Broadcom AI Revenue Outlook Why A Hold Still Works

Have you ever watched a company print numbers that look excellent on paper and still felt the room go quiet? That is the mood around Broadcom right now. The latest quarter was not a miss. Revenue jumped. Adjusted earnings jumped harder. Management even sketched a multiyear path for Broadcom AI revenue that would have sounded reckless two years ago. And yet the stock still treated the print like a lukewarm cup of coffee. I keep coming back to one simple question: when did “good” stop being enough for custom silicon?

What The Quarter Actually Delivered

Fiscal third quarter revenue, for the period ended in early August, climbed about 85 percent year over year to roughly $29.59 billion. That edged past a consensus figure near $29.36 billion. Adjusted earnings per share rose about 96 percent to $3.32, against expectations around $3.24. On a raw scoreboard, that is a beat. Not a blowout in the old-school sense, but a beat all the same.

Shares still wobbled in extended trading. At one point the move looked worse, down as much as 7 percent, before settling into a more modest decline. The year-to-date tape has been equally unsatisfying for anyone who bought the AI story at full volume. A gain of only about 6 percent for the year is not what most holders imagined when custom accelerators became the market’s favorite plot twist.

I’ve found that semiconductor stories now live or die on the next three years, not the last ninety days. The quarter is a receipt. The call is the movie trailer. Investors showed up for the trailer.

The AI Line That Matters More Than The Beat

Inside Semiconductor Solutions, the segment Wall Street actually obsesses over, revenue rose about 70 percent to $20.84 billion and cleared a consensus mark near $20.29 billion. The AI slice inside that machine is the real headline. AI semiconductor revenue surged about 221 percent year over year and about 51 percent sequentially to $16.7 billion. That landed above the $16 billion the company had pointed to a quarter earlier.

That AI bucket is not only custom silicon. It also includes networking gear that moves data between specialized processors and keeps clusters from turning into expensive space heaters. Ethernet switches do not get the same applause as a new accelerator name, but they are the plumbing. Without plumbing, the mansion floods.

Infrastructure Software, the other operating piece, rose about 29 percent to $8.75 billion and slipped just under a consensus figure near $8.82 billion. The call did not linger there. The brief comment that did land was about a private-cloud AI stack that lets enterprises run models next to existing applications without turning the whole campus into a public-cloud science project. Useful. Not the reason anyone stayed up for the print.

A Multiyear Map That Still Left People Hungry

Based on fiscal fourth-quarter AI revenue guidance of $21.7 billion, management now sees about $58 billion of AI revenue for fiscal 2026. That is a step up from a prior view near $56 billion. Fine. Necessary. Not enough to reset the mood on its own.

The market has been staring at fiscal 2027 for months. Last quarter, a simple reiteration of “at least $100 billion” in AI revenue for that year helped send the stock lower. This time the figure moved to $115 billion, and management framed it as supply constrained. On first listen, that raise felt small. A lot of people had mentally penciled in something closer to a $20 billion to $25 billion lift. You could almost hear the collective “that’s it?”

Then came fiscal 2028. Management said AI revenue could double again, to about $230 billion. That number sat well above a consensus view near $177 billion. That is the line that finally put some air back in the tape after hours. It is also the line that makes the whole debate more interesting, because it forces a grown-up conversation about land, power, shells, and memory, not just wafer starts.

The longer-term outlook is not a wish list. It is a schedule that already tries to price in chips, buildings, and megawatts.

On profitability, leadership said the company looks on track to exceed $30 of adjusted earnings per share in fiscal 2028. Consensus had been nearer $26.38. If you take that $30 figure at face value, the stock starts to look cheap on a simple multiple, somewhere around 12 times those earnings depending on where shares sit when you do the math. Cheap is not the same as loved. The market is clearly refusing to pay up for those earnings the way it once did for anything with an AI sticker on the box.

Customers, Gigawatts, And The Power Ceiling

The forecast math leans on a tight circle of frontier buyers. Alphabet remains the historic anchor through tensor processing units that Broadcom co-designs. That relationship has been poked and prodded for months after a rival design win elsewhere and chatter about share pressure. Management’s defense was blunt. The pitch was intellectual property, chip-to-chip interconnect, leading-edge high-bandwidth memory and static memory integration, plus packaging that actually ships on time.

Time to market from product definition to production is the unglamorous advantage that keeps a custom program alive when everyone else is still arguing over a slide deck.

Beyond that first customer, Anthropic is described as the likely largest buyer of specialized processors over the next two years, tied to an expected 5 gigawatts of deployments and a line of sight toward another 10 gigawatts. Computing gets measured in gigawatts now because electricity, not press releases, is the bottleneck. That still sounds slightly unreal when you say it out loud. It is also the most honest unit in the industry.

OpenAI is framed as the second-largest specialized-chip customer on this path. The near-term plan is about 1.3 gigawatts next year, consistent with earlier comments, with a longer line of sight toward 5 gigawatts of a new processor family and later generations. Meta sits on a separate track: three generations of its training and inference accelerator, with visibility to about 3 gigawatts through 2028. That Meta figure did not change. Stability can be boring. In this tape, boring is almost a relief.

Perhaps the most interesting aspect is how energy has become the real product roadmap. You can design a gorgeous die and still lose if the building has no power feed. I keep picturing a Formula One car that wins every qualifying session and then sits in the pit because the fuel truck is late. That is data-center AI in 2026.

The Financing Vehicle Nobody Wanted To Romanticize

To help two of those labs fund deployments, Broadcom set up a special purpose vehicle with large alternative-asset partners. The idea is to help stand up more than 20 gigawatts of compute infrastructure by the end of 2028. In plain English, the labs need steel and silicon faster than their current cash flow can comfortably write checks. The vehicle is a bridge. A similar style of arrangement has already shown up elsewhere in the AI supply chain with major banks in the mix.

This is where the after-hours cheer started to snag. Circular financing is an ugly phrase. It may be too crude for what is actually happening. It is also the phrase that will keep showing up in every skeptical note. Customer concentration among a handful of frontier labs sits right next to it. Add political pushback against new data-center sites, and you have a triple that can compress multiples even while revenue estimates grind higher.

In my experience, markets punish uncertainty more than they punish a slightly smaller raise. A $115 billion 2027 figure that is “constrained by supply” should, in theory, be bullish. It says demand is not the problem. The problem is whether the physical world can keep up. That is a harder story to underwrite with a rich multiple.

Fourth-Quarter Guidance Had A Split Personality

Total revenue for the current fiscal fourth quarter was guided to about $34.8 billion. That sat a touch under a consensus figure near $35.03 billion. The shortfall, if it is real, may live in software. AI semiconductor revenue is expected to accelerate to $21.7 billion. That would beat a consensus mark near $20.93 billion and imply growth on the order of 236 percent year over year. So the thing the market cares about is still speeding up. The thing the market shrugs at is the one that missed by a rounding error.

Adjusted operating margin for the quarter is expected around 66 percent, pointing to roughly $22.97 billion of operating income. Consensus had been closer to 66.5 percent and about $23.05 billion. Higher memory content was cited as a reason margins are not stretching further. That detail is easy to skip. It should not be skipped. Memory is not a footnote anymore. It is eating bill of materials, rack power, and patience.

Is a 66 percent margin supposed to disappoint anyone? In most industries that number would be a parade. In this one, investors have been trained to ask why it is not 67. That is the cost of being treated like a scarce AI asset. Every basis point becomes a character in the drama.

Why The Multiple Compressed Anyway

Using that longer-dated earnings target, the shares can look inexpensive. The catch is that inexpensive semiconductor stocks are often inexpensive for a reason the tape already knows. Group multiples across the chip complex have been sliding. Political noise around data-center buildouts has grown louder. Financing structures that help labs spend today and pay across a longer curve make credit desks nervous even when the engineering is sound.

I cut a mental position in half last week for exactly that last reason. A smaller sleeve makes it easier to wait. That is not a victory lap. It is an admission that patience has a price, and the price is size. A full-size holding in a name the market no longer wants to re-rate can wreck your week even if the 2028 math is honest.

A lower price target, from a prior $480 down to $430 in this framework, is not a verdict that the business got worse. It is a verdict that the multiple the market will pay for semiconductor earnings got worse. Those are different sentences. People mash them together because it is faster. Faster is how you blow a thesis.

• Demand visibility improved at the outer years, which is rare and valuable.
• Supply and infrastructure remain the binding constraints, which is less rare and less fun.
• Customer concentration is still high enough to keep nights interesting.
• Financing vehicles help the build, and they also feed the circular-money debate.
• Software is steady enough, but it will not rescue a multiple if AI sentiment slips.

Custom Silicon Is A Relationship Business Wearing A Lab Coat

People talk about custom chips as if they were catalog parts. They are not. A tensor unit or a specialized accelerator is a multiyear marriage between a model company and a design house. You argue over floorplans. You miss a tape-out. You sweat packaging. You pray the interconnect does not become the punchline. Then you do it again for the next generation because the model architecture moved under your feet.

That is why the Google defense on the call mattered more than a single quarter’s sequential pop. If the partnership holds because the design library, the memory integration, and the packaging cadence are genuinely hard to copy, the 2028 number has a spine. If those advantages turn out to be polite marketing, the spine turns into a question mark. I do not think it is only marketing. I also do not think any custom program is immortal.

Networking sits in the same relationship logic. Clusters fail in the spaces between chips. Switches, optics, and the unsexy work of making thousands of accelerators behave like one computer are where a lot of the durable margin can hide. When AI revenue is described as custom silicon plus networking, that second word is doing more work than the slideshow implies.

The Software Side Nobody Came To Hear About

It is easy to treat the software segment as leftover furniture. That would be sloppy. A private-cloud AI layer that lets a bank or a hospital keep sensitive workloads next to the systems they already run is not a moonshot. It is a door. Enterprises are late to frontier training runs. They are not late to the fear of putting everything in someone else’s building.

The slight consensus miss in software revenue will fade by next Tuesday. The strategic question will not. Can that stack become a second engine if custom-silicon growth ever cools from triple-digit fireworks to merely excellent? I would not build a whole thesis on it. I also would not donate it to the footnote graveyard.

Margins in a mixed hardware-and-software house always tell on themselves. Memory-heavy AI boards pull goods costs one way. Software mix pulls them another. The 66 percent operating-margin guide is the compromise those two forces just signed in public.

How To Sit With A Hold Without Fooling Yourself

A hold is not a hug. It is a pause that still requires work. If you keep the shares, you are betting that outer-year revenue and that $30-plus earnings marker are more durable than the current multiple implies. You are also accepting that the market may keep haircutting chip stocks until power permits, local politics, and financing optics look cleaner.

1. Separate the quarter from the decade. The beat was real. The debate is 2027 and 2028.
2. Track gigawatts the way you used to track units. Power is the scarce SKU.
3. Watch customer mix, not just aggregate AI sales. One delayed campus can move a year.
4. Treat financing structures as both a catalyst and a headline risk.
5. Resize the position so a stubborn multiple does not force a bad sale.

That last point is the one I keep repeating to myself. Conviction without sizing is just a speech. A half-sized book can wait for the physical world to catch the forecast. A full-sized book starts negotiating with itself after two ugly sessions.

What Would Make The Story Break

Every hold needs a kill switch. Mine would start with a visible crack in the Google program that management cannot explain with packaging or interconnect talk. It would include a slip in the Anthropic or OpenAI deployment calendar that is about demand, not transformers sitting on a dock. It would include political delays that turn “line of sight” gigawatts into “maybe in the next mayor’s term.”

A second break would be memory and advanced packaging staying tight enough that the 2028 double starts to look like a slide, not a plan. Supply constraint can be a badge. It can also be a ceiling that never lifts. Those two readings use the same words and lead to opposite portfolios.

A third break, quieter and nastier, would be the market deciding that all frontier-lab capex is one trade. If one lab stumbles, the multiple on every supplier can gap down together. That is not fair to a company with a broader networking and software mix. Markets are not hired to be fair.

What Would Make The Multiple Heal

Healing would look ordinary, which is why it is hard to pitch. Permits approved. Substations energized. HBM allocations that do not kidnap the bill of materials. A second or third custom program that is not just a press photo. Software that stops missing by a hair. A financing vehicle that funds steel without looking like a loop of IOUs.

It would also help if the rest of the semiconductor group stopped leaking valuation. Broadcom does not trade on an island. When the group’s price-to-earnings gravity drops, even a $230 billion AI year gets a colder multiple. That is the unsexy part of this whole episode. The company raised the destination. The market lowered the camera’s zoom.

I’ve found that the healthiest way to read a print like this is to ask whether management added information or just added adjectives. This call added information. Gigawatts. Customer ranking. A 2028 earnings marker. A margin explanation tied to memory. That is a real update. The stock’s shrug says investors wanted a bigger 2027 raise and a cleaner political backdrop. Wanting is not analysis. It is still how tapes move.

A Longer View On Custom Accelerators

Zoom out and the industry pattern is almost simple. General-purpose graphics processors proved the market. Custom silicon is trying to harvest the part of that market that can tolerate a longer design cycle in exchange for better performance per watt. Networking vendors are trying to make sure those custom islands can still talk. Memory vendors are collecting rent at every stop. Utilities and landowners suddenly sit in the same meeting as chip architects. That last sentence would have sounded like satire in 2021. It is the meeting agenda now.

Broadcom’s edge, if it holds, is the ugly middle of that stack: design services that ship, packaging that does not slip a generation, and interconnect that keeps a rack from becoming a latency museum. Those are not poster features. They are why a lab comes back for generation two and generation three. Recurring custom programs are the whole game. A one-off win is a press cycle. A three-generation win is a franchise.

That is also why Meta’s unchanged 3 gigawatt line through 2028 is not a yawn. Unchanged, in a year when every number is supposed to leap, can mean the program is simply on rails. Rails are good. Rails are how you sleep.

Valuation, Temperament, And The Temptation To Overtrade

A cheap multiple on 2028 earnings is only a gift if 2028 arrives looking like the slide. Plenty of “cheap” chip stocks have been cheap all the way through a cycle because the cycle moved. The temptation after a soft after-hours reaction is to trade the noise. Sell the dip that already happened. Buy the bounce that may last an hour. That is how a multiyear compounder becomes a week of fees and regret.

Temperament matters more than a new target price. If a 7 percent air pocket makes you rewrite the entire custom-silicon thesis, the position was too large. If a $15 billion raise in 2027 feels like an insult after a 221 percent AI growth print, you may be grading on a curve that no physical supply chain can hit every quarter.

None of this is a plea to ignore risk. It is a plea to name the risk correctly. The risk is not that this quarter failed. It did not. The risk is that the world of substations, local hearings, and high-bandwidth memory cannot sprint as fast as a conference call.

Hold checklist in one glance:
  Business beat: yes
  AI run-rate still rising: yes
  Outer-year guide higher: yes
  Multiple expanding: not yet
  Position size: smaller on purpose

The Human Read On A Very Technical Print

There is a human rhythm to these nights. First the headline numbers. Then the guide. Then the fight over whether the guide was “enough.” Enough compared with what? Enough compared with last quarter’s disappointment. Enough compared with a rival’s mythology. Enough compared with a chat thread that decided $20 billion was the minimum respectable raise. That is a strange way to allocate capital, and it is the way this market currently works.

I still come away mixed, and I think mixed is the honest word. Visibility improved. Outer-year targets moved up. The earnings marker for 2028 is ambitious in a useful way. The market will not pay a richer multiple until it trusts the physical build and the financing optics. Those two sentences can live in the same portfolio. They just cannot live in the same tweet.

So yes, I am willing to hold a smaller sleeve. Not because the after-hours tape was kind. It was not. Because the operating story still has a spine, the AI revenue path still points up, and the cheap-looking multiple is the market’s way of charging rent for uncertainty. Rent is annoying. Sometimes you still pay it if the building is the one you wanted.

Will the next print need another leap in 2027 to keep the peace? Probably. That is the new contract. Beat the quarter, raise the decade, and then raise the decade again. It is exhausting. It is also the job, at least until the power grid, the memory market, and the zoning board say otherwise.

If you came looking for fireworks, you got a flare and a map. Maps are less fun. They are how you avoid driving off the cliff while everyone else is still cheering the flare. That is enough reason, at least for now, to stay in the seat without stomping the accelerator.