The Focus List: Own the Receipts, Not the Frontier

The Federal Reserve raised interest rates for the first time in three years, and it did not hedge. The vote was 12 to 0, a quarter point. The target range is now 3.75% to 4%, and officials project one more increase before the year is out.

The 10-year Treasury yield closed just above 5%. If you have a savings account or a car loan, that number already reached you.

Then there is the other half of the week. Over the weekend, Dario Amodei, who runs the AI lab Anthropic, published an essay arguing that frontier development should slow down. Two rivals publicly agreed.

The tape answered first. On Monday the iShares Semiconductor ETF fell about 5.6% in a single session. Broadcom lost 4.8%, and Constellation Energy lost 7.1%. Netflix rose nearly 4%.

Two halves of one week, moving in opposite directions.

Every one of these four names is the same trade seen from a different seat. The operating numbers kept climbing while the multiple kept falling. Broadcom trades at roughly 17 to 18 times next fiscal year's earnings against a five-year average of about 32 times.

Constellation is about a third below its high with guidance raised. SoFi sits near 28 times this year's guided earnings after a third of its value came off since January. Netflix is the biggest de-rating of the four, at about 20 times forward earnings against a five-year average near 40.

The market spent the week arguing about whether it still wants to pay for the frontier. These four are how you get paid for the receipts instead: compute shipped, power contracted, loans originated, subscriptions collected.

This is the first edition of the Focus List, so there is nothing to score. Each entry has four parts: why the name is on the list, what it has to prove, the bear case, and the call with what would change it. Analysis, not advice.

Broadcom — the arms dealer of the custom-silicon era

Broadcom designs custom AI chips for very large technology companies that would rather not buy Nvidia's. It also sells the switches and optics that move data between them, plus infrastructure software.

The customer list is six names deep, including Google, Meta, ByteDance, OpenAI, Anthropic and Apple. Those six carry the entire AI line, and Broadcom holds roughly 70% of AI ethernet switching.

That is why it is on the list, and why it fell 4.8% that day. For Broadcom, the slowdown debate is a question about whether six customers can pay for what they ordered.

The company's own numbers have not slowed. Revenue in the quarter reported on Sept 2 was $29.6 billion, up 86% and a record. AI semiconductor revenue was $16.7 billion, up 221%. Operating margin was 66%, and guidance is $34.8 billion including $21.7 billion of AI revenue.

That guide came in about $300 million light against consensus. The shares fell roughly 3% after hours despite the record. Multi-year ambition does not buy a stock time. The marginal quarter does.

Management put fiscal 2026 AI revenue at $58 billion and fiscal 2027 at about $115 billion on the call. It also gave a first fiscal 2028 line of sight near $230 billion. Those are targets, not delivered revenue, and they are what the bears are attacking.

Anthropic is on track to become Broadcom's largest custom-chip customer in 2027 and 2028. OpenAI's next accelerator is scheduled for 1.3 gigawatts of deployment in 2027. Asked about the slowdown essay, chief executive Hock Tan said nothing had changed his revenue view.

The next print is expected in early December. There is no scheduled catalyst before it. Any lab capex revision moves this stock the same day, and so would a listing for Anthropic, which Tan has called an open secret.

The accountability line is specific. AI semiconductor revenue at or above the guided $21.7 billion, the $115 billion fiscal 2027 figure repeated on the December call, and the 66% margin held. If the company hedges either number, the slowdown debate stops being a sentiment story.

Here is the short case. Broadcom says it will ship roughly $350 billion of AI silicon across two years. Part of the enabling work runs through its own balance sheet. It stood up a financing platform with Apollo and Blackstone targeting more than 20 gigawatts of compute for OpenAI and Anthropic by the end of 2028.

The platform bridges the gap between those customers' cash flow and their upfront investment, a polite way of describing vendor financing. The customer list is six companies. Two are burning cash, and one is publicly arguing the industry should slow down.

Tan has said the open question is whether the chips get deployed on time, and that land and power, not demand, set the schedule. The second leg is the market's reaction to the last print. It raised a multi-year target and put $230 billion on 2028. The stock still fell, because the next quarter came in light.

What proves the bears right is a delay or restructuring in that financing platform. So is a capex cut at Anthropic or OpenAI, AI revenue missing the implied sequential ramp, or operating margin breaking below 66%.

The call is accumulate here. The compression has already happened, and the print that disappointed is behind it. What would change the call is a hedge on the $115 billion line, or a margin that slips below 66%.

Constellation Energy — the landlord of the AI buildout

Constellation runs the largest nuclear fleet in the United States. It added gas, hydro and renewable plants with its Calpine purchase. It sells that electricity into wholesale markets and, increasingly, to technology companies on 20-year contracts.

Around 55 gigawatts of already-licensed, always-on generation is the asset. Data centres need round-the-clock power where the grid is tight, and nobody else can sell it without waiting years to build.

The stock lost 7.1%, a fourth consecutive down day, and sits in the bottom sixth of its 52-week range. It trades about a third below its high and roughly 12% below its 200-day average.

Rising rates hit long-duration, capital-hungry businesses first.

The results have been mixed in a way that matters. Second-quarter revenue of $7.50 billion grew 23% and still came in short of the roughly $7.94 billion analysts expected. Adjusted operating earnings rose to $2.55 a share from $1.91, while GAAP net income fell to $1.42.

Nuclear output slipped to 44,160 gigawatt-hours from 45,170, and the capacity factor eased to 93.0%. Guidance was raised to $11.50 to $12.50 a share from $11.00 to $12.00.

The dated catalyst is not the earnings print, which is expected in early November. It is a market design fight. PJM, which runs the grid across much of the eastern United States, has asked regulators to accept a backstop procurement filing by Sept 29.

If accepted, generators bid from Sept 30 to Oct 21, with results due before the main capacity auction. PJM wants 6,831 megawatts, the exact shortfall its last auction left behind. No order has been issued yet.

The interesting detail is who can bid. Only new generation qualifies. The maximum willingness to pay is a capacity-weighted average of $555 per megawatt-day for terms up to 15 years. Constellation's existing fleet cannot bid.

The backstop would add competing supply at a price far above the $325 per megawatt-day cap that binds existing nuclear capacity. The auction that does price Constellation's fleet runs in early December. The last one cleared at the ceiling regulators set. It added only 525 megawatts of new supply and finished 6,831 megawatts short.

The third quarter must show adjusted operating earnings inside the raised corridor. That needs about $3.74, revenue that beats rather than misses, and both the 2027 Crane restart and the mid-2027 Meta contract on schedule. Not one dollar of its biggest AI contracts lands in 2026 results.

That last point is the bear case. The largest contract is a 20-year agreement with Meta for 1,121 megawatts at the Clinton plant in Illinois. It does not begin paying until June 2027, when a state subsidy programme expires.

The Microsoft contract for 835 megawatts starts when Three Mile Island Unit 1 restarts in 2027, a first-of-a-kind regulatory exercise accelerated from 2028. The newest agreements start in 2029.

You are therefore paying roughly 20 to 23 times earnings, depending on the year you use, for growth that is contractually scheduled and therefore allowed to slip. The arrangement also caps the upside: auctions clear at ceilings regulators set, and the higher backstop price is reserved for new entrants.

What proves the bears right: the December auction clearing at the cap again with little new supply. So is a Crane restart that slips past 2027, another revenue miss with nuclear output down, or a capacity factor below 93%.

The call is wait for a pullback. The contracts are signed and the fleet cannot be replicated, but the money arrives in 2027 and 2029. What would change the call is a Crane restart confirmed by regulators.

SoFi — the teller of the AI buildout

SoFi is a nationally chartered bank that lives in a phone. It takes deposits, makes personal, student and home loans, and sells brokerage and crypto to 15.8 million members. It also owns Galileo, the technology platform other financial firms build on.

It is on this list because of what the Fed did. SoFi funds itself with deposits, and it pays up for them to compete. When the policy rate rises, deposit costs move faster than a fixed-rate loan book reprices.

The shares have lost more than a third since January and sit in the bottom tenth of their 52-week range. They are roughly 48% below their high and about 14% below the 200-day average, the most stretched of the four.

The results themselves were strong. Second-quarter net revenue of $1.22 billion grew 43%, and GAAP net income of $156.6 million grew 61%, the eleventh consecutive profitable quarter. Originations set a record at $14.8 billion, the first $10 billion personal-loan quarter.

Guidance was raised to $4.75 billion to $4.85 billion of adjusted revenue and about $0.60 in adjusted earnings per share. Net interest margin was 5.98% in the quarter, just under the 6% line the bulls watch.

Full-year guidance of about $0.60, against $0.12 in the second quarter, means the second half has to run at roughly double that pace.

There is also the crypto plumbing. SoFi and Kraken announced a partnership on Sept 3. Kraken joins SoFi's settlement network for round-the-clock dollar transfers and lists the SoFi stablecoin. The same day, regulators approved SoFi's stablecoin-issuance subsidiary.

The catalyst is a collision. The next print is expected in late October. If the pattern holds, it lands the same day the Fed's next meeting opens, so a deposit-funded lender reports into a central bank that just hiked.

The accountability line sits where the gap is. Net interest margin back above 6%, fee revenue growing faster than the loan book, and Technology Platform revenue positive year over year again. Financial Services plus Technology Platform also needs to cross half of adjusted net revenue. Today that pair is 46%.

That is the bear case in one number. SoFi is priced as a platform and reports as a lender. Lending brought in $724.8 million, 59% of the company, growing 63% at a 56% contribution margin. Technology Platform, the segment carrying the AI-native label, fell 23% from a year earlier to $84.5 million at a 14% margin.

The smallest, slowest and least profitable piece is the one asked to justify a software multiple. Roughly 28 times this year's guided earnings is a lot for a consumer lender whose funding reprices faster than its assets, with margin under 6%.

The growth mix is the third leg. The loan platform business contributed $143.3 million on $3.1 billion of personal loans originated for third parties. Selling loans works beautifully in a benign credit market and closes overnight in a bad one.

Then there is the lawsuit. Citizens Bank sued SoFi in the District of Connecticut in early September over more than 30 mortgage hires. The complaint alleges SoFi induced those employees to breach non-solicitation agreements and pleads racketeering. SoFi rejects the allegations.

What proves the bears right is net interest margin below 5.75% or another decline in the technology platform. So is fee-based revenue failing to cross half of adjusted net revenue.

The call is wait for a pullback. The rate cycle just turned against a deposit-funded lender, and the platform story is currently a shrinking segment. What would change the call is a margin print back above 6% alongside genuine growth in Galileo.

Netflix — the escape hatch from the AI buildout

Netflix sells monthly subscriptions to series, films, live events and games in more than 190 countries. It also sells advertising against its cheaper tier. Its capital goes into shows, not silicon.

That is why it rose nearly 4% while the semiconductor complex fell about 5.6%. It is the one name here that gets paid if the capex debate resolves badly. The shares are about 38% below their high and roughly 11% below the 200-day average.

They are also the only shares on this list trading above the 50-day. Against a five-year average multiple near 40, roughly 20 times forward earnings is the biggest de-rating of the four, on record profit.

The catalyst is the only company-confirmed date among these four: results on October 20. Guidance into them is $12.86 billion of revenue, up 12%, at a 33.2% operating margin. The full year was narrowed to $51.0 billion to $51.4 billion at a 31.5% margin. Free cash flow guidance was raised to about $12.5 billion.

What it must show is that 12% revenue growth can coexist with 2% engagement growth. Revenue at or above the guide, the margin at 33.2%, and advertising tracking to about $3 billion. Management said on the call that first-half viewing hours rose about 2%, with hours per member softer.

The bull case therefore rests on price and advertising doing the work that engagement used to do. The bear case is simple: a growth multiple on a business whose engagement has stopped growing. Content spending is guided up about 10% this year. Slower hours plus faster content cost is deteriorating unit economics.

The answer is monetisation that has already been partly spent. Prices went up across all plans earlier in 2026, so the next step depends on advertising, where it competes with Amazon, Google and a Paramount-Warner Bros. combination that owns far more of the classic library.

The quarter that just closed had cracks of its own. Revenue of $12.56 billion slightly missed, and the stock fell more than 7% the day after the print. Free cash flow dropped to $1.53 billion from $2.27 billion, though that carries taxes on the $2.8 billion Warner Bros. termination fee, a one-off rather than a trend.

Two more things a short-seller would raise. Netflix stopped disclosing quarterly membership numbers, which removes the cleanest evidence that the model still works. Large holders also moved in opposite directions. Fourteen of 22 prominent investors added, while a filing showed a reported sale of more than 20 million shares in the second quarter.

The AI link shows up in a small way. Management said on the call that generative tools are in production across roughly 300 titles, mostly in post-production. One title used 17 minutes of AI-enhanced footage made about twice as fast at roughly half the cost.

What proves the bears right: full-year revenue landing at the bottom of the range, or an operating margin under 31.5%. So is advertising guided short of $3 billion, or a disclosure showing viewing hours per member still falling.

The call is accumulate here, and the reason is the size of the de-rating rather than a view on the slate. The multiple is about half its own five-year average on record profit, the date is confirmed, and the business improves if the AI buildout does not. What would change the call is a guide-down on advertising or margin on October 20.

Where the four fit together

If you own none of them, start with what each is a claim on. Broadcom is compute shipped to six customers. Constellation is power contracted for 20 years.

SoFi is the credit plumbing, and the least AI of the four. Netflix is the hedge: it earns more if the argument in the other three names gets worse.

Watch three things. Whether regulators accept the PJM backstop filing in time for the Sept 30 bid window, because that decides whether scarce power stays scarce. Whether the 10-year holds near 5%, because that is the discount rate on every long-dated promise here. And Netflix on October 20, the one hard date on the calendar.

Three of these four are priced as though the buildout stops. The operating numbers say something else. The frontier is priced as a promise, and the price of that promise went up. Compute shipped, power contracted, loans originated and subscriptions collected are receipts you can count.

Disclosure: The Signal holds no position in AVGO, CEG, SOFI or NFLX. Positions may change. This is not financial advice.