Is NVIDIA Still Worth It Now That Its Biggest Customers Build Their Own Chips?

The Numbers That Make This Story Real

Start with scale because the rest of this argument only makes sense once you have the numbers in front of you. NVIDIA's data center revenue went from $3.6 billion in fiscal 2022 to a quarterly run rate of over $35 billion in early 2026. Annualize that and you're looking at over $140 billion a year. To put it in context: the entire data center industrySemiconductors all the world's chipmakers combined generated about $527 billion in revenue in 2023. One company's data center business alone is approaching a third of that figure

NVIDIA's fiscal fourth quarter 2026 revenue was $39.3 billion up 78% year over year. Data center accounted for 88%. Gross margins came in at 73.5% which isn't exactly software company territory but close enough that you had to check you were reading a hardware company's income statement. Microsoft Google Amazon and Meta all fourhyperscalers that build most of the world's AI infrastructure committed more than $300 billion in combined capital expenditures between 2025 and 2026. Most of that money is going toward the purchase of NVIDIA chips

Why This Looks Different From the Fiber Optic Bubble

The most common argument about bears has nothing to do with NVIDIA. It's a historical analogy: this expansion of AI is similar to the fiber optic overbuild of 1999 to 2000 which flooded the world with fiber that no one used for a decade and buried the companies that installed it. I take that comparison seriously. It's the right instinct to reach for

But there is one structural difference that matters. Telecom overbuilding created capacity that sat idle turning on fiber without traffic waiting years for demand to recover. The AI investment cycle works differently in a specific sense: it builds infrastructure to meet the demand that already exists today. Every GPU that Google runs for Gemini inference is generating revenue right now. Every Azure GPU cluster serving Copilot every Meta recommendation system retrained with new silicon billsto a customer or generates advertising revenue from the day it is activated. Utilization is the whole game here and in that dimension this cycle does not rhyme with fiber

The Critical Caveat: Custom Silicon

This is the real bear case and it has nothing to do with whether the demand for AI computing is real. It clearly is. The question is who maintains the economics necessary to meet that demand over the next decade

Google has TPU v5. Amazon has Trainium2. Microsoft has Maia 100. Meta has MTIA. All four are custom accelerator chips designed specifically to reduce each company's dependence on NVIDIA for specific workloads. At hyperscale volume custom silicon can achieve a significantly better cost per inference for a model architecture it was designed around. That's not hypothetical. That's why these companies spent billions designing chips ininstead of simply buying more GPUs

Here's the part that should worry an NVIDIA bull more than the chips themselves: As the compute mix moves from training which is still primarily NVIDIA's game toward inference where custom silicon is stronger the hyperscalers' own chip investments become more valuable relative to NVIDIA's. My read is that the market bifurcates rather than NVIDIA losing out entirely. NVIDIA keeps training.Custom silicon consumes an increasing share of hyperscale commodity inference

Valuation

NVIDIA trades at about 35 times forward earnings. Said out loud that number sounds absurd. That said next to a company that grows revenue 70% to 80% annually on 73% gross margins with software that nothing else in the industry can currently replicate it's starting to look defensible

The comparison I find useful is not an abstract model of fair value. It is what the market has historically paid for a company that truly mastered a transformational technology cycle. Microsoft during Windows and Office. Cisco during the early development of the Internet. Intel during the PC manufacturing ramp. Each of them had a high multiple for years and in each case earnings growth eventually caught up to and justified the premium investors had already paid

On a five-year horizon long enough to capture the composition of the CUDA moat and the training compute cycle I think the answer is probably yes NVIDIA is worth it. On a one- to two-year horizon where gains in custom silicon stocks are the catalyst that actually moves the stock from quarter to quarter there is real downside risk built into what the market currently expects

What Actually Protects NVIDIA (It's Not the Silicon)

If you look past the ticker and ask what NVIDIA's position really stands for the honest answer is not the chip. Silicon will eventually be commoditized. It's the things that wrap around the chip that are really hard to copy

Start with CUDA the software layer that NVIDIA has been building for fifteen years. Every major AI research lab writes code against it. Hundreds of thousands of models papers and production systems take it for granted. A hyperscaler that builds its own accelerator still has to port that entire ecosystem to a new compiler stack a project that consumes years and burns engineers or accept that its custom chip only runs the limited set of workloads that its own team has manually optimized. ThatThe second path is exactly what TPU Trainium and Maia do today. They handle a company's internal workloads well. They don't run whatever some random researcher decides to try next Tuesday

The second piece is networking and I think it's underrated relative to CUDA. Training a frontier model is not one chip doing a job. It's tens of thousands of chips moving huge volumes of data around each other continually and the interconnect NVIDIA's NVLink and InfiniBand stack is what makes that cluster behave like a single machine instead of ten thousand separate machines. Copying the compute die is difficult. Copying the interconnect and systems softwarethat keeps it fed is in my reading more difficult because it is only shown at the scale of an entire training group not on a spec sheet

The third piece is simply difficulty. Building an accelerator for a workload that is stable and well understood is a solved engineering problem right now. Building one for the frontier where the architecture of the winning model could look different in eighteen months means designing hardware for a target that keeps moving. That's the trap that custom silicon on the cutting edge falls into and it's why I don't think the bear case is really about chips at all. It's about software and plumbing

A Worked Example: Buying Merchant Silicon vs Building Custom

Here's a worked example with made-up but clearly labeled numbers that shows why "design your own chip" only points in one direction

Let's say a hyperscaler needs 100,000 accelerator chips for a stable well-defined inference workload running the same type of model at the same type of scale for years. It's a matter of deciding between buying off-the-shelf commercial GPUs or designing a custom chip in-house

Option A buy commercial silicon. Let's call the price $25,000 per chip everything included.There is no design cost because someone else already built it. Total cost of 100,000 chips: 100,000 times $25,000 or $2.5 billion

Option B custom design. This requires a non-recurring engineering cost up front the design team tooling licenses test chips validation runs before a single production unit is shipped. Call it $500 million. Because the custom chip boils down to exactly what this workload needs and carries no supplier markup the manufacturing cost per unit is $15,000. Total cost: $500 million more100,000 times $15,000 which is $500 million plus $1.5 billion or $2 billion

Metric	Commercial Silicon	Custom Silicon
Total cost in 100,000 chips.	2.5 billion dollars	$2.0 billion
Effective cost per chip	$25,000	$20,000
Volume less than 50,000 chips	cheaper	more expensive

With 100,000 chips custom makes $500 million about 20% cheaper than buying commercial silicon. But that $500 million design cost has to be amortized based on volume and volume is exactly what makes or breaks this math. Equate the two options to find the break-even point: $500,000,000 divided by the $10,000 savings per chipwhich is $25,000 minus $15,000 equals 50,000 chips. Below 50,000 units you pay for the design without enough volume to recover it and merchant silicon is cheaper. Above 50,000 each additional chip is pure savings relative to the purchasing merchant

That's the whole decision about custom silicon in a single calculation. It only works if you have enough volume at a stable enough workload to get over that break-even line. It's also exactly why this move is available to Google Amazon Microsoft and Meta and not to almost anyone who buys compute. They are among the only companies buying at a volume where the arithmetic actually works

Case Study: Cisco and the Rise of the White Box Switch

To see a real-world version of this dynamic look at networking equipment not AI and specifically Cisco

For a long period Cisco was the default supplier of switches that move data around a data center sold as an integrated box: Cisco silicon Cisco software Cisco support all included at a premium price. Then hyperscalers came along with enough purchasing volume to do what smaller customers were never able to do. Facebook now Meta launched the Open Compute Project in 2011 sourcing its own data center hardware designs instead of purchasing equipmentfinished.Across the industry hyperscalers increasingly built what the industry calls white box switches: commercial switching silicon largely from Broadcom combined with their own networking software or open source alternatives like SONiC rather than a fully bundled Cisco or Juniper box

The mechanism should sound familiar. It's the same one that Google Amazon Microsoft and Meta are running at NVIDIA right now: take the expensive high-margin fully packaged product shrink it to the commercial silicon underneath and build your own software on top of it because you buy enough volume to make the fixed cost of doing so worth it

Cisco didn't collapse. It still dominates enterprise networking security and campus equipment companies with thousands of customers none of which have the scale to build their own switches. What specifically eroded was its position within hyperscale data centers the exact segment of customers with the volume to do this to a vendor

I think that's NVIDIA's honest interpretation. It doesn't have to mean that NVIDIA collapses. It may mean that NVIDIA continues to dominate everywhere except the small number of customers large enough to walk away from parts of the relationship which also happen to be their most important customers

The Counterargument: Where Custom Silicon Actually Wins

Let me properly analyze the bearish case because I don't think it is a weak argument disguised as a risk factor

Custom accelerators are not a scientific project. TPUs have done a significant portion of Google's internal training and inference since long before this AI cycle began and they work at Google volume on Google workloads. The worked example above is not hypothetical. If a hyperscaler has a stable well-understood high-volume inference workload the arithmetic for building custom silicon for it actually works favorably and should continue to work favorably as theseCompanies iterate their own chip designs over successive generations

There is a second effect that doesn't require the custom chip to completely replace NVIDIA GPUs and I think this one is underrated. The mere existence of a credible internal alternative changes the deal. A hyperscaler that can target its own fleet of running TPUs or Trainiums has a real alternative on the table even if it only runs a modest portion of its inference workload on it. That's influence on price on allocation during a supply shortage on contract terms regardless of whetherthe custom chip becomes more than a minority of the total compute. NVIDIA's customers don't need to leave to hurt NVIDIA's pricing power. They just need an exit credible enough that NVIDIA has to negotiate as if it existed

Where I would reject my own argument: If I'm wrong about CUDA being sticky it's probably because I'm underestimating how much of the frontier research workload eventually looks less like open experimentation and more like stable repeatable production inference exactly the case where custom silicon wins cleanly. That shift is already underway. It's just a question of how far and how fast it goes

How I Actually Think About This

So how do I actually use all of this instead of just describing it?

My honest read is that I don't treat NVIDIA as a binary bet on AI computing demand because I think part of the thesis is close to being resolved. The demand is real and is reflected in revenue not just ads. What I really look at is the change in the mix within that demand training versus inference because that is the variable that determines how much of the pie NVIDIA retains over time

The way I'd really track this if you followed the name closely is the hyperscalers' own capital spending disclosures and their comments on earnings calls on custom silicon not NVIDIA's numbers directly. When Google Amazon Microsoft or Meta talk about what proportion of their internal workloads run on their own chips versus NVIDIA GPUs that's the leading indicator. NVIDIA's own reported revenue is the laggard because the shiftof a hyperscaler to custom silicon appears in NVIDIA's order book next year not in this year's earnings statement

I also try to separate two questions that I used to combine: is the construction of AI computing real?and is NVIDIA specifically the right way to own it? I think the first is close to a yes. The second is the hardest and is really sensitive to the time horizon which is why I gave two different answers in the valuation section above instead of one clear number. That's not being indecisive. It's the honest form of uncertainty

None of this is a recommendation to buy or sell anything. It's the framework I'd like to have in my head before forming an opinion and it's the one I'd like to be able to defend line by line if someone on the other side of the table were to reject every sentence

The Bottom Line

NVIDIA's numbers are truly extraordinary: data center revenue increased from $3.6 billion to an annualized run rate of more than $140 billion 78% year-over-year growth in the most recent quarter gross margins of 73.5% and hyperscalers committed more than $300 billion in combined capital expenditures that mostly flow to NVIDIA. This is not a bubble in the fiber optics sensebecause the capacity that is being built is being used from day one

The real risk isn't demand. It's capturing the economics of meeting that demand as the mix moves from training to inference where Google's TPU Amazon's Trainium Microsoft's Maia and Meta's MTIA are designed to compete. What really makes NVIDIA's case is CUDA the interconnect and systems software wrapped around the chip and the enormous difficulty of designing custom silicon for a moving target like the frontier model architecture not the compute die itself. The example above showswhy custom silicon only gains real volume in a stable workload and the Cisco case study shows what happens to an incumbent's position when its largest customers get exactly that kind of volume: not a collapse but a real erosion in the specific segment where the customer has scale

At about 35 times forward earnings I think NVIDIA is probably worth it on a five-year outlook that allows CUDA to sustain itself and the training cycle to develop. On a one- or two-year outlook where share gains in custom silicon are the catalyst the downside risk is real and I wouldn't rule it out. Both of those things can be true at the same time and I think that's the honest answer not a hedge