AI Chip Stocks in 2026: The Bull Case and the Bear Case Are Both Right

The number that spooked people in June had nothing to do with an earnings report. It was a rental price. The hourly cost to rent an Nvidia B200 (the flagship Blackwell data center GPU, the chip everyone is actually fighting over) fell from $6.11 an hour on May 30 to $4.22 an hour by June 21. That is a 31% drop in three weeks.^[1] If you want a real-time read on whether AI compute is scarce, spot rental rates are about as close as a public investor gets, and they were telling you supply had started catching up to demand.

The tape reacted the way tape reacts. Nvidia fell 6.2% on June 5 and dragged the Nasdaq down with it, then ground lower through the month, closing around $192 in late June after a May closing high of $235.47.^[2][3]The word that got attached to the selloff was “sustainability.” Not the carbon kind. The can-this-possibly-keep-going kind.

Here is where most of the coverage goes wrong. It treats “is AI real?” and “are these stocks correctly priced?” as the same question. They are not even close. You can believe AI is the biggest platform shift since the smartphone and still think the semiconductor complex is priced for a version of it that will not happen. I hold both views, and the interesting work is in showing exactly where they meet.

Start with the part that is not in dispute

Nvidia's most recent quarter, reported in May, did $81.6 billion in revenue, up 85% year over year. Data center alone was $75.2 billion, up 92%. Networking inside that was $14.8 billion, up 199%. Gross margin came in at 74.9% on a GAAP basis. Guidance for the current quarter is roughly $91 billion.^[4] These are not the numbers of a company selling into a fake market. A fraud does not have 75-point gross margins on $75 billion of quarterly hardware while every buyer on earth complains they cannot get enough of it.

The demand underneath is real too. Google said at I/O in May that it is now processing over 3.2 quadrillion tokens a month across its surfaces (a token is roughly a word fragment, the unit an LLM reads and writes). That is 7x growth in a year, from about 480 trillion.^[5] Two years ago the number was 9.7 trillion. Whatever you think of Google as a narrator of its own story, a 300x move in two years is not a marketing deck. Somebody is running that inference, and inference runs on silicon.

So: AI is real. Token volume is compounding. Capex is being spent. Move on. The question is not whether the technology works. It is whether the cash flows land where the current prices assume they will.

What would actually have to be true

Forget sentiment. If you own the AI chip complex here, you are underwriting three specific things, whether or not you have written them down.

1. Token demand keeps compounding faster than compute gets cheaper.Every chip generation is 2 to 3x more efficient per watt. That means the same dollar of capex buys multiples more output each cycle. For revenue to keep growing, usage has to outrun efficiency. Google's 7x is comfortably above that bar right now. The question is what happens the first year it is not.
2. Inference has to actually be profitable at the application layer. OpenAI is running roughly $25 billion in annualized revenue in mid-2026 against a projected loss around $14 billion for the year, with long-term compute commitments reported in the hundreds of billions.^[6] That is the marginal buyer of a very large share of frontier compute. If application-layer economics never turn, somebody upstream eventually eats it.
3. The replacement cycle has to be a feature, not a bug. The bull case quietly depends on hyperscalers ripping out and replacing GPUs every few years forever. That is great for Nvidia and terrible for the return on invested capital of the people buying them. Both cannot be true indefinitely.

Notice that none of these require an AI bust. They only require the growth to decelerate to something merely excellent. That is a much lower bar for the bear case than most bulls admit.

The three bear signals that are actually load-bearing

Most bear takes are vibes. These three are not.

1. The money goes in a circle

Nvidia agreed to invest up to $100 billion in OpenAI, which planned to spend that money on Nvidia chips. The structure was later reworked, and reporting in early 2026 described the original plan as having stalled.^[7][8] New Street Research estimated that for every $10 billion Nvidia puts into OpenAI, it sees roughly $35 billion in GPU purchases or lease payments back.^[7] Nvidia has done versions of this across the ecosystem.

The charitable read is that this is strategic capital seeding an ecosystem, which is a thing chip companies have always done. The uncharitable read is that a vendor financing its own customers is the oldest way to make a demand curve look steeper than it is. Both reads are available. What is not available is pretending the revenue is indistinguishable from a bank walking in off the street with a purchase order.

2. The depreciation schedules

Michael Burry made this argument loudly in late 2025 and it has not gone away, because nobody has actually rebutted the math. Hyperscalers depreciate GPUs over roughly five to six years. Burry argues the real economic life is closer to two or three, given that each generation is meaningfully more efficient and old chips become uneconomic for frontier work fast. His estimate of the gap: about $176 billion of understated depreciation and therefore overstated profit across the industry between 2026 and 2028.^[9]

You can argue with the number. The inference-serving fleet genuinely does keep older chips busy longer than a training fleet would, and an A100 still earns its keep. But the direction is not arguable. The industry moved from assuming 3 to 4 year refresh cycles to depreciating over 5 to 6, and it did so during the exact period when the refresh cadence accelerated.^[9]That is the wrong direction, and it flatters reported earnings at every hyperscaler that buys Nvidia's chips.

3. The customers are building the product themselves

This is the one I think is most underrated, and it is the one that does the most damage to Nvidia specifically rather than to AI generally.

Custom silicon from the hyperscalers (Google's TPU, Amazon's Trainium, Microsoft's Maia, Meta's MTIA) is now estimated at 20% to 28% of the AI chip market. Custom ASIC shipments are projected to grow around 44.6% in 2026 against roughly 16.1% for merchant GPUs.^[10]Nvidia's AI accelerator share is tracking from roughly 87% in 2024 toward the mid-70s by the end of 2026, even as its absolute revenue keeps climbing, because the pie is growing faster than anyone can take share.^[10]

Hyperscale customers are still about 50% of Nvidia's data center revenue.^[4] Those are the same companies spending billions to stop being Nvidia customers. The strategic logic is airtight from their side: inference is the workload that is now most of the compute, it is the most predictable, and predictable workloads are exactly what custom ASICs are for. You do not need to beat Nvidia at everything. You need to beat it at the one thing you run a trillion times a day.

The multiple is not the problem, and that is the twist

Here is the thing that breaks most people's mental model. Nvidia is not trading at a bubble multiple. It is around 23x forward earnings, against a five-year average closer to 70x.^[14] At a $5.1 trillion market cap it is the largest company on earth and it trades at a market-ish multiple.^[3]

Read that again, because it changes the argument. The market has already derated Nvidia. It is not paying up for the growth. It is treating the earnings as cyclical, which is the polite way of saying it does not believe they are permanent. A stock up 5% while its revenue grows 85% is a market compressing the multiple in real time.

And this is exactly where the two questions people keep merging pull apart. AI being real is what justifies the $725 billion of hyperscaler capex.^[11] Nvidia being correctly priced requires something stronger: that Nvidia keeps capturing the majority of that spend, at 75-point gross margins, while its four biggest customers spend aggressively to make sure it does not.

Meanwhile the working capital keeps creeping. Nvidia's inventory went from about $10.1 billion in January 2025 to $21.4 billion in January 2026. Days sales outstanding pushed up to roughly 52.^[15]Neither is alarming in isolation. Big cloud customers pay slowly, and building inventory ahead of a Blackwell ramp is what you would do if you expected the demand you are guiding to. But if you are looking for the first place a demand air pocket shows up in the financials, it is here, not in the revenue line. Revenue is the last thing to break.

Where I actually land

I think AI is real, the capex is mostly rational, and the specific trade of “own Nvidia because AI is big” is finished. Not wrong. Finished. That trade had a five-year run and the market already told you it is over by taking the multiple from 70x to 23x while the business tripled.

The bear signals I take seriously are the ones about margin capture, not the ones about demand. Circular financing means some meaningful slice of reported demand is Nvidia's own balance sheet coming back around. Stretched depreciation means hyperscaler earnings are flattered, which means their capex appetite is being set with an optimistic view of what the last round actually cost them. Custom silicon means the most profitable, most repeatable workload in AI is being systematically walked out the door. Any one of those is survivable. All three at once is why a 23x multiple on a company growing 85% is not the free lunch it looks like.

The bull signal I take most seriously is the one nobody can fake: inference volume. 3.2 quadrillion tokens a month, 7x in a year. If that number is still compounding at anything above 2x in mid-2027, the entire bear case collapses into a timing argument, because demand that large eventually finds a way to pay for itself. If it decelerates toward 2x or below while capex is still running at $725 billion a year, we will have built a lot of very expensive, very efficient inference capacity for a market that grew into the last generation of chips instead of the next one.

So the considered position: the AI trade is broadening, not ending, and that is precisely what SOXX up 112% next to Nvidia up 5% is telling you. The value is migrating to the parts of the stack that every architecture needs regardless of whose logic chip wins. Memory. Networking. Foundry. Power. Whatever else changes, a TPU needs high-bandwidth memory the same way a B200 does, and a data center needs a substation either way. Nvidia is a great company that is now, correctly, being valued as a cyclical one. Treat it that way and you will make far better decisions than whichever side of the June selloff you were arguing.

The chips are not the bubble. The assumption that one company keeps 75% of the margin on a market its own customers are trying to commoditize, that is the bubble.