I kept staring at the same chart last week and could not shake the feeling that something was off. A company sitting at the center of the biggest computing buildout in a generation had quietly become the laggard of its own industry. Not a broken story. Not a fading franchise. A leader whose share price had simply stopped keeping pace with the business underneath it. That kind of gap is rare, and it is usually where the interesting decisions live.
People love a clean narrative. Hot theme, hot stock, easy headline. What they often miss is the quieter setup: a best-in-breed operator whose multiple has compressed while the order book has not. I started paying closer attention when the shares were still hovering near the $200 area back in July. I liked the asymmetry then. I like it more now, even with the price sitting closer to a 52-week high. Price and value are not the same conversation, and this one is a reminder of that.
A Leader On Sale While Peers Run Ahead
The disconnect is the whole story, at least at the start. Year to date, the shares have trailed the broad semiconductor index by roughly 55 percent. Stretch the window to a year and the gap widens toward 75 percent. That is not a rounding error. That is a market voting, loudly, that something else in the chip complex deserves the premium. Memory names, equipment makers, smaller design houses. Plenty of them have had a spectacular run. The company that still designs the accelerators almost everyone else is trying to catch has not.
I have found that relative underperformance is easy to misread. Sometimes it is a warning. Sometimes it is a gift wrapped in boredom. The difference usually shows up in the operating numbers, not in the price tape. Here the operating numbers have not rolled over. Revenue in the latest quarter was up about 106 percent from a year earlier. Management is still talking about something close to 70 percent revenue growth in fiscal 2028. Demand, by the company’s own account, is running ahead of supply through 2028. If that is a deteriorating franchise, I would like to see what a healthy one looks like.
Perhaps the most interesting aspect is how ordinary the share price action has started to feel. Up on the year, yes. Dominant on a multi-year view, obviously. But relative to the rest of the semiconductor complex, the stock has been the one left in the hallway while the party moved down the hall. Markets do this. They rotate. They get bored of the name everyone already owns. They chase the next supply bottleneck. None of that automatically means the original bottleneck has been solved.
What Best In Breed Actually Means Here
The phrase gets thrown around until it means nothing. In this case it still means something concrete. In server graphics processors, market share sits near 97 percent. That is not a lead. That is a near-monopoly in the product that turns electricity and capital expenditure into trained models. Customers are not buying a logo. They are buying a stack: the chip, the interconnect, the software layer that has become the default way serious teams build and run these systems.
Software is the part casual observers skip. The proprietary platform has settled into the role of an industry standard, the way a programming environment sometimes becomes the room everyone works in even when rival rooms exist. Switching costs are not theoretical. Retraining teams, rewriting pipelines, requalifying workloads, and accepting a performance hit while you do it is a real invoice. I have watched plenty of “good enough” challengers fail to dent that kind of habit. Good enough rarely wins when the workload is the most expensive thing a company runs.
A dominant share with a software habit attached is a different animal from a dominant share that can be swapped out on the next purchase order.
Newer platforms extend the same logic rather than reset it. Blackwell is already in the field. Vera Rubin is the next architecture on the roadmap, not a science project. The company is also pushing further into central processors, which matters because the modern AI server is no longer a single-chip story. It is a system. Owning more of the system is how a leader stops being a component vendor and starts looking like infrastructure.
Why The Lag Does Not Match The Business
Underperformance can come from three places. The business slows. The multiple was absurd and mean-reverts. Or capital simply prefers a different pocket of the same theme. All three can be true at once. In my experience, the third one is doing a lot of the work right now.
Memory suppliers have been telling a familiar story on recent calls: demand still outruns available supply as AI infrastructure spending keeps climbing. That is not a contradiction of the accelerator thesis. It is a cousin of it. You cannot fill a data hall with memory and leave the compute sitting in a warehouse. The whole stack is tight. When one link in that stack rallies harder than another, the laggard is not automatically the broken link.
Gross margins have also settled into a zone that looks less fragile than the scare stories suggested a year ago. Around 72 percent is not a startup margin. It is a mature, high-value product margin, and it has been described by bulls as largely derisked at that level. I would not treat any margin as permanent. Competition, mix, and pricing all move. But a business that can hold margins in the low seventies while growing at a rate most industrial companies would not recognize is not a business the market should be pricing like a cyclical also-ran.
The Valuation Chart Tells A Different Story Than The Price Chart
Here is the part that keeps pulling me back. The shares trade near a 52-week high, which is the kind of sentence that makes people close the tab. Highs feel expensive. They are not, by themselves. Expensive is a relationship between price and earning power. On that measure the stock is sitting at about 16.7 times forward earnings. The average over both the past five years and the past ten years is roughly 35 times. That is the cheapest multiple in at least a decade.
Read that again slowly. A company still expected to grow revenue at a rate most of the index would trade their playbook for is priced at less than half its own long-run earnings multiple. Some of the compression is deserved. Multiples that lived in the stratosphere were pricing perfection, and perfection is a bad base case. Some of the compression is the market doing what it always does after a parabolic run: it demands proof, then demands more proof, then demands a discount while the proof arrives.
I see a path to roughly $22 a share in earnings by fiscal 2028. At the current price, that lands near 10.7 times those earnings. Ten times for a franchise with this share, this software grip, and this visibility on demand is not a number I associate with euphoria. It is a number I associate with a market that has already moved on and forgotten to update the spreadsheet.
| Lens | Rough figure | What it suggests |
| Forward earnings multiple | About 16.7 times | Cheapest in at least a decade |
| Five- and ten-year average multiple | Roughly 35 times | Historical premium has vanished |
| Fiscal 2028 earnings case | About $22 per share | Implies roughly 10.7 times |
| Latest revenue growth | About 106 percent | Business has not cooled with the stock |
| Fiscal 2028 revenue outlook | Near 70 percent | Growth still extraordinary |
| Server GPU share | Roughly 97 percent | Leadership remains intact |
Tables flatten nuance, so treat that grid as a map, not a verdict. The point is the spread. Price performance and business performance have diverged. Valuation and history have diverged. When two divergences point the same way, I pay attention.
A Buyback That Is No Longer A Footnote
On September 28 the board added $150 billion to the repurchase authorization. The remaining total now sits around $235 billion, expected to be executed through fiscal 2028. That is the largest increase in company history. Set against the market value, it is roughly 4 percent. Similar in scale, as a statement, to the then-record authorization a major consumer-electronics company put in place in 2024.
Buybacks are not magic. A company can light cash on fire buying shares at a silly price, and plenty have. A company can also use a repurchase program as a quiet way of saying the internal model and the external multiple are not having the same conversation. At 16 times forward earnings, with a balance sheet that can fund both growth and returns, the second reading feels more honest. Four percent of the company over a defined window will not rewrite the story by itself. It does shrink the share count while earnings are still climbing, which is how a multiple gets even cheaper if the price stands still.
I like that combination more than I like a buyback announced at 40 times earnings with a press release full of adjectives. Capital return at a compressed multiple is a different instrument. It is management voting with the corporate checkbook, not with a slogan.
- Authorization increased by a record $150 billion.
- Remaining capacity near $235 billion through fiscal 2028.
- Roughly 4 percent of current market value.
- Executed alongside, not instead of, heavy investment in the next platforms.
Growth That Still Does Not Look Ordinary
Extraordinary is an overused word. Here it is just arithmetic. Doubling revenue in a single year, at this scale, is not a small-base trick. Expecting something near 70 percent the following fiscal year is not a company guiding to a soft landing. It is a company telling you the buildout has not peaked. Supply, not demand, is the constraint they keep pointing at through 2028.
Constraints cut both ways. They cap near-term shipments, which caps near-term revenue, which is why some investors get impatient. They also imply a backlog of work that has not been delivered yet. I would rather own a constrained leader than an unconstrained follower with a glossy roadmap and an empty order book. The former has a visibility problem. The latter has a relevance problem.
Customers are still writing checks for systems that will define their cost structure for years. Cloud operators, model labs, enterprises that do not want to rent every inference forever. The mix will shift. Training was the first wave. Inference, agents, industry-specific models, and on-premises deployments are the next ones. A platform that already sits in the training cluster has a head start on the inference cluster. That is not guaranteed share. It is a shorter sales cycle.
The Stack Is Wider Than One Chip Generation
A common worry is that the current generation is the peak and everything after it is a price war. Possible. Also incomplete. The roadmap is not a single product. Blackwell is shipping. Vera Rubin is the following architecture. Networking and interconnect sit beside the processor. The software layer keeps developers inside the ecosystem. Central processors are a newer front, and a logical one, because the server is a system and systems favor vendors who can tune the pieces together.
Think of it like a rail network rather than a single locomotive. Rivals can build an engine. Matching the timetable, the stations, the signaling, and the crews is the slower job. Some will do pieces of it well. A few will win niches. Replacing the default network is a decade-shaped task, not a quarter-shaped one.
How the moat actually stacks up: Chips — current and next architectures Software — the default build-and-run layer Systems — interconnect plus a push into CPUs Supply — demand still ahead through 2028 Capital return — record authorization at a low multiple
None of those lines is immortal. Custom silicon from the largest buyers is real, and it should be treated as real. Hyperscalers have the incentive and the talent to design accelerators for their own workloads. Some of that silicon will take share at the margin, especially on narrow tasks. The open question is how much of the general market it displaces, and how fast. So far the answer has been “some, not most.” I will update that view when the share numbers say otherwise, not when a keynote says otherwise.
What The Skeptics Are Actually Arguing
A fair article has to sit with the other side. The bear case is not silly. It runs something like this.
- Spending on AI infrastructure is cyclical, and cycles end.
- Customers are concentrating, so a pause at two or three buyers moves the whole income statement.
- Custom chips and rival accelerators will pressure both share and price.
- Export rules can close doors without warning.
- The stock already reflects a great company, so the upside is capped.
Point five is the one the valuation chart pushes back on hardest. A great company at 35 times earnings can still be a mediocre stock. A great company at 16 times, with earnings still rising fast, is a different wager. The first four points are live risks, not talking points. Concentration is uncomfortable. Policy is a variable nobody models cleanly. Competition will take something. I do not need those risks to vanish. I need them to be priced, and at a decade-low multiple they look more priced than they did when the shares were everyone’s favorite ticker.
Cycles are the risk I respect most. Capital expenditure waves overshoot. They always have. Data centers in the late 1990s, fiber, shale, solar manufacturing. The pattern is old. The difference this time, if there is one, is that the compute is being used, not just installed. Utilization and return on that spend will decide how long the wave lasts. If returns disappoint, budgets shrink, and every supplier feels it. That is the scenario where today’s multiple stops looking cheap and starts looking early. Worth holding in your head. Not the same as the base case.
A Personal Read On The Setup
I am not interested in owning every AI-adjacent name that rallies on a headline. I am interested in the rare moments when the obvious leader is the one the market is bored with. July, around $200, felt like one of those moments. The subsequent stretch has not broken the thesis. If anything, the buyback enlargement and the still-steep growth guide have added a floor under the argument.
There is a habit, in this kind of market, of waiting for a dramatic washout before calling something a buy. Sometimes that washout comes. Sometimes the stock grinds, the earnings arrive, and the multiple never revisits the panic low you were saving your cash for. I have missed more of those than I like to admit. The current price is not a panic low. It is a respectful price for a disrespectfully strong business. That is enough for me to keep adding, in size that assumes I can be wrong about timing.
Price near a high and value near a low can coexist. The chart of the multiple is the one I trust more than the chart of the ticker.
A working rule, not a forecast
How I Think About Position Size And Timing
None of this is a suggestion to mortgage a plan on one ticker. Concentration cuts both ways. A sensible approach, the one I actually use, is to scale in rather than to declare a single perfect entry. The July area was attractive. The current area is attractive for a different reason: the business has kept delivering and the multiple has not re-expanded to match. If the shares run ahead of earnings again, the case gets weaker. If they stall while earnings climb, the case gets stronger. That is a workable loop.
Time horizon matters more than the entry print. A fiscal 2028 earnings case is useless to someone who needs the trade to work by next month. It is useful to someone who can sit through a dull quarter, a policy headline, or a rotation into whatever the tape loves in November. I would rather be early and bored than late and chasing. Boring has been the underrated feature of this name for months.
Risk management here is mostly about what you refuse to assume. Do not assume 97 percent share is permanent. Do not assume 72 percent margins are a floor carved in stone. Do not assume every dollar of announced data-center spend converts on schedule. Do assume that a leader with a software standard, a full roadmap, and a record buyback at a mid-teens multiple has more ways to win than a momentum story with a story and not much else.
The Industry Backdrop Is Still Tight
One earnings call does not make a cycle, but the tone across suppliers has been consistent enough to notice. Memory demand continues to exceed available supply as infrastructure spending grows. Equipment makers talk about extended lead times. Cloud operators keep lifting capital plans, then lifting them again. You can argue about the end date. Arguing that the present is soft requires ignoring a lot of primary commentary.
Tight supply is a strange kind of good news. It limits the upside in any single quarter, because you cannot ship what you do not have. It also tells you the customer is not browsing. When buyers are willing to wait, pricing power tends to hold. That is part of why margins in the low seventies do not look like a fluke to me. They look like the outcome of a product people need more of than the industry can currently build.
There will be a quarter, eventually, when supply catches demand and the language shifts from shortage to choice. That quarter will rattle the whole complex. I would rather meet it owning the vendor customers already standardized on than owning a second-source hope. Standardization is sticky. Hope is not.
What Could Make The Cheap Multiple Correct
Cheap can be a trap. The question is what would have to be true for 16 times forward earnings to be the right price rather than a discount. A few paths qualify.
Growth could fall off a cliff after fiscal 2028, turning today’s earnings power into a peak. Share could crack faster than the roadmap can answer, if custom silicon and rivals take the general workload and not just the edge cases. Margins could compress into the 50s if price competition arrives before volume does. A policy shock could remove a meaningful slice of demand with no replacement. Any one of those, sustained, would justify a multiple that looks ordinary or worse.
I do not dismiss them. I weight them. The evidence in front of us, today, points the other way: demand ahead of supply, a growth guide that is still extreme, margins that have stabilized rather than cracked, and a buyback that management sized like they believe the shares are inexpensive. Evidence can change. Until it does, treating the decade-low multiple as a warning rather than an opening feels like fighting the last war, the one where the stock could do no wrong.
Earnings Power The Price Has Not Fully Met
The cleanest way I know to hold the thesis is to separate the company from the ticker. The company designs the default accelerator, wraps it in software customers already know, and has the next two architectures on a schedule the industry plans around. The ticker has spent a year losing a footrace to other semiconductor names. Those are compatible facts. They are not the same fact.
If fiscal 2028 earnings land near $22, and if the multiple merely returns to something like 20 times, not 35, the price implication is obvious without me spelling out a target. I am wary of targets. They age badly and they pretend precision the inputs do not have. The direction is the useful part. A business growing that fast, bought at a low-teens multiple on those future earnings, does not need a heroic rerating to work. It needs the earnings to show up. The rerating, if it comes, is extra.
That is the asymmetry I wanted in July and still want now. Downside exists, and it is not trivial, because a cyclical pause would hurt. Upside does not require a return to mania. It requires the market to stop pricing a dominant, still-accelerating franchise like a company the cycle has already left behind.
Simple frame: future earnings power × a sane multiple − what is already in the price = the gap worth underwriting.Why Now, If The Shares Are Not Washed Out
Waiting for a crash is a strategy. It is also a strategy that misses a lot of compounders who never crash on your schedule. The reason to look now is not that the tape is panicked. The reason is that the valuation chart and the price chart disagree, and the business is refereeing in favor of the valuation chart.
Near a 52-week high. Cheapest forward multiple in at least ten years. Record buyback just enlarged. Growth guide still in a zone that would have sounded absurd five years ago. Share in the core product still near total. If you need all of those to line up with a 30 percent drawdown before you are willing to care, you will often be waiting while the earnings do the work without you.
I still consider the shares on sale at this valuation. Not on sale versus the March low of some past scare. On sale versus the earning power, the franchise, and the company’s own history of what investors have been willing to pay. Sale is a relative word. This is the comparison that matters.
A Few Practical Filters Before Anyone Hits Buy
Conviction without filters is just mood. A short list I actually run before adding:
- Is the underperformance about the business, or about rotation? Right now it looks like rotation.
- Is the multiple low because earnings are about to fall? The guide says no.
- Is capital return a substitute for growth, or a companion to it? Companion.
- Can I hold through a dull year? If not, the setup is the wrong shape.
- Am I sizing it like a speculation or like a core holding? Those are different bets.
Fail any of those and the headline appeal should not override the portfolio math. Pass them and the case is sturdier than a viral chart. I have found that writing the filters down keeps me from confusing a good story with a good position. The story here is good. The position still has to fit.
What I Will Be Watching Next
Supply commentary is first. If the company stops saying demand exceeds supply and starts talking about choice, the cycle conversation changes. Margin cadence is second. A drift from the low seventies toward the mid sixties, without a volume offset, would tell me pricing power is slipping. Customer concentration is third. Any single buyer stepping back matters more here than it would at a diversified industrial. The buyback pace is fourth. Authorization is intent. Execution is evidence.
Roadmap timing sits alongside all of that. Blackwell has to keep landing cleanly. Vera Rubin has to stay a date, not a hope. The CPU push has to show up in design wins, not just in slides. None of this requires a daily quote check. It requires reading the quarters like an operator instead of like a spectator.
Rival product launches will generate noise. Some of it will be deserved. I plan to weigh shipped volume and developer adoption over announcement volume. The industry has a long tradition of keynote victories that never quite become purchase orders. Share is the scoreboard. Everything else is color commentary.
Putting The Pieces Back Together
Strip the ticker symbol off and the description is almost dull in its strength. Dominant share in the product the AI buildout actually runs on. A software layer customers treat as standard. A roadmap that extends past the current generation. Margins that have settled rather than cracked. Growth that is still extreme. A repurchase program enlarged to a record, sized at a few percent of the company, running through fiscal 2028. A forward multiple near 16.7 times against a long-run average near 35. A plausible earnings figure around $22 that would put the same price near 10.7 times.
Put the ticker back on and the market’s recent behavior looks like a shrug. Up on the year, badly behind the semiconductor complex, close to a high that does not feel like a high once you use earnings instead of memory as the yardstick. Shrugs are where patient capital has historically done its better work. Not always. Often enough that I would rather be early to the shrug than late to the rerating.
I am still buying. Not because the theme is fashionable. Because the leader is priced like the theme has moved on, and the order book says it has not. That gap will close one way or the other. Either the business slows and the multiple was right, or the earnings keep arriving and the multiple was the opportunity. I know which side I am underwriting, and I know I can be wrong. The price I am paying leaves room for that.
If you only remember one comparison, make it this. The shares have lagged the chip sector by a wide margin while the company has kept a grip on the most important product in that sector. Laggards with broken businesses are traps. Laggards with intact businesses and decade-low multiples are the setups I came into this market to find. This still looks like the second kind.
