OG | AI & Markets

Following the money behind AI. The hype, the numbers, and who pays for it.

[OG | AI & Markets @0xCapexOG]

Amazon says its servers last six years. Meta says eleven. Nvidia ships a new architecture every eighteen months.

Wall Street just started financing that hardware like real estate.

Wrote up what that one assumption is holding up. https://t.co/DXF0gD0uB9

[OG | AI & Markets @0xCapexOG]

AI doesn't need to take over the financial system.

The financial system already outgrew us.

02:44 - the bond market already sits past what a human brain can hold. You need enormous data and complex math just to follow it.
03:04 - so humans keep handing financial decisions to machines, one at a time.
03:11 - and AI will build instruments more sophisticated than bonds and stocks.
04:03 - the endgame isn't a robot uprising. It's advisers telling a president "markets are collapsing, the AI says do X, we don't know why, we have to believe it".
05:14 - trust doesn't evaporate. It migrates. From bankers to algorithms.
Now put that next to something that happened this month. BlackRock's CEO compared AI data center financing to the birth of the mortgage-backed securities market, and he meant it as praise.
MBS is the last time an instrument outran the people pricing it.

[OG | AI & Markets @0xCapexOG]

The bubble can be in the financing even if the demand is real.

Those two get argued as if they're the same question. They aren't.

Demand: Anthropic planned for compute to grow tenfold this year and saw revenue grow more than threefold in a single quarter. Oracle is sitting on $638B of contracted backlog.

Financing: that same Oracle is running negative $23.7B of free cash flow, funding capex with $43B of debt and roughly $40B more planned.

Real demand doesn't protect you from a structure that reprices faster than the contracts pay.

That's the actual trade. Not whether AI works.

[OG | AI & Markets @0xCapexOG]

The bubble can be in the financing even if the demand is real.

Those two get argued as if they're the same question. They aren't.

Demand: Anthropic planned for compute to grow tenfold this year and saw revenue grow more than threefold in a single quarter. Oracle is sitting on $638B of contracted backlog.

Financing: that same Oracle is running negative $23.7B of free cash flow, funding capex with $43B of debt and roughly $40B more planned.

Real demand doesn't protect you from a structure that reprices faster than the contracts pay.

That's the actual trade. Not whether AI works.

[OG | AI & Markets @0xCapexOG]

Amazon says its servers last six years. Meta says eleven. Nvidia ships a new architecture every eighteen months.

Wall Street just started financing that hardware like real estate.

Wrote up what that one assumption is holding up. https://t.co/DXF0gD0uB9

[OG | AI & Markets @0xCapexOG]

Amazon says its servers last six years. Meta says eleven. Nvidia ships a new architecture every eighteen months.

Wall Street just started financing that hardware like real estate.

Wrote up what that one assumption is holding up. https://t.co/DXF0gD0uB9

[OG | AI & Markets @0xCapexOG]

Ed Zitron just did two and a half hours on why the AI buildout breaks in 2027. A million people watched it before lunch.

He's right about the mechanism and wrong about the timing, and the difference matters.

What holds up: the buildout is debt-funded now. Oracle is running negative $23.7B of free cash flow while capex heads toward $70B, funded with $43B of debt and $40B more planned.

What doesn't: he treats the demand as fake. Anthropic planned for compute to grow tenfold this year and saw revenue grow more than threefold in a single quarter. That is not manufactured.

The honest version is narrower than either camp wants. The demand is real, the financing is fragile, and those two things can be true at once.

Which is why the number to watch isn't revenue. It's the gap between when contracts pay and when debt reprices.

[OG | AI & Markets @0xCapexOG]

Oracle is carrying $638B of contracted backlog against negative $23.7B of free cash flow.

[OG | AI & Markets @0xCapexOG]

Ed Zitron just did two and a half hours on why the AI buildout breaks in 2027. A million people watched it before lunch.

He's right about the mechanism and wrong about the timing, and the difference matters.

What holds up: the buildout is debt-funded now. Oracle is running negative $23.7B of free cash flow while capex heads toward $70B, funded with $43B of debt and $40B more planned.

What doesn't: he treats the demand as fake. Anthropic planned for compute to grow tenfold this year and saw revenue grow more than threefold in a single quarter. That is not manufactured.

The honest version is narrower than either camp wants. The demand is real, the financing is fragile, and those two things can be true at once.

Which is why the number to watch isn't revenue. It's the gap between when contracts pay and when debt reprices.

[OG | AI & Markets @0xCapexOG]

Amazon says its servers last six years. Meta says eleven. Nvidia ships a new architecture every eighteen months.

Wall Street just started financing that hardware like real estate.

Wrote up what that one assumption is holding up. https://t.co/DXF0gD0uB9

[OG | AI & Markets @0xCapexOG]

@nikolayxyz That's the whole thing in one line, better than I put it. And financing stress shows up in duration long before it shows up in defaults.

[OG | AI & Markets @0xCapexOG]

The odds are chasing the fundamentals here, not the other way around. Anthropic's annualized run rate went from $9 billion at the end of 2025 to over $65 billion by July, more than 7x in seven months. An October IPO target is already on the table, and prediction markets have had a track record of calling these milestones before the actual filings do.

[OG | AI & Markets @0xCapexOG]

This reverses Nvidia's own written policy - as of 2023, their governance filings explicitly said the company would never contribute to any PAC, direct or through intermediaries. They still managed $1.49 million in campaign contributions that year without one. Something changed enough in the last couple of years to flip a stated policy like that, and it's not hard to guess what.

[OG | AI & Markets @0xCapexOG]

Turns out the answer to "who pays for AI" is: anyone who needs to borrow money.

01:07 — over $1T of capex this year, and more than $2T by 2028
08:54 — why the money keeps coming: $6B of fab capex eventually generates over a trillion dollars of end AI revenue
54:34 - the bill through 2029 is around $11T, and roughly $5T of it has to be raised as debt.
57:06 - Meta borrows at 5-6% today. Patel: "I don't see why they wouldn't pay 8%." They'd happily pay it, because the return on the compute is enormous.
57:31 - but that 250 basis points doesn't stay inside AI. It reprices borrowing for every bank, telecom, packaged-goods company and mortgage in the economy.
58:45 - the precedent he reaches for is Volcker. When rates jumped in the 1980s, around 40 countries defaulted.
Everyone still reads AI capex as a technology story. The people actually building it are describing a rates story.
That's the part that eventually shows up in your life whether you use any of these models or not.

[Gary Marcus @GaryMarcus]

Abandon ship!

[OG | AI & Markets @0xCapexOG]

Turns out the answer to "who pays for AI" is: anyone who needs to borrow money.

01:07 — over $1T of capex this year, and more than $2T by 2028
08:54 — why the money keeps coming: $6B of fab capex eventually generates over a trillion dollars of end AI revenue
54:34 - the bill through 2029 is around $11T, and roughly $5T of it has to be raised as debt.
57:06 - Meta borrows at 5-6% today. Patel: "I don't see why they wouldn't pay 8%." They'd happily pay it, because the return on the compute is enormous.
57:31 - but that 250 basis points doesn't stay inside AI. It reprices borrowing for every bank, telecom, packaged-goods company and mortgage in the economy.
58:45 - the precedent he reaches for is Volcker. When rates jumped in the 1980s, around 40 countries defaulted.
Everyone still reads AI capex as a technology story. The people actually building it are describing a rates story.
That's the part that eventually shows up in your life whether you use any of these models or not.

[OG | AI & Markets @0xCapexOG]

Amazon says its servers last six years. Meta says eleven. Nvidia ships a new architecture every eighteen months.

Wall Street just started financing that hardware like real estate.

Wrote up what that one assumption is holding up. https://t.co/DXF0gD0uB9

[OG | AI & Markets @0xCapexOG]

The Fed's own Senior Loan Officer Survey from three weeks ago backs this up in an important way, lending standards stayed basically unchanged this quarter. That means this surge is demand-driven, not banks getting reckless, and that distinction usually decides whether a credit boom ends well or badly. One sector in particular has been quietly absorbing an outsized share of that new borrowing.

[OG | AI & Markets @0xCapexOG]

The $38B RWA figure isn't Solana-specific, it's the entire cross-chain tokenization market. Solana's own July ecosystem report puts its actual RWA value at $3.73 billion, real growth, but a fraction of that headline number. One asset category is quietly doing most of the real work behind that growth, and it's not the one most people assume.

[OG | AI & Markets @0xCapexOG]

@ArtificialAnlys The number that matters here is 51 vs 51, Qwen3.8-27B scores nearly identical to Kimi K2 on Artificial Analysis' agentic index, despite Kimi being a 2.8 trillion parameter model. That's a 100x parameter gap closing to basically zero on agentic tasks.

[OG | AI & Markets @0xCapexOG]

@KobeissiLetter The 90% problem: China buys the vast majority of Iran's oil, and sanctioning Chinese banks over it looks unlikely right before Xi's September visit. Rhetoric this sweeping usually has a carve-out nobody's saying out loud yet.

[OG | AI & Markets @0xCapexOG]

@unusual_whales The bigger story is the swing, not the tie, Democrats were priced at just 18% a year ago on these same prediction markets. Hartnett's team likes gold as a hedge no matter which way this goes.

[OG | AI & Markets @0xCapexOG]

@LizAnnSonders Gold's still near 17% below its January record of $5,602, this is a partial recovery, not a new high. And it's happening even as global gold demand hit its lowest level in nearly five years, this rally is running on policy narrative, not physical buying.