The Focus List: Four Invoices in a Market of Promises

Two things happened in the same week, and they contradicted each other.

The first was a record. The Nasdaq set all-time closing highs on September 21 and 22, and AMD crossed a $1 trillion market value for the first time. Meta ran about 11% in a single session that week on its new Muse agent.

The second was a warning about who pays. On September 24, Oracle sent a “force majeure” notice on a New Mexico data center reported at roughly $165 billion, citing power delays. Its shares fell about 3%.

The bond market then agreed with Oracle. The 30-year Treasury yield reached its highest since 2004, with the 10-year around the 5% line. A yield is the interest rate a government pays to borrow, and it sets the bar for every other promise of money. When it rises, cash due later is worth less today.

So the week’s argument was not really about demand. The argument is about the financing underneath it.

The four names below are receipts businesses in a promise market. Meta sells the ads that fund a frontier AI lab, and AMD designs the second source of AI silicon the largest buyers now refuse to do without. Zscaler sells the door every employee and every AI agent walks through, and Axon sells police departments a subscription they sign once and keep for a decade.

A promise is exciting and a receipt is boring, a distinction anyone who has tried to cancel a phone contract already understands. None of the four needs the bubble question settled to send an invoice next quarter.

The scorecard: one week on

Last edition set one test for each of four names. One week on, all four are pending, because no company has announced a date. The moves below are measured from the September 18 close.

Microsoft’s test is its fiscal first-quarter print, expected in late October, on Azure growth and quarterly capex. Neither has been restated. A dividend rise on September 15 and a reporting reorganization on September 2 both predate that edition. The stock is up roughly 4.5%.

Arista’s test is its third-quarter print, expected in early November, and revenue against the $3.3 billion guide. Nothing on the print has moved. But a critical flaw in its VeloCloud product was reported as actively exploited on September 22, with a short federal deadline to patch. The stock is up roughly 3.6%.

Vistra’s test is the print where guidance gets refreshed to fold in Cogentrix and the Meta contracts, expected in early November. It has not arrived. What did: $1.5 billion of hybrid notes sold on September 24, ranking ahead of common shareholders. The stock is down roughly 1.6%, and the $7.4 billion to $7.8 billion 2027 opportunity range is confirmed in the company’s own release.

RTX’s test is the Raytheon segment margin against the 12.6% it reported in the second quarter. The late-September confirmation last edition expected had not been issued by September 26. The stock is down roughly 2.4%.

1. Meta Platforms (META) — The Ad Budget

Picture a business with two bank accounts. One collects advertising money from millions of small businesses. The other pays for a frontier AI laboratory that builds models nobody has agreed to buy yet.

That is Meta: an advertising business funding an AI lab out of its ad budget. The question each quarter is whether the ad budget grows faster than the lab spends. In plain English, it sells advertising to roughly 3.6 billion people who open Facebook, Instagram, WhatsApp or Messenger on an average day. Advertisers pay; users do not.

The receipts keep arriving. In the quarter ended June 30 the company booked $60.8 billion of revenue, up 28%, and $59.4 billion of that was advertising, about 98% of the business. It guided the current quarter to $61 billion to $64 billion.

Here is the awkward part: Meta is neither cheap nor hated. At roughly 22 times forward earnings, the price you pay for next year’s expected profit, it is not expensive for 28% growth.

But it has run about 32% in a month and sits near the top of its 52-week range. The market is paying for Muse as a new revenue line while free cash flow has almost vanished.

The catch is where the money went. Meta spent $31.1 billion on property and equipment in the June quarter and kept $784 million of free cash flow, cash left after paying for buildings and gear.

Costs rose 55% while revenue rose 28%, and operating margin fell to 31% from 43%. Capital spending was narrowed upward to $130 billion to $145 billion, and its youth-related US trials may bring a material loss.

What we’re watching: third-quarter results, expected in late October. The proof is revenue inside the guided band with 2026 capital spending at the low end, so free cash flow re-accelerates off $784 million. The kill number is expenses above $169 billion, or capital spending at the $145 billion top with revenue at or below $61 billion.

Our take: Wait for a better entry. The operating story is intact and 22 times forward earnings is not a bubble multiple, but you would be buying after a 32% month. What changes the call is a pullback toward the 50-day average, or capital spending at the low end with free cash flow re-accelerating.

2. Advanced Micro Devices (AMD) — The Second Source

Every large buyer of AI chips has the same nightmare. It is needing gigawatts of computing and being able to buy it from exactly one company.

That is the case: AMD gets paid when the largest customers decide they cannot buy all their AI silicon from one supplier. In plain English, it designs the chips in AI servers not made by Nvidia, selling them as complete cabinets of computing called Helios. It also makes PC and console chips.

The demand is not theoretical. Second-quarter revenue was $11.54 billion, up 50%, and the data center segment was $6.70 billion, up 107% and 58% of the company. Gross margin rose to 54% from 40% on the official accounting and to 56% adjusted, and adjusted earnings per share of $1.66 beat expectations.

Management guided the current quarter to about $13 billion and expects data center sales to accelerate in the second half. Behind the guide sits a stack of contracts. Meta signed for six gigawatts of chips in a deal reported at up to $60 billion over five years. OpenAI signed for multiple gigawatts, and Anthropic for up to two.

This is not a hated stock, and it is not cheap. AMD closed the week at its 52-week high, up about 296% in a year. At roughly 40 times forward earnings the multiple requires the guidance to be right.

The catch is how those contracts get paid for. Meta’s agreement is partly settled in AMD shares, through a warrant on up to roughly 160 million shares. Those warrants vest in tranches as shipments scale and the share price clears rising thresholds. AMD is also putting up to $5 billion of cash into Anthropic, a customer that burns cash.

A Helios rack carries networking and cooling content a chip does not, so revenue can grow faster than profit. And Oracle’s question is AMD’s too: if a gigawatt promise can be delayed by power, AMD holds gigawatt promises of its own.

What we’re watching: third-quarter results, expected in early November. The proof is revenue at or above $13 billion, with data center growing at least as fast as it did in June and adjusted gross margin at 56%. The kill number is a sequential slowdown, or margin below 56% as the mix shifts from chips to racks.

Our take: Accumulate. This is the strongest verified momentum on the list, and the catalyst is growth rather than hope. What changes the call is a decelerating data center quarter or evidence that booked gigawatts are not turning into shipped silicon.

3. Zscaler (ZS) — The Toll Booth

The office network used to be a building. Everyone worked inside it, the walls did the security, and the internet came in through one door.

Then the building emptied, the applications moved elsewhere, and the door turned out to be the only part worth keeping. Zscaler sells that door as a subscription, a toll booth between every employee and every application, because the corporate network is the thing you no longer trust.

In plain English, it sells cloud-delivered security. Instead of routing traffic through a company firewall and a virtual private network, it checks every person, every machine and now every AI agent. Each one is sent only to the applications it may reach. Enterprises and governments pay per user per year.

The base is still compounding. Revenue in the July quarter was $898.2 million, up 25%, and annual recurring revenue, the subscription total that renews each year, was $3.771 billion, also up 25%. Adjusted operating margin hit a record 24% of revenue, adjusted earnings per share of $1.19 beat expectations, and there is roughly $1.8 billion of net cash.

The stock has been repriced. It trades at roughly 34 times forward earnings, down from about 100 times a year ago. It sits about a third of the way up its 52-week range, roughly 43% below its high. The debate is whether a slide from 25% growth to a guided 17% is the category maturing or an acquisition artifact.

The catch is that the guidance reads like a confession. Zscaler grew full-year revenue 25% and pointed the new year at 16.6% to 17.5%. Excluding the acquisition, net new recurring revenue grew 17%, below the 20% organic pace the company had been posting.

Free cash flow fell to 7% of revenue from 24%, as capital spending and capitalized internal software jumped to $218.5 million from $78.7 million. On September 24 the company said its chief revenue officer was stepping down.

What we’re watching: fiscal first-quarter results, expected in late November. The proof is revenue inside the guided $935 million to $939 million with net new recurring revenue re-accelerating organically above 17%. The kill number is recurring revenue tracking below the guided range, or another surprise in the sales organization.

Our take: Accumulate. The de-rating already happened, and the bear case is now the base case in the price. What changes the call is revenue below $935 million or organic net new recurring revenue below 17%.

4. Axon Enterprise (AXON) — The Evidence Locker

A police department does not choose software the way you choose a phone. Switching systems means moving years of video evidence and court records.

Axon sells the subscription a department signs once and keeps for a decade, because by the second year its evidence is already inside Axon’s cloud. In plain English, it sells the operating system for public safety: TASER devices, body and in-car cameras, drones, 911 software, and the cloud where departments keep the footage. Government agencies pay for it on multi-year subscriptions bundled with hardware.

The growth has not slowed. Second-quarter revenue was $904 million, up 35% and the tenth quarter in a row above 30%. Annual recurring revenue rose 39% to about $1.6 billion, and net revenue retention was 126%, meaning existing customers spent 26% more than a year earlier. Future contracted bookings, work signed but not yet delivered, reached $15.1 billion, up 41%.

None of that has helped the stock. Axon sits about 20% of the way up its 52-week range, roughly 43% below its high. It fell about 30% between late August and September 15, the day it announced a bond sale. At roughly 40 times forward earnings the market has stopped paying a software multiple for a hardware-and-software bundle.

The catch is the conversion. Across four quarters Axon produced about $130 million of free cash flow on a market value near $35 billion. It guided stock-based pay for 2026 to $590 million to $620 million.

Into that, on September 15 and 16, it priced $1.0 billion of convertible notes. Those are bonds that pay no interest until 2031 and can turn into shares, and the shares were priced roughly 48% above the market. And the company skipped a Senate hearing on camera surveillance on September 23.

What we’re watching: third-quarter results, expected in early November. The proof is an eleventh straight quarter of growth above 30% and a full-year rate at the top of the raised 32% to 34% guide. The margin needs to hold at or above 25.5%. The kill number is a margin below 25.5%, a software gross margin below 71%, or free cash flow at or under zero.

Our take: Wait for a better entry. The bookings number is the best evidence on the list that demand is not the problem. But you are paying about 40 times forward earnings for roughly $130 million of trailing free cash flow. What changes the call is a quarter where free cash flow turns positive and the margin holds.

The bottom line

The week asked whether the AI buildout is a bubble. We keep asking the quieter question: who sends an invoice while the argument runs?

Meta sold the ads that paid for the lab. AMD designs the second source the largest buyers will not do without. Zscaler guards the door every employee and every agent walks through. Axon signs the warrant that turns a camera into evidence.

Two of the four, AMD and Zscaler, are priced as if the doubt were permanent. The other two, Meta and Axon, still have to show that growth becomes cash.

Oracle’s notice and a 30-year yield at its highest since 2004 did not change what any of them sells. It did make the waiting more expensive. We will track all four and grade every call in public.

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