Cerebras Seeks $4.8 Billion as AI Compute Demand Lifts IPO Market

Ed Ludlow said Cerebras was seeking to raise as much as $4.8 billion in an upsized IPO, one-third more than previously planned. The company, described as both an AI chipmaker and data-center operator, had increased both the number of shares it planned to sell and the expected price range.

Bloomberg AI reporter Rebecca Torrence said the IPO had been “more than 20 times oversubscribed.” The prior price range of $115 to $125 a share had moved to $150 to $160, and at the top end Cerebras would be valued at more than $34 billion. Torrence said that would make it the biggest IPO of the year “by a long shot” and called the demand a clear sign of appetite for AI compute and infrastructure.

Caroline Hyde asked whether investors were showing concern about Cerebras’ ability to win market share or manage supply-chain risks. Torrence said she was seeing “less nerves than just pure unfettered excitement,” helped by pent-up demand for new listings and Cerebras’ partnerships with Amazon and OpenAI. She said the IPO was expected to price on Wednesday afternoon, May 13, with trading expected to begin Thursday.

The offering landed in a market already leaning toward AI exposure. Hyde said the Nasdaq 100 was on track for another record and a possible seventh straight week of gains, which would be its longest winning streak since October 2024. That was happening despite concern over the Strait of Hormuz, Iran, the United States and peace talks. In her framing, tech was moving higher through the geopolitical noise.

Carol Schleif argued that the AI trade was being supported by earnings as well as anticipation. Markets had risen, she said, but earnings had risen too: with the earnings season nearly complete, top-line and bottom-line growth were running at double-digit rates, and not only in technology.

Schleif did not portray clients as uniformly relaxed. BMO Wealth Management had “some nervous clients,” she said, who did not want to “get off the bus” but were asking how markets could sit at all-time highs while short-term risks remained so visible. Hyde separated those near-term risks — oil, helium and other materials moving through the Strait of Hormuz — from the longer-term demand for AI infrastructure. Schleif agreed with the distinction.

Her case was that capital infrastructure spending at this scale tends to support productivity and GDP performance. She also said the market’s rise was not detached from revenue, earnings or margin support. Well-managed tech companies were watching margins, headcount, costs and inputs. The main issue, she said, was not demand but capacity.

“It’s not like we’re going to have dark data centers,” Schleif said, with one exception: “Other than the fact that if you can’t get energy to turn them on.”

That made power central to the AI-infrastructure trade. Hyde pointed to nuclear power and Constellation Energy, whose shares were down about 4.4% even after first-quarter operating revenue beat expectations. The pressure, she said, came from nuclear refueling outages. For companies tied to data-center growth, execution on energy supply had become part of the investment case.

A Bloomberg Intelligence graphic sharpened the concentration risk: “AI is increasingly eating the global earnings cycle. From Silicon Valley’s hyperscalers to Seoul’s memory-chip giants and Taiwan’s semiconductor supply chain, a narrow group of AI-linked companies is doing much of the heavy lifting for global profit growth while large parts of the broader market struggle to keep pace.”

Schleif said the market action underneath the indexes supported that view. Even with risks around the Strait of Hormuz and oil flows to Southeast Asia, AI-linked supply-chain stocks were rallying because investors saw the theme as larger than the immediate shock. She described the moment as “grabbing the tiger by the tail and just hanging on for the ride,” but said it had fundamental underpinnings and reached beyond a small set of companies.

She also pushed back on the idea that the market’s strength came only from a narrow megacap group. Citing Bloomberg data, Schleif said eight of 11 sectors had double-digit revenue and earnings growth. Her advice was not to chase IPOs, but to rebalance strategically: trim some winners, maintain portfolio balance, and look for second- and third-order beneficiaries in energy, space, supply-chain rebuilding and reconstruction after conflicts.

Supply chains remained a geopolitical issue. Hyde cited an expected China meeting between President Donald Trump and President Xi, with executives including Tim Cook, Elon Musk, Micron’s Sanjay Mehrotra, Qualcomm’s Cristiano Amon, Illumina’s Jacob Thaysen, Meta’s Dina Powell McCormick and Cisco’s Chuck Robbins expected to join Trump. Schleif said clients had been watching a “geopolitical pivot” and “geo-economic pivot” for roughly a year and a half. In her account, the United States and China both have leverage and vulnerabilities: China around rare earths, the US around energy and materials.

Circle’s market move and quarter were presented through financial results first. Hyde said Circle shares were up about 9% after a 20% revenue surge, while net income declined. A Bloomberg Tech market graphic said first-quarter profit was hit by stock-based compensation costs. Another Bloomberg Tech graphic listed first-quarter total revenue and reserve income of $694 million, stablecoin market share at period-end of 28%, and USDC minted of $73 billion.

Jeremy Allaire used those results to make a broader strategic argument. He said Circle’s public-company strategy was to build “the world’s largest stablecoin network” and make USDC the “highest utility form of digital dollar money in the world.” According to Allaire, USDC had become the leading dollar digital currency. He said that, based on third-party data, Circle saw almost $30 trillion of on-chain USDC transactions in the first quarter, accounting for 80% of stablecoin transaction volume.

Allaire said Circle was also building platform pillars alongside the stablecoin business: CPN, its payments network, and Arc, which he described as an “economic OS” for a future in which transactions, financial services and broader economic activity are digitally and software-mediated. He tied that future directly to AI operating systems and agentic systems.

Asked when those investments would bear fruit, Allaire pointed to Arc. Circle had presold $220 million of Arc tokens, led by a16z crypto and joined by Apollo, BlackRock, Standard Chartered, Intercontinental Exchange and others, he said. Allaire said Circle expected to monetize Arc through its stake in the network, partner programs, validation, transaction fees and services built on top of the infrastructure. CPN’s annualized volume, he added, had risen about 75% since Circle last reported.

Ludlow pressed the vulnerability in Circle’s model: USDC circulation can lift revenue, but lower interest earned on reserve assets can pressure profit in a lower-rate environment. Allaire rejected the idea that lower rates necessarily weaken Circle’s growth. Since the yield curve began coming in during December 2023, he said, rates had fallen more than 40%, while USDC circulation had grown by “multiple hundreds of percent” and transaction volumes had grown sharply.

Allaire’s argument was that lower interest rates can increase money velocity and demand for money, supporting stablecoin usage. He cited comments from the Treasury Secretary about trillions of dollars of stablecoins in circulation and regulation such as the Clarity Act as signs that adoption would be driven by utility.

Bloomberg’s Stacey Smith said there are signs AI is boosting productivity, but the same numbers may also reflect job displacement. First-quarter 2026 productivity came in at 2.9% year over year, higher than expected. Smith called it a “good news, bad news situation”: productivity looks strong, while the labor market looks weak.

Ludlow asked whether the productivity data showed a clear AI contribution. Smith said it was difficult to know where the gains were coming from. She described a bet between Stanford economist Erik Brynjolfsson and Northwestern economist Robert Gordon over whether productivity would grow by an average of 1.8% between 2020 and 2030. The current trend looked likely to beat that level, she said, but the reason remained disputed.

Gordon, according to Smith, saw layoffs as a possible explanation. Productivity is output divided by hours worked. If companies reduce headcount, the denominator falls, and measured productivity can rise even without the technology-led efficiency gain AI optimists expect.

A displayed Bloomberg Businessweek excerpt added the caution that companies are investing heavily in AI, but productivity growth will likely remain modest until they redesign workflows around the technology. The same excerpt cited an MIT study from the prior year finding that, despite large AI investments, 95% of businesses reported no measurable return on investment.

Alphabet’s market capitalization was shown at $4.75 trillion, close enough to Nvidia to raise the possibility that Google’s parent could overtake the AI-chip leader as the world’s most valuable company. Ludlow said Nvidia still held that title at about $5.4 trillion. The day’s trading moved against the headline — Alphabet was down almost 2% while Nvidia was up more than 3% — but the gap had been closing.

Ryan Vlastelica said investors had developed a broader appreciation for Alphabet’s position across AI. He listed Gemini, accelerating growth at Google Cloud, AI-supported trends in Search and YouTube, Waymo in physical AI, and Alphabet’s TPU semiconductor business. Alphabet had said it might start selling TPUs to other cloud companies, he said, putting it more directly in Nvidia’s territory.

“When you add all of that together,” Vlastelica said, “it just seems like they are really the company of the AI era.”

Hyde noted that analysts remained positive even after a steep share-price rise, with zero sell ratings and 18 buys. Vlastelica said recent results reinforced the breadth of the story. Portfolio managers increasingly viewed even a stronger OpenAI model as evidence of rising AI adoption that could still benefit Google Cloud and Alphabet’s chip business.

That breadth was the contrast with Nvidia. In Vlastelica’s account, Alphabet has multiple AI-linked revenue sources that can offset weakness in any one area. He said skepticism from a year earlier — especially concern that AI competition would erode Search — had not really played out. Instead, Alphabet had shown accelerating growth and strength across markets. Its valuation was elevated versus its own history, he said, but not “dot-com priced for perfection.”

Bloomberg’s Margi Murphy said Google researchers believed they had uncovered the first-ever zero-day attack built by artificial intelligence. Hyde framed it as the kind of scenario the White House had warned about: AI tools being used to build serious attacks against widely used software.

Murphy said Google had identified a prominent criminal group using an LLM to find an exploit in a popular software tool. Google foiled it in advance and notified the software developer, which issued a patch. The case still mattered, she said, because it showed that AI-related security concerns were already real.

Ludlow asked for the basic definition of a zero-day. Murphy explained that it is a flaw in software the developer does not know about. Once a hacker identifies it, the developer has “zero days” to fix it. If LLMs can scan software and find previously unknown flaws quickly, the race between attackers and defenders changes.

The risk, Murphy said, is speed. AI-assisted vulnerability discovery could increase the number of exploits and, eventually, the number of hacks available to criminal or espionage actors.

Google did not name the LLM it believed was used. Murphy said Google was clear that it did not believe the tool was Mythos or Gemini, but was keeping the model’s name private to protect sources. Google also did not name the software developer or the criminal group. That made the case harder to assess in detail, Murphy acknowledged, but she said the broader point did not depend on which model was used: available LLMs are already being abused this way.

Rezolve AI CEO Daniel Wagner framed the company’s hostile bid for Commerce.com as part of a land grab in AI-enabled retail. Commerce.com’s board had rejected an earlier offer and, according to a Bloomberg Tech graphic, called the revised proposal — one Rezolve share for every two Commerce.com shares — “even less favorable” and at a “significant discount” to the company’s current market valuation.

Wagner said Rezolve had gone from nothing to $232 million of contracted annual revenue in one year, with $60 million of revenue in the first quarter versus $46 million for all of the prior year. He described Commerce.com as a stuck business with 60,000 merchants and a forecast growth rate of 1.5% over the next 12 months. Rezolve did not need Commerce.com, Wagner said, but could unlock faster growth from that merchant base if it acquired the company.

Asked how Rezolve fit with Circle’s stablecoin vision, Wagner distinguished tokens from infrastructure. Circle, he said, is a coin or token; Rezolve is trying to provide the infrastructure that lets tokens be used in everyday commerce. Stablecoins are not how a customer buys coffee at Starbucks today, he said, because crypto remains “frustrating and clunky.” Wagner said RezolvePay would make stablecoin payments as efficient for retailers, brands and consumers as credit cards.

Ludlow asked the competitive question: if whoever controls the agent layer wins in commerce, why would Amazon, Alphabet or OpenAI not simply do the same thing?

Wagner’s answer was that general-purpose generative AI remains too prone to hallucination for retail sales. He said OpenAI had done a highly publicized deal with Walmart five to five-and-a-half months earlier, and he claimed that the deal had been disbanded two or three weeks earlier because it made “terrible mistakes.” Wagner described generative AI as probabilistic and said it can guess wrong. In retail, he argued, wrong answers can create returns, mislead customers into buying the wrong products, or insult people.

Rezolve’s claim, as Wagner presented it, is that it spent 10 years preventing hallucinations in this vertical. Wagner said Microsoft and Google had endorsed Rezolve’s belief that it has the only reliable technology for merchants in this area. His formulation was that an AI system meant to act as “the best salesman on the planet” cannot hallucinate.

Apple’s role was narrower but still connected to the AI stack: some users are moving agent workloads onto local machines rather than sending every request to cloud servers.

Mark Gurman said Apple was planning a slight redesign for macOS 27 to smooth visual quirks introduced by macOS 26 Tahoe, especially shadows and transparency in the Liquid Glass interface. Gurman said Liquid Glass looked better on iPhone and iPad than on Macs using older LCD technology, and that macOS 27 was expected to clean up transparency, icons and shadows. That design issue was mostly a consumer-software matter, but it sat beside a more material AI hardware point: Apple’s machines are being used in ways the company had not necessarily centered in its own AI story.

Austin Carr said the Mac mini had become a go-to device for running standalone AI agents. Users were buying Mac minis and higher-end Mac Studios, installing agent frameworks such as OpenDevin, and using them as 24/7 personal AI systems at home. A Bloomberg Tech graphic summarized the appeal as around-the-clock operation, low electricity cost and the ability to perform tasks with little human oversight.

Carr called it a rare AI bright spot for Apple because the company had not really foreseen the trend. The hardware turned out to be well suited to local AI computing because Apple’s unified architecture and memory bandwidth can handle workloads that previously would have been less efficient locally. At the same time, he said, AI models are becoming smaller and more efficient.

The use cases were practical: opening a browser, checking files, interacting with business systems and apps, or running a “headless Mac” plugged into a wall as a local server. Carr said some people were using the setup as an AI developer assistant, while others were building household chief-of-staff or butler-like systems to monitor grocery lists and manage local files. He said the trend could grow if Apple eventually builds its own OpenDevin-like system.