Dan Ives AI Buildout Still Early: Top Five Stock Plays

I keep a battered notebook next to the screen for days when a market call sounds too neat. Wednesday was one of those days. A veteran technology analyst told anyone still listening that the artificial-intelligence buildout is only in the third inning of a nine-inning game, and that the spending attached to it is closer to four trillion dollars than to a fad. I had to read that twice. Third inning. Not extra innings. Not a victory lap. If he is even half right, a lot of portfolios are still positioned for a story that has barely left the dugout.

The note that landed this week initiated coverage on nearly fifty technology names and landed on a stance the author called very bullish. His top five equity plays are familiar enough to make a skeptic yawn: Nvidia, Microsoft, Palantir, Apple, and CrowdStrike. The quieter part of the work, the part I keep circling, is the claim that demand is about to fan out rather than pile higher on the same handful of buyers. That is a different argument from “chips go up.” It is an argument about who writes the next checks.

Why a Third Inning Changes the Whole Conversation

Baseball metaphors get abused on trading desks. This one earns its keep if you sit with it. A third inning means the starter is still on the mound, the bullpen has barely stirred, and the scoreboard is not the final story. Investors who treat every strong quarter as late-cycle evidence tend to sell the names that still have the longest order books. I have watched that movie more than once. It rarely ages well when the underlying contracts are multi-year and the customers are still learning what the tools can do.

The published view is blunt. Investors are still underestimating the scale and scope of that spending wave over the next few years, even with safety worries, regulatory noise from Washington, and the usual geopolitical jitters. Second, third, and fourth derivatives of the shift, the note argues, are only starting to show up across the technology landscape. That sentence is easy to skim. It is also the whole thesis. Hardware is the first derivative. Software that sits on the hardware is the second. Security, workflow redesign, and last-mile automation are further out on the same chain.

The scale of what is now committed is the evidence, not the press release.

A framing I keep hearing from long-only technology desks

Committed capital is a stubborn kind of proof. A pilot can be canceled over lunch. A data-center lease, a power interconnect, a multi-year cloud obligation, those stick. When commercial remaining performance obligations at one of the largest software platforms reached 678 billion dollars, and when growth excluding a frontier-model partner still ran at 25 percent, the backlog stopped looking like a single-customer story. Sequential growth, the analyst noted, came from customers outside the frontier-model companies. Breadth is the tell.

What “Underestimating the Wave” Actually Means

People hear four trillion and either shrug or reach for a chart of past bubbles. Fair. Big numbers have a way of numbing the room. The useful question is narrower. Who is still not in the order book? Sovereign programs that want their own stacks. Neoclouds that rent capacity the hyperscalers will not spare. Ordinary enterprises that finally have a budget line with the words artificial intelligence on it. Model developers who are not the two or three names everyone already owns. If demand is broadening past the largest cloud buyers, the third inning is not a slogan. It is a customer list getting longer.

I have found that the cleanest way to test a bull case is to write down what would falsify it. A stall in power availability. A sudden cut in enterprise software renewals. A regulatory shock that freezes cross-border chip shipments for longer than a quarter. None of those are science fiction. They are also not the base case in this coverage launch. The base case is that the build keeps spreading while investors keep pricing it as if the easy money already left.

The Five Names, Without the Cheerleading

Outperform ratings sit on the equipment builders and on the software names the desk wants closest to the spend. Nvidia carries a 300 dollar price target, about 25 percent above Tuesday’s close. Microsoft is marked at 615, roughly 16 percent of upside. Palantir is at 250, implying something near 30 percent. CrowdStrike is at 335, about 20 percent. Apple rounds out the five. Those are targets, not promises. Targets are opinions with a number stapled on. Still, the spread tells you where the conviction sits.

Notice what is missing from that table. There is no claim that these are the only stocks that can work. The coverage net is wide, close to fifty names. The five are the ones he wants in the window. Lesser-known ideas sit further down the page with larger percentage gaps to their targets. That is usually where the argument gets more interesting, and more fragile.

Nvidia and the Buyers Who Are Not Hyperscalers

For years the chip story collapsed into a single sentence. A few cloud giants buy almost everything. That sentence is no longer the whole picture, at least not in this note. Demand, the analyst wrote, is broadening beyond the largest hyperscalers as sovereign programs, neoclouds, enterprises, and model developers build dedicated infrastructure. Read that list slowly. Four buyer types. Only one of them is the group everyone already models.

Sovereign programs are the awkward cousin at the table. Governments do not buy like software startups. They buy late, they buy in bulk, and they care about where the boxes sit. A national AI stack is not a vibe. It is land, power, export licenses, and a political calendar. If even a slice of that demand is real, the order book stops being a pure function of three American cloud budgets. I am not sure markets have priced that shift with any precision. They have priced the brand.

Neoclouds are the other wrinkle. These are operators who exist because the biggest platforms cannot, or will not, hand capacity to every customer who asks. They are a pressure valve. They are also a new set of purchase orders for the same accelerators. Enterprises sit behind them, slower, pickier, and suddenly willing to fund pilots that last longer than a quarter. Model developers who are not household names still need clusters. Put those flows together and the 300 dollar target is less a cheer and more a bet on a wider funnel.

• Hyperscalers remain the anchor buyers, not the only buyers.
• Sovereign stacks add a political buyer with multi-year horizons.
• Neoclouds absorb overflow demand the giants will not serve.
• Enterprises convert experiments into line items, slowly.
• Smaller model shops still need dedicated iron.

None of that erases cyclical risk. Accelerator generations turn over. Customers pause to digest. A single supply hiccup can swing a quarter. The third-inning claim simply says the pause, if it comes, is a pause inside a longer build, not the end of the build. Perhaps the most interesting aspect is how little of that distinction shows up in daily commentary, which still treats every guidance tweak as a verdict on the entire decade.

Microsoft and a Backlog That Is Getting Wider

Microsoft’s role in this framework is distribution plus obligation. Backlogs and outstanding contractual commitments are growing, and the customer base under those commitments is widening. Commercial remaining performance obligations at 678 billion dollars is a number large enough to make eyes glaze. The detail that matters is the exclusion. Remaining performance obligations grew 25 percent even after stripping out the best-known frontier-model relationship, and all of the sequential growth came from customers outside that circle.

That is a breadth comment dressed up as an accounting line. If the only engine were one lab writing enormous checks, the story would be concentrated and brittle. A backlog that expands without that engine is a different animal. It suggests ordinary commercial customers are signing longer deals. It suggests the platform is collecting obligations from buyers who do not make headlines. I have sat through enough software earnings calls to know how rare that sentence is. Most “AI revenue” slides are still a mix of reclassified seats and hope.

The 615 dollar target implies a mid-teens climb, not a moonshot. That restraint is part of why the call is usable. A modest upside on a mega-cap is a statement about durability, not about a hidden double. You do not need the stock to re-rate like a small software name. You need the obligation book to convert, and you need the conversion to survive a normal enterprise budget argument. Those are slower wins. They also tend to be the ones that compound.

Palantir, Apple, and Two Different Kinds of Reach

Palantir at 250 dollars, with implied upside near 30 percent, is the software bet that the data layer becomes the place work actually happens. The company has spent years selling the idea that messy operational data can be queried, governed, and acted on. Agentic tools make that pitch louder. If little pieces of software start taking initiative inside a business, someone has to decide which data they may touch and which actions they may attempt. That is not a chatbot feature. That is an operating system argument, and it is why a security-minded buyer and a data-platform buyer can end up in the same meeting.

Apple sits in the five for a reason that is easy to underweight. Consumer devices are still where most people will meet these tools, whether or not they ever open an enterprise console. On-device inference, tighter privacy defaults, and a distribution pipe that already lives in pockets change the shape of demand at the edge. I do not think every handset cycle is an AI cycle. I do think a platform with that much installed hardware cannot be left out of a buildout that is supposed to broaden. The note treats it as a core play, not a footnote.

Put the two next to each other and the portfolio logic gets clearer. One name lives inside government and corporate workflows. The other lives in the consumer graph. A third-inning build that only rewards data-center landlords would be a narrow build. A build that also pays the workflow layer and the device layer is the version this coverage is underwriting.

CrowdStrike and the Attack Surface Nobody Budgeted For

Here is the line I underlined. Agentic software, those small programs that act as users and carry out tasks, has expanded the attack surface faster than budgets. Frontier labs have moved inside the security stack. The expectation in the note is that security spending lifts from roughly 5 percent of information-technology budgets toward 10 percent over the next three years. Double the share. Not double the poetry. A budget share.

CrowdStrike’s 335 dollar target, about 20 percent of upside, sits on that claim. The logic is uncomfortable and, to my ear, coherent. Frontier model companies are increasingly positioning themselves as security vendors even as high-profile breaches keep landing in the news. That encroachment threatens older franchises. It also shoves more money toward the category, because every new agent is another identity, another permission, another log someone has to trust. The analyst’s phrase for it, offered in a midweek conversation, was close to a golden age for cybersecurity companies. Strong language. The mechanism underneath it is plain.

Agents do not just answer questions. They take actions. Actions need permissions. Permissions are where breaches begin.

Think about a clerk who can read a file. Now think about a software clerk that can read the file, draft the reply, call an internal tool, and schedule the follow-up without a human in the loop. The second clerk is useful. The second clerk is also a new employee you did not interview. Security budgets that were sized for human mistakes are being asked to cover machine initiative. That gap is the 5-to-10 percent argument in street clothes.

Cybersecurity stocks, in this view, are still early in a heavy growth cycle. Early is a dangerous word. It has excused a lot of expensive stock. Used narrowly, it means the budget share has not moved yet, so the revenue is not fully in the models. If the share does move, the category stops being a steady compounder and starts being a reallocation story inside every chief information officer’s spreadsheet. I would rather underwrite a budget-share shift than a vague “AI tailwind.” Budget shares are things finance teams argue about in November.

The Quieter Calls: Equipment Conversion and Sidewalk Robots

The five headliners will get the clicks. The note also puts a 24 dollar target on cloud equipment provider Boost Run, suggesting upside around 58 percent, and an 8 dollar target on Serve Robotics, implying something near 67 percent. Larger gaps. Thinner crowds. That combination is either where research earns its fee or where optimism outruns evidence. Both can be true in the same paragraph.

On Boost Run, the write-up treats second-quarter 2026 results as the first meaningful sign that contracted demand is turning into reported scale. Contracted demand is a promise. Reported scale is a print. The distance between those two is where equipment stories go to die or to re-rate. If the conversion is real, a 58 percent gap is the market still treating the company like a hope rather than a supplier. If the conversion stalls, the target is a souvenir. I have no special edge on that print. I do know the question to watch is conversion, not commentary.

Serve Robotics is framed as an early leader in autonomous last-mile delivery, the sidewalk kind, not the highway kind. Sidewalk robots are easy to mock until you price a human courier in a dense city and then price the robot that replaces the third trip of the afternoon. Last-mile is a brutal cost line for anyone who promises things will arrive today. An early leader is not a monopolist. It is a company that has permits, routes, and a machine that does not quit at the curb. The 67 percent upside is a statement about how little of that option is in the price, not a statement that sidewalks are already solved.

How the note ranks conviction, in plain language:
  Core five: Nvidia, Microsoft, Palantir, Apple, CrowdStrike
  Wider net: nearly fifty initiated names
  Higher-upside side bets: Boost Run, Serve Robotics
  Shared claim: spending broadens before it peaks

Agentic Tools, and Why the Second Derivative Matters

Agentic AI is the phrase that will get overused until it means nothing. Strip the jargon and it is simple. Software that does not wait for a click. It plans a step, takes the step, and reports back. For a consumer, that might be a trip booked while you are in another app. For a business, it might be an invoice matched, a ticket escalated, a security alert closed. Each of those steps consumes compute. Each of them also creates a log, an identity, and a failure mode.

That is why the note ties agents to both chip demand and security budgets. The first derivative is the cluster that trains and serves the model. The second is the software seat that lets an employee delegate. The third is the control layer that decides the delegate cannot wire money to a stranger. Investors who only own the first derivative are betting the build stays concentrated in hardware. Investors who own the later derivatives are betting the build changes how work is staffed. Those are related bets. They are not the same bet.

In my experience, the market pays up for the first derivative early and argues about the later ones for years. That lag is the third-inning setup in another costume. If agents really do take initiative, security does not stay a 5 percent line item out of habit. Habit is what gets rewritten when a breach is traced to a permission nobody reviewed. The frontier labs moving into the security stack accelerates the argument. Incumbents can lose share and still grow if the pool doubles. Both outcomes can show up in the same sector chart, which is why single-stock stories inside cybersecurity will diverge even if the category call is right.

Safety, Washington, and the Jitters Already in the Price

The bull note does not pretend the objections are fake. Safety concerns, regulatory calls, geopolitical jitters: they are listed in the same breath as the four-trillion claim. Listing them is not the same as discounting them. A third-inning investor still has to live with export rules, with hearings, with a model that says something it should not say in a hospital system. Those are not rounding errors. They are reasons a multiple can compress even while revenue compounds.

What the coverage rejects is the idea that those risks have already ended the spending cycle. Commitment, in this telling, is the counterweight. You can regulate a product and still pour concrete for the building that hosts it. You can argue about safety standards and still sign a three-year cloud obligation because the alternative is falling behind a rival who signed last spring. I find that tension more honest than a note that waves the risks away. The risks are the cost of admission. The order book is the reason to stay in the room.

1. Separate a product scare from a capex cancellation. They are not twins.
2. Watch backlog breadth, not just the headline total.
3. Track whether new buyer types show up in supplier commentary.
4. Treat security budget share as a measurable claim, not a slogan.
5. Size smaller names as options on conversion, not as core ballast.

How a Careful Reader Might Use the Map

None of this is a shopping list. Price targets age. Tuesday’s close is already a historical object by the time a blog post circulates. The usable piece is the map. Equipment at the center. A cloud platform whose obligations are widening. A data-software name. A device platform. A security franchise levered to a bigger attack surface. Then two higher-octane ideas tied to conversion and to physical automation.

A portfolio that mirrors that map does not need equal weights. It needs a reason for each sleeve. The equipment sleeve is a bet that buyer types keep multiplying. The backlog sleeve is a bet that obligations convert without a single-customer crutch. The security sleeve is a bet that budget share moves from 5 toward 10. The sidewalk-robot sleeve is a bet that last-mile cost curves bend in cities that will tolerate a machine on the pavement. If you cannot say the bet in one sentence, you are renting someone else’s enthusiasm.

There is a temperament test buried in the third-inning line. Late-inning investors want confirmation and then an exit. Early-inning investors tolerate messy quarters because the customer list is still forming. Most people say they are the second type and trade like the first. I have done it. The notebook helps. So does writing down, before the next print, what would actually change the inning count.

Where the Analogy Breaks

Baseball games end. Capital cycles sprawl. A nine-inning frame can lull you into thinking there is a scheduled seventh-inning stretch and a knowable final out. There is not. Power constraints can skip an inning. A breakthrough in efficiency can make yesterday’s cluster look oversized. A buyer strike among enterprises can turn “broadening demand” into a slide title that did not survive contact with procurement. The analogy is a pacing tool, not a calendar.

Still, pacing matters. If you believe the committed spend is early, you fade panics that are really digestion. If you believe the spend is late, you sell strength and call it discipline. Those two investors can look at the same 678 billion dollar obligation figure and walk out of the room with opposite trades. The figure does not settle the argument. The composition of the figure might. Growth that survives the removal of one celebrity customer is composition. Growth that vanishes without that customer is concentration wearing a backlog costume.

A simple filter: backlog growth excluding the largest related party, plus a new buyer type in supplier commentary, plus security mentioned as a budget share rather than a feature. Three checks. None of them require a slogan.

What I Will Be Watching Next

First, whether supplier calls keep naming sovereign programs and neoclouds as more than color. A single mention is marketing. A repeating mention with a dollar figure is a channel. Second, whether commercial obligation growth at the large platforms keeps arriving from outside the frontier-model circle. Third, whether security vendors start talking about agent permissions as a product, not as a threat slide. Fourth, whether contracted equipment demand at the smaller supplier actually shows up as reported scale, the way the Boost Run note hopes the latest quarter began to show. Fifth, whether sidewalk autonomy remains a pilot or starts to look like a route.

Those five watches line up, roughly, with the five headliners and the two side bets. They are also boring on purpose. Boring checks are how you keep a vivid thesis from becoming a poster. The poster version is “AI is early, buy the leaders.” The working version is a set of customer types, a budget share, and a conversion from contract to print. If those hold, the third-inning language will look conservative in hindsight. If they crack, it will look like a metaphor that outlived the data.

I do not need the entire four trillion to arrive on schedule to take the map seriously. I need enough of it to arrive from buyers who were not in the 2023 slides. That is a lower bar than the headline, and a harder one to fake. Breadth is difficult to counterfeit for long. A single whale can paint a quarter. A widening list of ordinary customers is how a buildout stops being a trade and starts being an economy.

A Last Pass Through the Risks Worth Keeping

Concentration has not vanished just because a note says demand is broadening. The largest buyers still dominate shipments. A pause from one of them can overwhelm a year of sovereign press releases. Valuation on the leaders already assumes a long runway, which means “early” can be true and the stock can still be expensive. Those two facts live together more often than commentators admit.

Security is not a free call option on fear. If frontier labs really do plant themselves inside the stack, incumbents can lose the very budgets the category call is counting on. A rising pool with a shifting share is a stock-picker’s market, not an index hug. Apple’s place in the five depends on devices mattering to inference and distribution. If the useful work stays in distant clusters, the handset becomes a remote control and the thesis thins. Palantir’s upside assumes enterprises will pay for a governed data layer rather than stitch together cheaper tools. That assumption has been debated for a decade. Agents raise the stakes. They do not end the debate.

Smaller names carry a different failure mode. A 58 percent gap and a 67 percent gap are not gifts. They are the market’s way of saying the evidence is incomplete. Treating them as core holdings because a well-known analyst published a target is how research gets misused. Treating them as tracked claims, with a kill-switch tied to conversion and route density, is how the same research becomes useful. I prefer the second habit. It is less fun to describe at dinner.

Pulling the Thread Without Forcing a Moral

So where does that leave a reader who does not want a slogan? With a pacing claim and a customer claim. The pacing claim says committed spending is closer to the third inning than to the ninth, and that safety debates and political noise have not canceled the contracts. The customer claim says the next checks come from sovereign programs, neoclouds, enterprises, and model developers, not only from the original cloud whales. Nvidia is the equipment expression of that claim. Microsoft is the obligation expression. Palantir and Apple are two kinds of reach. CrowdStrike is the bill for a wider attack surface. Boost Run and Serve Robotics are the higher-variance expressions, one in contracted equipment, one on the sidewalk.

You can disagree with every target and still steal the checklist. Who is buying. Whether the backlog grows without its celebrity account. Whether security is gaining share of the information-technology wallet. Whether contracted demand becomes a print. Whether a small robot is earning routes or earning pilots. Those questions will outlast the Wednesday note. They are also, if I am honest, more interesting than another argument about whether the trade is “over.” Over is a feeling. Innings are a way of admitting you do not know the feeling yet, and of forcing yourself to name the evidence that would change it.

I closed the notebook with the five names written in the margin and the two quieter ones underneath, smaller. Not as a dare. As a reminder that the buildout, if the breadth shows up, will not look like the last two years repeated. It will look like more buyers, more permissions to police, and a few physical workflows that finally get cheap enough to automate. That is a less cinematic ending than a bubble poster. It is also the version a third-inning scoreboard is actually describing.