The Conglomerate That Stopped Being a Holding Company
For most of its history, SoftBank Group described itself, and was understood by markets, as an investment holding company. It owned stakes in technology businesses, the Vision Fund deployed capital across hundreds of startups, and Masayoshi Son’s reputation rested on his record as an investor who occasionally made spectacularly successful bets, Alibaba being the canonical example, alongside spectacularly unsuccessful ones, WeWork being the canonical cautionary tale. Whatever the merits or failures of any individual position, the underlying business model was legible. SoftBank bought equity, held it, and eventually sold it or watched it go to zero.
What has happened to SoftBank over the past eighteen months does not fit that model anymore, and the company itself has stopped describing its strategy in those terms. SoftBank now holds an equity stake in OpenAI reported to be worth in excess of sixty billion dollars on paper, following its participation in a financing round that valued the company at three hundred billion dollars and its subsequent commitment of an additional thirty billion dollars in three instalments through Vision Fund II. It controls Arm Holdings, the chip architecture licensing firm whose designs underpin a significant share of the processors being deployed across the global AI infrastructure buildout. It owns SB Energy, a power generation and data centre development subsidiary that has become one of the primary infrastructure partners for OpenAI’s Stargate project in the United States. And in April 2026, it became the domestic infrastructure anchor, alongside Sakura Internet, for a ten billion dollar Microsoft investment designed to build sovereign AI compute capacity inside Japan.
Each of these positions, examined individually, looks like a conventional strategic investment of the kind that large technology conglomerates make routinely. Examined together, they describe something categorically different: a company that holds equity in the AI model layer through OpenAI, controls intellectual property at the chip layer through Arm, owns physical infrastructure at the energy and data centre layer through SB Energy, and is now building the domestic cloud and enterprise distribution layer in its home market through SB OpenAI Japan and the Microsoft partnership. SoftBank is not investing in the AI value chain from the outside. It is attempting to become the AI value chain, end to end, across multiple geographies simultaneously.
The question this long read examines is not whether this strategy is ambitious. It self-evidently is, and no other company globally has assembled a comparable combination of positions across model ownership, chip architecture, energy infrastructure, and sovereign cloud distribution under a single corporate structure. The question is whether the integration that makes this strategy potentially powerful, the way each piece reinforces the others, is the same integration that makes it dangerous, the way stress in any single piece could propagate through the entire structure at a moment when the underlying demand assumptions have not yet been tested by a genuine downturn.
Mapping the Empire, Layer by Layer
The Model Layer: An Eleven Percent Stake That Became the Centre of Gravity
SoftBank’s relationship with OpenAI began as a venture investment and has evolved into something closer to a strategic merger of interests, even though the two companies remain legally and operationally separate. SoftBank’s ownership stake, reported at approximately eleven percent following its participation in OpenAI’s most recent major funding round, represents far more than a passive financial position. S&P Global’s analysis in early 2026 estimated that OpenAI now accounts for roughly thirty percent of SoftBank’s total investment portfolio value, a concentration level comparable to Arm Holdings’ share of the same portfolio, meaning that two positions, OpenAI and Arm, together represent the majority of what SoftBank’s equity is now worth.
The thirty billion dollar additional investment that SoftBank committed to OpenAI, structured to be delivered in three instalments through April, July, and October of 2026 via Vision Fund II, represents one of the largest single commitments any investor has made to a private AI company. The scale of this commitment, combined with the existing eleven percent stake, means that SoftBank’s financial fortunes are now directly and substantially tied to OpenAI’s trajectory in a way that goes well beyond the diversified portfolio approach that characterised the original Vision Fund’s strategy of spreading bets across many companies in the hope that a small number of outsized winners would compensate for the inevitable failures.
The relationship is reciprocal in ways that extend beyond capital. SoftBank and OpenAI jointly operate SB OpenAI Japan, a venture marketing enterprise AI products under the Cristal Intelligence brand to Japanese corporate customers, giving SoftBank a distribution channel for OpenAI’s technology in its home market while giving OpenAI a localisation and enterprise sales partner with deep relationships across Japanese industry. This is not simply an investor watching its portfolio company’s revenue grow from a distance. It is an arrangement in which SoftBank’s commercial success in Japan is partly a function of OpenAI’s product capability, and OpenAI’s revenue in Japan is partly a function of SoftBank’s distribution effort, creating a feedback loop between the two companies’ commercial outcomes that did not exist in SoftBank’s prior portfolio company relationships.
The OpenAI position has also become, in a separate and consequential way, a source of liquidity for SoftBank itself. In June 2026, reports emerged that SoftBank had been in talks to use its OpenAI stake as collateral for a loan, an arrangement that stalled when lenders proved unwilling to extend the scale of financing SoftBank sought against the position. The analysis that emerged from this stalled negotiation was pointed: if a stake reportedly worth over sixty billion dollars on paper could not be converted into a loan of even a fraction of that value, the signal was not primarily about AI sentiment but about how little private company valuations count as collateral when lenders demand assets they can actually recover in a default scenario.
The Chip Layer: Arm’s Architecture as the Hidden Connective Tissue
If OpenAI represents SoftBank’s bet on the model layer of the AI stack, Arm Holdings represents its bet on the layer beneath, the chip architecture that determines how efficiently AI workloads can be processed across the data centres that Stargate and similar projects are building. Arm does not manufacture chips. It licenses the instruction set architecture and processor designs that chip manufacturers, including the custom silicon being developed for major AI infrastructure projects, build upon. SoftBank’s majority ownership of Arm, maintained since the company’s acquisition years before the current AI boom and reinforced through Arm’s 2023 initial public offering, has positioned SoftBank to capture value from a layer of the AI stack that operates somewhat independently of which specific AI model or which specific cloud provider ultimately wins the broader competitive race.
The strategic logic connecting Arm to the rest of SoftBank’s AI positions is best understood through what one analysis described as a virtuous cycle, though the description applies regardless of whether the cycle proves virtuous or vicious in practice. SoftBank’s capital funds Stargate’s infrastructure buildout. That infrastructure incorporates processors built on Arm’s architecture. Arm’s licensing revenue and market valuation benefit from the scale of this deployment. Arm’s rising valuation strengthens SoftBank’s balance sheet and borrowing capacity, the same balance sheet and borrowing capacity that funds further investment in Stargate and OpenAI. Each component of the cycle depends on the others continuing to perform, and the cycle’s strength as a flywheel during a period of AI infrastructure expansion is structurally identical to its vulnerability as an amplifier during any period of contraction.
Arm’s share price performance through 2025 and into 2026, with gains reported in the range of thirty-eight percent over the period, has been a major contributor to the overall surge in SoftBank’s market valuation, a surge significant enough that SoftBank briefly overtook Toyota as Japan’s most valuable company by market capitalisation in 2026, a milestone that received substantial attention precisely because of how dramatically it illustrated the market’s repricing of SoftBank from a recovering post-WeWork conglomerate to the most direct large-cap proxy for AI infrastructure exposure available to investors. The margin loan structure that SoftBank has used against its Arm position, increasing from approximately thirteen and a half billion dollars as of March 2025 toward a reported total approaching eighteen and a half billion dollars with additional facilities under discussion, represents a direct mechanism through which Arm’s valuation is being converted into cash that funds the rest of SoftBank’s AI strategy, making Arm not just a strategic holding but an active financing instrument for the broader empire.
The Infrastructure Layer: SB Energy and the Physical Build-Out of Stargate
The least visible component of SoftBank’s AI empire to general audiences, but arguably the most operationally significant, is SB Energy, the SoftBank subsidiary that has transformed itself over recent years from a renewable energy and storage developer into one of the primary execution partners for the physical infrastructure that Stargate requires. The partnership between OpenAI and SB Energy, announced in January 2026, involved a combined one billion dollar investment, five hundred million dollars each from OpenAI and SoftBank, directly into SB Energy to support its growth as what both companies described as a leading development and execution partner for AI data centre campuses and the energy infrastructure that supports them.
The scope of what SB Energy is building is substantial in absolute terms. The company was selected to build and operate OpenAI’s previously announced 1.2 gigawatt data centre site in Milam County, Texas, one component of a Stargate buildout that, across its various sites including the flagship Abilene campus, was targeting nearly seven gigawatts of planned capacity and over four hundred billion dollars in investment commitments as of late 2025. SB Energy’s own trajectory from a company that had previously secured eight hundred million dollars in redeemable preferred equity from Ares Infrastructure Opportunities funds to one now central to a multi-gigawatt national infrastructure programme illustrates how rapidly SoftBank’s infrastructure ambitions have scaled within a compressed timeframe.
The arrangement also created a reciprocal commercial relationship structurally similar to the one between SoftBank and OpenAI at the corporate level. SB Energy became an OpenAI customer, gaining full API access and committing to deploy ChatGPT across its own workforce, meaning that the infrastructure builder is also a customer of the technology whose infrastructure it is building, a closed loop that exemplifies the broader pattern of circular financing that has drawn scrutiny across the AI sector. The Bank for International Settlements, in its 2026 analysis of AI sector financing, identified this pattern explicitly, noting that circular financing structures within the AI ecosystem, where investors provide capital while simultaneously becoming customers and infrastructure partners of the companies they are funding, raise concerns about systemic spillovers if expected returns on AI investment fail to materialise, particularly given that some of these financing structures may mask the true scale of leverage by keeping it off balance sheet.
The physical infrastructure that SB Energy is building does not exist in isolation from the chip layer that Arm occupies or the model layer that OpenAI represents. The data centres SB Energy constructs are designed around the processor architectures that Arm licenses, run workloads generated by the models OpenAI develops, and represent capital expenditure that SoftBank itself has partly financed through the leverage built on its Arm and OpenAI positions. Each layer of SoftBank’s empire is, in a literal physical and financial sense, built into the others.
The Japan Pivot: Bringing the Empire Home
Sakura Internet and the Domestic Sovereignty Play
Until April 2026, SoftBank’s AI empire was overwhelmingly an American story, with OpenAI, Arm’s customer base, and SB Energy’s data centre campuses all oriented toward the United States market and the Stargate buildout specifically. The Microsoft announcement that month represented the first major component of the strategy explicitly oriented toward SoftBank’s home market, and the structure of the deal reveals how SoftBank is replicating, in Japan, the same multi-layer integration approach it has built in the United States, with itself occupying a different but analogous position in the stack.
The arrangement positions Sakura Internet, a relatively small domestic cloud infrastructure provider, as the entity providing physical GPU capacity inside Japan, while SoftBank Corp, the telecommunications subsidiary, provides network infrastructure and customer relationships, and Microsoft provides the Azure software layer that lets enterprise and government customers access this domestically-hosted capacity while remaining within the Azure ecosystem they may already use globally. The political framing of the deal, emphasising that this allows companies and government agencies to keep sensitive data within Japan’s borders while still accessing Microsoft’s cloud services, aligns directly with Prime Minister Sanae Takaichi’s stated technology and economic security priorities, giving SoftBank’s commercial positioning a degree of government alignment that reinforces its strategic value beyond pure market economics.
Sakura Internet’s financial position illustrates both the opportunity and the difficulty inherent in this domestic build-out. The company’s revenue from GPU infrastructure services reached over eight billion yen in the fiscal year ending March 2026, growing more than twenty percent year over year, and the company secured a roughly 3.8 billion yen order from a national institution for generative AI workloads running through March 2027. At the same time, Sakura’s overall financial results for the same fiscal year showed an operating loss, a reversal from operating profit the prior year, driven substantially by the capital expenditure required to build out the GPU infrastructure that its new role demands. Net cash used in investing activities nearly tripled year over year, reflecting equipment purchases tied directly to generative AI service expansion.
This is the pattern that recurs across every layer of SoftBank’s empire: substantial revenue growth in the specific AI-related business lines, accompanied by capital expenditure and leverage that outpaces the revenue growth in the near term, on the expectation that the revenue growth trajectory will eventually overtake the cost of the capital deployed to build the capacity that revenue depends on. Sakura’s selection as a provider for Japan’s Government Cloud services, a designation that provides recurring revenue and regulatory credibility for high-security public sector contracts, gives the company a demand floor that pure commercial GPU rental businesses lack, but the fundamental arithmetic of building expensive infrastructure ahead of the demand that will eventually pay for it is identical to the arithmetic facing SB Energy’s American campuses and, at the corporate level, SoftBank’s overall position in OpenAI and Arm.
Cristal Intelligence and the One Million Engineer Bet
The talent and distribution dimension of SoftBank’s Japan strategy operates through two parallel commitments that, together, represent an attempt to create the demand-side conditions that the supply-side infrastructure investment requires. Microsoft’s commitment to train one million AI engineers and developers in Japan by 2030, delivered in partnership with NTT Data, NEC, Fujitsu, and Hitachi alongside SoftBank and Sakura, addresses what multiple analyses have identified as Japan’s most significant constraint on AI adoption: not insufficient compute capacity or insufficient enterprise willingness to adopt AI tools, but an insufficient pool of workers with the skills to build, deploy, and maintain AI systems at the scale that the infrastructure investment is designed to support.
SB OpenAI Japan’s Cristal Intelligence brand represents the demand-generation side of this equation, marketing enterprise AI products directly to Japanese corporate customers in a market where adoption indicators are already notably strong by global standards. Microsoft’s own data indicates that ninety-four percent of Nikkei 225 companies have already deployed Microsoft 365 Copilot, an extraordinarily high penetration rate that suggests Japanese enterprise readiness for AI tools is not the binding constraint that infrastructure-focused analyses sometimes assume. A Microsoft survey separately found that approximately one in five working-age adults in Japan is already using generative AI tools, ahead of the roughly one in six global average, indicating that consumer-level familiarity with AI tools is also relatively advanced.
What this combination suggests is that SoftBank’s Japan strategy is not primarily attempting to create AI demand where none exists, but rather attempting to ensure that the demand that already exists, evidenced by high enterprise adoption rates and above-average consumer usage, is served by domestically-hosted infrastructure rather than infrastructure located outside Japan’s borders. This is a meaningfully different strategic problem than the one SB Energy faces in Texas, where Stargate’s campuses are being built partly in anticipation of demand that does not yet exist at the scale the infrastructure is designed for. In Japan, SoftBank’s bet is closer to a market capture play, redirecting existing and growing AI demand toward infrastructure in which SoftBank holds a direct stake, than a market creation play, building capacity in advance of demand that may or may not materialise.
This distinction matters for how the overall empire’s risk profile should be assessed. The American infrastructure positions, SB Energy’s campuses and the broader Stargate buildout, carry genuine demand risk: if AI usage growth slows, the multi-gigawatt capacity being built may exceed what the market actually requires, at significant cost to the companies that financed its construction. The Japan positions carry comparatively less demand risk, given the demonstrated adoption levels already present in the market, but carry execution risk tied to whether Sakura Internet and SoftBank Corp can actually deliver the domestic infrastructure capacity at the pace and reliability that would let this existing demand be redirected domestically rather than continuing to flow to infrastructure located in the United States or elsewhere.
The Leverage Question Nobody Can Avoid
What the Balance Sheet Actually Shows
Every layer of SoftBank’s AI empire has been financed, at least in part, through debt, and the cumulative scale of that debt has reached levels that credit rating agencies and equity analysts are now treating as a primary variable in evaluating the company, separate from and in some respects more important than the underlying quality of the assets the debt has purchased. As of the end of 2025, SoftBank reported standalone interest-bearing debt of approximately sixteen point three trillion yen, equivalent to roughly one hundred and four billion dollars, a figure that does not include the additional thirty billion dollar OpenAI commitment being delivered through 2026, the forty billion dollar bridge loan secured in March 2026 specifically to fund OpenAI-related investment and general corporate needs, or the margin loan facilities secured against Arm shares.
S&P Global’s response to this accumulation of leverage has been to revise SoftBank’s credit outlook downward, citing liquidity and asset quality concerns stemming directly from the scale of the leveraged investments the company has made into OpenAI and the broader AI infrastructure ecosystem. The specific concern that credit analysts have articulated is not that SoftBank’s AI positions lack value, Arm’s market valuation and OpenAI’s reported valuation are both, by any conventional measure, substantial, but that the liquidity of these positions does not match the liquidity of the debt obligations SoftBank has taken on to acquire and expand them. SoftBank’s own shares trade daily on a liquid public market, but the underlying assets, OpenAI equity in particular, are private company holdings that cannot be readily sold to meet near-term obligations, creating what analysts have described as a liquidity mismatch between SoftBank’s asset base and its liability structure.
The stalled loan negotiation involving SoftBank’s OpenAI stake, reported in June 2026, is the clearest illustration of this mismatch in practice. A position reportedly worth over sixty billion dollars on paper, based on OpenAI’s most recent funding round valuation, could not be converted into a loan facility of the scale SoftBank had sought, with lenders apparently unwilling to extend financing against the position at terms SoftBank found acceptable. The analysis that followed framed this outcome with appropriate caution: the failure to secure the loan does not necessarily indicate that OpenAI’s valuation is inflated, but it does indicate that private company valuations, however large, do not automatically translate into the kind of collateral value that lenders require when extending billions of dollars in financing, particularly in an environment where lenders are increasingly attentive to exactly the kind of circular financing structures that characterise SoftBank’s broader AI strategy.
Nvidia’s own posture toward further OpenAI investment adds another dimension to this picture. Reports in early 2026 indicated that Nvidia’s chief executive signalled the company’s existing thirty billion dollar commitment to OpenAI might represent the full extent of its investment, suggesting a pullback from a previously discussed larger commitment, a development that, combined with SoftBank’s stalled loan talks, paints a picture of an AI financing ecosystem in which even the largest and best-capitalised participants are becoming more cautious about the scale of further commitments, even as SoftBank’s own exposure continues to deepen through contractually committed instalments that cannot easily be unwound.
The WeWork Comparison That Son Wants Everyone to Forget
The comparison that hangs over SoftBank’s current AI strategy, whether or not company executives wish to engage with it directly, is the company’s history with WeWork, the office-sharing company that SoftBank’s Vision Fund poured billions of dollars into at valuations that subsequently collapsed amid governance failures and a business model that did not withstand scrutiny once growth-at-all-costs capital stopped flowing freely. The WeWork episode remains the reference point that analysts reach for when evaluating Masayoshi Son’s risk appetite, not because the AI sector necessarily shares WeWork’s specific business model flaws, but because the pattern of concentrated, highly leveraged conviction bets, often made at valuations set during periods of maximal investor enthusiasm, recurs across both episodes.
What distinguishes the current AI empire from the WeWork position in ways that could matter significantly is the diversity of the underlying assets and the degree to which they are operationally interconnected rather than simply co-located in a single portfolio. WeWork was a single bet on a single business model. SoftBank’s AI empire spans model ownership, chip architecture, energy infrastructure, and domestic cloud distribution across two continents, assets that, even if any single component underperformed, would not necessarily transmit that underperformance to the others in the same direct way that a WeWork failure was contained to the WeWork position itself.
What does not distinguish the current situation favourably from WeWork, and arguably makes it more concerning, is precisely the integration that this long read has described as the empire’s strategic novelty. WeWork’s failure, however painful, was a write-down of a single investment. A scenario in which OpenAI’s growth trajectory disappoints relative to the valuations at which SoftBank has invested would not simply impair SoftBank’s equity stake. It would impair the collateral value underlying the margin loans against Arm shares that have been used to fund that stake, reduce the demand projections underlying SB Energy’s multi-gigawatt data centre construction, weaken the commercial case for SB OpenAI Japan’s Cristal Intelligence offering, and potentially affect Sakura Internet’s ability to monetise the GPU infrastructure it has built specifically to serve Azure customers running OpenAI-derived workloads. The integration that makes the empire strategically coherent during a period of AI demand growth is the same integration that would transmit stress across every component during a period of AI demand disappointment.
Gil Luria’s characterisation of the strategy, that SoftBank has made itself into a highly leveraged bet on AI carrying significant upside as well as risk, captures the symmetry that the empire’s structure produces. The question that remains genuinely open, and that will not be answered by anything short of the AI sector’s actual demand trajectory over the coming several years, is whether the integration SoftBank has built represents a sophisticated hedge, in which strength in one layer compensates for weakness in another, or a sophisticated amplifier, in which the layers move together more than they move independently, leaving SoftBank with concentrated exposure to a single underlying variable, the pace of global AI adoption, dressed in the appearance of diversification across model ownership, chip design, energy infrastructure, and sovereign cloud distribution.
An Empire Built on a Single Bet, Wearing Many Disguises
The most accurate way to describe SoftBank’s transformation over the past eighteen months is not that the company has diversified into AI infrastructure across multiple layers of the technology stack, even though that description is technically true of the individual transactions involved. The more precise description is that SoftBank has taken a single, enormous bet, that the current trajectory of global AI demand growth will continue at something resembling its current pace for long enough to justify the capital that has been deployed across model equity, chip architecture, energy infrastructure, and sovereign cloud distribution, and has expressed that single bet through a portfolio of positions that look, on the surface, like a diversified industrial strategy.
The strategy’s coherence is genuinely impressive as a piece of corporate construction. Few companies in the world could assemble eleven percent of OpenAI, majority control of Arm, a multi-gigawatt data centre development arm, and a sovereign cloud partnership in Japan’s domestic market within the same period, and fewer still could articulate, as SoftBank’s executives have, a logic connecting these pieces into something resembling a flywheel. The flywheel framing is not wrong. It is, however, incomplete, because every flywheel that accelerates in one direction can decelerate in the other, and the mechanisms that connect SoftBank’s positions, the margin loans against Arm, the bridge loan tied to OpenAI, the reciprocal customer relationships between SB Energy and OpenAI, the capital expenditure commitments at Sakura Internet, do not distinguish between acceleration and deceleration. They simply transmit whatever momentum exists across the entire structure.
Whether this empire represents the most ambitious and ultimately successful AI infrastructure strategy assembled by any single company, or the most concentrated and ultimately costly demonstration of what happens when conviction outpaces the verification that only time and an actual demand cycle can provide, is not a question that the structure of the empire itself can answer. It is a question that depends entirely on variables, the pace of enterprise AI adoption, the trajectory of AI model capability improvement, the willingness of capital markets to continue financing infrastructure built ahead of confirmed demand, that exist outside SoftBank’s control and that no amount of strategic integration across the layers of the AI stack can hedge against if they move in the wrong direction simultaneously. Masayoshi Son has built something that has never existed before. Whether that is the empire’s greatest strength or its defining vulnerability is the question every layer of it now depends on answering correctly.
Sources: https://www.cnbc.com/2026/06/04/softbank-openai-liquidity-crunch-arm-son.html | https://www.bis.org/publ/bisbull120.pdf
