IBM just told you who pays for AI

iPrompt Signals

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

Issue 19 · Friday 17 July 2026 · R. Lauritsen

On Tuesday morning, inflation posted its biggest monthly drop in six years — the cheaper money the levered end of AI has been begging for all year. The same morning, IBM had its worst day since its records began in 1968. On Thursday, TSMC reported the biggest quarter in its history — and chips fell. Everything the AI trade asked for arrived this week. It sold anyway. That's the trade splitting in two — and the next three weeks of earnings decide how wide the split gets.

Weekly Scoreboard

Bottom line: cheaper money, proven demand, falling tape. When good news stops working, the question isn't the news — it's who's paying for the party. This week produced a candidate answer.

Prices are Thursday 16 July closes, rounded; weekly moves vs Thursday 9 July. SKHY shown against its $149 issue price. Sources: exchange data via S&P Global Market Intelligence.

Our Investing Angle

Everyone spent the week arguing about whether the AI trade got good news or bad news. Wrong question. The better one: why did the best news week since April produce a falling tape?

The hypothesis — and we're calling it a hypothesis, because right now it rests on one dramatic Tuesday: the AI trade is becoming a reallocation. Capital isn't just flowing into tech; it's moving around inside it. This week supplied the three exhibits. Money got cheaper (story 3 below). Demand got proven (story 2). And IBM offered the first hard evidence of a funding source nobody had put a number on: clients pulling money out of traditional software and services to buy AI hardware (story 1). If that's a sector pattern — and 22 July starts answering that — then the marginal AI dollar is someone else's renewal, licence or consulting contract, the sector average stops mattering, and dispersion becomes the trade.

If it holds, the collectors are the choke points: TSMC, the HBM memory oligopoly, Nvidia. The donors: IBM, repriced this week at a cost of roughly $67bn in market value; the consulting layer — Accenture, Cognizant — where “reprioritisation” lands next; and the maintenance-heavy end of legacy enterprise software. Uncomfortable disclosure: our own Global layer holds SAP. Nobody's exempt, including us.

And if it doesn't hold, we'll know fast: should the August prints arrive and the donors don't confirm — no echo from Accenture's bookings, no SAP guide-down — then IBM was one data point wearing a trend's clothes, and we'll say so in this section.

The full working — how substitution would have to flow, who sits where on the donor map, what the 1993 and 2015–19 precedents say, and a five-question test for your non-AI holdings — is this week's deep dive: The Donor Economy 😀

Top Headlines

Five stories — each one evidence for, or against, the hypothesis above.

1. The AI boom sent IBM the bill

Forget the 25% crash for a second — the reason matters more than the size. IBM pre-announced a Q2 miss on Tuesday ($17.2bn revenue against $17.9bn expected) and CEO Arvind Krishna said the quiet part: clients “dramatically reprioritised” spending in the final weeks of June — out of traditional software and IT projects, into AI servers, storage and memory. The donor hypothesis's founding exhibit — one company's worth of evidence. The 22 July full print decides whether the software cut is broad-based or just IBM losing deals.

2. TSMC proved the demand — and got sold for the cost of serving it

Record revenue ($40.2bn, +34% — its largest quarter ever), 67.7% gross margin, AI packaging sold out more than a year out, full-year AI outlook raised above 40% growth, another $100bn pledged to Arizona. The stock fell 2.3% on the day and 6% on the week — because 2026 capex guidance jumped from $52–56bn to $60–64bn. The market's message to the strongest balance sheet in the trade: we no longer pay for demand, we charge for the capex that serves it. Free cash flow is the new scarce thing.

3. Last week's first tripwire fired — dovishly

June CPI fell 0.4% on the month — the steepest drop since April 2020 — taking annual inflation to 3.5% against 3.8% expected (BLS). September hike odds fell from 75% to 63%; one 25-point hike is still priced for 2026. This was the reading last week's funding-squeeze thesis feared: the debt channel got its reprieve, and Oracle's borrowing window got cheaper. One caveat — energy made this print (−5.7% in June), and energy can un-make it.

4. SKHY did a 30% round trip — and the tourists got the bill

The week-old ADR ran from its $149 issue price to $194.80, then gave nearly all of it back — a 13.7% collapse on Thursday to $152.31. Fundamentals didn't move in either direction: leveraged 2x single-stock ETFs launched within days of the listing, Korean retail piled in on margin, the KOSPI dropped 6% on Thursday, and Seoul's regulators are discussing stability measures. The equity window is open — but hot-money instruments on a week-old listing is how appetite looks right before it gets tested. The autumn raise calendar, Anthropic's October IPO included, queues behind that $149 line.

5. Physical AI found its state sponsors

Two, in one week. Japan launched what Nvidia calls the world's first national AI infrastructure: a 140-megawatt “AI factory” running 27,500 of Nvidia's newest GPUs, with 21 Japanese industrial partners — Fanuc and Yaskawa building robot intelligence on Nvidia's Cosmos models, Toyota extending its physical-AI work. Beijing's industry ministry, meanwhile, wants 10,000 humanoids in factories by year-end, backed by ¥100bn (≈$14.8bn) of funding this year — more than the previous five years combined, per tracker ITjuzi. The robot capital cycle is starting the way the chatbot one did in 2023: with sovereign money arriving before the revenue does.

Three Ideas to Research This Weekend

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

Lead idea — Don't catch IBM. Map the donors instead.

Why now: the worst crash in IBM's history as a listed company just lit up every value screen on Wall Street — and the dip-buyers showed up within 48 hours (+3.7% Thursday).

The case: the productive question isn't whether IBM at $219 is cheap. It's which other names share the disease. Budget substitution is a list, not a stock — and the market has only repriced the first name on it.

The risk: IBM may fix IBM — cost cuts, a services pivot — while you're busy extrapolating its warning to companies that don't deserve it.

Tripwire: the 22 July full print. Software guided down while AI-adjacent hardware holds = substitution confirmed. Everything guided down together = a company losing deals, not a sector confessing. Cross-check within a fortnight: Accenture's consulting bookings — services are where reprioritisation shows up first.

How to research: IBM's 22 July call transcript; then the IGV-versus-SMH ratio — one chart that draws the donor/collector spread daily.

Quick follow — Japan robotics, now with sovereign money

I keep a rule about partnership press releases: they're worth nothing until a purchase order shows up. Japan is about to test it at national scale — Fanuc and Yaskawa building on Nvidia's Cosmos robot models, backed by a state AI factory, with China running the mirror-image programme. Robot makers priced at industrial multiples, physical-AI stories priced at venture multiples; the gap is the idea. The risk: “partner” is not “customer”, and Japanese industrials monetise slowly. Tripwire: Cosmos-related order numbers in Fanuc's or Yaskawa's October half-year results. None = logo exchange; move on. How to research: Fanuc (6954.T), Yaskawa (6506.T); BOTZ holds both for the diversified version.

Quick follow — SKHY, a second look at the issue price

Look — last week's lead idea round-tripped 30% in five sessions, and I won't pretend that was the plan. The honest read: the setup improved. Same 60% HBM share, same Nvidia anchor, a forward earnings multiple near 5 — and the hot money that could leave mostly just did. The risk: memory cycles end suddenly, and CXMT's $8bn Shanghai IPO is aimed at exactly this market. Tripwire: tonight's close against $149. Held = the equity window survives its first stress test. Broken = the autumn raise calendar reprices. Second fuse, carried a third week because it's still the real answer: July DRAM contract prices, early August. How to research: SKHY; the memory end of SMH.

AI Investment Framework

6 layers · Updated weekly · Not financial advice

What changed, layer by layer (the dashboard shows where each layer stands; this table is why):

One thing to notice: the two highlighted lines — Platforms and Infrastructure — crossing in opposite directions on a record-demand week. That's the split, drawn. The average of those lines tells you nothing; the gap between them is the story.

WHAT WE'RE WATCHING

CHANGES THIS WEEK: Physical AI conviction DEVELOPING → MEDIUM (state capital on two continents). Applications risk 3 → 4 (budget cannibalisation moved from theory to a reported number). Everything else holds.

Quick Glossary

CoWoS — TSMC's advanced packaging step, where memory gets bonded to AI processors. Sold out for more than a year — the physical reason AI chip supply can't just be turned up.

Dispersion — The gap between winners and losers inside one sector. It can widen violently while the average stands still — this week's S&P was flat while IBM lost a quarter of its value.

Forward P/E — Price against next year's expected earnings. SKHY's ~5 means the market pays five years of profits — and doesn't believe year six.

Pre-announcement — Warning investors about results before the scheduled date. Companies only do it when the miss is too big to sit on — which is why the market treats it as a confession, not an update.

Your Move

One task this weekend — just one. Run the donor test on the non-AI side of your portfolio. For each legacy tech name you own, ask how much of its revenue is the traditional IT budget that AI hardware is now competing for — the deep dive turns that into five scoreable questions. That's the whole assignment. The calendar above grades it for you: 22 July names the donors, 29 July prices the platforms.

And one word by reply: is IBM's warning a company problem or a sector confession? We'll print the split next Friday, next to the 22 July answer.

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. If you did last weekend's homework — labelling each AI holding by the channel that funds its build-out — this week was the payoff: the equity-channel name round-tripped 30%, the debt-channel name got a cheaper window, and the cash-flow names barely noticed. The labels work.