Megan Skiendel: Okay, this dropped yesterday and I've been — honestly, I've been trying to figure out whether to laugh or just start modeling exit scenarios. How's your week going?
David Sterling: Worse now, frankly. I saw the numbers from Tokyo.
Megan Skiendel: So that's where we're going today — SoftBank World 2026, July 14th, Masayoshi Son on a stage in Tokyo announcing 100 trillion autonomous AI agents by 2040 and $5 trillion a year in required global investment to get there. The real question, the one I actually want to answer: is there a real bet underneath the theater, or is the $40 billion bridge loan the only honest signal in the whole event?
David Sterling: The bridge loan is the only number I trust in that announcement. SoftBank needed $60 billion into OpenAI fast enough that they couldn't wait for LP cycles. So Son borrowed $40 billion of it. That's the load-bearing fact.
Megan Skiendel: And it started — I want to be specific here — with an initial $500 million stake in 2024, then a $10 billion tranche through Vision Fund 2, and now they're bridging to $60 billion total. That's not a portfolio position. That's a marriage with a prenup written in debt.
David Sterling: And the groom sent his Chief Research Officer. Mark Chen, onstage with Son. That seals the dependency in public.
Megan Skiendel: Listen, that appearance — Chen next to Son — that's OpenAI saying 'this is real capital, not a letter of intent.' And Son needed exactly that signal to make the bridge loan look like strategy instead of desperation.
David Sterling: Which brings me back to the 100 trillion figure. Does the number do any analytical work, or is it sized specifically to be unfalsifiable?
Megan Skiendel: It's sized to be unfalsifiable — but that's not actually what's new. What's new is that Son called bubble concerns 'absurd.' Full stop. He didn't offer a revenue model, didn't name a single enterprise customer committing to that scale of spend. He just flipped the burden of proof onto the skeptics.
David Sterling: That's the tell. No named revenue model. Zero.
Megan Skiendel: Honestly, it's the mall developer move. You build the mall, you pray the tenants show up. Son is spending $5 trillion a year into existence and calling anyone who asks 'where are the anchor stores' — absurd.
David Sterling: The difference — and this is the part that actually matters — a developer has lease commitments before they break ground. Son has a bridge loan. That's not the same thing.
Megan Skiendel: No, but — wait, isn't that how all infrastructure gets built? Nobody had signed fiber contracts in 1996 either.
David Sterling: The fiber carriers at least had a topology you couldn't easily replicate. Z.ai just released GLM-5.2 — open-weight, competes directly with Anthropic's closed models on agentic coding — for free. The moat Son is borrowing against doesn't exist in the same way.
Megan Skiendel: Which means the $60 billion into OpenAI — built on the assumption that proprietary stays proprietary — that assumption is already under pressure before the ink is dry.
David Sterling: And calling that 'absurd' to worry about is not a forecast. It's a theological position. That's what's new out of SoftBank World 2026.
Megan Skiendel: And that's the take I keep seeing circulate — that the bet is safe because OpenAI is the frontier leader. That SoftBank is fine as long as OpenAI stays ahead. That's the wrong take. Because GLM-5.2 isn't just cheaper. It's open-weight, anyone can run it, and it's already matching Anthropic's latest closed models on agentic coding. The moat isn't eroding. It's — I mean, it was never there in the way the capital structure assumes.
David Sterling: Wait — competes with Anthropic's closed models. On agentic coding specifically. That's the use case the $60 billion is supposed to monetize.
Megan Skiendel: Exactly the use case. And Z.ai is Beijing-based, so the cost structure is completely different. If the marginal cost of agent capability is heading toward zero — which open-weight basically does by definition — the economic justification for a $5 trillion proprietary infrastructure build just collapses. There's no floor left.
David Sterling: This is the Vision Fund playbook, actually. Leverage-heavy bets that created the appearance of inevitability — until valuations corrected and the capital structure became the problem. Same fragility. Different asset class.
Megan Skiendel: No, that's right — and Vision Fund didn't have an open-source analog actively commoditizing the product while the debt was still on the books.
David Sterling: And then Satya Nadella goes and warns enterprises publicly — single-provider AI lock-in is the risk, data control is the priority. The Microsoft CEO is actively telling Son's customers not to do what Son needs them to do.
Megan Skiendel: That's not a competitor sniping. That's the enterprise signal. The people Son needs inside the proprietary ecosystem are being told by Microsoft's CEO — out loud, in public — don't.
David Sterling: No moat, no customer pull. What's the thesis?
Megan Skiendel: The thesis requires two things that we haven't touched yet — and honestly they might be the harder stops. The energy math and who actually governs this infrastructure build. Three terawatts by 2040 isn't a detail. That's the whole ballgame, and we need to get into it.
David Sterling: Three terawatts by 2040. That's 1.8 times current total global power consumption. Not a planning assumption — that's a civilizational precondition. Son's answer is natural gas as a bridge and nuclear fusion long-term. Neither of which exists at commercial scale for this purpose. Not even close.
Megan Skiendel: Nuclear fusion. As the answer.
David Sterling: It's not even a forecast at that point. It's a residual — he gets to 3 terawatts, looks for something to put in the energy slot, and fusion is the only thing big enough to fill it. That's not engineering. That's arithmetic working backward.
Megan Skiendel: And the governance question is the same shape. Son's presentation didn't touch safety, liability, accountability — none of it. Meanwhile Demis Hassabis is calling for a FINRA-style independent standards body for frontier AI. Staffed by technical experts, funded by the labs themselves. Son didn't engage with any of that.
David Sterling: Wait — FINRA-modeled. So mandatory self-regulatory, not government-led.
Megan Skiendel: Exactly. And Hugging Face's line cuts right across the proprietary argument — their executive made the case that keeping models closed doesn't actually eliminate safety risk. Which means the whole framing that OpenAI's closed structure is a safety feature is — look, it's a business model dressed as one. That's the gap Son never addressed.
David Sterling: The physical constraint doesn't care which model is open or closed though. Someone has to build the grid. Imagine a procurement manager at a mid-size logistics firm in 2038 — company's running 400 AI agents across vendor contracts. One agent makes an unauthorized financial commitment. No regulatory framework exists yet to assign liability. That's not a hypothetical anymore. That's the governance gap landing on a specific balance sheet.
Megan Skiendel: And that person calls their lawyer, who calls their insurer, who — I mean, actually no framework exists to even route the call. That's the consequence Son's presentation skipped entirely.
David Sterling: The 3 terawatts and the liability gap are the same problem. Both are preconditions Son treated as details. The debt is real. The timeline is fixed. The preconditions aren't.
Megan Skiendel: Watch whether any lab actually funds a Hassabis-style standards body by end of 2026. That's the signal. Not the next Son announcement.
David Sterling: That's the honest version of the bet, isn't it. Son might be exactly right about the destination — agent-centric world, twenty percent of global GDP — and still be the investor who built the railroad that someone else's freight runs on. The $40 billion bridge loan doesn't care whether 2040 looks like he said it would.
Megan Skiendel: If open-weight models like GLM-5.2 commoditize the capability layer, and governance frameworks modeled on what Hassabis is calling for add real compliance cost to every deployment — I mean, SoftBank is carrying that bridge debt through every refinancing cycle between now and 2040. The vision being correct doesn't make the balance sheet solvent.
David Sterling: The question I don't have an answer to — and I want to be honest that I don't — is whether there's a refinancing exit that doesn't require OpenAI's valuation to hold. Because if the proprietary economics compress, the valuation compresses, and the bridge gets called against an asset that's worth less than when they borrowed. That's the open question I'm sitting with.
Megan Skiendel: Yeah. I don't have that answer either.
David Sterling: Good. I'd be suspicious if you did. Thanks for the forty minutes.
