In partnership with
Why Most AI Tools Fail at B2B Customer Service
B2B customer issues rarely resolve in a single interaction. They move across teams, systems, and account history that most AI tools were never designed to track.
A new briefing paper from Harvard Business Review Analytic Services, sponsored by Front, looks at why AI built for high-volume consumer support creates new coordination gaps when it meets the multi-team reality of B2B service, and what leaders should ask before investing in their next AI tool.
Get the free briefing paper to see the three questions that separate tools built for your environment from tools that will create more work.
D E E P D I V E
ISSUE 34 COMPANION · FRIDAY 23 OCTOBER 2026
The Four Wallets
Own cash, customers, lenders or shareholders: who pays for each company's AI build, what each route really costs, and a test you can rerun on your own holdings.
Three announcements this week, three routes to the same money. In Berlin, Nvidia said it is paying suppliers in advance for capacity: $1.1bn to KYEC and a multibillion-dollar sum to Ibiden. In the bond market, CoreWeave's 2031 notes yielded 15.58%. And Meta created a non-voting share class whose stated uses include raising capital, though it has announced no sale.
Behind all three sits a fourth wallet that usually goes first: a company's own cash. Alphabet and Meta both spent more than their operations brought in during their latest quarter, while holding $242.5bn and $82.45bn of cash and securities. Reserves can cover gaps like that for a long time. How long the gaps last is the real question.
This series has asked whose balance sheet stands behind each payment, how long the payment is promised for, and what the money costs. This week: whose wallet, and what each one really charges.
Nvidia · Supply-chain release · 21 Oct | Benzinga · CoreWeave 2031 notes · 20 Oct | Meta · Q3 results · 21 Oct | Alphabet · Q2 results · 22 Jul
FOUR THINGS THAT GET BLURRED
Most confusion about AI funding comes from treating four different measures as one.
Measure | What it tells you | What it doesn't tell you |
Cash shortfall (negative free cash flow) | Spending exceeded cash from operations in a period | Whether reserves cover it, or what will fund the next one |
Financing cost | What new money costs: interest for debt; for equity, the return new owners expect | Whether the project is worth doing |
Ownership dilution | Your percentage after new shares are sold for cash | Whether you end up better or worse off |
Value created or destroyed | Whether the project's cash flows, timed and risk-adjusted, beat the return owners require | The short-run effect on earnings per share |
A company can have the first without the other three, as Alphabet did in Q2, or dilution that still creates value. Keep them apart and the headlines get easier to read.
THE WALLET MAP
Start with the suppliers and work towards the spenders. At each step: whose money funds the next build, and on what terms?
Company and layer | Main wallet for the next build | Terms, where public | This week |
Ibiden; Samsung Electro-Mechanics · substrates | Customer money: an Nvidia advance (Ibiden), co-investment (SEMCO) | Amounts and pricing not disclosed | Ibiden to triple its Ono plant; SEMCO adds Sejong and Busan lines |
Lam Research · chip tools | Its customers' capex, which is its revenue | No new financing needed for the build | Record $8.22bn quarter; supply and clean-room limits into 2027 |
Tesla · robot fleet, compute | Own cash; $30bn of credit lines signed in September | 2026 capex now more than $20bn | Q3 free cash flow $1.53bn |
SpaceX · compute landlord | Profits and about $90bn of cash and securities | Q3 capex $19.6bn against $14.6bn of revenue | Net income $4.4bn |
Alphabet · cloud, models | Own cash ($242.5bn), then bonds | Q2 free cash flow −$5.86bn; buybacks paused | Reports 28 October |
Meta · models, ads | Own cash ($82.45bn), then debt or new Class C equity | Q3 free cash flow −$2.12bn; no buybacks in 2026 | Class C distributed after 19 January 2027; no sale announced |
CoreWeave · GPU landlord | Lenders: debt and leases of about $51.6bn | 2031 notes yield 15.58% in the market | −10.1% to its lowest close of 2026 |
Free cash flow and capex use each company's own definitions; Meta's include finance-lease principal. 'No new financing needed' means the build needs none, not that the company carries no debt.
Nvidia · Supply-chain release · 21 Oct | Samsung Electro-Mechanics and Nvidia · Release · 21 Oct | 24/7 Wall St. · Lam results · 21 Oct | Tesla · Q3 10-Q · 21 Oct | Electrek · Tesla credit lines · 22 Oct | investingLive · SpaceX Q3 · 20 Oct | Crypto Briefing · SpaceX cash · 20 Oct | Alphabet · Q2 results · 22 Jul | Meta · Q3 results · 21 Oct | Meta · Class C 8-K · Oct | CNBC · CoreWeave debt total · 22 Oct | Benzinga · CoreWeave 2031 notes · 20 Oct
My reading: the higher a company sits in the chain, the more often someone else pays for its growth. Suppliers get paid out of their customers' capex, and Nvidia, the richest customer of all, now pays some of them before they build. The spenders at the bottom are still drawing on deep wallets of their own. What changes, if the gaps persist, is the mix behind them.
WHAT META'S NEW SHARES DO, AND DON'T
In January every Meta holder receives one Class C share per existing share. That is a stock dividend, economically a two-for-one split: twice the shares, the same ownership, no new money raised. Meta lists capital raising among Class C's uses, so later sales for cash would dilute existing holders. None is announced. Yahoo Finance reports that the filing expects Class C to trade at some discount to Class A, which would make any future sale a little dearer for existing owners. Meanwhile the share count roughly doubles in early 2027, so adjust for that before measuring dilution. Ibiden's two-for-one split on 1 October needs the same treatment.
Meta · Q3 results · 21 Oct | Meta · Class C 8-K · Oct | Yahoo Finance · Class C discount, advisers · 22 Oct | FilingReader · Ibiden split · 4 Aug
THE DILUTION TEST
When a company sells new shares for cash, two hurdles matter, and they are not the same.
The accretion hurdle: does earnings per share rise? It does if the project's after-tax accounting return beats the earnings yield on the shares sold, which is earnings per share divided by the sale price. When shares are highly valued, that bar is low. At 25 times earnings, the earnings yield is 4%.
The value hurdle: does the project beat the return owners require for the risk? Nobody publishes that number; we assume 9% to 10%. Comparing an accounting return with it is only a screen. A proper test needs the cash flows: when they arrive, what must be reinvested, how long the assets last and what they're worth at the end.
Between the two hurdles sits the trap: a project can lift earnings per share and still fall short of what owners require. Expect management to talk about accretion. Check the cash flows.
For scale, Meta closed on Thursday at $648.85, about 24 times trailing adjusted earnings of $26.55 as compiled by MarketBeat. That's an earnings yield near 4.1%. An earnings yield is not a cost of equity. It sets the accretion hurdle, nothing more.
WORKED EXAMPLE ONE: SAME $100BN, DIFFERENT RETURNS
All figures hypothetical. A company earns $100bn a year on 2.5bn shares, or $40 a share. The shares trade at $1,000, a P/E of 25. It sells $100bn of new stock at $950, a 5% discount, adding about 105.3m shares, 4.2% more. The money goes into data centres. Here's the steady-state result at different returns.
After-tax return on the new $100bn | Earnings | EPS | Change in EPS | Clears an assumed 9% required return? |
2% | $102bn | $39.15 | −2.1% | No |
4% | $104bn | $39.92 | −0.2% | No |
6% | $106bn | $40.69 | +1.7% | No |
10% | $110bn | $42.22 | +5.6% | On this screen, yes |
15% | $115bn | $44.14 | +10.4% | On this screen, yes |
Per year, once the project is fully running. EPS breakeven is about 4.21%. Ignores fees and ramp-up.
The shaded wedge: earnings per share rise, yet the assumed required return is missed.
Three caveats. Data centres rarely earn their full return in year one, so even a good project usually dilutes early and pays back later. The hurdle moves with the share price: if the stock falls 20% before the sale, the company needs more shares and the EPS breakeven rises to about 5.3%. And 'yes' in the last column is a screen, not a verdict: front-loaded cash flows clear it more easily than the same average return arriving late.
WORKED EXAMPLE TWO: ONE $10BN BUILD, FOUR WAYS TO PAY
Hypothetical again. $10bn of new capacity. The table separates three kinds of cost that are easy to blur: cash a company must pay, the effect on ownership, and obligations that show up in contracts rather than in the cash flow statement.
Wallet | Contractual cash payments | Ownership effect | Commercial and other obligations |
Own cash | None. Interest income forgone: about $0.4bn a year at roughly 4% short-term rates | None | A thinner cushion for the next gap |
Customer advance or co-investment | None in cash; an advance is repaid in delivered capacity | None | Reserved capacity, price commitments or exclusivity; undisclosed for Ibiden and SEMCO |
New debt | Interest at the rate set at issue: $0.6bn a year at 6%, $1.5bn at 15% (illustrative rates); principal at maturity | None | Covenants, refinancing risk and higher leverage, all borne by shareholders |
New shares | None required; dividends are discretionary | For Meta, $10bn sold 5% below Thursday's close is about 16.2m shares, roughly 0.6% of its diluted count | Owners gain only if the project beats the return they require |
Rates are illustrative. A traded bond's yield, such as CoreWeave's 15.58%, is a guide to what new borrowing might cost, not a quote and not the coupon the company pays today.
Nvidia · Supply-chain release · 21 Oct | Benzinga · CoreWeave 2031 notes · 20 Oct | Yahoo Finance · Meta close · 22 Oct | Meta · Q3 results · 21 Oct
No single number ranks these. Debt has the most visible cost and the most rigid one. Shares have no required payment and a permanent claim on future profits. Customer money can look free and carry its cost in price or flexibility. Own cash is cheapest until the cushion is needed for something else. So 'which wallet is cheapest?' depends on the terms, and the terms are usually the least disclosed part.
IF THE GIANTS KEEP SPENDING: THREE PATHS
Party | Cash from operations catches up | Gap funded mainly with debt | Gap funded partly with new shares |
Meta and Alphabet holders | Buybacks can restart; per-share growth holds | Same percentage; more interest, refinancing risk and leverage | Smaller percentage; outcome depends on returns |
Equipment and substrate suppliers | Orders as before | Orders as before, while lenders stay open | Orders as before, while buyers keep raising |
Neoclouds | Demand intact | Compete with giants for the same lenders | Compete with giant share sales for investors |
Model labs (Anthropic) | IPO window open | IPO competes with giant bond deals | IPO competes with giant share sales |
A conditional scenario, not a forecast. The path will show in the cash flow statement before it shows in the share price.
WHAT TO WATCH
When | Checkpoint | What it would tell us |
23 Oct (reported, unconfirmed) | Anthropic public S-1 | How much of a model lab's compute bill is fixed, and which giants it's owed to |
28 Oct | Alphabet Q3 | A second cash-negative quarter with a higher 2027 plan means reserves and debt carry more of the build |
28 Oct | SK Hynix Q3 call; Fed decision | Lower 2027 memory investment would be the first test of the suppliers' queue |
End Oct (unconfirmed) | Ibiden interim; SEMCO Q3 (29 Oct, tentative) | How the Nvidia money is booked, and whether its terms are disclosed |
After 19 Jan 2027 | Meta Class C distributed | A first sale of Class C for cash would be the dilution event; the distribution is not |
TURN THE ANALYSIS INTO A DECISION
The bull case for the spenders: revenue is growing faster than anyone modelled, Meta's Q4 guide reaches $76bn, and if operating cash catches capex within a few quarters no outside wallet is needed. Nothing rules that out yet.
The case for caution strengthens if Alphabet reports a second cash-negative quarter on 28 October with a higher 2027 plan, if Meta sells Class C for cash before its Q4 report, or if the giants' bond spreads widen. It weakens if operating cash overtakes capex, or if gaps are funded from reserves while they stay small.
Four weeks, four questions. Whose balance sheet stands behind the payment. For how long. At what price. And now: whose wallet.
YOUR WEEKEND REVIEW: THE FOUR-WALLET CHECK One holding, about 20 minutes. 1 Own cash. Free cash flow for each of the last four quarters, and cash and securities today. A negative quarter means the gap came from reserves, borrowing or new shares. Find out which. 2 Lenders. Debt raised in the last four quarters (the financing section of the cash flow statement), and the market yield on its longest bond as a rough guide to new borrowing costs. 3 Owners. Diluted share count for the last eight quarters, adjusted for splits and stock dividends, next to buybacks and stock-based pay. A count that still rises after adjustment means owners are funding part of the build. 4 Customers and tools. Search the latest filings for 'co-invest', 'prepayment', 'share class', 'at-the-market', 'convertible' and 'shelf'. The tool usually arrives before it's used. 5 Screen, then check. Divide trailing EPS by today's price: that's the accretion hurdle for a share-funded project. Then ask what the company's projects plausibly return in cash, and when. Verdict: if the adjusted share count is rising and free cash flow has been negative for several quarters, treat dilution as a live cost. Whether it's a bad one depends on returns, so that's where your research goes next. |
This companion supports Issue 34 of iPrompt Signals, the Friday briefing on AI and robotics investing. Forwarded this? Subscribe at [SUBSCRIBE LINK] to see Wednesday's Alphabet result scored in next week's Tripwire Tracker.
Information and education only, not financial advice. Worked examples are hypothetical and rates illustrative; the 9% to 10% required return is an assumption. The expected Class C discount comes from a press account of Meta's filings. iPrompt Signals is not a registered investment adviser.
iPrompt Signals Deep Dive · Issue 34 · 23 October 2026 ·