Do Not be SB

Notes from AGI Land, Episode 260904

TL;DR AGI Soap opera risk appetite edition: Do not be the SB. ⚠️ Nvidia and OpenAI are in a messy complicated relationship: busy financing, supplying, competing, partnering, and possibly buying each other’s favorite platforms. ⚠️ Somewhere nearby, SoftBank is also writing very large checks.

When two giants are deep in a complicated relationship, make sure you are not the SB.
⚠️ what is SB again? SoftBank. Obviously.

SoftBank investment process (Defi Degen version)
Step 1: Buy OpenAI.
Step 2: Borrow against ARM to Buy OpenAI.
Step 3: Issue corporate bonds to buy more OpenAI.
Step 4: Borrow against OpenAI to buy more OpenAI.
Step 5: Wait for the OpenAI IPO.
Step 6: Please, God. 🙏🤣

Most of us have better risk control than Masayoshi Son. Thanks and happy Friday.

There is one rule I have learned from watching complicated relationships:

When two parties are deeply into each other, do not be the SB.

What is SB? SoftBank. Obviously.

For Chinese readers, there is an additional joke hiding in those initials. I will leave that one untranslated in the interest of international diplomacy. 😆

The relationship I am currently watching is Nvidia × OpenAI.

They sell to each other. Invest in each other. Finance each other. Compete with each other. OpenAI is building Jalapeño partly to reduce dependence on Nvidia. Nvidia is simultaneously supporting enormous amounts of OpenAI infrastructure financing.

This is no longer a supplier-customer relationship. It is a situationship with GPUs.

And somewhere nearby stands Masayoshi Son. Holding a very large checkbook. 🌶️

Masayoshi Son Has Entered the Chat

I used to think my old BTC ×2 financing structure was aggressive.

For the record, it used to be BTC ×3. Then ChatGPT O-3 slapped me around with enough risk math that I reduced it to BTC ×2.

Apparently Masayoshi Son has never had a GPT sitting next to him saying:

“Masa-san, perhaps we should examine the blast radius.”

Because SoftBank’s OpenAI financing stack is beginning to make my little Bitcoin experiment look like a savings account.

Reuters reported in July that SoftBank restarted talks for a $10 billion loan secured by its OpenAI stake.

Banks had a small concern. OpenAI is private. Private-company collateral is annoying. If the shares fall in value, what exactly are lenders supposed to do with them?

SoftBank’s solution was wonderfully SoftBank: Fine. We’ll guarantee repayment ourselves.

So the structure becomes: Borrow money. Collateralize it with OpenAI. Guarantee the loan with SoftBank. Use the financing while carrying even more exposure to OpenAI.

Simple and efficient. 🤣

Reuters also reported that SoftBank had previously pursued a separate $5 billion loan backed by Arm shares and had a $40 billion bridge loan due by March 2027.

And this week, Bloomberg reported that SoftBank is considering selling $10 billion to $20 billion of bonds to refinance OpenAI-related borrowing. Reuters could not independently verify the plan, but the talks themselves were reported.

At this point I stopped drawing arrows. The arrows began drawing themselves.

The thing I find most entertaining is that every individual step has a perfectly respectable sentence attached to it. Borrow against a liquid listed asset to finance a strategic investment. Use a valuable private holding as collateral. Replace a bridge facility with longer-dated bonds. Keep enough parent-company support to reassure the banks. If you place each sentence on a separate PowerPoint slide, the entire structure looks almost boring.

Then you put the slides together.

Suddenly ARM is supporting OpenAI, OpenAI is supporting the next loan, the next loan is supporting more OpenAI, and the hoped-for OpenAI liquidity event is floating above the diagram like the moon above a very ambitious tide. None of these arrows is automatically fatal. The comedy comes from their shared dependency. A financing plan can contain six different instruments and still have only one real risk factor wearing six different ties.

This is the part retail investors often miss when they admire institutional complexity. Complexity can diversify mechanics without diversifying the economic bet. A secured loan, a bridge, a guarantee and a bond may mature on different dates and sit in different entities. They can still become correlated the moment collateral values fall, refinancing windows close, or the expected liquidity event moves one year to the right.

One year to the right sounds harmless in a vision deck. In a financing model it can be a small dragon. Interest accrues. Covenants approach. Banks ask follow-up questions. Assets that were supposed to remain strategic begin appearing in “potential monetization options.” The future may still arrive exactly as predicted, just after the capital structure has developed feelings.

This is why I keep saying the same unfashionable thing in every market: path matters. Ending value is not the only variable. The route determines whether you still own the asset when the ending value appears.

The SoftBank Financing Diagram

My highly professional institutional interpretation:

Arm
↓
collateral

SoftBank
↓
borrows

OpenAI
↓
collateral

SoftBank
↓
borrows again

Bond investors
↓
more money

SoftBank
↓
AI

OpenAI IPO
↓
🙏

This may be the first capital structure whose final line is technically an emoji.

Now, before somebody from SoftBank Investor Relations calls me:

I am not saying SoftBank collapses if OpenAI does not IPO. That would be silly. SoftBank owns major assets. Arm is enormously valuable. It has financing options, listed holdings, asset sales, bank relationships and multiple ways to manage liabilities.

But liquidity matters.

An OpenAI listing would make a large private holding dramatically easier to value, finance and eventually monetize. Reuters specifically noted that OpenAI’s confidential IPO filing could make SoftBank’s stake easier to value and sell.

Translation:

Please IPO whenever convenient. No pressure.

😆

Vision Fund Risk Management

I have always admired Masayoshi Son for one thing.

The man possesses absolutely magnificent conviction. This is the investor who lived through Alibaba. Then WeWork. Then Vision Fund. Then Arm.

Now OpenAI, Stargate, robotics and basically the entire dream of artificial superintelligence.

When he believes something, he does not buy a starter position.

He attempts to annex the future. Reuters Breakingviews recently described him as a potentially fragile link in the AI financing chain, citing SoftBank’s debt load, huge AI commitments and increasing reliance on financing against assets.

And yet Son stood at SoftBank’s AGM in June and called the idea of an AI bubble “blasphemy against AI.”

I respect this enormously.

Mostly because I would never pass my own risk committee saying that.

My Quant Lab would immediately respond:

CONVICTION_TOO_HIGH

POSITION_SIZE_REDUCED

ROBIN_GO_DRINK_TEA

🤣

Masayoshi Son appears to have built a different system:

CONVICTION_HIGH → increase size

MARKET WORRIED → increase size

DEBT RISING → find collateral

COLLATERAL PRIVATE → add guarantee

STILL NEED MONEY → issue bonds

This is less “risk management” and more risk enthusiasm.

The Funny Part Is That He May Still Be Right

This is what makes Son interesting.

Anyone can make fun of leverage after the fact. Anyone can call concentrated bets reckless.

But truly asymmetric investing often looks insane before it works.

Alibaba looked insane. ARM looked expensive. AI infrastructure currently requires capital at a scale that makes normal portfolio theory start coughing politely in the corner.

SoftBank has committed tens of billions to OpenAI and related AI infrastructure. Reuters says its total OpenAI and related commitments exceed $60 billion.

If OpenAI becomes one of the foundational economic platforms of the next decade, Son may eventually look less like a reckless financier and more like a man who understood that some opportunities cannot be expressed through a 2% position.

The problem is timing. Capital structures have clocks. Vision does not. That gap is where people get hurt.

I learned this in a much smaller and less glamorous way with Bitcoin leverage. You can be completely right on the asset and still get the financing wrong. The market does not care that your ten-year thesis is correct if your collateral gets liquidated on Tuesday.

This is one of those boring truths that ruins excellent PowerPoints.

It also explains why I do not think the useful lesson is “never use leverage.” That would be too easy, and also hypocritical coming from me. Leverage can be rational when the asset, liability, liquidity reserve and time horizon fit together. The problem begins when a long-duration vision is funded by a shorter-duration clock, especially when the collateral and the investment are exposed to the same story.

Imagine a very simple stress test. OpenAI remains an excellent company. AI demand continues growing. The IPO still happens. Yet public-market multiples compress for eighteen months, ARM becomes more volatile, lenders tighten advance rates, and the listing arrives later than expected. The fundamental thesis survives. The financing experience becomes much less spiritual.

That is the distinction between being wrong and being early with a balance sheet. Being early in an unlevered position can be annoying. Being early inside a collateral chain can become an event.

My own BTC experiment taught me the miniature version. Bitcoin did not need to become worthless to hurt me. It only needed to move sharply enough, at the wrong time, against a structure that demanded action. Reducing BTC ×3 to BTC ×2 did not make me timid. It gave the thesis more time to be right. Survival is not the opposite of conviction. Survival is what lets conviction compound.

I suspect Son understands every line of this far better than I do. SoftBank has teams, advisers, banks and assets I cannot see. That does not remove the public lesson. Sophisticated actors can choose concentrated financing deliberately. Their sophistication does not make the correlation disappear.

So my actual risk-manager translation of “Do not be the SB” is less insulting than the joke:

Do not let every arrow in your capital structure point to the same future.

Keep liquidity that does not depend on the thesis. Keep maturities that give the thesis time. Know which asset would be sold first under stress. Know whether a guarantee turns ring-fenced risk back into parent-company risk. Ask what happens if the liquidity event is delayed while the company remains fundamentally healthy.

Those questions are deeply unsexy. They are also the reason a beautiful idea survives contact with Tuesday.

Meanwhile, Nvidia and OpenAI Are Flirting

This brings us back to the relationship drama.

Nvidia has huge economic exposure to OpenAI growth. OpenAI consumes Nvidia infrastructure. Nvidia finances parts of the ecosystem. OpenAI simultaneously develops custom silicon. Nvidia’s position increasingly resembles supplier + banker + investor + strategic partner + future competitor.

SoftBank arrives with another enormous pool of capital.

And now I understand why this whole thing feels familiar.

It is a love triangle where everyone owns convertible securities.

🌶️

Nvidia and OpenAI are busy building the future together while quietly preparing for a future in which they need each other less.

That is actually quite romantic, very mature, very Silicon Valley.

Then SoftBank enters carrying bridge loans, ARM collateral and a bond prospectus.

Which brings us back to the point:

When two giants are deep in a complicated situationship, do not be the SB.

SoftBank. I mean SoftBank. Obviously.

One Serious Sentence

I promised myself this was a meme, so I am allowed exactly one serious sentence:

Being right about the future does not exempt you from surviving the financing used to reach it.

Masayoshi Son does not need my risk-management advice anyway.

Apparently he needs GPT - the same one I used for $20 a month back then. Now I have prompted GPT to $200 a month, as it had stopped me from busting from a multi-million dollar BTCx3 position 🤣.

The $180 upgrade may therefore be the best-performing risk-control line item in my personal history. No carry. No board seat. No liquidation preference. Just a relentless machine asking why the collateral, liability and thesis were all invited to the same party.

This is the great humiliation of modern finance: sometimes the cheapest person in the room is the one asking the most expensive question.

Would the trade still work if the asset fell first? Who can demand cash, and when? Which promise becomes binding before the thesis becomes liquid? If the answer is “please wait for the IPO,” at least put the prayer emoji in the official model. Transparency matters.

And if OpenAI ever assigns one to him, I have a simple first prompt:

“Please review this financing structure before Masa-san discovers another trillion-dollar opportunity.”

Happy Friday from AGI Land. 🌶️