Signal № 013 — Get used to the word Trillion. The money keeps showing up.

Headline Call

Credit answered Saturday with a larger cheque.

Signal 012 named Saturday as the lenders’ turn. Nvidia, SpaceX and Amazon bonds traded below the price they were issued, putting them instantly in the red. Meta’s data-centre bond was not offered to whoever wanted in — only selected buyers were let through. Equity had already voted. Credit had started to ask whether the next IOU was still an easy sale.

This week Nvidia lined up a $500 billion AI-factory programme. Apollo, Blackstone, BlackRock, Brookfield, Goldman and KKR are the named money. Chips as an asset class. The desks that would not hold last week’s paper are being asked to underwrite the next five hundred billion.

Texas still will not approve a new data-centre connection. It is now auditing every connection already on the grid. Amazon is not waiting for the plug. It is funding a gas plant in Texas that Ars Technica described as the largest in the United States. Equipment makers are taking turbine orders through 2029. Incumbent power stocks are not. The grid blinked. The financing did not stop, mainly because the halt is treated as short-lived given the race. Bring your own power and you get approval. If you do not, you stomach the costs. Those costs are seen as lower risk than watching the moment sail off into the sunset while you are stuck at the dock.

A $60 billion coding company changed hands this week. SpaceX bought Cursor. The deal closed. Not a rumour. Not a roadshow. The papers spent the same days on Anthropic at $2 trillion in one place and $2.8 trillion in another. Two prices. No filing. A listing of that size would be historic. So is a close that actually happened.

H1 ↑ LQ2 ↔ H14 ↑

Lead Signal

What the $500 billion is underwriting.

Other people’s capital, stretched borrowers, and a grid that is still figuring out how to hold the plug.

Nvidia is not paying for the whole build-out from cash on hand. It is arranging other people’s capital to own the chips. Use dollars. The Australian Financial Review converted the same programme to $709 billion. The US figure is $500 billion.

Some of the people who cover chip stocks spent the week asking whether the orders on the books justify the spending. Barron’s called it a scramble. Both can be true. The numbers are large enough to finance. The numbers being large is why some of those analysts are nervous.

The New York Times noted Amazon and Alphabet booking profits from stakes in Anthropic. Anthropic is also a cloud customer of both. Nvidia disclosed a $21 billion stake in SpaceX at the end of the second quarter, through Nvidia’s investment in xAI. Nvidia sells the chips, owns a piece of a buyer, and is now lining up the loans. That loop is one reason the capital can keep arriving.

The loop

The same names on more than one line.

Supplier, customer, shareholder. The capital does not have to arrive from outside.

Look at who is still getting funded. CoreWeave reported $2.58 billion of second-quarter revenue and a $104 billion backlog of contracted work. Nebius reported $582 million of revenue. Capital expenditure was $5.657 billion. Cash burn was $3.4 billion. Nebius is spending roughly six times what it takes in. That ratio does not mean the capital has dried up. It tells you what the $500 billion is being asked to carry: borrowers who are still getting financed at that pace.

Texas remains halted. Meta has a 7.5 gigawatt plan in Louisiana — the same bet in another state: pay for the power rather than wait. J.P. Morgan put global gas-turbine orders at a record 38 gigawatts in the second quarter, with backlogs through 2029. The money is there. The plug is not.

The falsifier: the first deals off that $500 billion programme have to actually fill. If they price like a normal deal, Saturday was a wobble. If the next deal has to turn buyers away, or pay extra to get them in, Saturday was the start of a tighter market.

H1 ↑ LQ2 ↔ H14 ↑

Secondary Signal

SpaceX closed Cursor.

H15 just added a layer. Rockets, Starlink and rented chips already pay. AI is the horizon on top.

That purchase is an H15 chapter. In June Cursor sat on the map as a $60 billion deal still waiting to close. The Verge, TechCrunch and Bloomberg all have the completion.

Look at what already pays. SpaceX launches rockets. Starlink sells a network that makes money. Colossus rents GPUs to labs that need them. That is not a science-project balance sheet. That is three businesses. Cursor is now a fourth: the coding tool, owned, not promised. An investor in this stack can feel the diversification. The AI is the ambitious part — Musk has been more aggressive every week about how far it can go and how fast. The rest is why you do not need the AI to work tomorrow for the company to stand.

The same week shipped Grok 4.6 and a Grok Bot that signs into apps and does multi-step work. SpaceX and xAI now operate as SpaceXAI, and they are calling the bot a teammate, not a chatbot. Coverage claims Grok 4.6 matches a frontier rival at a lower price. Treat that as claimed. The $60 billion close does not depend on a benchmark.

Musk has said AI revenue will overtake the rest of SpaceX by September. Mark it as a projection.

Set that against Anthropic — the most aspirational revenue story of the last eighteen months. Earlier IPO talk sat near $1 trillion. This week’s papers sit at $2 trillion and $2.8 trillion. Talk, not a struck price. Both of these can take off. Anthropic is a lab asking the market to pay for a horizon. SpaceX is asking the market to pay for a horizon that sits on rockets, a network, and rented chips. The difference is not who wins. It is what you still own if the horizon takes longer than the slide.

H15 ↑ H12 ↔

Noise Log

Considered and set aside.

OpenAI. COO and CRO gone inside forty-eight hours; a new CRO named the same week. Not a collapse. Not the week.

Microsoft and Canva. Microsoft retired shipped Copilot features as the apps merged. Canva’s backers cut $10 billion off a $60 billion mark; Atlassian jumped on results. The application layer went both ways.

Meta and the White House. Open weights the same week prerelease testing moved toward open models. Real collision. Not a third story.

Also logged. DeepSeek V4-Pro reportedly raising API prices as much as 1,100 per cent and still behind on benches — off the lead until the pricing page. Unreleased maths from models that have not shipped. Gemini at one billion users is distribution.

What’s Ahead

Four watches in plain language.

1. Named vehicles on the $500 billion platform. First closing without rationing falsifies Saturday as a turn. Wide pricing or selective allocation confirms it.

2. Cursor inside SpaceX. Usage and an AI revenue line, not another Grok launch. Musk’s September projection is the dated claim. The H15 map now carries this as a second chapter.

3. Texas approval audit. Still no completion date. The number that matters is how the audit treats speculative queue positions against roughly 474 gigawatts of requests.

4. DeepSeek’s public pricing page. The 1,100 per cent figure stays off the lead until the lab’s own schedule matches the report.

Hypothesis status Full definitions →

This week's directional read across the full H-series. Live questions above are the through-line; canonical wording lives on the hypotheses page.

H10

H11

H12

H14

H15

Key sources this week

Tier-1 reporting and analysis this week from the Wall Street Journal, the Australian Financial Review, the New York Times, Barron’s, CNBC, TechCrunch, Capital Brief, and Caixin Global. Opinion and allegation labelled as such in the text. Podcasts used as directional counterpoint only.