AI Demand Is Rewriting Tech Financing From Hyperscalers to IPOs

Alphabet’s planned $80 billion equity raise was treated as a signal about the financing burden of AI infrastructure, not just a single-company transaction. Ed Ludlow described the package as one of the largest equity deals of all time and framed it as a capital-expenditure story: Alphabet has already told investors its AI spending will rise, and Bloomberg Intelligence is modeling substantially larger spending ahead.

The proposed raise has three parts: a $10 billion Berkshire Hathaway investment, $30 billion in public offerings, and a $40 billion at-the-market program that would allow Alphabet to sell shares from time to time beginning in the third quarter. Caroline Hyde noted that the structure appeared to distribute the issuance in novel ways, perhaps with dilution in mind. Alphabet shares were modestly lower intraday, down about 0.9% when the transaction was being discussed.

Robert Schiffman of Bloomberg Intelligence called the raise “wildly bullish,” not because equity is conventionally cheaper than debt, but because of what it suggests about Alphabet’s confidence in returns on AI infrastructure. Schiffman said the decision runs against what he learned in “math camp,” where weighted-average-cost-of-capital logic would normally point a company toward issuing more debt. In his view, choosing equity instead implies Alphabet believes the return on AI investments can justify the scale of the raise.

Schiffman connected the move to a much larger funding cycle among hyperscalers. He said the biggest hyperscalers are spending close to $800 billion cumulatively this year, that spending will be “well over a trillion” next year, and that cumulative spending could reach $5 trillion over the next five years. The result, he said, is a global search for capital across public debt, private debt, equity, and currencies.

The financing choice also matters for creditors. Bloomberg Intelligence’s analysis said Alphabet’s equity raise was a positive for bondholders because it avoids incremental debt and could reshape financing decisions among the sector’s largest issuers. Schiffman said Oracle had already moved in a similar direction earlier in the year with a planned $25 billion equity issuance.

The tradeoff is buybacks. Schiffman said companies are unlikely to issue $80 billion of stock and then quickly buy back shares. He said buybacks have already slowed and expects more slowing over the next year as spending rises and companies preserve balance-sheet capacity for AI infrastructure. The underlying reason, in his view, is not simply fear of missing out. He said confidence in AI monetization is “finally coming to fruition,” citing first-quarter results.

A chart using consensus estimates and reported financials put that pressure in financial form: AI capex among Amazon, Alphabet, Microsoft, Meta, and Oracle was shown rising while free cash flow faded through 2027 estimates. The point is not that every hyperscaler will finance itself the same way. It is that AI infrastructure has become large enough to alter how the largest technology companies think about balance sheets.

Anthropic has confidentially filed a draft S-1 to go public, putting it ahead of OpenAI in an increasingly explicit IPO race. Shirin Ghaffary stressed that the filing remains short on details: no public financial statements, no price range, and no offering terms. But the act of filing matters because OpenAI is also expected to move toward a filing, while SpaceX is already preparing for a listing.

Ghaffary said OpenAI CEO Sam Altman had reacted by saying the company would go public when it wants to and not before it is ready. That left uncertainty around OpenAI’s timing, while Anthropic has already taken the formal confidential-filing step.

What is known, according to Ghaffary, is that Anthropic’s revenue has been growing quickly while costs remain high. She said Anthropic had reached an annualized run-rate revenue projection of $47 billion and that the company’s CEO had described an 80x increase in that projection in the first quarter on an annual basis. The claim was presented as evidence of extraordinary growth, not as a full picture of the company’s economics: Ghaffary paired it with the reminder that costs remain high.

Hyde asked whether it is useful to treat these listings as a race. Ghaffary said the dynamic is hard to deny: SpaceX, OpenAI, and Anthropic are all either going public, discussing it, or planning it within a span of months, with some filings potentially occurring within weeks. Going first may help shape the market narrative and may let earlier issuers absorb more of the available investor demand for AI stocks. But Ghaffary also noted the other side: public investors have had limited access to “core AI-only” generative-AI companies, so pent-up demand may be large enough that the first-mover advantage is uncertain.

That uncertainty became sharper in the PitchBook analysis. Emily Zheng said the numbers are in “uncharted territory.” Over the past 10 years, she said, total U.S. VC-backed public listings amounted to $1.5 trillion. SpaceX alone would exceed that decade-long figure by valuation. On proceeds, she said SpaceX is looking at $75 billion, while OpenAI and Anthropic together are looking at $100 billion, again exceeding all IPO proceeds raised over the past decade.

PitchBook estimated U.S. VC-backed IPO values at $3.807 trillion for 2026, including estimates for SpaceX, Anthropic, and OpenAI. Prior years in the same comparison were far smaller: $178.9 billion in 2020, $518.4 billion in 2021, $6.7 billion in 2022, $26.1 billion in 2023, $41.4 billion in 2024, and $105.3 billion in 2025.

Zheng said the timing cannot be assumed simply because filings exist. SpaceX is expected to price first; Anthropic’s confidential filing is necessary but not a guarantee of a listing in the next few months or even the next year. She said all eyes will be on SpaceX because if its IPO “flops,” Anthropic may wait for volatility to ease.

Hyde raised the circularity problem in AI markets: Anthropic’s compute spending and its reliance on providers such as Google or Amazon will become clearer when filings are public. Zheng said the AI cycle is highly interconnected, with roughly 30% of U.S. VC-backed startups AI-native or AI-adjacent. That makes the durability of AI valuations and pricing a central question for venture capital, because so much of the market depends on the same technology progressing and eventually becoming profitable.

Asked by Ludlow whether there is precedent for three companies trying to occupy the same lane at this scale, Zheng said she did not think there was a similar public-market precedent for such a transformative technology.

SpaceX is negotiating to pay one of the lowest underwriting fees ever for what is expected to be the largest IPO ever. Katherine Doherty said the company is aiming for less than 0.75% on a planned $75 billion raise. She stated that, split among more than 20 banks, the fees would be $500 million, with Morgan Stanley and Goldman Sachs leading the transaction and receiving the largest share.

The leverage sits with SpaceX. Doherty said banks want to be attached to the listing because it is a historic transaction and because it may lead to future IPO assignments and fee opportunities. Ludlow said people on the SpaceX side frame it bluntly: Elon Musk is a very good negotiator. But even if the final fee lands around 50 or 60 basis points, he suggested, the banks will still “do fine” on a $75 billion raise.

SpaceX is also the first major test of the IPO wave that Anthropic and OpenAI may follow. The displayed terms said SpaceX had filed for a Nasdaq IPO, aims to raise a record $75 billion, targets a valuation of at least $1.8 trillion, would be the largest IPO of all time, and that the filing reveals Musk’s voting control at 85%.

Another set of reported estimates, attributed on screen to Bloomberg News and The Information, grouped SpaceX, Anthropic, and OpenAI into a “2026 IPO boom.” It listed SpaceX at $75 billion in June, Anthropic at $60 billion in the fourth quarter of 2026, and OpenAI as “TBA” in the fourth quarter of 2026.

Doherty said Wall Street has been preparing for this moment after repeatedly talking about “green shoots” and strong pipelines. The pressure now is to move beyond pipeline language into actual deals and billions of dollars coming to public markets. The SpaceX listing, in that framing, is not just a fee event. It is a test case for whether the AI and space IPO wave can maintain momentum.

Tom Mueller presented Impulse Space as a company that starts where launch ends. SpaceX Falcon and other launch vehicles get payloads to low Earth orbit; Impulse wants to move them within orbit, to other orbital planes, to geosynchronous orbit, to the moon, to Mars, and beyond.

Impulse has raised $500 million at a $4 billion valuation to scale spacecraft that can haul satellites between orbits. Mueller described the company’s role as building the “highways” of the space economy. Impulse was described as having three active Mira flights, a planned Helios debut in 2027, and a new 240,000-square-foot production hub.

The key technical claim concerned speed and energy. Mueller said satellites headed to geosynchronous orbit, about 22,000 miles out, typically get dropped into a lower orbit or transfer orbit and may take months to reach the final destination. Helios, he said, can get them there in one day. He described it as “basically a rocket on a rocket”: a large propellant tank and a high-energy pump-fed engine that performs a couple of burns to reach high-energy orbit. He said the same system can take much more payload to the moon, Mars, or the outer planets.

The SpaceX IPO, in Mueller’s view, would have an industry-wide effect. Hyde asked about the “SpaceX mafia” that could be created by liquidity from a listing. Mueller said the moment is good for building the space economy and argued that “the true space age is starting now.” He pointed to the possibility of building megastructures in space, including “a million data servers,” and using lunar resources, a theme he said he had discussed since before founding Impulse. He also said NASA’s announced permanent moon base aligns directly with what Impulse wants to enable.

Asked to estimate the economic impact of SpaceX’s listing on the industry, Mueller called it “gigantic.” He said he grew up expecting a Star Trek-like future and still believes the world will get there. In that imagined economy, space is a major share of global activity. Today, he said, the space economy is only a single-digit share of the global economy, but he expects it to grow faster than any other sector by that measure.

On defense applications, Mueller was more restrained. Hyde asked about administration spending on a “Golden Dome” and interceptor work with Anduril. Mueller said Impulse is there to provide advanced solutions, including helping the government move around and protect assets. When Ludlow asked whether a space-based interceptor demonstration could occur within 12 to 18 months, Mueller said he could not discuss program specifics.

Perplexity’s announcement with Intel at Computex centered on what Ludlow called “the world’s first hybrid local server agentic inference orchestrator,” a phrase he joked sounded AI-generated. Aravind Srinivas accepted Ludlow’s simplified description: the orchestrator decides whether an AI workload, or part of one, should be handled locally on device or at the edge, or sent to cloud servers for more powerful compute.

Srinivas said the premise is that not all compute should be centralized in giant servers and run through the largest frontier models. He pointed to concern over token costs and argued that customers need “efficient token value per watt per user.” That requires simultaneously managing privacy, accuracy, intelligence, and cost.

Perplexity’s proposed answer is a hybrid system that routes across models, files, tools, chips, and servers. Srinivas said the software must decide when to use a local file system, a local subagent model, a local LLM, or a frontier model, depending on the task, the prompt, and the confidentiality or sensitivity of the files and applications involved. He compared it to an operating system balancing objectives at once.

Although Perplexity demonstrated the system with Intel, Srinivas said the company is chip-agnostic. It works with Intel and Nvidia RTX, he said, and Perplexity intends to be chip-agnostic in the same way it has been model-agnostic.

That model-agnostic posture shaped his answer on competitive threats. Hyde asked what it means for Perplexity if Anthropic, OpenAI, and SpaceX all become public companies. Srinivas said Perplexity “loves” Anthropic, OpenAI, xAI, and other frontier labs because every improvement in their models improves Perplexity’s unified system, which routes across them. He said Anthropic’s model improvements since the beginning of the year had helped Perplexity’s own revenue triple in five months.

Hyde pressed him on the number. Srinivas said Perplexity had crossed $500 million around mid-April, after reports that the company was at about $450 million in March, but said the company was not announcing newer numbers.

On user behavior, Srinivas rejected the idea that Perplexity tries to maximize time spent on the platform. If a user receives an accurate answer on the first turn, he said, they may not continue in that same chat. The company instead looks at whether the same user returns for more research tasks. He pointed to subscription mix as evidence of power-user demand: at the beginning of the year, the split between Perplexity’s $200-a-month Max plan and Pro plan was about 9 to 91; by the time of the interview, it was closer to 30 to 70. He said that shows users paying roughly $2,000 a year out of pocket for better research, orchestration, and accuracy, separate from enterprise spending.

Hyde also raised copyright litigation, including a lawsuit from CNN alleging violations of federal copyright law. Srinivas said Perplexity is confident in its position and will let the legal process decide the specific situation. He added that “nobody has any copyright over truth and facts.”

At Computex, the AI hardware discussion turned to a memory bottleneck rather than only to accelerators. Ian King said SK Hynix’s pledge to double capacity matters because the company is the second-largest maker of computer memory chips and data centers require enormous amounts of memory. More capacity could help sustain the AI growth story by easing shortages.

King added an important caveat: SK Hynix’s chairman, Chey Tae-won, has previously warned that the company could be losing money tomorrow and has been careful about the memory market outlook. King said the lack of a specific timeline for doubling capacity matters because it would reveal more about how quickly relief might arrive.

Rene Haas of Arm, speaking with Stephen Engle, said AI compute demand is unlike prior cycles because it touches every industry and every person will interact with AI either directly or indirectly. He framed the opportunity by reference to global GDP of $130 trillion and a knowledge-work or white-collar market of perhaps $30 trillion, which he said remains largely untapped.

Haas would not predict whether memory remains short until 2030, but he said he believes demand could continue much longer than historical patterns would suggest. He also said the memory companies’ caution is understandable: Micron, Samsung, and SK Hynix were losing money in 2022 and 2023, and if they overbuild they risk being exposed when the supply crunch eases.

The solution Haas sketched was not simply more government help, though he acknowledged U.S. government support for Micron and efforts to encourage Samsung and Hynix to expand in the U.S. He said the industry may need more innovative business models, including hyperscaler investment in fabs and equity partnerships to secure supply. The larger principle was shared risk: asking public memory companies to carry all investment risk when hyperscalers are the large consumers is “very, very hard to sustain.”

King also noted that Nvidia, Intel, Qualcomm, and others were aligned on another Computex message: humanoid robots are not arriving as fast as the industry would like. Their pitch is to provide hardware, software, and tools to bring robots to market faster. The significance, according to King, is that these companies see robotics as a next major market that could bring AI into homes, offices, and factories and help make AI pervasive enough to support current spending levels.

HPE’s stock move gave a concrete example of AI infrastructure demand reaching enterprise suppliers, not only hyperscalers. The company’s shares were up more than 21% intraday after second-quarter results beat expectations and HPE raised its full-year forecast. Hyde said sales could rise by as much as 33%, and that the stock move added $13 billion in market capitalization.

Antonio Neri called it an exceptional quarter, citing record results across key metrics and demand across networking, cloud, and AI. He said demand is durable enough for HPE to raise its current-year guide and provide next-year guidance six months early. He pointed to record backlog and a pipeline that remains multiples of that backlog.

Ludlow asked whether the outlook was a volume story or a pricing story. Neri said it was a volume story with disciplined pricing and execution. He said HPE is effectively pulling the 2028 outlook it had previously given at its securities analyst meeting forward by two years into 2026. At the midpoint, he said, earnings per share would be $3.40, a dollar higher than the previous guidance.

The growth drivers Neri listed were broad. Campus and branch orders were up almost 30%. Data-center switching orders were up close to 20%. Server demand was up triple digits. Storage had its sixth consecutive quarter of triple-digit growth. Private cloud, including “AI factories for enterprise,” continued to grow.

Hyde pressed on margins and cost pressure, especially given rising DRAM and NAND prices. Neri said the portfolio mix matters: HPE now has an $11 billion networking business, which changes the gross-margin profile. He cited a record gross margin of 36.9% and said the company is ahead of plan on Juniper and Catalyst milestones and synergies, which are helping cost of sales and operating expenses. He acknowledged memory cost increases but said demand is the core factor because customers need access to AI technology.

Ludlow asked whether HPE’s results depended on one large customer or reflected a broader customer base. Neri said it was not one customer. HPE has been selective in AI at scale, he said, because profitability and working capital matter, and the company has prioritized debt paydown and profitable growth through networking, cloud, and AI in enterprise, sovereign, and inferencing markets.

Neri said many customers are bringing AI infrastructure on-premise for compliance, governance, data privacy, and security reasons. He used HPE itself as an example: the company has 1,200 AI use cases internally, 250 of which are in production. HPE uses a mix of proprietary, closed, and open models under stringent governance, and runs that work on-premise.

Asked whether this is simply an enterprise supercycle or a deeper structural shift driven by agentic AI, Neri chose the latter. He said agentic AI is transforming business processes and workflows and making companies more agile and efficient. Enterprise adoption remains early, in his view, but customers do not want to be left behind. His productivity thesis was direct: AI’s value to enterprises is that it helps them become more competitive.

On labor, Neri did not say HPE needs fewer employees. He said the types of roles companies hire for are changing, and that skill sets must evolve. HPE has 65,000 employees, and Neri said he reminds them to use AI to become more productive. He argued that having “a minor on AI” and using the technology will become a competitive advantage in every role across the enterprise.

The White House released an AI executive order focused on government access to frontier models and cybersecurity risks. Michael Shepard said the order was the directive President Donald Trump had been set to sign about two weeks earlier before pausing it over concerns that elements of the plan could impede innovation.

Shepard said the order calls for the government to work with AI developers voluntarily to gain access to cutting-edge frontier models and determine whether they could pose cybersecurity risks. The word “voluntary,” he said, appears several times. The approach is cooperation rather than dictation.

The order also calls for a “classified benchmarking process” for AI, according to on-screen text. Shepard said determining whether a model is advanced enough to be covered by the order would require experts from various government departments, with classified clearances, working in a classified top-secret setting. The process would evaluate whether a model has cybersecurity capabilities significant enough to worry the government and broader industries.

The order lays out a process through which federal agencies, state and local authorities, and critical-infrastructure operators could gain access, again voluntarily through companies, to models deemed so cutting-edge that they may pose cyber risks or could be used to test networks for vulnerabilities. Hyde noted that the order arrived the same day Anthropic was releasing Mythos to an additional 150 organizations around the world.