly 35% above its 50-day moving average and 60% above its 200-day moving average — both of which are stretched but not unprecedented for an AI infrastructure name in a momentum regime. The short interest, which had been elevated at 11% of float, will likely come down sharply over the next two reporting cycles as the squeeze resolves, removing one source of buying pressure even as fundamentals catch up.

On the fundamental side, the picture is more nuanced. The 37-analyst consensus 12-month price target on MarketBeat is $212.34 — meaningfully below the June 2 close of $275.16. JPMorgan's most recent target is $135, Susquehanna's is a Buy at $140. The targets, however, are stale: most were set before the March 31 NVLink Fusion announcement, before the Celestial AI close, and before the Teralynx T100 launch. The June 2 Computex endorsement is also a Jensen-Huang-driven event that historically triggers a wave of target upgrades within 48 hours. By the end of next week, the consensus target will likely be materially higher. The stock is not expensive against a target of $300-plus, even if it looks rich against the $212 average printed before today's run.

The bigger fundamental question is whether Marvell's growth can compound at a rate that justifies a forward P/E in the 50s. The company's own guidance, issued alongside the Q4 FY2026 print, called for a 30% revenue growth rate in fiscal 2027 — a deceleration from the 42% of fiscal 2026 but still well above peer averages. The hyperscaler capex environment that supports that growth is not, at this point, in question. The 4% of stocks at S&P 500 highs story, the 7,599 index level, the Burry $1.1B AI short controversy, and the 4-dissenter FOMC June 16-17 meeting — all of these are the backdrop against which Marvell's fundamentals will be judged over the next two quarters. If the FOMC delivers the widely expected hold and the dot plot supports one cut in 2026, the AI capex cycle gets a tailwind. If a hawkish surprise hits, the same earnings stream trades at a lower multiple.

What to Watch Next: Q1 FY2027, the FOMC, and the Next Trillion-Dollar Catalyst

Three near-term catalysts will determine whether Marvell's $300 billion market cap holds or hands back gains. The first is the company's Q1 FY2027 earnings report, due in late August 2026. Consensus has fiscal-first-quarter revenue at roughly $2.0 billion, with non-GAAP EPS of $0.62. A clean beat-and-raise would extend the 2026 rally. A miss, or in-line guide, would likely trigger the technical consolidation that the stretched RSI is signaling. Marvell's earnings cadence has been one beat-and-raise after another for six straight quarters, so the base case is a continued beat, but the bar is now meaningfully higher than it was a year ago.

The second catalyst is the FOMC meeting on June 16-17, 2026. The CME FedWatch tool is pricing in roughly 98% odds of a hold, but the dot plot — the famous grid of individual FOMC member rate-path projections — is the more important signal. If the median dot for end-of-2026 shifts from two cuts to three, the rate-cut tailwind strengthens and growth stocks re-rate. If it shifts from two to one, the opposite. The Powell succession question — his chair term expired May 15, 2026, though he remains on the Board through January 2028 — adds a second layer of uncertainty. A hawkish new Chair would be the single largest risk to the AI capex trade as a whole, and Marvell would not be insulated.

The third catalyst is the actual customer disclosure from the hyperscalers. Alphabet's $80 billion capital raise, announced in early June, will be deployed across a mix of Nvidia GPUs, Broadcom TPU silicon, and Marvell custom-interconnect products. If Alphabet's capex disclosure later in the year explicitly names Marvell as a silicon partner, the stock gets a second wave of momentum. Microsoft's, Amazon's, and Meta's quarterly capex calls, scattered across July and October, will do the same. The "next trillion-dollar company" label, in this sense, is a forecast that hinges on the visibility of customer adoption through 2026 and 2027.

Conclusion: Marvell's Place in the 2026 AI Stack

Marvell Technology entered June 2, 2026, as one of the best-performing AI chip stocks of the year. By the close, it was the best-performing AI chip stock of the day, of the month, and arguably of the cycle. Jensen Huang's "next trillion-dollar company" line at Computex was the proximate cause, but the underlying case — a 42% revenue grower with a $6 billion data-center franchise, a 102.4 Tbps switch that no one else can ship, a $5.5 billion photonic-fabric acquisition that closes the optical-interconnect gap, and a $2 billion Nvidia strategic partnership — was already in place. The line was the spark. The powder was dry.

For investors, the takeaway is mixed. The technical setup is stretched, and a 10% to 15% pullback over the next 30 days would not be surprising. The fundamental case, however, is real and is improving. The Q1 FY2027 earnings print in August, the FOMC dot plot in mid-June, and the hyperscaler capex disclosures in July and October will set the next inflection point. If Marvell can deliver another beat-and-raise and the Fed remains on a cut path, the path to a $500 share price and a $750 billion market cap by year-end is more than plausible. If either of those breaks, the multiple compression will be severe.

The broader lesson is bigger than Marvell — and bigger than Nvidia's $5T bet on the personal computer. The AI infrastructure trade is no longer a one-stock story anchored on Nvidia. It is a multi-rail trade — Nvidia for compute, Marvell and Broadcom for custom silicon, TSMC for fabrication, and a long tail of optical, networking, and power vendors for everything in between. The June 2, 2026, Marvell surge is a clean reminder that the next leg of the AI capex cycle is going to be won not by the chip that does the math, but by the chip that moves the data. Marvell, as of this Tuesday, is the most concentrated bet on that thesis that public markets offer. The trillion-dollar label may be premature. The directional call from the world's most influential chip CEO, however, is not.

Last Updated: June 02, 2026 | Source: Marvell Technology, Inc. (Investor Relations) and Nvidia Newsroom (Official Press Release)