SoftBank Group (Vision Funds + Arm + OpenAI)
Masayoshi Son (founder/chairman/CEO); Vision Funds 1 & 2 are the investment vehicles, with OpenAI held via Vision Fund 2.
SoftBank is the most concentrated AI mega-bet among the corporates: an ~$64.6B cumulative position in OpenAI (~13% ownership) via Vision Fund 2, ownership of ~90% of Arm (the dominant compute-IP licensor), and a new 'AI Computing Segment' bundling Arm, Graphcore and Ampere. It is also financing OpenAI's data-center buildout, taking on loans backed by its Arm stake.
- Not open — Not an open-ended vehicle — analysis only, not an available allocation.
Top holdings
~$64.6B cumulative (~13%) — Held via Vision Fund 2; valuation rose from $157B (Oct 2025) to $852B (March 2026 round co-led by SoftBank). Second-largest external holder behind Microsoft.
~90% owned. Dominant CPU/compute IP licensor; anchor of the new AI Computing Segment. Used as collateral for OpenAI-funding loans.
long (acquired). Arm-based server CPUs; folded into AI Computing Segment.
long (acquired). AI accelerator chips; part of AI Computing Segment.
SoftBank has held and traded an Nvidia position; size not currently disclosed.
Recent moves
Feb 27, 2026: agreed to $30B OpenAI follow-on via Vision Fund 2 (first tranche executed April 2026), lifting cumulative to ~$64.6B / ~13%. Racing to fulfill a $22.5B funding commitment by year-end; exploring a ~$40B loan for OpenAI. Created the AI Computing Segment in FY2025 reporting. Sources: SoftBank IR, CNBC, Reuters.
Loading disclosed positions…
Our take
The highest-beta way to own the AI trade through a listed parent: 9984.T is effectively a leveraged, marked-to-private bet on OpenAI plus near-total ownership of Arm's royalty stream. Edge: scale and conviction no other public vehicle matches. Caveats are severe — the OpenAI mark is private and volatile (the June IPO-delay drawdown shows the downside), the position is debt-financed against Arm, and as a Japanese listco it trades at a persistent holdco discount to NAV. Genuinely mission-critical concentration risk.
Blended cost basis is what the headline gain obscures.
Through 31 Mar 2026, $34.6B was in against a $79.6B carrying value — ~2.3x, struck off the $852B post-money the Feb 2026 round established. The follow-on adds $30B at a $730B pre-money ($10B each in April, July, October 2026; two tranches executed). Fully deployed, roughly half of the $64.6B carries almost no embedded gain, so blended MOIC compresses sharply unless OpenAI climbs from an already-record level. Second, the mark is partly self-referential: SoftBank is a price-setter in the primary rounds that establish the carrying value it books as profit — legitimate under IFRS fair-value accounting, still weaker evidence than an arm's-length exit. Third, the financing is duration-mismatched: the bridge matures 25 Mar 2027 while the listing reportedly slipped toward 2027, and a $10B margin loan sought against the OpenAI stake at ~SOFR+425bp — near triple its 2018 Alibaba-backed spread — had not closed as of July 2026.
Thesis
SoftBank has converted a diversified tech holdco into a levered, single-name proxy for OpenAI. Vision Fund 2's OpenAI position carried a $79.6B fair value against $34.6B cost at 31 Mar 2026, and the $45.0B unrealized mark on it produced essentially all of the ~$46B fiscal-year Vision Fund gain. The critique that matters is not whether OpenAI is a good asset — it is that the reported gain sits on the cheap early money while the majority of committed dollars ($30B of a $64.6B cumulative total) is being deployed at a $730B pre-money, on borrowed funds.
Corporate holdco pursuing vertical AI integration: a ~13% OpenAI stake (on completion of the third tranche) at the model layer, ~90% of Arm plus Ampere and Graphcore at the compute-IP layer, and balance-sheet participation in OpenAI's data-center buildout. Funded increasingly with debt — a $40.0B unsecured bridge facility signed 27 Mar 2026, maturing 25 Mar 2027 — rather than realized proceeds.
Assessment
- Conviction and scale no other listed vehicle matches: ~13% of OpenAI plus ~90% of Arm is genuine vertical exposure, not a passive minority stake.
- Vision Fund 2 is SoftBank-funded, so there is no external LP base to force gates, redemptions, or a fire sale at a bad mark.
- Early entry was real: $34.6B cost against a $79.6B carrying value at 31 Mar 2026 reflects timing, not just index beta.
- Owning Arm gives a saleable, listed, liquid asset that can be monetized independently of the private OpenAI position.
- The reported gain is unrealized and set largely by rounds SoftBank itself anchors — mark-setter and mark-holder are the same party.
- Incremental $30B goes in at a $730B pre-money, so the marginal dollar has little margin of safety versus the legacy position's ~2.3x.
- A $40B unsecured bridge maturing 25 Mar 2027 against a private asset whose listing reportedly slipped to 2027 is a refinancing-timing problem.
- A $10B margin loan sought on the OpenAI stake at ~SOFR+425bp — near triple the 2018 Alibaba-backed spread — and reportedly unclosed: lenders discount the collateral more than the balance sheet does.
- Persistent holdco discount: NAV/share ¥13,000 at 23 Jun 2026 against a share price roughly half that, even as NAV nearly doubled from ¥40.1T (31 Mar) to ~¥74T (23 Jun).
Record
FY2025 (ended 31 Mar 2026): Vision Fund segment gain ~$46B, of which $42.1B sat in Vision Fund 2 and was attributed to OpenAI; the OpenAI position alone contributed a $45.0B valuation gain, carried at $79.6B against $34.6B cost. Group NAV rose to ¥40.1T from ¥25.7T a year earlier, then to ~¥74T by 23 Jun 2026. Attribution is therefore almost entirely one private mark, not diversified selection skill — the rest of the Vision Fund book is a rounding error against it. The fragility showed on 26 Jun 2026: a report that OpenAI's IPO could slip to 2027 knocked SoftBank shares down 12–14% in Tokyo, a reported ~$38B of market value in one session, with no change in any operating fundamental. A book whose annual result is one unlisted valuation prints violently in both directions.
Risks & fit
- Refinancing: the $40B unsecured bridge matures 25 Mar 2027; without an OpenAI listing or asset sales, repayment leans on Arm/T-Mobile monetization at whatever price is available.
- Mark reversal: OpenAI carried near a record valuation — a down round or a flat 2027 listing turns the FY2025 profit into a reported loss with no cash having moved.
- Key-person: strategy, sizing, and the concentration decision all trace to Masayoshi Son; succession is undisclosed and unhedged.
- Reflexivity: SoftBank's own capital helps set OpenAI's price, so a funding pullback could compress the mark it needs to fund itself.
- Structural discount: even a correct AI call may not reach shareholders if the ~50% NAV discount widens as leverage rises.
The leverage-and-mark critique fails if OpenAI completes a listing at or above the reported ~$1T target and SoftBank actually converts paper into cash or non-recourse collateral — specifically: the 25 Mar 2027 bridge termed out or repaid from realized proceeds rather than new borrowing, the $10B stake-secured margin loan (or a successor) actually closing, and the discount to NAV narrowing rather than widening as debt rises. Conversely the critique is confirmed if the third tranche is funded with more bridge debt, the listing slips past March 2027, and the carrying value is written down while the maturity stands.
Readers analyzing how a listed holdco turns private-mark accounting into reported profit, how a marginal dollar's cost basis diverges from a headline MOIC, and how a holdco discount behaves under rising leverage. Also a case study in single-name concentration inside a nominally diversified vehicle. Note 9984.T is listed but Vision Fund 2 itself is SoftBank-funded and not open to outside capital.
No external management or performance fee — Vision Fund 2 is SoftBank-funded, so public holders pay no 2/20. The economic equivalents are the discount to NAV (¥13,000/share NAV at 23 Jun 2026 vs a share price roughly half that) and bridge-facility interest.