CNBC Halftime Report — Time to Be Cautious on Stocks?

Vertex owns the treatment of cystic fibrosis almost outright — a genetic lung disease with a small patient population but essentially no competing therapy, which is why the business throws off dependable cash. Morgan Stanley reinstated coverage at overweight with a $665 target.

The whole investment question for a company like this is what happens after the one franchise. Lebenthal's answer is that the science travels: they have moved into kidney disease and, more importantly, non-opioid pain treatment, which is a vastly larger market than cystic fibrosis. Terranova adds the other route — the $10 billion acquisition announced Monday, which buys diversification rather than waiting for the labs to produce it.

Note what he is careful to say: they are not buying because they have to. A company forced into deals by a decaying core is a different (and worse) proposition from one buying from strength. Lebenthal's caveats are the honest ones — it is not cheap, and single-product biotech always carries the risk of one bad trial headline.

Lilly makes the obesity and diabetes drugs (the GLP-1 class) that have been the biggest earnings story in pharma this decade. JP Morgan reiterated overweight with a $1,400 target, praising a product portfolio that "appears increasingly differentiated from its peers."

Terranova's argument is not about GLP-1 volumes at all — it is about what Lilly is doing with the money. The company has committed roughly $20 billion to deals so far in 2026, buying its way into new therapy areas while the cash is flowing. He calls it "a large cap biotech company more than anything else," meaning: value it on a pipeline built by acquisition, not on one blockbuster drug franchise.

The comparison he asks for on air makes the point cleanly. Novo Nordisk sells into the same GLP-1 boom and its chart looks nothing like Lilly's. If the category alone explained the returns, the two would move together. They do not — so what is being rewarded is the reinvestment, not the drug class.

Wynn runs luxury casino resorts — Las Vegas, Macau, Boston, and a new one being built in the United Arab Emirates. Wolfe Research initiated it at outperform with a $128 target, about 40% above the price, on the argument that in a "K-shaped" economy (where high earners keep spending while everyone else pulls back) the premium end of gaming keeps working.

The segment is worth more than the call because Lebenthal, who owns the stock, argues against his own position first: "what an awful chart," and an unprompted account of his record — trimmed half at $130, bought back at $106, "my mistake so far." His single explanation is the UAE property, which sits on the Strait of Hormuz; with a war under way, the market is refusing to pay for a resort it is not sure will open on schedule.

Wapner and Terranova make the test that any single-cause story has to survive: was the stock already breaking down before that cause existed? It was — the casino group peaked in November, and Las Vegas Sands, which has no Gulf exposure at all, looks just as bad. That points to a sector problem (Las Vegas room pricing having overshot after the post-pandemic boom) rather than a Wynn-specific one. Lebenthal accepts the sector diagnosis and argues the pricing reset is what brings demand back.

Taiwan Semi does not design chips; it manufactures them for everyone who does — NVIDIA, Google, Apple, AMD, Broadcom. A "fab" is a chip factory, and building one costs tens of billions of dollars and takes years, which is why almost no designer wants to own one. That is the whole business.

Weiss's claim is that this makes it structurally better than owning any single designer: "a better must own, frankly, than NVIDIA, because Nvidia's getting their chips made there. So is Google, so are all the others." Whoever wins the AI chip race, the wafers come from the same place — so you are buying the outcome of the competition rather than a bet on one competitor.

He also flips his usual caution about capacity expansion. Normally a manufacturer adding capacity is a warning sign, because new supply arrives just as demand fades. Here he argues the reverse: demand is already booked, and the physical build takes years, so the new capacity will be full before it exists. Worth noting his one trim was for position size after a run, not a change of view — a distinction that matters when reading anyone's sells.

The useful thing here is not the stance but the question Lebenthal poses, because it resolves most of the argument about AI semiconductors in one line: NVIDIA says the industry cannot make enough chips; Broadcom's results imply it is losing business to rivals. Both cannot be the main story.

If NVIDIA is right and the constraint is supply — particularly memory — then competition is close to irrelevant: everything anyone can build gets sold, and a competitor winning a design does not cost the incumbent a sale it could have filled anyway. If Broadcom is right and the constraint is demand, then every rival win is a real loss. Lebenthal sides firmly with supply, which is why he does not treat AMD's or Broadcom's Alphabet partnership as a threat.

His second point is the durable one: CUDA, NVIDIA's software layer. Chips get faster every year, but the code customers have already written for a particular chip is expensive to rewrite — that switching cost is the moat, and it is not something a rival fixes with better silicon. Wapner's caution stands alongside all of it: NVIDIA's own outstanding quarter did not lift the rest of the AI complex, so a good print no longer guarantees a good tape.

Oracle has spent heavily to turn itself into a supplier of rented AI computing power to companies that need it. It is Lebenthal's final trade, and he is holding it through a bad stretch — he describes it as a name in the AI CapEx build-out where "the stock has done terribly."

His thesis is one sentence: "we are compute constrained and Oracle will benefit." In plain terms, there is more demand for AI computing than there is hardware to run it on, so anyone who owns capacity can charge for it. The read-through he uses is Dell's raised forecast that morning: if the companies selling data-centre hardware are guiding higher, the spending that Oracle's plan depends on is still coming.

The risk is worth naming, because he does not: this is the most leveraged version of the AI-CapEx bet. The spending is Oracle's own, upfront and debt-funded, and the revenue arrives later. If the build-out slows for any reason — including the political backlash discussed at length elsewhere in the same episode — the cost is already sunk.

Adobe sells the standard software for design and creative work, on subscription. The market has spent the year worrying that generative AI makes those tools optional, and the stock fell hard — from about $200 at the low six weeks ago to around $285 now, still cheap at roughly 10 times earnings.

Lebenthal's argument has two legs. The first is evidence rather than opinion: his own firm — 1,800 people, $200 billion under management — just renewed its Adobe contract. There were price concessions, which is the honest part, but a large enterprise did not leave. Big organisations do not replace the software their whole workflow runs on because a cheaper tool exists; that inertia is what a subscription business is.

The second leg is an event, not a fundamental: the CEO seat is empty, and naming one removes an overhang. He goes further and floats the possibility of a chief executive appointed to sell the company — at 10 times earnings, an acquirer could pay a large premium and still get it cheaply. Wapner correctly labels this what it is: owning a stock for a catalyst rather than for the business, which has a different risk profile, since the catalyst can simply not happen.

Dell builds and sells the servers that fill AI data centres. Its results the previous night beat and it raised its forecast; Citi and Bank of America both moved their targets to $600. Nobody on the desk owns it, which is what makes it useful evidence rather than a talking point.

The argument being settled is whether AI capital spending is about to stop. A server maker is one of the earliest places that would show up, because orders are placed months before data centres open. Guiding higher is a statement about the order book, not about sentiment — which is why Lebenthal uses it as the direct answer to "AI is a bubble, the CapEx is going to turn off."

Read it against the prior session's page, where the same stock was written up at $434 with $500 as resistance and $400 as the line where a trader should walk away into this exact print. The event resolved to the upside — a reminder of why setting the level in advance is the whole point of that discipline.

Alphabet appears on both sides of the ledger in this episode, which is unusual and worth reading carefully.

The bear half comes from an owner: nine months ago Google had the leading AI model, and since then Anthropic has taken the lead in usage while OpenAI has not fallen behind either. The stock has gone from roughly $400 to $330. Lebenthal's point is that in this market the model ranking is not a vanity metric — it determines who wins enterprise contracts — so the Gemini 3.8 Flash release this week actually matters.

The bull half is Weiss's, and it rests on the same volatility: leadership in AI "cycles," so a company that has lost the crown once can take it back, and at this price you are not paying for the crown. He also gets a clean regulatory win the same day — a judge ruled Google does not have to sell its advertising-technology business, rejecting most of the government's proposed remedies, which removes a large and long-running tail risk from the valuation.

Valero is a refiner: it buys crude oil and turns it into gasoline, diesel and jet fuel. Refiners do not primarily make money from the oil price — they make it from the spread between what crude costs and what finished fuel sells for, which is why they can do well in exactly the environment (rising oil) that hurts most other businesses.

Terranova's final trade is a sector instruction — "stay with the refiners" — and it is the single-name expression of the map he drew earlier in the hour: if the AI trade does not restart, the momentum factor rotates into energy, healthcare and financials, exactly as it did in 2022. Naming a specific instrument for a factor call is what makes the call testable rather than decorative.

Liz Thomas's final trade is gold, and she repeats it after being wrong on it: down 3% since she last picked it, and she "likes it even more now." (No instrument was named on air; GLD, the largest gold ETF, is used here as the row's proxy.)

The reasoning sits in her main segment rather than the final-trade slot. Her description of the environment is a Treasury and a Federal Reserve sending conflicting signals, yields higher than a generation of investors has ever dealt with, and inflation nobody has a confident model for. Gold pays no income and produces nothing, so it does badly when real interest rates are attractive and does well when people distrust the institutions setting them. Buying it after a fall, on an unchanged thesis, is the coherent version of holding a hedge — the alternative, dropping it because it dipped, means you never owned a hedge in the first place.

Anthropic is a private AI lab, so there is no way to own it directly — but it moves two public positions on this show, which is why it is tabled.

First, it is named as the reason Alphabet's stock is down: an owner of Alphabet says Anthropic "has clearly taken over in terms of the most popular, most used large language model," and that Google needs its new release to land. Second, Weiss lists it alongside OpenAI and Alphabet as a source of new cybersecurity tools, which is part of why he thinks the security vendors' valuations are unsafe.

The general lesson is worth carrying: a private company you cannot buy can still be the thing determining what your public holdings are worth. Tracking the frontier-model leaderboard is not a technology hobby if you own Alphabet or a software vendor — it is position monitoring.

Broadcom designs custom AI chips for the big cloud companies (notably Alphabet) alongside its networking business. It reports after the close, and Wapner sets it up as a live read on whether the momentum trade is still alive.

Terranova's account of what broke is precise and dated: on June 4th the company missed its AI revenue number by $1 billion — $16 billion instead of $17 billion — and the stock fell 12% the next day. The reason a small percentage miss did that much damage is that the number was the evidence for the whole story; when the evidence wobbles, the multiple people were willing to pay collapses. It has: 21 times forward earnings now, against 32 times in June, and a ten-year average of 18. Cheaper, but not cheap.

His conclusion is the part most worth copying, because it is the option almost nobody exercises on television: "sometimes there's no trade." He is not bullish and not bearish; he is saying the setup does not offer him an edge, so he will not manufacture one. And he attaches a falsifiable test to it — if 200% earnings growth and 85% revenue growth is not enough to move the stock, then the AI momentum trade is genuinely dead for now, and that tells you something about every other name in the group.

Palantir turns up as the analogy someone used to talk Terranova into Broadcom: a heavily-sold stock that reported strong earnings and bounced hard. "Why can't this be the next Palantir?"

His answer is an admission — he sat that one out — and then a correction to the analogy. Palantir's bounce happened because the stock was washed out going into a strong print, and it is down 7% today, which is the useful detail: the pattern produced a bounce, not a new trend. Borrowing a setup from a name where it "worked" requires knowing what actually happened afterwards, not just on the day.

SanDisk makes flash memory, and it was, in Terranova's words, one of the standout momentum trades of the year. It appears here as a lesson rather than a call.

David Tepper sold his position in the second quarter — in what was fundamentally the best quarter the company had ever had. That is the whole point. Selling on strength, when every fundamental number is at its best, is nearly impossible to justify from a spreadsheet and is exactly what Terranova means by calling investing "tacit knowledge": some of what an experienced manager does cannot be written down as a rule, which is also why copying the rule rather than the judgment usually fails.

AMD appears once, as the named competitive threat to Broadcom in the custom-AI-chip market. Lebenthal's argument is that the threat does not matter if the industry is supply-constrained — when nobody can build enough, a rival winning a design does not take away a sale you could have made. No view was given on AMD itself.

The EWZ is a fund holding the largest Brazilian stocks. It landed in the twenty most-active options names in the entire US market — extraordinary for a single-country fund — with six times normal volume and over $30 million of premium traded, 94% of it in calls.

A call is the right to buy at a set price later. Ninety-four percent call activity sounds like a wall of bullishness, and this is where Renick's detail earns its place: about half of that premium was being sold, not bought. Selling a call is the opposite bet — it pays you now and profits if the move stalls. So the flow is not a crowd piling in; it is a genuine two-sided fight over whether a rally that has already happened continues, with implied volatility (the price of the options themselves) elevated because of it.

The most-bought single contract makes the speculative flavour concrete: a November call needing another 20%-plus move to pay anything at all. Possible drivers named: surging commodity prices, and next month's Brazilian election.

From the news update rather than the desk: Southwest is building airport lounges in Austin, Baltimore, Nashville and Honolulu — at least eleven in total — alongside a premium credit card next year. For an airline built for decades on a single cabin class and no frills, this is the continuation of a deliberate move up-market, chasing the higher-margin traveller and the card-fee revenue that comes with them. No committee view was offered.

Vertiv makes the power and cooling equipment inside data centres. Weiss names it as evidence that the political backlash is already showing up in prices — it "hasn't done really much of anything" while the build-out story it depends on stays intact.

One caveat should travel with this row: the automatic transcript renders the name as "Verdict," and a third name in the same sentence is unrecoverable garble. The reading rests on the prior session's page, where Terranova named Vertiv and GE Vernova as his two positions facing a stretch of intense data-centre headlines. Treat the identification as probable rather than certain.

OpenAI is private, but it drives the episode's central segment. Sam Altman says he is comfortable with his own company's compute plans and worried about everyone else's — specifically "random new neoclouds" (small firms renting out AI computing power) announcing enormous build-outs without the customers or revenue to pay for them.

Weiss's response is the one to remember, because it applies to every CEO comment about an industry: check the incentive. OpenAI has already locked in the capacity it needs. A wave of new supply would lower the price of the thing it has bought, and slow the build-out of rivals it competes with. That does not make Altman wrong — a lot of unfunded capacity really is being announced — but it does mean his warning is not disinterested, and Liz Thomas independently reaches the same read: "spoken like a true CEO."

OpenAI is also the reason two other trades move here: its cyber-tool releases are part of Weiss's bear case on the security vendors, and the administration filed a brief supporting it in the New York Times copyright case, arguing that training AI models is generally fair use.

Palo Alto is the largest cybersecurity company. It reported record results — a quarter, in Terranova's words, "they have never had before" — and the stock fell 10.5%, its worst day in about two and a half years, dragging CrowdStrike, Okta, Fortinet and Datadog down with it.

The explanation is positioning, and it is the single most useful idea on the page. A stock's price already contains what people expect. When everyone who wants to own a name already owns it — "positioning is full" — there is nobody left to buy on good news, and the only people who can act are sellers. So a genuinely excellent quarter produces a fall, and no amount of fundamental analysis would have predicted it, because the fundamentals were never the variable.

Two things follow. First, Terranova's honesty about duration: he holds the stock, he is not selling, and he expects this to take time — "I don't think it's a one day event either… I'm going to sit long and painfully watch it." Crowding unwinds over weeks, not sessions. Second, Weiss's separate and more serious charge: the AI labs themselves are now shipping security tools, so the competitive position — not just the positioning — may have changed, against valuations he thinks are indefensible either way.

CrowdStrike sells endpoint security — software that watches laptops and servers for attacks. It fell with the group on a day when it reported nothing at all, purely because a peer's results triggered a de-crowding of the whole sector.

That is worth internalising if you hold anything in a popular theme: the reason your stock moves on a given day is frequently not about your stock. The corollary is Terranova's: this kind of unwind "will get worked off over the course of time," so a one-day drop in a crowded group is rarely the end of the adjustment.

Okta handles identity — the logins and access permissions behind corporate software. It is named in the same list of cybersecurity names down across the board on Palo Alto's print, with no company-specific news of its own. Same mechanism as the rest of the group: full positioning, expectations that could not be beaten, and an adjustment that takes time rather than a session.

Fortinet is Palo Alto's closest comparison — both sell firewalls, the equipment and software that police traffic entering a corporate network. When the leader posts a record quarter, the natural expectation is that the read-through lifts the peer. It did the opposite, which is the clearest illustration on the page that the selling was about who already owned these stocks, not about how the businesses are doing.

Datadog is an observability platform — it monitors whether software and infrastructure are running properly, which is adjacent to security but not the same business. It gets named in the same breath as the three security vendors and falls with them.

That is itself the useful detail. Factor-driven selling follows how stocks are grouped in investors' baskets, not how the companies actually make money. If your holding sits in a popular basket, it will trade with that basket on the days the basket is being sold, whatever the business does.

GE Vernova makes the turbines and grid equipment that power data centres. Weiss uses it as the tape's evidence that the political backlash against data centres has stopped being a story and started being a price: the stock is down, and the pressure is concentrated in "those stocks that power the data centers or sell into the data centers."

What makes it hard is the mechanism, which he states plainly: local residents object to a data centre near them, and they blame it for rising electricity bills. That is a grievance politicians can act on, at the state and county level, without any federal legislation — which is why Liz Thomas calls the regulatory risk unpriced. The demand for the equipment is not in question; the question is whether the projects get permitted.

Read against the prior session, this is a harder line. Terranova framed the same pressure as roughly sixty days of noisy headlines around positions he intends to keep. Weiss is describing something that could last into 2028.

PG&E is the California utility. Lebenthal sold the entire position, and his reason is the cleanest sell discipline on the page: "what I wanted was a plain vanilla utility, and what I got is something that's a piñata in California politics."

A regulated utility is supposed to be a boring bond substitute — a monopoly with allowed returns set by a regulator. What he owns instead is a company that may be made liable by legislation for fires it did not cause, and which announced a strategic review that morning. Note the argument he does not make: he is not claiming the market is wrong or that the liability is unfair. He is saying the asset no longer resembles the one he bought, so the reason he owned it has gone.

The timing is the earlier method applied to himself. He said at the top of the hour that the right September move is not to sell the market but to cut the idiosyncratic holding you no longer believe in, and hold the cash as dry powder. This is that trade, and Wapner's "took the L, move on" is the appropriate attitude to it.

Las Vegas Sands is the Macau-weighted casino operator. Terranova confirms he sold out, and uses the name as the control experiment against Lebenthal's Wynn story.

The logic is worth borrowing whenever someone explains a stock's fall with one company-specific cause. Wynn's decline is blamed on its Gulf resort and the war. Las Vegas Sands has no Gulf exposure, and it looks just as bad — and both peaked back in November, before the war. When a comparable company without your explanation falls the same way, your explanation is probably not the driver. Lebenthal's counter is a real one, though: Macau is recovering, so Sands and Wynn are not perfect comparisons, and the shared problem is Las Vegas room pricing rather than either company.

Novo Nordisk is the Danish maker of the other leading obesity and diabetes drugs — Lilly's only real competitor in the GLP-1 category. It appears here purely as a contrast: Terranova asks for its chart on air to show Lilly's "significant outperformance."

No case was made for or against the company. But the comparison carries an argument: two companies selling into the same booming category have produced very different returns, which means the boom itself is not what the market is paying for. What it is paying for, on Terranova's account, is Lilly's reinvestment of the proceeds into a broader pipeline.