$29bn walked in. A credit rating walked out. Same week, same trade, opposite verdicts. The build-out just met its bill.

iPrompt Signals

AI & ROBOTICS INVESTING — EXPLAINED SO YOU CAN ACTUALLY ACT ON IT

Issue 18 · Friday 10 July 2026 · R. Lauritsen

On Thursday evening, SK Hynix priced the largest foreign listing in American history — $29 billion of new equity to build more memory plants. Hours earlier, S&P had cut Oracle to one notch above junk, citing the cash-flow strain of its data-centre build. Same trade, same day, opposite verdicts. The question of 2026 is quietly changing from who builds the biggest cluster to who gets to borrow — and at what price. Below: how to tell which side of that line your holdings are on.

Weekly Scoreboard

Bottom line: The market spent the week ignoring a war and repricing how the build-out gets paid for. That's the issue, in one line.

Prices are Thursday 9 July closes, rounded; weekly moves vs Thursday 2 July (US markets were shut Friday 3 July). Sources: exchange data via S&P Global Market Intelligence.

Top Headlines

1. SK Hynix just sold Wall Street the supply wave

The HBM leader — roughly 60% of the high-bandwidth memory feeding Nvidia's and Google's chips, on industry estimates — priced $29 billion of new ADRs at $149 on Thursday night: the largest foreign share sale in US market history, and reportedly about seven times oversubscribed. Proceeds fund the Yongin mega-fabs and an Indiana packaging plant. The story prices the equity channel: Wall Street just funded the supply wave — eagerly.

2. Apple wrote Broadcom a $30bn cheque

Apple committed $30 billion to Broadcom on Wednesday for wireless chips built in US factories — a purchase commitment, note, not cash upfront — and AVGO rode it to an 11% week. Micron's $3bn US fab expansion followed a day later. The story prices the customer channel: demand guaranteed in writing before the fab is warm, no dilution, no coupon. How much risk actually transfers depends on the contract terms — the deep dive separates commitments from true pre-payments.

3. Meta's cloud grew a product page

Last week's $150bn market move came on a Bloomberg report. This week Meta shipped something: Muse Spark 1.1, an agentic coding model priced aggressively low, sold through a new Meta Model API. Our read — and it's a read, not a fact: selling model access looks like the first step in making $115–135bn of capex pay for itself. The 29 July earnings call will tell us how much of that is real.

4. The market un-priced the panic early

Here's the awkward one. CoreWeave, down 14% on the Meta report, finished this week up 12% — on a Gartner “Visionary” badge and an ARK dip-buy, not on any new contract or any denial. Nebius round-tripped to flat. Price action has detached from evidence — specifically in the levered corner of the trade. Worth remembering when the real data lands.

5. The bill arrived — and the Fed is split over what it costs

Consensus analyst estimates for 2026 hyperscaler capex crossed $750bn this week, per Bloomberg's tally — with Oracle's capex near 86% of estimated sales (the ratio S&P cited in Thursday's downgrade to BBB−) and Morgan Stanley modelling Amazon at roughly $17bn of negative free cash flow. That last figure is a bank's estimate, not company guidance. The same afternoon, the June Fed minutes showed participants split 9–8 on whether 2026's next move should be a hike — a count of stated preferences in the minutes, not a formal policy vote. The credit channel — the one the non-self-funding end of the build-out leans on — just got more expensive to be wrong in.

Our Investing Angle

Everyone watched missiles and the neocloud bounce this week. The smarter bet was watching the funding pipes.

The thesis — stated as a hypothesis we can test: the binding constraint on AI is shifting from chips to the cost of money. This week put all three external funding channels on public display at once. Equity: headline 1 — $29bn, seven times oversubscribed. Customer pre-payment: headline 2 — $30bn, in writing. Credit: headline 5 — a downgrade and a hawkish set of minutes in the same afternoon. One channel is euphoric, one is gold, one is creaking.

If the hypothesis is right, the next stretch of AI returns gets decided by balance sheets more than benchmarks. The self-funders and the pre-paid can keep building through a rate scare; the borrowers — Oracle and CoreWeave are the clearest public examples — don't get that luxury. (Rate-sensitive software multiples are a related but separate story: different mechanism, slower fuse. The framework handles it below.)

What we're not doing here is pretending the sorting is easy. Most AI names don't sit cleanly in one channel. Amazon spends enormous cash flow and has flagged it may raise debt and equity. CoreWeave's debt is secured on contracts from customers who are building exits. SK Hynix is raising equity and collecting pre-payments. Which channel actually owns each name — the balance-sheet numbers, the hybrid cases, what each channel costs, and what happened the last three times a funding window slammed shut — is this week's deep dive: Who pays for the machines?

And how we'll know if the thesis is wrong: Tuesday's CPI, the $149 line, and the 29 July call. All three sit in the framework's watch-list below, with the specific readings that would weaken it — or reinstate it.

Three Ideas to Research This Weekend

Not recommendations — starting points for your own research. One lead idea, two quick follows.

Lead idea — SKHY: the supply wave, offered to you at $149

Why now: it starts trading on the Nasdaq today — $29bn of fresh AI-memory paper, reportedly seven times oversubscribed, priced (per Barron's) cheaper than Micron while sitting closer to Nvidia.

The case: roughly 60% HBM share, Nvidia and Google as anchor customers, and a valuation gap to its US peer that the ADR structure exists specifically to close.

The risk: you'd be buying the memory cycle's hottest name in the very week it raises $29bn to build the capacity that ends memory cycles. The crowd inside is the size of the crowd that can leave.

Tripwire: Friday 17 July's close vs the $149 issue price. Held above = US appetite for AI paper intact. Below issue inside a week = the marginal AI dollar is tapped out, and every raise queued behind it gets harder. Second fuse, carried from Issue 17: July DRAM contract prices, early August.

How to research: SKHY from today, or the memory-heavy end of SMH. Read the ADR prospectus section on use of proceeds — it's the supply wave's construction schedule, itemised.

Quick follow — Oracle: watch the bonds, not the stock

I keep coming back to one detail from Thursday: S&P cut Oracle to the last rung of investment grade and the stock rose 3%. Equity investors shrugged at the exact thing this issue argues matters most — and finding out who's wrong costs you nothing. Oracle is the test case for whether the debt channel stays open, because it's the biggest AI builder that can't self-fund. Tripwire: Oracle's next benchmark bond sale (it borrows most quarters). Priced smoothly at a modest concession = the credit window is open and the bear case is deferred. Pulled, shrunk, priced wide — or an S&P cut to BB+ — means the window is closing and every levered AI name reprices with it. How to research: ORCL plus its listed bond yields (FINRA's site shows them free). The 10-year spread over Treasuries is the single number to track.

Quick follow — Hyundai (filed under: nothing to do with this week's thesis)

Too interesting to sit on, so it gets a paragraph. Hyundai paid SoftBank $325m this week for the last 9.65% of Boston Dynamics — full ownership of the Atlas humanoid at an implied value near $3.4bn, a fraction of what private humanoid startups are fetching, with its own factories (including the Georgia Metaplant) to deploy into. The risk: robotics is a rounding error in Hyundai's revenue, and Korean conglomerate discounts die hard. Tripwire: Atlas on a Hyundai production line with published unit numbers by the Q3 call in late October = the re-rating case opens. Another demo video = trophy asset; move on. How to research: Hyundai Motor (005380.KS); BOTZ for the diversified version.

AI Investment Framework

6 layers · Updated weekly · Not financial advice

What to notice first: Infrastructure's V — last week's cliff, retraced and then some in five sessions. The market has already voted that the buyers' exits were noise. 29 July decides whether it keeps the money.

WHAT WE'RE WATCHING

CONVICTION CHECK — INFRASTRUCTURE (HIGH, UNCHANGED)

The tempting move was ↑ after an 11% Broadcom week and a freshly funded memory build-out. We held. The layer's open question — was the memory drop positioning or fundamentals? — gets answered by contract prices in early August, not by a bounce. Conviction that moves on price action isn't conviction; it's momentum wearing a lanyard. Everything else holds.

Quick Glossary

Capex (capital expenditure) — Money spent building long-lived assets: data centres, fabs, machines. Analysts expect the big cloud builders to spend ~$750bn of it in 2026.

Free cash flow — The cash a business generates after paying for its capex. Negative free cash flow means the build-out is funded from outside — equity or debt.

HBM (high-bandwidth memory) — The stacked memory chips bolted to AI processors. SK Hynix makes roughly 60% of the world's supply, on industry estimates.

Oversubscribed — When investors ask for more shares than are on offer. SK Hynix's book was reportedly about 7x covered: $7 chasing every $1 available.

Your Move

One task this weekend — just one. Label every AI name you own by the channel that funds its build-out: own cash flow, customer pre-payment, fresh equity, or debt. If you can't confidently label a holding, that is the finding — the deep dive has a five-question stress test for exactly that case. While you're at it, put two dates in the calendar: 14 July prices the money, 29 July prices the plan. One label per holding, two dates. That's the whole assignment.

Stay curious — and stay qualified.

— R. Lauritsen

Editor, iPrompt Signals

Know someone building an AI position? Forward this — they'll thank you by Friday.

P.S. Last week's lead idea said the neoclouds were two different balance sheets wearing one label. The market agreed violently — CoreWeave +12%, Nebius flat — before any evidence arrived. If you did the homework, the bounce told you something. If you didn't, it just felt like relief. Bounces aren't answers.