SFTBY — SOFTBANK GROUP CORP.

SoftBank Group holds one of the finest assets in technology — ARM Holdings' semiconductor IP monopoly — alongside a $110 billion stake in OpenAI, all available at roughly a 35% discount to net asset value. The discount is rational: a $40 billion unsecured bridge loan maturing in March 2027 has restructured the entire equity story around a single external event, the OpenAI IPO. Interesting until the IPO arrives; potentially dangerous if it doesn't.

---

The past three years have witnessed something unusual in capital markets: a fundamental rerating of technology infrastructure investments, driven by the conviction that whoever controls AI compute will capture an outsized share of global economic value for decades. Hyperscaler capital expenditure has roughly quadrupled since 2022. Semiconductor valuations have disconnected from cyclical norms. Private AI companies are raising capital at valuations that would have seemed satirical as recently as 2022 — OpenAI at $852 billion, Anthropic at $61 billion, xAI at $80 billion. The consensus narrative is that the AI compute supercycle is not a cycle at all but a structural shift, and that latecomers will be permanently disadvantaged.

Into this environment, SoftBank Group's stock rose 112% in the twelve months through April 2026, then collapsed more than 50% from its peak of $22.50, and now sits at approximately $11.16 — somewhere between a deeply discounted AI infrastructure vehicle and a leveraged bet on a single asset's IPO timeline. The company has not materially changed. The narrative around it has swung violently in both directions, which is typical for holding companies whose underlying assets reprice faster than the holding structure can communicate value.

What has changed is the balance sheet. In March 2026, SoftBank disclosed a $40 billion unsecured bridge loan from JPMorgan, Goldman Sachs, Mizuho, SMBC, and MUFG — the largest such facility in corporate history — to fund its third tranche of investment in OpenAI. The loan matures in twelve months, and its repayment depends, practically speaking, on OpenAI accessing public markets within that window. Standard & Poor's responded by downgrading SoftBank's credit outlook to negative. In April 2026, the company raised $3.56 billion in unsecured bonds at the highest interest rates it has ever paid. This is no longer a strategic optionality story. It is a ticking clock.

Technology holding companies occupy a peculiar and frequently misunderstood position in the investment universe. Unlike operating businesses, their value is derivative — they own pieces of businesses whose fortunes are determined elsewhere. The structural problem is well-documented: holding companies trade at persistent discounts to the sum of their parts because investors rationally assign a cost to the management layer that sits between them and the underlying assets. That discount reflects both the overhead of the holding structure and the historical tendency of holding company management to destroy value through overactive capital deployment. The academic literature on conglomerate discounts suggests the typical penalty is 15-30% of intrinsic value; SoftBank has historically traded at a steeper discount of 40-60%, reflecting the market's specific skepticism about Vision Fund-era capital allocation.

The semiconductor IP licensing industry that anchors SoftBank's primary asset is structurally different from most technology markets. The product — intellectual property governing how chips process instructions — is non-depletable, infinitely reproducible, and commands royalties on every chip shipped using the licensed architecture. The global market for semiconductor IP licensing was approximately $8 billion in 2025 and is projected to reach $13.5 billion by 2030. This is a small market by technology standards, but it is extracted from the much larger semiconductor market of $600+ billion annually. The IP licensor sits at the apex of the value chain: every chip design decision by every chipmaker is constrained by which architectures are worth licensing, and the switching costs of changing architectures mid-product-line are prohibitive. There are no inventory cycles, no commodity price risks, no manufacturing defects. The economics resemble a tollbooth on an essential road, except the road is embedded in billions of devices that cannot be replicated without paying the toll.

AI infrastructure investment has created a new and particularly valuable on-ramp to that tollbooth. Global AI chip market revenue was $61.8 billion in 2025, growing at roughly 33% annually toward a projected $295 billion by 2030. Data center capital expenditure by the four major hyperscalers — Amazon, Microsoft, Google, and Meta — will exceed $320 billion in 2026 alone. Every server chip, every inference accelerator, every edge AI device incorporates processor IP that must be licensed from someone. The structural forces determining who wins in this market are the same forces that determined who won in mobile: ecosystem depth, software compatibility, and the sheer weight of accumulated developer tooling. These forces heavily favor ARM.

SoftBank Group operates through four segments, two of which matter for an investor's purposes. SoftBank Corp — the Japanese wireless carrier, listed separately on the Tokyo Stock Exchange with SoftBank Group owning approximately 43% — is a stable, oligopolistic telecom business generating approximately ¥6.9 trillion ($46 billion) in annual revenue, growing at 8% per year, and paying consistent dividends to the parent. It is a cash machine. In an industry where Japan's three major carriers control approximately 97% of the domestic market, pricing is rational and subscriber attrition is low. This telecom base provides the operational underpinning that allows the parent to pursue speculative investments without the catastrophic liquidity crises that have historically felled overleveraged holding companies.

The Vision Fund division — once SoftBank's primary identity — has been quietly retired as an active investment thesis. Vision Fund 1 and 2 together manage $166 billion in committed capital across 332 portfolio companies. The combined unrealized portfolio value stands at $51.8 billion, reflecting substantial write-downs from peak 2021 valuations. Vision Fund 2 generated 0.2% IRR on $56 billion in committed capital — a figure that describes capital misallocation at an extraordinary scale. The February 2026 earnings presentation officially marked the transition to what management calls the Izanagi era, focused on hardware, infrastructure, and AI software integration rather than late-stage venture capital. The Vision Fund era is over. What remains is the question of what comes next.

The two assets that define SoftBank's current investment case are ARM Holdings and OpenAI. ARM is the company SoftBank acquired in 2016 for $32 billion — a price that seemed aggressive at the time — and chose not to sell when NVIDIA offered $66 billion in 2022. That decision to hold through the NVIDIA acquisition attempt is, in retrospect, the defining capital allocation choice of Son's recent tenure. OpenAI is the company into which SoftBank has now committed $64.6 billion, $40 billion of which was funded through the bridge loan announced in March 2026. ARM generates real revenue, real profits, and a real moat. OpenAI generates $25 billion in annualized revenue growing at rates that make its $852 billion private valuation mathematically defensible if the trajectory holds. The question is whether the vehicle holding these assets — SoftBank Group itself — can maintain the balance sheet stability to allow their full value to accrue to shareholders.

ARM Holdings has the most defensible position in the semiconductor industry that does not involve manufacturing. The company licenses processor architectures and collects royalties when chips based on those architectures ship to end markets. It controls 41% of the semiconductor IP licensing market — more than triple the share of its nearest competitor, Synopsys. Every chip that powers a smartphone uses ARM architecture; 99% of the world's smartphones run on ARM-based processors; 325 billion ARM-based chips have shipped since the company's founding. These numbers describe market penetration so complete that displacement requires not just a better technology but the coordinated abandonment of billions of dollars in existing software, toolchains, and chipmaker R&D commitments accumulated over 35 years.

Company	Gross Margin	IP Market Share	Revenue Model	Primary Switching Cost
ARM Holdings	97.4%	41%	Per-chip royalty + licensing	35 years of software ecosystem, ISA compatibility
Synopsys	~80%	13%	Licensing-only	EDA tool integration
Cadence Design Systems	~85%	~8%	Licensing-only	EDA tool integration
RISC-V (open source)	N/A	~10% (growing)	Royalty-free	None — competitive threat in cost-sensitive segments

The moat is best understood not as brand loyalty but as ecosystemic gravity. When a chipmaker chooses ARM architecture, they inherit a universe of compatible software libraries, reference designs, developer tools, and chipmaker certifications built over three decades. RISC-V, the open-source competitor, is growing at 60% annually in IoT and embedded systems and has become a strategic priority for Chinese chipmakers seeking independence from Western IP. China represents 19% of ARM's FY2025 revenue, growing at just 7.5% year-over-year — the slowest region — as local alternatives gain share. This is a real, structural pressure, but not a near-term displacement: RISC-V has 20 billion cores in operation worldwide versus ARM's 325 billion shipped, and lacks the software ecosystem depth that makes ARM's position self-reinforcing in high-complexity computing.

The more nuanced risk is that ARM is now moving from pure IP licensor to chip designer. In March 2026, ARM launched its AGI CPU — the company's first in-house chip — with launch customers including Meta, OpenAI, Cloudflare, and Cerebras. Management projects $1 billion in chip revenue by 2028, scaling to $15 billion by the late 2020s. This creates a potential channel conflict: Apple, Qualcomm, NVIDIA, and Amazon all design their own ARM-based chips and pay ARM royalties for the privilege. If ARM becomes their direct competitor, some of these customers may accelerate RISC-V investment to reduce their dependence on an architecture whose owner now competes with them. Whether the chip revenue upside outweighs the licensing relationship risk is the central uncertainty in the ARM investment case — and ARM management argues, with some credibility, that its Compute Subsystems (CSS) products already generate 2x the royalty of standard Armv9 licensing precisely because customers prefer paying a premium for ARM's pre-integrated designs.

SoftBank's reported financials require careful interpretation. In the nine months ended December 31, 2025, the company reported revenue of ¥5.2 trillion growing 8% year-over-year, operating income of ¥884 billion, and net income of ¥485 billion. These figures suggest a smoothly profitable conglomerate. They are not the primary story. The income statement is dominated by investment gains and losses on mark-to-market portfolio positions — in Q2 FY2025, the Vision Fund recorded a $19 billion unrealized gain driven by OpenAI's valuation step-ups. These gains are real in the sense that they reflect genuine appreciation in portfolio holdings, but they are not cash until realized, and they are not stable. In FY2022, the same mechanism produced a ¥3.5 trillion ($27 billion) reported loss when portfolio valuations collapsed. The ¥485 billion in net income currently being reported should be understood as a snapshot of portfolio marks, not an indication of normalized earning power.

The balance sheet is the real financial story. SoftBank's parent-level LTV — net debt divided by the fair value of portfolio holdings — is 20.6% as of December 2025, up from 12.2% in September 2024. The self-imposed policy ceiling is 25%, with management signaling willingness to temporarily breach that ceiling for the OpenAI commitment. With 4.4 percentage points of headroom remaining and $40 billion in debt recently added, there is limited capacity for further investment or for asset value decline before the LTV covenant becomes binding. ARM Holdings generated $4.0 billion in total revenue in the fiscal year ended March 2025, with $2.17 billion in royalties growing 20% and $1.84 billion in licensing growing 29%. Non-GAAP operating margin was approximately 41%. The difference between GAAP and adjusted figures at ARM relates primarily to stock-based compensation; the adjusted figures more accurately reflect ARM's cash economics.

Masayoshi Son owns 34.2% of SoftBank Group — enough skin in the game to ensure his personal fortune rises and falls with the company's share price. His capital allocation record divides cleanly into two categories that are equally large in absolute dollar terms. The $20 million Alibaba investment in 2000, which eventually generated over $60 billion in realized gains, stands as one of the most profitable venture capital bets in history. The $14 billion cumulative investment in WeWork, which Son pushed forward over the objections of his own advisors, produced a loss of equivalent magnitude. Vision Fund 2 deployed $56 billion across 300+ companies and generated a 0.2% IRR. Son has been unusually candid about these failures — describing WeWork as "a stain on my life" and the Vision Fund losses as a source of personal shame — which is at least evidence that he learns from disasters, if not that he avoids repeating them.

The OpenAI bet has structural similarities to the Vision Fund era: massive conviction, concentrated position, external leverage used to fund a privately valued asset. What is different is the quality of the underlying company — OpenAI at $25 billion in annualized revenue growing explosively is categorically different from WeWork at $47 billion in private valuation with no path to profitability. Son also holds personal stakes in Vision Fund 2 and the SB Northstar trading vehicle that create potential conflicts of interest between his personal economics and those of public shareholders. These are governance imperfections worth noting, but they are secondary to the primary alignment mechanism: with 34.2% of SoftBank's equity, Son's financial interests are overwhelmingly tied to the company's performance. He cannot afford for this bet to fail.

The right metrics for understanding SoftBank's growth trajectory are ARM's royalty revenue, Armv9's share of that royalty mix, the parent company's LTV, and the NAV discount to market cap. ARM royalty growth proves the semiconductor moat is compounding. Armv9 penetration is the forward indicator for royalty growth — it shows how much of the royalty ramp is still ahead. LTV indicates whether the balance sheet can sustain the asset base without forced sales. NAV discount measures whether the market is recognizing or discounting the underlying value.

Period	ARM Royalty Rev (ann.)	Armv9 % of Royalties	SBG NAV (¥T)	SBG LTV (%)
FY2023 (Mar 2023)	~$1.68B	<5%	~¥15–16T (est.)	~25%
FY2024 (Mar 2024)	~$1.81B	~15%	~¥22T (est.)	18.7%
FY2025 (Mar 2025)	$2.17B	~31%	~¥33T	18.0%
H1 FY2026 (Sep 2025)	~$2.48B (run rate)	~38%	~¥33T	20.6%*

*LTV as of December 31, 2025

ARM royalties have grown from an estimated $1.68 billion in FY2023 to a $2.48 billion run rate — a 48% increase in three years — while Armv9 penetration has moved from negligible to approximately 38% of the royalty mix. The mechanism is straightforward: Armv9 chips command 2x the per-chip royalty of the prior Armv8 generation. Management projects Armv9 will reach 67-70% of the royalty mix at maturity, implying that the blended royalty per chip shipped is still roughly 40-50% below what it will be at full architectural transition, even with no increase in the number of chips shipped. This is the ARM compounding story: the installed base of devices does not need to grow for ARM's royalty revenue to grow. The chips already being designed are getting more expensive per unit of compute, and Armv9 captures a higher fraction of that value.

The data center dimension accelerates this argument. ARM has grown from approximately 5% of hyperscaler server market share in 2020 to approximately 50% in early 2026 — a 10x expansion in five years, driven by Apple Silicon's demonstrated performance and energy efficiency advantages and Amazon Graviton's economics for cloud workloads. The addressable data center chip market is $242 billion in 2025, growing toward $1.2 trillion by 2030. ARM captures royalties on server chips at rates reflecting the much higher selling prices of data center silicon versus mobile chips. At current penetration of roughly 50% of hyperscaler servers, ARM has captured less than a third of the data center chip TAM, and AI inference chips — where ARM architecture is beginning to compete directly — represent the fastest-growing and highest-ASP segment. The penetration runway is real, though the timeline for capturing it is uncertain.

The holding company has captured approximately 65% of its $207 billion NAV in two positions — ARM at roughly $115 billion and OpenAI at roughly $111 billion. The remaining 35% is spread across SoftBank Corp, Vision Fund holdings, and various direct investments. At SoftBank Group's market cap of approximately $134 billion, the investor is paying roughly 65 cents for each dollar of NAV — receiving the ARM moat and OpenAI's growth trajectory at a substantial discount, but accepting the holding company overhead, the leverage risk, and Masayoshi Son's capital allocation judgment as part of the bargain.

At SFTBY's price of approximately $11.16 as of early April 2026, SoftBank's market capitalization is approximately $134 billion against an enterprise value of approximately $220 billion. Against an adjusted NAV of roughly ¥31-33 trillion ($207-220 billion at current exchange rates), the stock trades at a 35-37% discount. The ARM stake alone — approximately 90% of ARM's public market capitalization — represents $103-117 billion, essentially equivalent to SoftBank's full market cap. Everything else — the OpenAI stake, the Japanese telecom business, the Vision Fund portfolio, and the net cash position — is available at no additional cost if the balance sheet can be sustained.

This arithmetic is the source of the investment case's appeal and its trap. The appeal: the NAV discount is real and the assets behind it are genuinely excellent. The trap: the discount has already narrowed substantially (from 60% in 2021 to 35% today), so the investor is not discovering an overlooked situation but rather inheriting a risk that the market has explicitly priced. The intelligent bear on this position argues that SoftBank is essentially a leveraged OpenAI bet at an $852 billion private valuation — that the ARM stake is the fig leaf that makes it look like an investment vehicle rather than what it is — and that Masayoshi Son's track record of mistaking capital availability for insight is directly on display in the $64.6 billion commitment to a company he does not control, funded partly by a $40 billion loan that expires in twelve months. The answer to this objection is specific: the Armv9 royalty ramp is an operating phenomenon, demonstrated in quarterly filings, independent of AI narrative and of OpenAI's trajectory. Even a bear who dismisses the entire OpenAI position as overleveraged hype must explain why ARM's royalty per chip is objectively increasing as Armv9 penetration rises. The bear's best argument is about the capital structure, not the underlying semiconductor asset.

The conclusion is that SoftBank at current prices is interesting but requires a specific catalyst to be actionable. That catalyst is the OpenAI IPO, which would crystallize the $111 billion stake value, enable repayment of the $40 billion bridge loan, and remove the balance sheet uncertainty that justifies the current 35-37% NAV discount. An investor with high confidence in an OpenAI IPO within the next twelve months would find the current price attractive. An investor who requires proof of the catalyst before committing capital should wait: the Q2 2026 and Q3 2026 earnings calls will provide evidence of whether OpenAI's IPO timeline is advancing or slipping, and the LTV trajectory will confirm whether the balance sheet is stable or deteriorating.

What would change this verdict? A meaningful deterioration in ARM's royalty growth — particularly if the AGI CPU launch alienates existing licensees and accelerates customer migration to RISC-V — would make the underlying asset less compelling regardless of holding company discount. A bridge loan refinancing at prohibitive terms, or an LTV breach above 25% forcing asset sales, would convert the discount from an opportunity into a distress situation. Conversely, a decline in SFTBY's price that pushes the NAV discount back above 50% would create a margin of safety sufficient to be compelling even without the OpenAI catalyst — the ARM business alone, bought at a 50% discount to its public market value through the holding company structure, would represent an attractive entry regardless of the other portfolio positions.

The assets are real, the discount is real, and the catalyst is visible on the calendar. Whether the calendar holds is the only question that matters.