Listen to the audio version of this article (generated by AI).
Forty-thousand dollars — that’s roughly the price of one Nvidia (NVDA) H100 GPU. And Big Tech needs millions of them.
Every quarter, Alphabet (GOOGL), Amazon (AMZN), Microsoft (MSFT), and Meta (META) are writing enormous checks to the same address in Santa Clara, Calif.
For a while, they had no choice. Nvidia had built the dominant chip architecture now deeply embedded in the AI research ecosystem. Nvidia’s market share in AI accelerators peaked around 90% in 2022 — 90% in an industry where 20% makes you a power player.
And Nvidia knew it. Its gross margins on AI chips ran somewhere between 75% and 80%. In other words, for every $40,000 chip sold, approximately $30,000 was pure profit. The cost to manufacture? Somewhere around $3,000 to $5,000.
So the world’s most sophisticated technology companies were essentially paying a private tax, to one company, that they had little leverage over.
Until the economics became too obvious to ignore — and revealed the next great vertical in the AI supply chain.
How the GPU Shortage Pushed Big Tech Toward Custom AI Chips
The AI boom exploded faster than anyone, including Nvidia, had predicted.
Demand for H100s rocketed. And Nvidia, constrained by Taiwan Semiconductor‘s (TSM) manufacturing capacity, simply could not build chips fast enough to meet it.
The waitlists stretched for months. Amazon told investors it had “more demand than it could fulfill” because of chip supply constraints. Anthropic‘s ability to train Claude was throttled by how many Nvidia chips it could get its hands on.
The companies with the most capital, the best engineers, and the most urgent need in the most important technology race in the world were suddenly capacity-constrained by one supplier.
But that constraint turned out to be one of the most important developments in the AI hardware market.
Because Google, Amazon, Meta, and Microsoft don’t respond to problems by complaining about them. They build their way out.
(The “build around the bottleneck” impulse isn’t exclusive to chips — it’s why we’re also tracking X Money and . But one disruption at a time.)
The Custom Silicon Push Was Years In the Making
The AI chip shortage of the past few years threw gasoline on Big Tech’s custom silicon fire. But it didn’t originate it. This move has been years in the making.
Google began developing its Tensor Processing Unit — a chip designed to run Google’s AI workloads with extraordinary efficiency — back in 2015. At the time, it was a ‘skunkworks’ project, barely mentioned externally for two years. And it has now become the foundation of the most mature custom AI chip program in the world.
Amazon followed with Inferentia (2019) and Trainium (2021). In 2023, Meta launched MTIA, and Microsoft announced Maia. By the end of that year, every major hyperscaler had at least one proprietary AI chip in production.
The breadcrumbs were hiding in plain sight. The financial press mostly missed them, focused instead on Nvidia’s stock price while its biggest customers were quietly spending billions to reduce their dependence.
Then came Project Rainier, Amazon’s internal AI computing cluster. It was built entirely from its own Trainium2 chips and contains 500,000 custom processors in a single system. For context, OpenAI used approximately 25,000 Nvidia GPUs to train GPT-4. Project Rainier is 20 times that, all on Amazon’s own custom silicon.
The April 2026 Custom AI Chip Deal Wave
Rainier’s launch signaled a major stress fracture in the AI chip space. And six months later, in April 2026, the dam broke open.
- April 6: Broadcom (AVGO) files an 8-K with the SEC disclosing a new five-year TPU partnership with Google — extending through 2031 — plus a deal giving Anthropic access to 3.5 gigawatts of Google TPU computing capacity starting in 2027. That nearly quadrupled Anthropic’s compute footprint.
- April 15: Meta and Broadcom jointly announce an extension of their custom chip partnership through 2029, with Meta committing to pay over $2 billion per year in design fees and deploy one full gigawatt of its own custom MTIA processors.
- April 19: The Information reports that Google is in active talks with Marvell Technology (MRVL) to develop two additional custom AI chips — a Memory Processing Unit and an inference-optimized TPU. Google, already in a five-year deal with Broadcom, is beginning a second design partnership for two more custom chips.
- April 20: Amazon announces an investment of up to $25 billion into Anthropic. In return, Anthropic commits to spend $100 billion on Amazon Web Services over the next 10 years — with Trainium custom chips at the center of that spend. Around the same time, Amazon CEO Andy Jassy says the company’s custom chip business has crossed a $20 billion annual revenue run rate, up from $10 billion earlier this year.
Four deals, two weeks, over $200 billion in long-term commitments, all pointing in the same direction: custom silicon.
Nvidia’s data center compute share has already dropped from roughly 90% in 2022 to around 75% in late 2025. Over half of internal hyperscaler inference workloads now run on custom ASICs. Custom chip sales are growing at 45% annually — nearly three times the 16% growth rate for GPUs.
The shift is underway. The only question now is who’s positioned to benefit from it.
Where the Custom AI Chip Toll Roads Are Forming
In any major technology transition, there are two kinds of companies worth owning. The destination companies — the ones competing to rule the new era — and the toll road companies — the ones collecting a fee at every step of the journey, regardless of who wins.
In the PC revolution, Apple (AAPL), Compaq, Dell (DELL), IBM (IBM), and hundreds of others all fought ferociously for market share. Intel (INTC) didn’t pick sides. It supplied the chips to all of them.
Every PC sold, regardless of brand, had Intel inside. Intel collected its toll on each sale. From 1981 to its dominance peak, INTC delivered gains of over 10,000%.
The AI chip revolution has destination companies too. OpenAI versus Anthropic versus Google DeepMind. They make great headlines. And if you pick the right one, fantastic.
But the toll road play is more interesting. Because the toll road doesn’t need a prediction. It just needs the revolution to keep happening.
And as of April 2026, the revolution already has roughly $200 billion in signed commitments pushing it forward.
So, where does the toll road money flow?
The Demand Engine: Hyperscalers and AI Labs
Google, Amazon, Meta, Microsoft, OpenAI, and Anthropic are the demand engine. Together, they’re spending more than $500 billion on AI infrastructure in 2026 alone. But these are still the Destination Companies, competing with one another for models, users, and market share. The more interesting layer sits underneath them — the companies getting paid no matter which hyperscaler wins.
The Architecture Toll Road: Arm
Arm Holdings (ARM) supplies core architecture used across much of the modern computing world — and increasingly, across the systems surrounding custom AI silicon. Every time a hyperscaler deploys Arm-based CPUs, control processors, or AI-adjacent silicon inside these systems, Arm collects royalties on every chip shipped.
The Designers: Synopsys and Cadence
Before a chip can be built, it has to be designed. And every chip designer in the world relies on either Synopsys (SNPS) or Cadence Design Systems (CDNS). More custom chip programs means more licenses, more upgrades, more revenue. And because these platforms sit deep inside chip-design workflows, they are difficult to replace once engineering teams standardize around them.
The Custom Silicon Builders: Broadcom and Marvell
This is the center of the custom silicon trade: the companies translating hyperscaler demand into physical AI chips.
Broadcom dominates the layer. It designs custom AI accelerators and networking silicon for companies like Google and Meta, giving it direct exposure to the rapid expansion of proprietary AI infrastructure. Its AI-related revenue has surged from roughly $2 billion in 2023 to a run rate now measured in the tens of billions annually — with management projecting continued growth through the end of the decade.
Then there’s Marvell Technology — smaller, quieter, and increasingly difficult to ignore. Marvell helped develop Amazon’s Trainium chips and Microsoft’s Maia processors, while also supplying the data-infrastructure silicon those AI systems need to move information at speed. Several of its largest hyperscaler programs are expected to ramp simultaneously over the next few years.
As Big Tech moves away from off-the-shelf GPUs and toward proprietary silicon, Broadcom and Marvell are becoming two of the most important engineering partners in the AI infrastructure stack.
The Foundry Toll Road: Taiwan Semiconductor
Every custom AI chip ultimately has to be fabricated somewhere — and at the bleeding edge of semiconductor manufacturing, Taiwan Semiconductor Manufacturing is effectively the only game in town. Whether it’s Google’s TPU, Amazon’s Trainium, or Nvidia’s GPUs, nearly all advanced AI silicon eventually flows through TSM’s fabs.
The Networking Layer: Arista, Credo, and Coherent
The networking layer may be the least appreciated part of the AI stack. Companies like Arista Networks (ANET), Credo Technology (CRDO), and Coherent (COHR) build the optical interconnects, switches, and high-bandwidth networking infrastructure that make AI clusters function. As compute scales from thousands of chips to millions, the bottleneck shifts from raw compute to moving data between those chips fast enough to keep them useful.
Why the Custom AI Chip Trade Deserves a Second Look
Much of the AI trade in 2023 and 2024 was speculative — bets on revenue that had not yet materialized across the stack, capabilities that might or might not arrive, and a competitive landscape that nobody could map with confidence.
The custom silicon trade of 2026 is different in an important way: it’s already in the numbers.
Broadcom has now delivered 11 consecutive quarters of AI revenue growth. Amazon’s custom chip business doubled in just months. More than half of hyperscaler inference workloads already run on custom ASICs. And over $200 billion in customer commitments has already been signed across the ecosystem.
This is not a bet on the future. It’s a bet on a structural shift that is already underway, already represented in earnings reports, and — based on the contract terms being signed — contractually committed through the end of the decade.
The toll roads are already being built. The contracts are already signed.
The only variable is whether you’re positioned on the right side of it.
The obvious toll roads in this cycle are semiconductors, networking, and compute infrastructure.
But if AI becomes the foundation of the next digital economy, will emerge around payments, transactions, and financial coordination.
That’s why we’re paying close attention to what Elon Musk is building inside X.
.
