SIGNAL / NOISE

LTV-positive. Cash-flow negative. Forever.

A year ago Microsoft (NASDAQ: MSFT), Alphabet (NASDAQ: GOOGL) and Meta (NASDAQ: META) threw off $60.7 billion of free cash flow in a single quarter. This summer: $7.4 billion. They didn't sell less. Revenue grew 19%, 25% and 26%. The bill grew faster: $125.6 billion of chips and buildings in three months. Alphabet alone burned $9.1 billion, its second negative quarter in 22 years as a public company, and on Thursday it lost $500 billion of market value, its worst day since the 2004 IPO.

The worse number is underneath. Search queries hit a five-year high. Search revenue grew 12%, down from 19% in the first quarter, because, per Philipp Schindler, AI Mode earns less per query. For twenty years usage and revenue moved together at Google. This is the first quarter they split. Google did what Kodak refused to do, cannibalized its own franchise on purpose, and got punished anyway, because it did it with no cash cushion.

I've sat through this pitch a hundred times. We're LTV-positive. The second we stop spending on growth, the cash gushes. Anthropic's S-1 makes the same case with better numbers: $1.70 of revenue for every dollar of compute. The page it skips is the other one. At least $114.5 billion committed to Google Cloud through 2033 and $110 billion to Amazon (NASDAQ: AMZN) through 2036, take-or-pay, against a product whose price falls 13x a year by Epoch's count. Sridhar said it to a million people in one line: buy compute for years, sell intelligence by the token. The S&Ls borrowed short and lent long and it killed them. This is the mirror image, and the growth has to run forever to cover it. Netflix made it through that cycle. MoviePass didn't. Management never decides when the race ends. The market does, by closing the window.

So where does the money go? Not out. Nvidia (NASDAQ: NVDA) and Broadcom (NASDAQ: AVGO) take the rake on every hand, and the startups paying the labs are spending LP money: Chamath says 1% of companies buy 90% of the tokens, and 80% of that is unprofitable venture-backed. Microsoft booked $10 billion from OpenAI last quarter, 11% of revenue. OpenAI is now handing its own subscribers $500 a month in API credits. The chips stay on the table.

The only money that can actually leave the table belongs to the consumer, and look where it's coming from. Microsoft offered buyouts to 60,000 people, 27% of the company, to protect the capex line. Nadella said it two years ago: people costs down, GPU per researcher up. It isn't only the builders. I've watched it on companies I follow closely that own no GPUs at all: headcount down, cash flow flat, and the savings pushed into reaching past the next layer of the chain for margin. AI is collapsing the margin that used to live at every layer (maker, distributor, retailer, broker) and everybody is moving up or down the chain to grab what's left. Every layer eliminated is a payroll eliminated. A wage is a cost to one company and revenue to a thousand others. SightBringer asked what nobody asks on an earnings call: what gives a person a claim on the output when the output no longer needs them?

The tractor did this to the farmhand, and the farmhand's kids adapted. It took seventy years and the invention of the American high school. And the tractor didn't come with a take-or-pay.

Count the chips that leave the table. Those are the only real ones.

At COAI today: the full Signal/Noise, with the S&L math, the Google query-revenue split, and the margin-stack map, is live at getcoai.com.

Which layer of your margin AI collapses first, whose payroll your revenue actually is, and how much of your cash flow is real once the spending stops.

ONE — A NUMBER THAT SUMMARIZES THE DAY

$60.7 billion to $7.4 billion. That's the combined quarterly free cash flow of Microsoft, Alphabet and Meta, one year apart. Revenue grew at all three. The chips ate the difference, and the chips are ten-year commitments against a product that gets 13x cheaper every year. Every player can show you LTV-positive math. None of them can stop while the others build. Nvidia takes the rake either way. The only money that leaves the table is the consumer's, and the consumer's paycheck is what's being cut to fund the next hand.

THREE — ACTIONS TO TAKE TODAY

For the investor — underwrite the window, not the unit economics. Anthropic's $1.70 per compute dollar is real. So is the $224 billion floor it owes Google and Amazon. The question that matters isn't whether the math works, it's how long outside money keeps the race funded. Gartner says $3.5 trillion of infrastructure is exposed to impairment by 2030. Until free cash flow turns, own the rake (chips, power, land) or own the window's timing. Don't own the pot.

For the business owner — find your layer today. If your margin exists because you sit between two parties, AI is coming for it, and everyone around you is already moving up or down the chain to take yours. Decide which direction you move before someone decides for you. Then do the uncomfortable exercise: list your ten biggest customers and ask whose payroll pays them. If the answer is "a company cutting heads to fund AI," your revenue is on the table too.

For the parent — your kid doesn't get seventy years. The farmhand's kids had three generations and a new school system to find the next job. Thariq spotted the paradox this week: technical workers fear agents will devalue them, while non-technical workers want more technical skill. Both are right. Teach your kid to be the one who runs the agents and judges the output, because the layers in between are the ones getting eliminated.

FIVE — STORIES TO KEEP YOU INFORMED

Friday, October 30

  • Google's queries hit a five-year high. Its stock had its worst day since 2004. Search grew 12% on record usage because AI Mode monetizes less per query. Cloud doubled to $30.7 billion, capex hit $51.4 billion, free cash flow went negative for a second straight quarter. The fix, per Schindler: more ads inside AI Mode. (Full analysis above.)

  • Microsoft offers buyouts to 60,000. That's not an overhiring correction. Take the full 27% and headcount lands below pre-pandemic levels. Lemkin on 20VC: "to pay for the GPUs we gotta get rid of the people." The depreciation on this year's chips hits the P&L in 2027, so this is the first round, not the last. (Full analysis above.)

  • ServiceNow grew AI contract value 65% and lost 24% in a day. Worst session in the company's history, on revenue that matched estimates. Wall Street has decided SaaS margins were a twenty-year accident of zero marginal cost. Enterprise software now gets no credit for AI revenue and full punishment for AI cost.

  • Aon's CEO says data centers may be too concentrated to insure, and uninsurable means unfinanceable. Business interruption "measured in millions of dollars per minute." Greg Case's line: "You can't access the debt markets." The build-out's binding constraint may not be chips or power. It may be the actuary.

  • New York signed the first law governing what a chatbot can say to a child. No pretending to be human, no claiming a relationship, no asking unprompted emotional questions, $25,000 per violation, and the attorney general can demand destruction of models trained on unlawfully collected data. Limits apply to every user until the company proves they're an adult. Other states will copy the text.

MARK TO MARKET

Where the cycle caught up to us this week.

  • Inference turns software margins into a cost-of-goods business. AI is "a metered fire" that converts SaaS's 72% gross margin into a 30 to 50% COGS business (us, Oct 13) → "For twenty years these were the best businesses ever invented. You wrote the code once and almost every new dollar fell to the bottom line. AI doesn't work like that. Every answer runs on a chip someone had to buy" (Martin Varsavsky on X, Oct 29).

The tape doesn't lie. We just read it early.

— Harry and Anthony

Sources:

  • Martin Varsavsky (@martinvars), X, Oct 29, 2026 — big-3 FCF $60.7B → $7.4B; capex $125.6B/quarter

  • CNBC, "Alphabet earnings miss reflects first real sign that AI is taking a bite out of search," Oct 29, 2026

  • Ars Technica, "Google admits AI search makes less money, but it's fixing that with more ads," Oct 29, 2026

  • The Independent, "Google stock price crashes and wipes $500 billion off company," Oct 29, 2026

  • Implicator.ai Morning Briefing, Oct 29, 2026 — Cloud $30.7B, capex $51.4B, FCF −$9.1B; Microsoft $10B from OpenAI

  • The Verge, "Microsoft's AI future has fewer jobs," Oct 29, 2026 — 60,000 buyouts; Nadella "GPU per researcher" quote

  • 20VC, "Anthropic Files S1 Ahead of $2TRN IPO," Oct 29, 2026 — Lemkin on buyouts and depreciation

  • Prof G Markets, "Microsoft Rises, Google Tanks," Oct 29, 2026 — headcount below pre-pandemic; "what's the difference?"

  • Tom's Hardware, "Anthropic's IPO filing shows most of its compute goes to R&D," Oct 29, 2026 — $1.70 per compute dollar; compute split

  • The Motley Fool, "Alphabet, Amazon, SpaceX, Microsoft, Broadcom, and AMD Are Counting on Anthropic's IPO," Oct 29, 2026 — $114.5B and $110B minimum commitments

  • Sridhar Ramaswamy (@RamaswmySridhar), X, Oct 27, 2026 — "buy compute for years, sell intelligence by the token"

  • Epoch AI via The AI Corner / State of AI Report 2026, Oct 29, 2026 — 13x annual cost decline

  • 24/7 Wall St., "Chamath Says Just 1% of Companies Drive 90% of AI Spending," Oct 29, 2026 (All-In, Oct 23)

  • OpenAI Devs (@OpenAIDevs), X, Oct 29, 2026 — $500/month API credits for Pro and Business subscribers

  • SightBringer (@_The_Prophet__), X, Oct 28, 2026 — wages as cost and purchasing power

  • Thariq (@trq212), X, Oct 28, 2026 — the skills paradox

  • Gartner, "$3.5T in AI infrastructure spending could be at risk by 2030," Oct 2026

  • CNBC, "ServiceNow shares plunge 24% after earnings revive AI disruption fears," Oct 29, 2026

  • CNBC, "Aon CEO warns AI data center insurance risks could threaten financing," Oct 28, 2026

  • The Neuron / Governor Hochul's office, New York chatbot-and-minors law (S.9051B/A.10379C), signed Oct 28, 2026

Reply

Avatar

or to participate