AEACTS OF EVOLUTION

THE AI ECONOMY PULSE

The Nine Signals That Will Decide the AI Boom

The issue prints . Posted publicly on . Original post

The AI Economy Pulse, by Acts of Evolution · Issue 01 · September 25, 2026

Everyone has an opinion on whether AI is a boom or a bubble. Very few people show their work.

I built Acts of Evolution to show mine. It’s an independent research project: a map of 144 AI companies and 543 sourced relationships across 13 layers of the stack, from power plants and chip fabs to the labs and the apps. It also has nine trackers that test the conditions the build-out depends on. Each tracker is built on a condition someone close to the money has stated publicly (all nine come from Altimeter’s Brad Gerstner at the All-In Summit on September 15). Each one is checked against dated, sourced events.

This newsletter is where I’ll read those signals every week, out loud. This first issue lays out the framework: what I’m watching, why, and where each signal stands today.

Why these nine

The AI boom is a chain. Money gets spent on data centers. That money has to turn into chips, and the chips have to get plugged into power. All of it has to be financed. Then someone has to use what gets built, at a cost that works, and pay enough for it to cover the bill. If that holds, it should eventually show up as productivity across the wider economy. Meanwhile, a large share of the stock market is riding on the outcome.

Every one of the nine signals is a link in that chain. A boom needs all of them to hold. A bust only needs a few to break.

A note on labels. The trackers on the site use their own verdicts (Holding up, Mixed, Under pressure, Not enough evidence yet) and a strict rule: analyst estimates and leaked projections never move them. In the Pulse I’ll give a weekly read of direction: accelerating, mixed, or deteriorating. Where the evidence is too thin to call, I’ll say baseline being set. Seven of the nine signals line up with a tracker on the site. Chip supply and model economics don’t have their own tracker yet, so for now I read those two straight from company filings.

The scorecard

Overall signal this week: Mixed. Spending and demand are still accelerating. The physical and financial limits on the build-out are tightening.

1. Infrastructure spending

What it measures: how much the largest companies are pouring into data centers, and whether contracted demand stands behind it.

Why it matters: this is the top of the chain. Everything downstream, from chip orders to power contracts, is sized to it.

Boom looks like: spending rising, with signed customer commitments behind it. Bust looks like: guidance cuts, cancelled projects, or spending that outruns any paying customer.

Where it stands: accelerating. After July earnings, Amazon, Alphabet, Meta and Microsoft were guiding to more than $735 billion of capex in 2026, up from more than $640 billion in January. Meta narrowed its range to $130–145 billion. On its earnings call, Amazon said it now expects about $220 billion, with “the higher cost of memory” pushing the number up, and “we will still not have enough capacity to meet all the demand we have in 2026”. On the demand side, Oracle reported remaining performance obligations of $664 billion, up $209 billion in a year, alongside negative $5 billion of free cash flow for the quarter. The spending is real. Whether it pays back is a separate question, which is why the site’s tracker on it reads Mixed.

2. Chip supply

What it measures: whether the chips, memory and networking gear can be made fast enough, and at what price.

Why it matters: a data center without chips is an expensive shed. It also works the other way: a sudden glut would be the first sign that orders are running ahead of real use.

Boom looks like: output ramping to meet demand without prices choking buyers. Bust looks like: either a shortage that stalls deployments or a glut with orders cut and inventory piling up.

Where it stands: mixed. Output is enormous. NVIDIA reported quarterly revenue of $96.2 billion, with data-center revenue up 117%, and guided to $108 billion next quarter. Broadcom’s AI semiconductor revenue grew 221% to $16.7 billion. Memory is the pinch point. Micron’s non-GAAP gross margin went from 39.0% a year earlier to 84.9%, and it guided to about 86%. That’s what pricing power looks like when supply is tight. Akamai is pre-buying roughly $1.7 billion of components such as memory to serve its new Anthropic contract. Supply is flowing, but the scarcity is adding to what everyone pays.

3. Power availability

What it measures: whether new AI compute actually gets connected to electricity, permits and a grid, on schedule.

Why it matters: power is the hardest part to speed up. The site’s tracker tests whether about 25 GW of new AI compute actually comes online in 2027.

Boom looks like: power contracts signed, plants approved, and rules that add cost without stopping construction. Bust looks like: permit freezes, delayed fuel supply, and projects waiting on a grid connection.

Where it stands: deteriorating. On September 21, Texas Governor Abbott directed the state’s environmental regulator to halt data-center permits until grid and water audits are complete. The regulator confirmed the pause. Virginia, California and Maryland also acted, making it four states in six days. In New Mexico, the gas pipeline for Project Jupiter, a roughly 2.45 GW Stargate site, slipped nearly six months to February 1, 2027, and Oracle sent a force majeure notice. Oracle says the project remains on schedule. New supply is still being signed, including a 20-year Vistra contract for up to 207 MW, but that power isn’t expected until Q3 2027.

4. Financing

What it measures: what it costs to borrow for the build-out, and whether lenders and investors stay willing.

Why it matters: data centers are built with borrowed money. When long-term rates rise, every project costs more, whatever the demand looks like.

Boom looks like: the 10-year Treasury staying below about 5.5% (the tracker’s threshold), oil easing, and capital still lining up. Bust looks like: rates pushing past that line, and credit stress at flagship projects.

Where it stands: deteriorating. The Fed raised rates on September 16 to 3.75–4.00%. The 10-year Treasury yield was 5.18% on September 24, above the 5% warning level I set in advance on the site but still under 5.5%. Bloomberg reported that Project Jupiter’s $18 billion construction loan trades below 90 cents on the dollar, according to one person. There are offsets. Oil eased over the week, with WTI going from $101.44 to $96.41, and NVIDIA announced partnerships aimed at mobilizing over $500 billion of third-party capital, subject to definitive agreements.

5. Enterprise adoption

What it measures: whether businesses are actually using AI, measured by what they report rather than what vendors claim.

Why it matters: the build-out assumes usage that doesn’t exist yet. Adoption is where that assumption gets tested.

Boom looks like: a broad, steady rise in business use, plus companies disclosing real usage and paid conversion. Bust looks like: a plateau, or pilots that never reach production.

Where it stands: accelerating, from a low base. In the Census Bureau’s business survey, 23.8% of U.S. employer businesses said they used AI in the prior two weeks (late August to early September), up from 17.3% last November on the same question. That’s still fewer than one in four. Meta’s Muse agent had an estimated 2.8 million installs in its first 12 days, but that’s a third-party estimate, and Meta hasn’t disclosed figures. So the site’s agents tracker still reads Not enough evidence yet, and that’s deliberate.

6. Model economics

What it measures: whether building and running AI models costs less than the models earn.

Why it matters: revenue that costs more than it brings in isn’t a business yet. It’s a subsidy.

Boom looks like: unit costs falling faster than prices, and cash burn narrowing. Bust looks like: losses widening as usage grows.

Where it stands: baseline being set. The labs don’t publish audited numbers. The FT reported that an OpenAI presentation projects about $278 billion of cumulative negative free cash flow from 2026 to 2030, against about $856 billion of compute and infrastructure spending. That’s a leaked internal forecast, not a result. The clearest fix would be a public filing. Anthropic has only confidentially submitted a draft S-1, which is why the IPO tracker matters to this signal.

7. Revenue conversion

What it measures: whether AI spending turns into revenue at the companies selling AI.

Why it matters: almost everything else is being financed against the expectation that this number grows very fast. The site’s tracker tests whether the top three labs get from about $100 billion of combined annualized revenue in July to at least $180 billion by year-end.

Boom looks like: disclosed run-rates climbing toward that line. Bust looks like: growth stalling while spending keeps rising.

Where it stands: baseline being set. The New York Times reported that Anthropic is on track to exceed $100 billion in annualized revenue by year-end, up from $65 billion in July. That’s reported by people familiar with its finances, not disclosed by the company. Nothing company-disclosed has landed yet.

8. Productivity

What it measures: whether AI shows up in what ordinary companies report, as margins and output per worker.

Why it matters: this is the difference between AI as a cost and AI as a return. The tracker tests whether AI lifts annual margin expansion from roughly 38 basis points toward 100.

Boom looks like: companies growing revenue without growing headcount, and saying AI is why. Bust looks like: heavy spending with nothing to show in margins.

Where it stands: baseline being set. U.S. nonfarm productivity rose at a 1.4% annual rate in Q2 and 2.2% from a year earlier. Nothing in that release ties the gain to AI. I haven’t yet found a company that attributes margin gains to AI in its reported results.

9. Market concentration

What it measures: how much of the stock market rides on a handful of AI-linked companies, and whether earnings or rising valuations are carrying them.

Why it matters: if the boom stumbles, concentration decides how many people feel it. That includes plenty of people who think they own “the whole market.”

Boom looks like: gains driven by earnings, with leadership broadening. Bust looks like: gains driven by rising valuations, with leadership narrowing further.

Where it stands: mixed. The earnings are real. NVIDIA’s quarterly net income was $59.7 billion, up from $26.4 billion a year earlier, which is why the site’s tracker reads Holding up. But the ten largest S&P 500 companies made up 37.8% of the index at the end of August, with NVIDIA alone at 8.1%, according to S&P Dow Jones Indices data. Even the condition’s own author flagged that semiconductors were 70% of the Nasdaq’s return, “both good and bad.”

How the weekly Pulse will work

Every week you’ll get the same structure, so changes are easy to spot:

  1. The overall signal: accelerating, mixed or deteriorating.

  2. The nine-signal scorecard, with any change in reading explained.

  3. Three developments that mattered, and which signals they moved.

  4. Who’s affected: the companies and supply-chain layers touched, linked to the map.

  5. The case for boom and the case for bust, each built from the strongest sourced evidence that week.

  6. What to watch next week.

  7. Every source, linked.

Next week already has real tests. Micron and Jabil report on September 30, and Accenture reports on October 1. That’s memory demand, server build-out and enterprise AI services in three days.

If you want to see the machinery behind the readings, the trackers and the company map are open to everyone.

Subscribe if you want this in your inbox every week. And reply to this email with the one thing about the AI economy you most want to understand. I read every reply, and the best questions will shape what I dig into next.

This is educational research, not investment advice.