AI Demand Pushes Beyond Nvidia Into Power, Memory, and Compute Markets

The pressure around Nvidia’s earnings was not only about whether the company could beat estimates. It was about whether the infrastructure trade around artificial intelligence still has enough supply, margin, power, and geopolitical clarity to justify the scale of investor expectations.

Caroline Hyde framed the market backdrop as a second straight day of pressure for technology shares, with the Nasdaq 100 down roughly half a percent early in the program and the PHLX Semiconductor Index down more than 2%. The chip index, she said, had lost about 6% over two trading days. By later in the program, the Nasdaq 100 was down around 0.8%, the semiconductor index around 2.4%, Nvidia about 1.3%, and Bitcoin was below $77,000, with Hyde pointing to broader risk aversion tied to geopolitics and rising oil.

At Dell Technologies World in Las Vegas, Ed Ludlow said the market needed to reconcile Nvidia’s China story before earnings. In March, Jensen Huang had said Nvidia was ramping its supply chain for H200 after U.S. license approvals. But President Trump, Ludlow said, had told reporters aboard Air Force One that H200s came up with Xi Jinping and that China did not want them. Ludlow described Nvidia as being at the center of a market concern that nothing concrete had emerged from the China visit, while also pointing to broader supply-chain questions after the “melt up” in AI infrastructure shares.

The headline expectations remained strong. Ludlow cited earnings-per-share growth above 80% and revenue growth just below 80% for Nvidia’s Wednesday report. Nvidia shares were up 19.22% year to date and roughly 65% over one year in the market data presented, even as they traded lower intraday.

Mandeep Singh said the core demand signal was still clear: Nvidia’s top-line growth was accelerating and chips remained in short supply. His caution was on margins, particularly from rising memory prices. Nvidia, he noted, had talked about maintaining gross margins in the mid-70% range, and the question was how much memory inflation would pressure that.

The second question was mix. Singh said Nvidia had “almost had a monopoly” in training workloads, but the shift toward inference and reasoning changed the competitive and technical dynamics. Anthropic’s traction with coding agents made it important to know whether Nvidia would still capture the bulk of chip usage in those newer workloads.

Hyde asked about Nvidia’s positioning in inference, referring to “the acquisition of Groq with a Q, not a K,” and what might be said about that partnership and buildout. Singh said Nvidia had discussed Groq as a large inference-side driver, “north of $20 billion” in connection with the acquisition and integration, while also noting that Anthropic had partnered with Nvidia. Given the way the exchange was phrased, the safer reading is not a settled corporate transaction history but the market question Bloomberg was putting to Singh: whether Nvidia’s inference strategy can be large enough to matter as workloads move beyond training.

The inference shift, in Singh’s telling, broadens the discussion beyond GPUs. Agentic AI requires more CPUs, which helps explain why Intel and AMD had been doing well. Nvidia’s report, then, was less a clean demand referendum than a test of how the company is positioned across GPUs, CPUs, memory costs, supply-chain noise, and inference architecture. Bloomberg Intelligence’s earnings setup was similar: fiscal first-quarter results were expected to modestly exceed estimates on Blackwell strength, but attention was moving to incremental upside drivers including Rubin supply-chain noise, Groq LPX, and Vera.

The most concrete sign that AI demand had moved beyond chips was the $67 billion all-stock deal for NextEra to buy Dominion Energy. Hyde called it the largest power deal in history and tied it directly to the AI data center boom.

The on-screen “Big Number” graphic made the same point plainly: “$67B NextEra to Buy Dominion to Form Utility Colossus.” The deal was presented as creating a utility spanning from Florida to Virginia’s technology hubs, with Virginia’s Data Center Alley central to the industrial logic.

Liana Baker said she was surprised it had taken this long for a major utilities deal to emerge. Demand for power, she said, was at levels not seen since after World War II. Utilities needed scale to pay for the buildout required by AI.

Baker said Dominion’s footprint could make regulatory approval more manageable because the target was apparently based in only three states. Analysts she cited did not view the review as trivial, especially with other power and nuclear regulators involved, but she said it did not look like an impossible “uphill battle.” Under the Trump administration, she added, now was the time for companies to “go bold and go big” on dealmaking. The estimated timing she gave was roughly one year to 18 months.

The tension is consumer cost. Hyde noted the political pressure around energy prices, especially because of AI buildout. Baker said the companies were already emphasizing credits for Dominion customers, a signal that they understood the risk of consumers viewing the transaction as a path to higher utility bills. Whether prices ultimately rise, she said, remained to be seen.

If power is one constraint, rentable compute is becoming another. CME Group and Silicon Data plan to launch what Hyde described as the world’s first futures market for AI computing power later this year, pending regulatory review.

Carmen Li said a compute futures market is necessary because the AI infrastructure stack has become commodity-like: energy, colocation, servers, and GPUs all have spot-market characteristics. Banks underwriting trillions of dollars of loans need ways to manage risk exposure and hedge future volatility as compute comes online. Li compared the need to oil futures, which she said help users hedge oil-price fluctuations.

The likely participants fall into two broad groups. First are those naturally long GPUs: data centers, neo-clouds, design houses, and fabs. They need ways to hedge through short futures or put options. Second are compute buyers, including AI startups, whose GPU and token costs have become major line items. Li contrasted that with traditional SaaS businesses, where people were historically the main cost line.

Silicon Data already operates Compute Exchange as a spot market for GPU resources, including forward and reserve contracts. Li said its participants range from AI startups to neo-clouds in Europe, the U.S., and Southeast Asia. The practical need is rate certainty: customers want to lock in longer-term contracts because on-demand GPU pricing can show “40% daily volatility,” according to Li. Her distinction was straightforward: buyers needing physical delivery or reserved GPUs should use Compute Exchange; those hedging financial exposure should use CME futures in the future.

The Silicon Data H100 Index shown in the program tracked GPU rental costs in dollars per GPU per hour from June through May and showed prices rising in recent months, from roughly the low-$2-per-hour range toward the high-$2 range. Li said prices had been going up since December. The idea that compute prices would simply fall toward zero, she said, had not materialized because the supply-demand curve and forward expectations keep shifting. Last year the bottleneck was fabs; this year it was memory; in the future it may be colocation space.

Li also emphasized that GPU pricing is not simple enough for a raw average. GPUs are not homogeneous: H100s can differ by memory configuration and location, and reserved contracts can run three to five years. Silicon Data’s role, as she described it, is to normalize those differences into indices that reflect the prices users actually pay. The company launched GPU indices on the Bloomberg Terminal last year, covering prices outside China and normalizing geolocation across Europe and the U.S. into one data point.

On regulation, Li said Silicon Data and CME were working with the CFTC and other bodies. She described the product as a traditional data futures product, not an exotic structure, and said she did not see a major concern.

The scarcity of compute is not only a market structure problem. It is also an internal corporate problem for AI labs. Hyde introduced Google’s AI bottleneck as a case in which computing power has become a prized resource inside the company, with priority increasingly going to major projects like Gemini and paying cloud customers. She said the bottleneck had frustrated some employees and prompted AI researchers to leave for startups.

Julia Love described a shift from the earlier culture at Google Brain, where AI research was more egalitarian: researchers received allocations of compute and could combine them for larger projects. That pool still exists, she said, but contracts during periods of high demand. Researchers whose work is not a company priority often cannot get meaningful research done with their allocations.

The causes, according to Love, are overlapping. Google’s generative-AI consumer products require large amounts of compute. The cloud business is taking off, and Google wants to meet that demand. At the same time, cutting-edge AI research has become much more compute intensive. “There’s a lot of mouths to feed,” she said before the segment was interrupted by a technical issue.

Hyde pointed to researchers leaving to found companies because they could not get what they wanted done internally. Love said that in a startup, those researchers have more doors to knock on, rather than depending on an internal allocation process that may favor Google’s highest-priority products and customers.

Recursive, a startup led by Richard Socher, emerged from stealth with more than $650 million raised and a $4.65 billion valuation. Its round was co-led by GV and Greycroft, with backers including AMD and Nvidia.

Socher described the company’s goal as building “recursive, self-improving superintelligence that automates knowledge discovery.” His core claim was that because AI is code and can code, it can be put into a closed loop: generate ideas, implement those ideas, validate them, and improve itself.

He argued that Recursive’s edge is organizational as much as technical. The company was built from the ground up for open-endedness, with the entire company focused on letting AI build the next better version of AI. The founding team, he said, includes people from Google DeepMind, OpenAI, Salesforce Research, Meta, and other labs. Existing frontier labs, in his view, started before this kind of AI-driven experimentation infrastructure was possible.

Compute is central to the thesis. Socher said Recursive is trying to prove a “big new scaling law”: more compute leads to more inventions and improvements. Compute is one of the company’s largest costs, which is why he said investment from Nvidia and AMD was important.

Hyde pressed on safety, given the open-ended ambition. Socher said safety was a “huge concern” and pointed to co-founders who had worked on “rainbow teaming,” where AI improves the safety of large language models. He said Recursive believes its approach is the fastest path to superintelligence and to what he called a “Eureka machine” that could invent everything else afterward.

He also addressed the practical question of time. Socher is CEO of you.com and involved with AIX, a venture firm. Asked how he divides his time, he said, “I work all the time,” supported by both AI agents and “incredible humans.” On compensation, he said Recursive shares equity broadly because founders should care more about the overall potential of the company than their own slice.

At Dell Technologies World, Daniel Roberts of IREN gave the infrastructure operator’s version of the same constraint. IREN is building AI cloud infrastructure across geographies, with Dell as a key server partner and Nvidia GPUs central to its systems.

Roberts said the public debate often underestimates the physical nature of the bottleneck. “You can’t implement a software update to bring on power,” he said. “You can’t code your way to an AI gigawatt of compute.” For IREN, the focus is power and “time to compute.”

He argued that IREN avoided some of the backlash facing data-center builders by choosing regional communities with abundant renewable energy rather than metropolitan areas. The strategy, begun eight years ago, created a structural cost advantage and, in his telling, reinvigorated local communities in British Columbia, Canada, and Texas. Roberts said IREN is located where communities want it to do business.

Demand, however, is outpacing supply. Roberts described the industry as still being in the “dial-up era of AI.” Users still wait for prompts, carefully craft them, and sometimes wait 15 to 40 seconds for complex answers. As response times fall, he argued, demand will accelerate further, which will drive more compute and more use cases.

The limiting factors are steel, concrete, power, and permitting. If a company started today to build a gigawatt AI factory, Roberts said, it would be looking at the 2030s before getting the first compute online. More specifically, it could take 18 to 24 months just to get a utility’s attention and determine whether a piece of land has power. He did not blame utilities directly, saying they run complex networks and are asked to underwrite 24/7 guaranteed power when they connect large loads. That makes them naturally risk-averse, especially while facing a flood of requests.

Roberts said IREN has a strong relationship with both Dell and Nvidia. In his words, IREN had “announced a partnership with Nvidia for 5 gigawatts of compute” and was working on Nvidia’s reference architecture for a gigawatt factory in Sweetwater. Ludlow noted that Nvidia had also made an investment; Roberts said that gave Nvidia exposure to IREN’s upside as IREN grows and deploys Nvidia GPUs. The way Roberts framed it, the point was the broader buildout ecosystem: land, steel, concrete, servers, GPUs, and end users all have to line up.

Asked where the best geography is for data centers, Roberts said each region differs, but the majority of IREN’s capacity is in Texas, which he said had been “pretty good.” Asked what he would ask Jensen Huang and Michael Dell, Roberts said the industry is moving from a world in which silicon is the choke point to one where the questions are how to solve steel, concrete, and kilowatts.

The same AI hardware demand that is pushing up memory prices is also creating labor and political pressure in South Korea. Peter Elstrom said Samsung was in tense negotiations with its union, which had threatened an 18-day strike as soon as Thursday. Samsung shares rose nearly 3% intraday after signs the union was open to talks and a local court moved to limit the threat of a potential strike.

Elstrom said the timing matters because Samsung is benefiting from a boom in memory-chip prices, and the AI industry needs more memory from Samsung and SK Hynix. A labor disruption could therefore affect not only Samsung and South Korean companies but the AI industry more broadly.

He also stressed that this level of labor conflict is unusual for Samsung. For decades, he said, the company had no real conflict with its union. A smaller strike a couple of years ago was not as serious. This situation had progressed further, though talks had resumed.

Hyde broadened the issue to the politics of AI profits, saying South Korean markets had lost “hundreds of billions of dollars in just 90 minutes” the prior week as investors questioned how AI profits might be shared with the country more broadly. Elstrom described a wider debate in Korean society over what to do with enormous AI-related profits and how governments should prepare for a world in which AI changes employment. A policy adviser to the president, he said, had floated the idea of a citizen dividend funded by profits from large companies. The possible uses included retraining, support for startup founders, and other social-stability measures. Elstrom characterized the proposal as an attempt to start a debate over how government should use an AI profit windfall.

Several shorter segments pointed to the same broad market question: how investors, companies, and product builders are repositioning around AI, even when the topic was not directly chips, power, or data centers.

Apple’s revamped Siri was described as an attempt to differentiate in a crowded chatbot market through privacy rather than maximal retention. Mark Gurman said Apple has treated privacy as a differentiator for more than a decade and presents it as a “North Star” whenever it releases new products, operating systems, software, or services. The new Siri, expected at Apple’s June 8 developer conference, will be marketed against rival chatbots by retaining less information and less “memory.” One feature Gurman reported was auto-deletion of chatbot conversations: users could save chats forever, delete them after a year, or delete them after 30 days. He said Apple plans to retain memory for a shorter time than many competitors, which could make Siri useful in the short term but less useful in the long term.

SpaceX was presented as a potentially historic listing story. Elon Musk said virtually at Samsung’s International Smart Mobility Summit in Tel Aviv that SpaceX needed to get its IPO going “pretty soon.” Hyde said Musk was back in Texas working on plans for the IPO and that a filing could come as soon as the week of the program. SpaceX had also notified investors it was executing a 5-for-1 stock split, lowering the fair market value of each share to about $105 from roughly $527, according to sources cited by Bloomberg.

The corporate updates were also framed around AI demand. Baidu’s sales decline was less than analysts expected as it shifted toward automated agents and AI-powered cloud units. Publicis agreed to buy LiveRamp for about $2.5 billion in cash, with the stated aim of expanding data assets and accelerating AI-driven marketing. Kioxia shares rose after forecasting an $8.2 billion quarterly operating profit, which Hyde described as another sign of AI hardware demand and a shortage in conventional memory chips.

In public-market positioning, Hema Parmar said new 13F filings showed divergence among hedge funds and money managers around large technology names. Bill Ackman took a Microsoft stake worth more than $2 billion, making it his fourth-largest holding. Parmar said Ackman appeared to view Microsoft as stronger and more resilient than investors believed, with Microsoft 365, Word, and Excel among the most valuable enterprise-tech franchises.

Other technology-focused funds moved the opposite way. Tiger Global and Coatue cut Microsoft stakes by more than half, while Lone Pine exited entirely. Parmar noted that Microsoft fell 23% in the first quarter before rebounding somewhat. Amazon was also a popular sell: Tiger Global trimmed its stake, Coatue cut Amazon by about a fifth, Lone Pine exited, and Berkshire Hathaway also sold out.

The filings, Parmar cautioned, do not show when in the quarter funds sold. The first quarter was volatile, especially March, and some funds lost money, which may have influenced positioning. But the pattern showed that even within the AI and big-tech trade, major investors were not moving in one direction.