The Nvidia Margin: the Business Logic Behind Google's and Amazon's Custom AI Chips
Photo: N43 and Hermes AIMerchant GPU margins are so large that building your own silicon became rational for anyone with scale — and in 2026 custom AI chips stopped being a hedge and became a product.
Source video: How Nvidia GPUs Compare To Google’s And Amazon’s AI Chips · CNBC · approximately 2.2M views observed via yt-dlp on November 19, 2026. Analysis by N43 and Hermes AI. This article is independent analysis and is not affiliated with or endorsed by any company named.
01 The margin that funds everything
Nvidia's data-center business has operated at extraordinary profitability, with GAAP gross margin climbing from 56.9 percent in fiscal 2023 to 72.7 percent in fiscal 2024 and 75.0 percent in fiscal 2025 by the company's own annual reports. Margins like that are unusual anywhere in hardware and nearly unheard of in merchant silicon, where mid-range pricing power is normally the ceiling.
The arithmetic is simple and ruthless: hardware that sells for roughly three to four times its build cost generates a pool of margin so large that avoiding the purchase price has itself become a strategic project. Every dollar of Nvidia margin a hyperscaler pays is a dollar that did not go to its own engineering payroll. That is the subsidy logic behind every custom-chip program in the industry: Nvidia's pricing strength finances its challengers.
This is the margin-arbitrage lens this analysis applies throughout. The question is not which chip wins a benchmark, but whether the spread between merchant prices and build costs is wide enough to justify designing, building, and operating your own silicon — and for buyers at hyperscale, it clearly has been.
02 The ASIC playbook
The custom-silicon playbook is consistent across Google's TPU, Amazon's Trainium, Microsoft's Maia, and Meta's MTIA. Each trades away general-purpose flexibility — the ability to run any workload well — for better cost and better power per token on a fixed, well-understood workload that the owner controls end to end.
That trade works because the owners are also the operators. A hyperscaler that controls its own models, scheduler, and rack design can strip out generality it does not need: narrower instruction sets, fewer caches, and interconnects tuned to one known topology. The chip becomes a component inside a vertically designed machine rather than a product that must satisfy many buyers with conflicting demands.
The limits are equally consistent. Each generation must be re-designed against Nvidia's moving target, and the moment workloads shift faster than the ASIC roadmap, the flexibility that was sold away becomes expensive to lack again.
03 From hedge to product
The 2026 development is that custom chips stopped being only a hedge and became a product. Google now sells TPU capacity externally: Anthropic has committed to make use of large quantities of TPU compute, and Apple has trained models on Google's TPUs as well. Silicon originally built to escape merchant prices now earns external revenue from other companies' workloads.
That changes the economics twice. Internal savings amortize development across the owner's own fleet, and external revenue adds a second return on the same engineering spend. A chip program that once had to justify itself purely as cost avoidance can now be evaluated like any other product line — a much easier case to make at budget time, and a reason to keep investing across generations.
It also changes the competitive picture. Nvidia now faces chips that are not merely substitutes hidden inside rivals' data centers but actively marketed alternatives, with their own customers, their own pricing, and their own delivery commitments.
04 Co-design: the real moat is the rack
The durable advantage in custom silicon is not the chip alone but co-design: accelerator, networking, and scheduling software designed together against known workloads. Google's TPU pods — thousands of chips connected by dedicated interconnects and scheduled as one logical machine — are the clearest example of what that alignment produces.
Pod-scale systems are hard to copy because the difficulty is distributed. The chip, the interconnect fabric, the failure handling inside the scheduler, and the compilers that partition models across the mesh must improve together, and no merchant purchase replicates that alignment. It is the same lesson the rack-scale GPU market teaches from the other direction: integration, not silicon, is where advantage and margin concentrate.
For buyers, the corollary is that comparing custom chips to merchant GPUs part-by-part misses the point. The comparison that matters is pod against rack — priced, cooled, and scheduled end to end over the life of a deployment.
05 The software gap that keeps Nvidia alive
What keeps Nvidia alive against all of this is software inertia. CUDA represents roughly two decades of kernels, compilers, debuggers, and — more decisively — trained engineers who reach for it first. Every custom stack must reproduce enough of that surface to hold production workloads, and each reproduces it unevenly, with gaps that surface unpredictably under load.
The gaps show up in specific, unglamorous places: model support on release day, kernel maturity on new architectures, tooling depth, and the porting cost of moving a tuned training pipeline off Nvidia parts. Hyperscalers bridge those gaps with their own large engineering organizations, which is precisely the cost most non-hyperscaler buyers cannot pay — and why custom silicon remains a scale game rather than an open market.
The honest summary is narrower than the headlines suggest: custom chips win where their owners control the entire stack, and Nvidia retains essentially everyone else.
06 Broadcom and the custom-silicon supply chain
The custom-chip boom has its own supply chain, and Broadcom sits at its center of gravity. Google co-designs TPUs with Broadcom, and Broadcom has described a growing roster of custom-accelerator customers and rapidly rising AI revenue from that business. Merchant ASIC design partners are the layer that turns hyperscaler architectures into manufacturable products at leading process nodes.
This layer matters for the margin story because it lowers the barrier to entry. A company with scale and a well-defined workload no longer needs an in-house silicon organization of thousands to attempt a chip; it needs enough engineers to specify and co-design one. The existence of a merchant design industry is what turns margin arbitrage from theory into an executable plan — and it is why more custom programs, not fewer, should be expected from here.
07 Limits and what to watch
The risks are cyclical and structural at once. Capex cycles: the tens of billions in announced 2025 capital plans (see chart after Section 05) are plans, not guaranteed spend, and any pause in AI infrastructure would strand depreciation schedules built on optimistic utilization assumptions.
Depreciation risk follows directly: custom chips carry shorter useful lives when each generation steps up performance sharply, and owners must recognize that cost whether or not token demand arrives on schedule. The strategic question to watch is whether custom chips remain second sources that discipline Nvidia's pricing, or become primary platforms for major workloads — the external TPU business suggests the second has already begun. Watch the margin itself too: if Nvidia's gross margin compresses under competition, the subsidy that funds the entire custom-chip thesis shrinks with it.
References
- Wikipedia: Nvidia — company background and data-center GPU business.
- Wikipedia: Tensor Processing Unit — Google's custom AI ASIC and its history.
- Amazon Web Services: AWS Trainium — vendor page for Amazon's custom training accelerators.
- Wikipedia: Broadcom Inc. — background on the leading custom-silicon design partner.
- Source video: How Nvidia GPUs Compare To Google's And Amazon's AI Chips (CNBC, ~2.2M views, observed November 19, 2026)
By N43 and Hermes AI for DutyStation News.





