The Donor Economy: who's actually paying for the AI build-out

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

Deep Dive · Issue 19 · 17 July 2026 · R. Lauritsen

The Donor Economy: who's actually paying for the AI build-out

At 9:47 on Tuesday morning, a 114-year-old company worth a third of a trillion dollars began losing a quarter of its value — not to a fraud, not to a product failure, but to a paragraph. IBM's pre-announcement said clients had “dramatically reprioritised” their spending in the final weeks of June: money out of software licences and IT projects, money into AI servers, storage and memory. By the close it was IBM's worst day since its records began in 1968, and roughly $67bn of market value was gone.

Nobody stopped buying technology. They stopped buying it from IBM.

This week's Signals floated a hypothesis: the AI trade is becoming a reallocation — capital moving around inside the tech economy rather than into it. A hypothesis, not yet a law; it currently rests on one dramatic Tuesday. This piece is the working: how substitution would actually have to flow, who sits where on the donor map, what history says about budget migrations, and how to test your own holdings before the market does.

The arithmetic — and what it does and doesn't prove

Consensus estimates put 2026 hyperscaler capex above $750bn — Bloomberg's tally of analyst models, the figure that anchored last week's piece. Start with what that number does not mean: most of it is funded exactly the way last week's funding map described — internal cash flow, equity, debt, customer pre-payments. Hyperscaler capex does not need to raid anyone's IT budget to exist, and nothing in this piece claims it does.

The donor channel is a different, smaller pipe: the AI spending of ordinary enterprises — the servers, storage, memory and inference capacity that CIOs buy directly, plus the cloud AI services they rent. That spending lives inside corporate IT budgets that grow single digits in a good year. When a budget growing at that pace has to absorb a category growing several times faster, only three things can happen: the budget expands (the productivity-dividend case), the AI spend gets deferred, or something else in the budget gets cut. IBM's Tuesday is direct evidence for the third: $17.2bn of revenue against $17.9bn expected, the shortfall concentrated in exactly the categories a CIO raids first — conventional software, infrastructure refresh, consulting hours. And the detail worth keeping: Krishna said the shift happened in weeks, not quarters. Budget substitution moves at the speed of a purchase order.

What would separate substitution from simple addition, if you were watching for it: donors missing while collectors beat (dispersion, not direction); consulting bookings turning down before licence revenue does; and CIO surveys showing flat totals with reshuffled line items. One of those three went on the record this week. The other two get tested this earnings season.

The same week supplied the collector's side of the ledger. TSMC's record quarter — $40.2bn of revenue, roughly two-thirds of it now high-performance computing — is what the collected dollar looks like when it lands. And TSMC's own capex raise ($60–64bn, from $52–56bn) shows the chain doesn't stop there: even collectors reinvest the dollar rather than keep it. Nobody in this economy banks the money. It flows uphill to whoever owns the current chokepoint.

The donor map

Three columns, classified — like last week's funding map — by the marginal dollar: whichever column a company's next customer dollar moves through, not its whole identity. Two rules before you use it. First, columns leak: Dell and HPE sell the very servers the reprioritised money buys while their legacy lines donate — one company, two columns, and the mix decides the stock. Second, the map describes dominant exposure at company level, today; segments within a company can and do sit in different columns.

The donor column is not a short list — by count, it's most of the enterprise-tech index. Which is precisely why substitution matters at index level: a cap-weighted tech fund owns both sides of the transfer and nets out the trade you thought you had.

What history says about budget migrations

This has happened twice at scale, IBM was present both times, and both episodes carry the same two caveats: the migration took a decade, and donors were survivable — sometimes even decent investments — at the right price. The drama is in the repricing, not the extinction.

Mainframe to client-server, 1991–93. Enterprises didn't cut IT spending; they moved it to PCs, servers and packaged software. IBM — the chokepoint of the old architecture — posted what was then the largest loss in American corporate history ($8.1bn, 1993) while total industry spend kept growing. A budget migration can be lethal to the donor while looking like a boom in aggregate.

On-prem to cloud, 2015–19. Gentler, same physics. Legacy licence and hardware vendors — IBM again, Oracle's licence base, HPE — spent half a decade flat-to-down in the market while AWS and Azure compounded inside the same customer budgets. Nobody's revenue collapsed; their multiples did. Donor status is a valuation event years before it's a revenue event — the market reprices the terminal value first and argues about the quarters later. That's what makes this week's question urgent rather than academic: IBM was priced as a compounder at Monday's close. It isn't quite, at Friday's.

The counter-argument worth respecting

The bull case for the donors says this is a timing gap, not a death sentence: AI eventually expands what software can do, IT budgets grow as a share of company revenue, and today's raided project budgets come back bigger. The cloud precedent half-supports it — total IT spend did accelerate, eventually. But note what the argument concedes: between the raid and the dividend sits a multi-year window in which donors report soft quarters into a market primed to IBM them. You can believe the dividend arrives and still not want to hold donors through the window. Different bets, different clocks.

The five-question donor test

Scope first: this is a heuristic for the non-AI enterprise-technology names in your portfolio — companies that might be donors without having confessed yet. Applied to a collector it returns a trivial five out of five (Nvidia passes by definition), so its job isn't to rank AI winners; it's to catch quiet donor risk in the rest of your tech sleeve. Five questions, one point each, answerable from an annual report and one earnings call. Higher is safer.

1. Is less than 40% of revenue maintenance, renewal or support on pre-AI products? (Yes = 1. Renewal revenue is what CIOs raid first — it's the money already in the building.)

2. Does the company report AI revenue as a number rather than an adjective? (Yes = 1. “AI-powered” in the deck is worth zero; a disclosed figure with a growth rate is worth one.)

3. Does its product hold budget that AI infrastructure can't reach — regulatory, safety-critical, or physically embedded spend? (Yes = 1.)

4. Has it raised prices in the past year without losing seats? (Yes = 1. Pricing power is the cleanest evidence customers can't substitute you — yet.)

5. Were the last two earnings calls free of “elongated sales cycles”, “deal slippage” and “reprioritisation”? (Yes = 1. Management vocabulary leads the numbers by about two quarters.)

Scoring: 4–5, the raid mostly can't touch it. 2–3, watch the language on every call — on public disclosures, this is roughly where IBM sat in April. 0–1, the market will eventually price it as a donor whether or not the revenue has confessed; the only question is which earnings date does it. The interesting scores are the 2s and 3s sitting in portfolios wearing “digital transformation” badges.

What to watch, in order

22 July — IBM's full print. The pre-announcement gave totals; the print gives mix. Software down with hardware-adjacent lines holding is the substitution signature, confirmed from inside the donor.

Late July — Accenture and the consulting tape. Consulting bookings are the earliest donor indicator anywhere in the economy: projects get cancelled before licences lapse. A second major services firm citing reprioritisation upgrades the hypothesis from anecdote to sector.

29 July — Meta, and the conduit question. The standing tripwire from Issues 17 and 18: does Model API revenue get a number? A conduit turning collector is the bull case for the entire middle of the map.

August — the software multiple, in aggregate. Watch IGV against SMH. If the ratio makes new lows through earnings season, the market is pricing the donor economy faster than the analysts are writing about it. It usually does.

The bottom line

Last week's piece traced the four channels that fund the build-out: cash flow, customers, equity, debt. This week the market showed a possible fifth — taking the money from whoever can't defend it. If the hypothesis survives its verdict dates, the build-out's quietest funding source is the donor economy: the licence renewals, refresh cycles and consulting hours that stopped being safe the week CIOs discovered they could trade them for GPUs.

The headline said IBM missed a quarter. The tape said something bigger: the same dollar can only be spent once, and the market has started asking, name by name, whose dollar it was. Run the five questions before it gets to yours.

Sources, by claim: IBM's Q2 pre-announced figures, the record single-day decline and the ≈$67bn market-value change — company statement and press coverage, 14 July. TSMC revenue, margin and capex guidance — company Q2 release and earnings call, 16 July; “largest quarter in its history” is measured against its own reported results. $750bn 2026 hyperscaler capex — Bloomberg tally of analyst consensus estimates, not company guidance. June CPI — BLS release, 14 July. Anthropic's $47bn is an annualised revenue run-rate reported in press coverage of its IPO preparations, not audited revenue. IBM's 1993 loss — company FY1993 results. Cloud-era performance comparisons — public market data, 2015–19. Where a figure is an estimate, the text says so.