High-Bandwidth Memory Repricing Pushes SK Hynix and Micron Past $1 Trillion

The market milestone for SK Hynix and Micron was not a generic semiconductor rally. The narrower claim was that investors are revaluing memory because high-bandwidth memory has become a critical constraint in the AI infrastructure build-out.

Ed Ludlow framed the number directly: SK Hynix and Micron had reached a combined market capitalization of more than $1 trillion. Ryan Vlastelica said the move reflected “a huge and growing appreciation” of how central high-bandwidth memory is to AI infrastructure. Demand, in his description, is now showing up almost directly in revenues and stock prices.

Micron’s revenue, Vlastelica said, nearly tripled in the latest quarter, which he described as the company’s fastest growth pace since the 1990s. Micron had risen more than 70% in May alone, which he said was its biggest one-month jump since December 1987. Caroline Hyde added that SK Hynix and Micron had both risen about 70% in May and more than 200% year to date, while both were posting quarterly revenue increases above 200%.

That combination — huge equity gains alongside huge revenue growth — created the valuation tension that ran through the memory discussion. Hyde noted that some analysts still described the stocks as “fundamentally cheap.” Vlastelica said Micron was trading below 10 times forward earnings, which he called “extremely low” for a company growing this quickly. But because memory has historically been cyclical, he said investors have to ask whether that low multiple is really a bargain or a warning that earnings are near a peak.

The old memory-market logic matters because it was built around commodity cycles. Ian King put the issue plainly: memory has historically been a commodity business in which one company’s chip could often be swapped for another’s, creating a market driven by spot prices, supply, and demand. The biggest end markets used to be PCs and then smartphones — mostly consumer devices. Long-term supply planning paired with short-term consumer-demand swings was, in King’s words, “a recipe for disaster.”

AI may be changing that cyclicality, but the claim was not treated as settled. Vlastelica stated the question directly: whether AI is changing memory’s cyclical nature enough to create a new paradigm in which Micron is genuinely as cheap as it appears. Ludlow answered “yes” on air, arguing that high-bandwidth memory is not merely adjacent to AI accelerators but embedded into the system design around GPUs and chiplets. In his view, the GPU shortage is not necessarily about Nvidia being unable to obtain enough GPU die; the missing part is the corresponding high-bandwidth memory.

That distinction explains why analysts are revisiting multiples. Vlastelica said UBS had raised its Micron target substantially — he corrected Ludlow’s phrasing by saying UBS had “tripled” its price target — and estimated Micron could eventually become a $1.8 trillion company. The note, as Vlastelica summarized it, argued that Micron deserved a multiple closer to Nvidia’s: roughly 15 times estimated earnings instead of the five-times multiple UBS had historically considered fair. That is not a small adjustment to a price target; it is a proposed change in the way the business should be valued.

King’s explanation sharpened why that claim is controversial. There are only three major providers left in high-bandwidth memory, Hyde said, describing the market as resembling an oligopoly. King’s answer was that the industry consolidated not because it was always attractive but because it had been “a horrible market for years.” The remaining players survived an expensive, volatile business in which large bets were required and bad years were common. They are thriving now, but King resisted treating that as the industry’s permanent baseline.

Still, the supply constraint was described as durable. Hyde said analysts broadly believed the bottlenecks would last through 2027, and she noted that neither Micron nor SK Hynix had a sell rating despite their year-to-date gains. King said Wall Street’s forward estimates reflected a belief that trillions of dollars will be spent on AI equipment and that memory is a fundamental base layer if the economy changes the way Nvidia CEO Jensen Huang has argued it will.

The market’s intraday pause did not undermine that framing. The PHLX Semiconductor Index was down roughly 2.7% during the discussion after a five-day gain of about 13%, while Micron remained slightly positive and SK Hynix had surged 10.86% in South Korea trading. The point was not that semiconductors could not sell off. It was that the market had begun to distinguish the memory bottleneck from the broader AI hardware trade.

The same AI hardware scarcity that is lifting memory names is also intensifying enforcement questions. Hyde reported that Taiwanese prosecutors suspected three individuals of smuggling Nvidia AI chips to China through Japan, citing sources. The case centered on Supermicro servers containing Nvidia hardware.

Michael Shepard said the suspects were believed to have falsified documents stating that Japan was the final destination for the devices when, according to authorities, the ultimate buyer was somewhere in China. Authorities seized about 50 servers before shipment, Shepard said, but at least one shipment appeared to have gone through, based on reporting by Bloomberg colleagues Mackenzie Hawkins and Debby Wu.

The notable detail was the route. Shepard said previous Nvidia hardware-smuggling cases had typically involved Southeast Asian economies such as Thailand and Singapore. Japan had not usually featured in the same way. He said Japan has instead been seen as a place where Chinese companies can legally rent computing power through data centers, a practice he described as within the permissible bounds of U.S. export controls. Instead of acquiring restricted hardware, Chinese companies can access AI compute by paying a rental fee.

This case, as Shepard described it, involved a different allegation: servers physically routed through Japan under allegedly false destination documents. Neither Nvidia nor Supermicro was accused of wrongdoing. Shepard said Nvidia did not respond to a request for comment on the story. Asked about Jensen Huang’s comments while in Taiwan, Shepard said Huang had been asked whether Supermicro needed to do more to address chip-smuggling concerns. Huang’s response, as Shepard summarized it, was that Supermicro may need to “tighten up the ship” on compliance and customer oversight — a rare public remark from Nvidia’s CEO about a partner.

Shepard said Supermicro has insisted its compliance is up to par and that it is doing more to strengthen compliance after earlier charges described in his account. The case, as presented, underscores the practical difficulty of export controls when restricted AI chips may be embedded in complete server systems and when authorities are scrutinizing stated destinations against alleged end buyers.

Nancy Tengler did not reject the AI trade. She called the current market a “productivity-driven bull market” and said she expected it to continue for some time. But her investment posture was not chase-at-any-price optimism. She said Laffer Tengler was “ready for a correction” because markets tend to get one about once a year and the current move had become “pretty frothy.”

Her discipline was position management. Tengler said her firm had added to Micron at 366 only a few months earlier and that she felt late at the time. Given the move since then, the lesson she drew was not to stop owning the winners but to trim gains, wait, and reallocate toward names that can continue to benefit. She described recent market rotations from hardware to software and back to hardware as hedge-fund-driven shifts that long-term investors should not chase indiscriminately.

When Ludlow showed the Nasdaq 100 rising sharply from about 22,500 at the end of March to nearly 30,000, Tengler said her firm would be trimming Micron and likely adding back later. Her warning was behavioral: “Don’t confuse that with being a genius because what the market giveth, it will also taketh away.” The move higher had been rapid and violent, and she argued investors needed to remain nimble.

That did not mean abandoning software. Hyde raised ServiceNow, one of Tengler’s long-term ideas, and asked whether Salesforce’s results could show that AI-disruptable software names could hold up. Tengler said the right question was which platforms would win. She said she was more concerned about Salesforce, which her firm did not own, and noted that the firm had exited Adobe earlier “for obvious reasons.” But she argued that if a company owns the platform, it can be a beneficiary of AI. ServiceNow, in her view, remained well positioned despite being difficult to own.

The next catalyst, Tengler said, may be less about earnings and more about the market’s attention shifting. After Nvidia and the broader earnings season, she expected investors to refocus on the Federal Reserve, inflation commentary, and geopolitical risks. She specifically mentioned whether the Strait of Hormuz is open. Until then, she said, markets would likely see hand-wringing amplified by algorithms and hedge funds. If the Strait of Hormuz opens, she said, investors could see a “melt-up.”

Her economic view stayed constructive. Earnings growth had been “amazing,” guidance had been raised, and she said her firm had remained steadfast in the tech trade for three years while others wrote it off. The correction she anticipated was therefore not a thesis break. It was, in her words, potentially another opportunity like DeepSeek or the first-quarter pullback.

SpaceX’s planned IPO was framed around more than liquidity for a rocket company. Ludlow introduced two linked points: the latest Starship test was a major milestone for the next-generation rocket program, and SpaceX’s IPO filing made clear that Elon Musk, Starship, governance, and AI infrastructure are central to the investment case.

A Bloomberg Intelligence graphic shown during the segment said SpaceX’s IPO filing revealed “significant governance concerns,” including near-absolute control by Elon Musk, limited ability for outside shareholders to seek change, and a $779 billion CEO compensation award. The same graphic said fear of missing out was likely to spur interest among would-be investors. Ludlow described that tension directly: governance concerns could be overshadowed by FOMO around what could be the biggest IPO ever.

Peter Diamandis, XPRIZE founder and early SpaceX investor, argued that the Starship test should be judged by the full-system nature of rocket testing. He said it was a brand-new vehicle with brand-new engines, the most advanced engineering system humans had built, and that rockets cannot be tested “a little bit at a time.” The goal, as he described it, is incremental improvement until the entire vehicle is reusable, refuelable on orbit, and able to serve as a platform for the Moon, Mars, and beyond.

Diamandis’s more expansive claim was that SpaceX should not be understood only as a launch provider. Ludlow pointed to the IPO filing’s discussion of SpaceX’s current compute limitations and future needs, including orbital AI data centers. Diamandis said Musk is building “a vertically integrated satellite network, global broadband, sovereign communications, AI compute and ultimately off-planet infrastructure.” In his view, this is a hyperscaler and infrastructure company, not merely a rocket company.

That claim connected space, AI, and economic competition. Asked by Hyde about a U.S.-China space race, Diamandis said humans are driven by competition and compared the current moment to the U.S.-Soviet race, with one difference: this is not just political, but economic. He pointed to potential revenue engines such as mining the Moon and asteroids for resources and building global compute infrastructure in Earth orbit. He also asserted that the idea of orbital compute had moved quickly into the hyperscaler conversation, before making the further claim that compute from orbit would be needed.

A Bloomberg News expectation card said SpaceX was poised to choose Nasdaq as its listing venue, that the filing could include financial details such as revenue and net income, and that the company was targeting as much as $75 billion in its listing at a valuation above $2 trillion.

Hyde pressed the democratization point: if SpaceX is already so valuable, is much left for retail investors, and how does this solve problems for more than a few early shareholders? Diamandis answered through an AI-abundance thesis rather than a valuation argument. He said the most powerful human asset is intelligence and that AI could increase effective intelligence not by 10 or 100 times, but by millions or billions of times. He connected that to breakthroughs in physics, chemistry, biology, materials, food production, healthcare, education, and longevity.

Diamandis acknowledged that this sounds “techno-utopian,” but argued that current lives already look “godlike” compared with those of parents and grandparents. He cited conversations with Musk in January and March in which Musk, according to Diamandis, talked about double- and triple-digit GDP growth driven by AI and humanoid robotics, with AI and robots creating so much product and availability that humans “could not want enough.” Those were presented as Diamandis’s and Musk’s outlook, not as demonstrated economic forecasts.

The most concrete corporate prediction came when Ludlow asked about speculation that a post-IPO SpaceX could merge with Tesla. Ludlow said he had reported on January 30 that talks had occurred before the xAI transaction. Diamandis said such a merger made sense because it would consolidate Musk’s control. He referred to SpaceX’s super-voting structure and inside owners’ voting control, contrasted that with Musk’s position at Tesla, and said combining the companies would allow Musk to operate across ground and space infrastructure more efficiently. He described a future infrastructure stack including cybercabs, Tesla vehicles with compute and power capacity, and SpaceX assets in orbit.

Diamandis put his prediction plainly: not if, but when. Because both companies would be valued public companies, he said, a merger would be easier than combining private companies and arguing over valuation.

Cognition’s new financing was described in large numbers: more than $1 billion raised at a $25 billion pre-money valuation, or a $26 billion valuation as Ludlow phrased it. Lux Capital, General Catalyst, and 8VC led the round, with a graphic also listing Founders Fund, BCV, and 137 Ventures among investors.

Scott Wu said the raise served two purposes: it let Cognition keep growing aggressively by scaling compute and hiring, and it helped the company remain independent. Independence was not an incidental talking point. It was his answer to a market where AI labs, coding-agent startups, and strategic acquirers are increasingly intertwined.

Wu said Cognition’s growth showed that “AI is doing real work at real companies everywhere.” He said Cognition works with the top five health insurers in the United States, banks, the Treasury, and NASA. His examples emphasized enterprise software work rather than consumer experimentation: health insurers building more tooling for care providers, banks delivering software to customers, and public-sector use cases.

Asked by Ludlow to update Cognition’s revenue trajectory, Wu said the company was close to a $500 million revenue run rate. Ludlow noted that earlier appearances had put the company at a few million dollars of run-rate revenue, and about $37 million a year earlier. A Cognition-sourced graphic gave a revenue run-rate figure of $492 million and said Devin enterprise usage had grown 11 times in six months and 75 times in 12 months.

Wu framed Devin as a compound system that uses multiple models. That was the basis for his “Switzerland” claim. Cognition works closely with OpenAI, Anthropic, Google, xAI, and others, he said, and can choose the best model for a given use case. In his view, customers benefit from a neutral intermediary that is not incentivized to push a single model family.

That neutrality was contrasted with the possibility that major platforms or labs acquire coding tools. Hyde asked about further M&A after disruption around Windsurf assets, IP, brand, and employees. Ludlow then raised a hypothetical: if SpaceX were to acquire Cursor, would that change the field? He noted that Nvidia engineers liked Cursor because they could swap underlying models depending on the coding objective, and he tied that flexibility to questions of data and platform allegiance. Wu did not validate the SpaceX-Cursor premise. He answered more generally that the coding ecosystem was broad enough for first-party products from labs and independent neutral players like Cognition.

The strongest product claim concerned how much code Devin is now producing. A Cognition-sourced chart said the share of code committed by Devin had risen from 13% in December 2025 to 89% in May. Wu said that internally at Cognition, more than 90% of the code the company writes is written by Devin, including work on Devin itself. He said customers use coding agents both on new code and, more commonly, on existing codebases where companies add features and maintain systems.

On labor, Wu rejected a simple displacement framing. Hyde asked what the productivity gains meant for the number of software engineers needed. Wu said the job would evolve and skill sets would change, but he expected far more people building software and products, not fewer. His preferred comparison was historical: there are now roughly 30 million to 35 million software engineers globally, versus under 1 million around the start of the century. The rise in engineers, in his argument, came alongside a massive increase in software produced. If AI makes software creation more efficient, he expects the world to produce more software rather than require fewer builders.

For the rest of the year, Wu gave one target: Cognition intends to cross $1 billion in revenue run rate and then continue expanding to as many companies as possible.

The software debate turned on whether AI agents are creating new revenue quickly enough to offset fears that AI will disrupt application-software incumbents. Salesforce was the test case because its results were due after the close and because, as Brody Ford put it, “Salesforce is the SaaS company.”

Ford described investor anxiety as “SaaS-pocalypse” fears. The only real way Salesforce could push back, he said, would be to show revenue acceleration coming from new AI offerings. He said that had not yet been seen, and investors were hoping for evidence in the back half of the year.

Hyde noted that Agentforce had been discussed as an $800 million revenue stream, while also pointing out that some expected Salesforce revenue growth was tied to Informatica and therefore inorganic. Ford said $800 million a year for Agentforce is “nothing to scoff at,” but the market is punishing application companies whose revenue is decelerating. The debate, in other words, is not whether Salesforce has an AI product. It is whether AI can change the growth trajectory of a mature software company.

Ford said a good Salesforce print could change the narrative for the whole sector because Salesforce stands in for SaaS. If investors believe AI is displacing software incumbents and that innovation is no longer happening at companies such as Salesforce or Adobe, then evidence to the contrary would matter. But the burden of proof is high. Hyde noted Salesforce had hit a three-year low the prior month, had not recovered much, and was down 30% year to date.

Ford’s explanation for slow adoption was practical rather than dismissive. He said the technology is real, but implementation at large corporations takes a long time. Individual users can experiment with ChatGPT-style workflows, but enterprise-wide deployment is far more difficult. Hyde suggested that the same compliance burden may also make Salesforce sticky, because regulated workflows are hard to replace casually.

The contrast with Cognition was implicit in the use cases. Wu described AI coding agents doing production work inside existing codebases. Ford described the difficulty of turning agent capability into enterprise-wide deployment and revenue acceleration at a mature application-software company. Both claims can coexist: AI systems may be doing more work inside companies, while incumbent software vendors still have to prove that agent products can restore the growth investors want.

Marvell appeared as the adjacent infrastructure name reporting after the close. Hyde pointed to photonics, optics, networking, and investor expectations around Marvell. Ludlow said the stock was down sharply ahead of earnings despite being up about 100% for the year. He described Marvell’s XPU strategy as similar to Broadcom’s custom-silicon model: rather than simply selling a standard chip, Marvell partners with hyperscalers or technology companies to build custom silicon. If the world remains in a compute deficit, Ludlow said, Marvell is a likely winner.

The AI infrastructure question extended from chips and software to electricity. The premise presented in Bloomberg’s Primer excerpt was straightforward: rich countries had not had to think much about their grids because electricity demand had been roughly flat since the 2000s. AI data centers, electric vehicles, and broader electrification are changing that.

A narrator said the world is predicted to use twice as much electricity by 2050, roughly equivalent to adding “a whole new USA’s worth of electricity every five years.” A chart attributed to the IEA World Energy Outlook 2024 showed electricity growth rising toward roughly 50,000 terawatt-hours by 2050. Another visual, attributed to Ember and Our World in Data, said China’s power generation is up sevenfold since 2000.

Jason Thomas of Carlyle was shown saying that four companies intended to spend more than $300 billion this year. The excerpt compared grid infrastructure to broadband: costly to build, but capable of enabling trillions of dollars of value. A graphic connected broadband to the valuations of Meta, Apple, Alphabet, and Amazon, all above $1 trillion in the data shown.

The relevance to the day’s chip and AI-market arguments is that compute demand cannot be separated from energy capacity. The memory story was about a component bottleneck. The SpaceX argument was Diamandis’s view of launch, communications, and potentially orbital compute as infrastructure. The Primer excerpt put the same AI build-out into the language of generation, transmission, and grid modernization.

Dan Murtaugh summarized the strategic frame in one sentence: “The battle to build the best grid is a battle to win the future.”

UBS Asia Pacific President Iqbal Khan, interviewed by Bloomberg’s Stephen Engle, focused less on frontier models and more on process redesign inside financial institutions.

Khan said AI’s opportunity is to simplify and speed up processes without cutting corners. His example was documenting an individual client’s source of wealth, which he described as complex. If AI can contextualize that work and create consistency, he said, it can fundamentally improve the process beyond what employees do today.

He said UBS has implemented Copilot across the firm. Khan said he had used it more in the prior six months both personally and professionally and had become more efficient and effective. Over time, he said, “everybody will become an AI-native,” though that depends on adoption and application.

The banking application he described was client onboarding. When a client comes through the door, UBS must determine whether the client is eligible for a service or solution from a compliance and regulatory perspective and whether that service is valuable to the client. Today, Khan said, that process is curated, semi-automated, manual, and people-driven. AI can create capacity by automating and improving parts of it.

Engle pressed the jobs question directly: does top-line job growth slow, or do banks cut back to become more efficient? Khan’s answer was conditional. If UBS and the industry can use the added capacity to serve clients better, gain more share of wallet, and grow faster, then the impact on costs and jobs will be smaller. If firms cannot turn productivity into growth, he said, AI will have ramifications and implications for costs and jobs.

That was the most restrained version of the AI labor argument. Wu argued software creation would expand and require more builders. Diamandis argued AI and robots could drive extraordinary growth and abundance. Khan argued AI can increase productivity and capacity, but whether that capacity protects jobs depends on whether firms can convert it into growth.