The Focus List: Own the Meter, Not the AI Argument

Your subscription bill is boring. It arrives every month, it barely changes, and the company on the other end can see most of next year’s revenue already. That boredom is the point.

Start with the week. On September 16 the Fed raised its benchmark rate a quarter point to 3.75%–4% — a first hike in more than three years, and unanimous. The 10-year Treasury yield topped 5%.

A rate is the discount on every promise of money arriving later, which is why long-dated plans got harder to believe. Oil held above $100 on the Iran war and squeezed budgets the same way.

The scoreboard looked calm anyway: the S&P 500 slipped 0.1% across the week, the Nasdaq added 0.7%, and the Dow lost 1.7% for its worst week since March. Underneath, the market was having a shouting match.

It started when Anthropic’s chief executive published an essay arguing for a slower pace of frontier-model development. The iShares Semiconductor ETF fell 5.6% in one session and Broadcom dropped 4.8%. A philosophy essay moved tens of billions of dollars before lunch.

The bounce was stranger than the drop. Chips gained 3.4% in the one strong session of the week — Intel up 7.7%, AMD 6.5%, Arm 8.6% — while software fell and utilities slipped about 1.4%.

Netflix gave the cleanest signal, falling about 10.6% across the week on a single catalyst: Wells Fargo’s Steven Cahall became the only sell-rated analyst on the Street. He estimates engagement fell to roughly 1.6 hours per subscriber per day in the first half, down 8% from 2023. Disney and Warner Bros. Discovery barely moved.

So the week left you with a familiar question: if nobody agrees on whether the AI buildout is a bubble, why bet on it either way? Our answer: stop betting on the argument and start owning the meter, because somebody else’s spending arrives as recurring revenue.

The scorecard: how last edition’s four are holding up

Edition #1 put Broadcom, Constellation Energy, SoFi and Netflix on the list and wrote down what each had to prove. Here is the grade.

Broadcom has added 5.3% since we wrote it up, though it gave back 1.2% across the week, about 28% below its 52-week high. Its test was the December print: repeat the $115 billion 2027 AI revenue target and hold the margin. On September 14 chief executive Hock Tan said the AI-slowdown push had not changed the targets, and the $230 billion 2028 figure stands. On track — December is still the exam.

Constellation is 10.5% lower on the week and about 38% below its high. A Form 144 shows an officer lining up a sale of roughly 5.41 million shares under a preset plan. Neither test — the Crane restart or the December auction — got new information. Under pressure: price, not business.

SoFi is down 2.1% on the week and about 48% below its high. Its tests were a net interest margin above 6%, Galileo growing again, and fee revenue clearing half the business. The last reported margin was 5.98%, two basis points short of the bar. Nothing yet — both tests land in late October.

Netflix is 10.6% lower and about 43% below its high. Its test was October 20: revenue at or above $12.86 billion and a 33.2% operating margin. The date is unchanged. Under pressure — the bear case now attacks the engagement growth it rests on.

1. Microsoft (MSFT) — The Meter

Picture the office you used to commute to. Somebody rented you the desk, somebody else billed you for electricity, and a third company supplied the coffee machine. Microsoft is all three, and it reads the meter.

In plain English, it sells companies the three things they cannot easily run without: the email, spreadsheets and video calls that keep an office moving. Then it rents them the cloud computing that stores their data, and the AI assistants on top.

The meter is running. In the quarter ended June 30 it booked $90 billion of revenue, up 18%, and $40.6 billion of operating income, also up 18%. Azure grew 43% and passed $100 billion of annual revenue for the first time. Copilot is past 30 million paid seats.

The number most people skip is the best one: $678 billion of commercial remaining performance obligation, which is signed work not yet delivered, up 84% in a year. That is as close to a receipt as software gets.

So why does it trade like a utility? At roughly 21 times forward earnings, which is next year’s expected profit, it sits almost exactly on the S&P 500’s own multiple of about 20. Its five-year average is nearer 29 to 33, and the stock is 11% below its 52-week high.

Here is the catch. This growth is bought with a capex bill rising faster than the revenue it creates. Capex — money spent on land, buildings and gear, mostly data centers — nearly doubled to $115.9 billion in fiscal 2026. Free cash flow fell to $19.6 billion from $25.6 billion.

The accounting is getting friendlier too. Profit includes a $3.2 billion gain on the Anthropic stake, and the adjusted figure strips out OpenAI investment losses. From fiscal 2027, the assumed life of a data center stretches from 15 years to 25.

What we’re watching: fiscal first-quarter results, expected around October 28 after the close, with confirmation usually two weeks ahead. Two numbers decide it: Azure growth in constant currency against the roughly 45% guide, and whether quarterly capex crosses $50 billion.

Our take: Accumulate. You are paying index money for the best meter in software, and the market is treating $678 billion of signed work as an afterthought. If Azure misses its guide while capex keeps climbing, that stops being investment.

2. Arista Networks (ANET) — The Traffic Cop

An AI data center is a factory floor where a hundred thousand machines are all trying to talk at once. Arista sells the traffic-control system, the switches and software that keep them from jamming every exit.

In plain English, Arista builds the high-speed networking gear that moves data between AI chips inside a data center. Its customers are a handful of giant cloud companies, and it has taken share from Cisco for a decade.

The business is compounding. Second-quarter revenue came in at $3.036 billion, up 37.7% and above its own $2.8 billion guide. Adjusted operating margin was 49.9%, and adjusted earnings per share rose almost 40% to $1.02.

Management raised the 2026 outlook for the third time this year, to about $12.6 billion of revenue and 40% growth. AI networking is now expected to bring in $3.5 billion or more, against roughly $1.5 billion last year.

It is not cheap, so the argument is not price. The stock trades near 39 times forward earnings against the index’s 20, and it is up 52% this year. The question is whether the compounding lasts.

Here is the catch, and it has a name: Nvidia. IDC’s tracker for the first quarter has Nvidia as the number one vendor in data-center Ethernet switching, with $2.1 billion of revenue, 21.5% share and 192.7% growth. Arista’s share is 20.7%.

The margin leak is the second problem. Gross margin fell to 62.9% in the first half from 65.2% a year earlier, which the 10-Q blames on more sales to the largest customers. Two customers were about 42% of fiscal 2025 revenue.

The growth is also pre-financed: purchase commitments went from about $3.6 billion a year ago to $9.7 billion. Brilliant if AI spending holds through 2028; painful if one hyperscaler pauses.

What we’re watching: third-quarter results expected around November 3 after the close, though no announce-date release had been issued as of September 19. Revenue against the roughly $3.3 billion guide, and whether that forecast gets a fourth raise. The kill number is gross margin, guided at 62% to 64%.

Our take: Accumulate. Supply, not demand, has been the constraint all year. Two quarters of gross margin below 62%, or one hyperscaler pausing orders, and we would step aside.

3. Vistra (VST) — The Electric Company

Vistra is a landlord for electrons. It owns the buildings — nuclear, gas, coal and solar plants — and it collects rent two ways.

In plain English, it sells electricity. Some of the money is a capacity payment: rent for agreeing to be available when the grid is tight. The rest is the metered bill for power customers actually use.

Second-quarter adjusted EBITDA from ongoing operations was $1,767 million, up 30%, with generation contributing about $994 million against roughly $593 million a year earlier. The fleet ran above 97% availability through Texas and PJM heat waves.

Two deals carry the story. Vistra signed 2,609 megawatts of 20-year nuclear contracts with Meta in January, and a PPA is a long-term contract to sell electricity at a set price. It is also buying Cogentrix, ten gas plants and about 5,500 megawatts, for roughly $4 billion.

It trades like something is wrong. At about 13.6 times forward earnings it is roughly a 29% discount to the S&P 500, and it sits at the bottom of its 52-week range after falling about 33% over twelve months.

The catch is that the core business is shrinking while the good news sits in deals that have not closed. Second-quarter GAAP earnings came in at $0.76 against $1.66 expected, a 54% miss.

Management also said the Texas power curves in the forward market sit below the ones underpinning its 2027 guide, and that grid payments and nuclear tax credits do not fully cover that gap. PJM’s last capacity price actually fell 2.5%, to $325 per megawatt-day.

Debt is $20.5 billion. On September 10 Vistra priced $1.5 billion of junior subordinated notes that rank ahead of common shareholders, and the dividend yields under 1%.

What we’re watching: third-quarter results expected around November 5. This is the print where management said it would refresh 2026 and 2027 guidance to fold in Cogentrix and the Meta contracts, which sit outside every guide so far. The range is $7.4 billion to $7.8 billion of 2027 EBITDA, with the core at the low end and roughly $700 million more from those deals.

Our take: Wait for a better entry. If the update puts 2027 clearly above $8 billion, the story is proven. If it lands in the low $7 billions and PJM’s December auction clears below its $337.83 cap, it is broken.

4. RTX (RTX) — The Quartermaster

Think of a quartermaster in a war film: the one who draws the kit, replaces what got used, and signs for what comes back. RTX sells the missiles, engines and seats, and gets paid again when they need servicing.

In plain English, it runs three businesses. Raytheon makes missiles and radar for governments, Pratt & Whitney makes jet engines for Airbus and Boeing, and Collins Aerospace makes the seats, avionics and landing gear inside those jets.

Second-quarter sales were $24.7 billion, up 14% reported and 16% organically, with adjusted earnings per share of $1.89, up 21%. The backlog reached $289 billion, split $170 billion commercial and $119 billion defence, up 22% from a year earlier.

Raytheon was the engine of the quarter: sales up 18% to $8.27 billion and margin at 12.6%. Book-to-bill was 2.42, which means it booked $2.42 of new orders for every $1 of revenue. Its backlog is $86 billion, and 48% of it is international.

Yet the stock fell about 12% in a month while the index fell under 1%. That is what priced in looks like: management raised full-year guidance in July and the market sold it anyway. At about 24.7 times forward earnings, a 29% premium to the S&P 500, you are paying up for 8% organic growth.

The catch is the mix. Of that $289 billion backlog, $170 billion is commercial aviation, tied to airline traffic and airframe build rates rather than Pentagon budgets. And the profit engine people assume, Pratt & Whitney, runs an 8.3% margin against Collins’ 16.7%.

Tariffs became a live cost line on September 18, when the FAA cited roughly $100 million of them on air-traffic-control modernisation. The dividend yields just above 1.5%, under its five-year average of 2.08%.

What we’re watching: third-quarter results expected around October 20, with formal confirmation usually arriving in late September. The one number is the Raytheon segment margin: 12.6% last quarter, guided to the mid-12% range for the year. Hold 12.6% or better on rising volume and the margin story is real.

Our take: Wait for a better entry. Slip back toward 11%, last year’s level, and you own a 25-times-earnings industrial with no margin story and a commercial backlog. We want the pullback to reach the multiple.

The bottom line

The market spent the week asking whether the AI buildout is a bubble. We asked a smaller question: who gets paid on a schedule while the argument runs?

Microsoft rents the enterprise, Arista wires the clusters, Vistra sells the electrons, RTX restocks the arsenal. Not one of them needs the frontier question settled to bill next quarter.

The Fed’s hike raises the bar for all four, because every long-dated promise is worth less against a 5% Treasury. We would buy the meters that pay soonest and wait on the two that pay later.

We will grade every call in the next edition, in public, and tell you when our thinking changes.

Analysis, not advice. We may own these names and we may sell them at any time.