SoftBank Group’s efforts to raise at least $6 billion through a margin loan backed by its large stake in OpenAI have stalled, just weeks after the Japanese conglomerate scaled back its initial target from $10 billion, Bloomberg reports, citing people familiar with the matter.

The setback comes as SoftBank weighs various fundraising alternatives and continues to navigate the complexities of one of the largest single-company bets in corporate history. While the company had secured around $5 billion in commitments before the discussions paused, it remains unclear whether those were firm or preliminary.

SoftBank could still revive the margin loan later, but for now, it is exploring other options, the sources said.

The investor’s shares fell as much as 9.7% on Wednesday.

The loan was intended to help finance SoftBank’s more than $60 billion commitment to OpenAI, a position that has drawn increasing scrutiny amid rapid advancements by rivals like Anthropic and questions about long-term valuation sustainability. The company’s exposure has been a point of internal anxiety for some officials, even as OpenAI’s progress has driven significant paper gains for SoftBank.

OpenAI’s confidential filing for a U.S. IPO on Monday, with Goldman Sachs and Morgan Stanley leading the effort and a potential listing as soon as this fall, had earlier improved sentiment among some potential lenders. However, challenges around valuing an unlisted company with explosive but unproven growth prospects contributed to hesitation.

Hua Cheng, head of Asia credit research at AllianceBernstein, offered a perspective on the situation.

“The margin loan is just one piece of a much larger puzzle, and unless we see a clear deterioration in their ability to raise funds this way, we don’t view it as a standalone red flag. The best-case scenario is an OpenAI IPO this year, with SoftBank offloading part of its stake to pay down debt. That would be consistent with what credit investors want,” Cheng said.

A successful OpenAI listing could provide SoftBank with an exit route or collateral flexibility, easing pressure on its balance sheet. However, any delays or valuation disappointments in that process could complicate matters further.

Debt Pressures and Fundraising Alternatives

Looming over SoftBank is a $40 billion bridge financing tied to its OpenAI investments, which must be repaid by March 2027. The company has indicated it plans to address this through asset sales and other financing measures. Beyond the margin loan, SoftBank has several potential avenues, including issuing additional bonds or borrowing against other listed holdings such as Arm Holdings and Intel, both of which have seen explosive gains this year (up 197% and 192% respectively amid the AI boom).

Despite Wednesday’s decline, SoftBank’s own stock has risen about 45.5% year-to-date, recently overtaking Toyota as Japan’s most valuable company by market capitalization. The company reported a sharp jump in quarterly profit last month, largely driven by valuation gains on its OpenAI stake. Credit markets have also shown improving sentiment, with credit-default swaps narrowing by about 61 basis points to 307 from a recent peak above 367 on May 20.

SoftBank has been aggressively expanding its AI ambitions. Late last month, it announced plans to invest up to €75 billion ($86.6 billion) to build AI data center capacity in France, positioning the country as a potential European hub. This fits into a broader strategy of leveraging its Vision Fund and direct stakes to become a dominant player across the AI value chain.

Yet the heavy concentration in OpenAI has raised questions among some investors about diversification and risk management. The stalled margin loan talks reflect the practical difficulties of monetizing even high-profile stakes in private companies, particularly when lenders demand clarity on valuation, liquidity, and exit paths.

The broader market debate around AI valuations has intensified in recent months. Breakthroughs by competitors like Anthropic have prompted some to question whether OpenAI’s lead is as durable as previously assumed, potentially complicating financing efforts tied to the stake.

For SoftBank, the current pause is not necessarily fatal. The company retains flexibility and multiple levers to manage its liabilities. A successful OpenAI IPO later this year would significantly ease pressure and validate SoftBank’s vision.

In the meantime, analysts expect management to balance its aggressive AI expansion with prudent capital management. But that is a delicate act for a conglomerate whose fortunes have long been tied to bold, high-conviction technology investments.

SpaceX is pushing aggressively into space-based artificial intelligence infrastructure, with executives telling investors the company aims to launch initial demonstration missions of orbital computing systems by late 2027 — ahead of the “as early as 2028” timeline outlined in its IPO filing.

The orbital-compute initiative sits at the heart of SpaceX’s long-term growth narrative as it prepares for what could be the largest public offering in history. In presentations to investors ahead of the IPO, President Gwynne Shotwell and CFO Bret Johnsen described early deployments as technology demonstrators designed to validate the concept before any larger-scale commercial rollout, according to two people familiar with the meetings.

According to sources cited by Reuters, SpaceX has requested regulatory approval to launch up to 1 million space-based data-center satellites, a scale that would dwarf its current Starlink constellation and position the company as a potential leader in off-world computing. While the IPO filing highlighted 2028 as a possible start for deployments, the updated internal timeline shared with investors suggests growing confidence in execution, though it still leaves room for Starship-related delays.

One attendee interpreted the more conservative public timeline as providing management flexibility amid the technical complexities of reusable heavy-lift vehicles and satellite manufacturing at unprecedented scale.

SpaceX is scheduled to begin trading on the Nasdaq on Friday under the ticker SPCX, with shares targeted at $135 each for a $1.75 trillion valuation.

Musk’s Vision: Overcoming Earth’s Compute and Energy Limits

The push into orbital AI data centers aligns with Elon Musk’s broader philosophy that humanity must expand beyond Earth to solve some of its most pressing constraints. In IPO materials and recent comments, SpaceX argues it is “the only company with a commercially viable path to building orbital AI compute at scale.” The company claims this could unlock a potential $23 trillion market opportunity by addressing limitations in terrestrial electricity generation and computing capacity, particularly as the U.S. lags behind China in scaling large infrastructure projects.

In a video released Monday, Musk emphasized that building orbital data centers is not an insurmountable engineering challenge, noting that much of the required technology already exists within the Starlink network. The first version of these AI satellites is expected to use Nvidia chips, with computing power equivalent to an Nvidia GB300 rack.

This strategy represents a logical evolution for SpaceX. Starlink has already proven the viability of large-scale satellite constellations for global connectivity. Extending that expertise into high-performance computing in orbit could solve critical bottlenecks: access to abundant solar power in space, reduced latency for certain applications, and relief from Earth-based energy and land constraints.

Starship as the Critical Enabler — and Risk

The entire vision hinges on Starship, SpaceX’s fully reusable heavy-lift rocket. While Starship has achieved several successful test flights, it remains years behind Musk’s original timelines and has yet to demonstrate the rapid reusability and high launch cadence needed for economic viability at the scale required for massive orbital data centers.

Michael Monaghan, partner and portfolio manager at Founder ETFs, noted that while many of Musk’s projects involve open-ended challenges, orbital data centers have clearer technical boundaries.

“I think that orbital data centers, while a difficult problem, have some bounds on it, which to me gives greater confidence that the timelines laid out will be hit,” he said.

Still, any meaningful delays in Starship development could push back the orbital compute roadmap, a risk implicitly acknowledged by the more conservative public timeline in the IPO filing.

The orbital AI pitch is central to justifying SpaceX’s ambitious valuation. With Starlink as the current cash cow, the company is asking investors to underwrite a future where space infrastructure becomes foundational to global AI capabilities. This narrative has clearly resonated, with the IPO drawing roughly $150 billion in demand — double the $75 billion targeted.

The all-primary offering structure, significant retail allocation (up to 30%), and Musk’s 366-day lock-up further signal confidence. Proceeds will fund the expansion of both Starlink and AI-related infrastructure.

Growing Investors’ Interest

Investors are increasingly showing faith that SpaceX’s move into orbital computing could reshape the AI infrastructure landscape. By bypassing terrestrial constraints, the company aims to offer massive, scalable compute capacity with potentially lower marginal energy costs. This positions it uniquely against traditional hyperscalers facing power shortages, regulatory hurdles, and land constraints on Earth.

Besides challenges ranging from regulatory approval for such a large constellation, technical reliability in the harsh space environment, latency for certain applications, and competition from other players exploring space-based solutions, SpaceX appears good to go.

The strong demand for the IPO suggests many are willing to look well beyond near-term profitability and toward a future where orbital assets become critical infrastructure for the AI economy.

Artificial intelligence startup Anthropic is making its most advanced AI technology available to a broader audience for the first time, while imposing strict safeguards designed to prevent the system from being used for potentially dangerous cybersecurity activities.

The company on Tuesday unveiled Claude Fable 5, describing it as the most capable model it has released for general use. The launch marks a significant milestone for Anthropic as it seeks to expand commercial adoption of its technology following the global attention generated by its experimental Mythos model earlier this year.

Mythos stunned governments, cybersecurity experts, and technology companies after Anthropic disclosed that the system had identified thousands of previously unknown software vulnerabilities, raising both excitement about its capabilities and concern about how such powerful tools could be misused.

Until now, access to Mythos has been restricted largely to about 200 organizations participating in Anthropic’s Glasswing program, including U.S. government agencies and selected cybersecurity institutions.

The wider rollout comes at a crucial moment for Anthropic. The company has emerged as one of the most valuable businesses in the artificial intelligence sector, carrying an estimated valuation of $965 billion and increasingly competing head-to-head with rivals such as OpenAI, Google DeepMind, and xAI as the race toward public listings accelerates.

Balancing Power and Safety

Anthropic’s challenge has been how to commercialize a model capable of sophisticated cybersecurity reasoning without enabling malicious actors to exploit those same capabilities. The company said Claude Fable 5 includes extensive safeguards that block users from employing the model to search for software vulnerabilities, conduct offensive cyber operations, or perform other high-risk activities.

Dianne Penn, Anthropic’s head of product management, research, and labs, explained how the restrictions would work in practice.

“Let’s say I’m a college student asking the model like help me find cyber vulnerabilities on X package or code. The model would refuse and Fable 5 will fall back to Opus 4.8 for a response,” Penn told Reuters.

The approach reflects a shift across the AI industry, with companies separating frontier research systems from commercially available products. Rather than releasing raw capabilities directly to consumers, firms are embedding policy controls that limit access to sensitive functions.

Anthropic said it conducted extensive testing to ensure users cannot easily manipulate or jailbreak the system into performing restricted actions.

The original Mythos preview generated attention because it demonstrated capabilities that went beyond traditional coding assistants.

According to Anthropic, the model was able to uncover large numbers of software vulnerabilities, highlighting how advanced AI systems may soon become powerful tools for both cybersecurity defense and cyber offense.

That dual-use nature has become one of the most contentious issues in artificial intelligence development.

Governments view advanced AI models as potential national security assets. The U.S. government has expanded cooperation with Anthropic through initiatives such as Glasswing, while countries including South Korea, Japan, and several European allies have sought access to frontier AI systems for cybersecurity purposes.

The debate has intensified as researchers warn that future generations of AI could autonomously discover and exploit vulnerabilities faster than human experts. Anthropic’s decision to keep unrestricted Mythos access confined to trusted organizations appears aimed at reducing those risks while still enabling legitimate research and security applications.

Anthropic recently confidentially filed for an initial public offering, placing it among a growing list of AI giants preparing to enter public markets. Investors are focused on whether frontier AI developers can convert technological leadership into sustainable revenue growth.

By releasing Claude Fable 5 more broadly, Anthropic gains an opportunity to expand enterprise adoption while preserving safeguards around its most sensitive capabilities. The company said customers already using the preview version of Mythos will be able to upgrade to Claude Mythos 5, while broader access will gradually expand through what it described as a more systematic trusted-access program.

This tiered approach allows Anthropic to monetize cutting-edge technology without fully relinquishing control over its most powerful capabilities.

Cost Efficiency Becomes a New Battleground

Beyond performance, Anthropic is also emphasizing efficiency. Penn said early users have reported that while Fable 5 is positioned as a premium model, it completes tasks using fewer tokens, potentially lowering overall costs for customers.

As enterprises spend billions of dollars deploying AI systems, attention is increasingly moving away from raw model capability toward economics. Companies are now evaluating not only which models perform best, but also which deliver the lowest cost per completed task.

Competition on efficiency is becoming important as businesses scrutinize soaring AI budgets and seek ways to manage spending without sacrificing performance. Anthropic has priced both Claude Fable 5 and Claude Mythos 5 at $10 per million input tokens and $50 per million output tokens.

The artificial intelligence industry’s most advanced models are becoming powerful enough to influence national security, software infrastructure, and critical systems. Yet the same capabilities that make them valuable also create concerns about misuse. Anthropic’s solution is effectively a compromise: make the technology more widely available while limiting access to its most sensitive cybersecurity functions.