Sign in Send a tip

AI and models · 18 Oct 2026

Why Big Tech Companies Are Building Their Own AI Chips

Google, Amazon, Microsoft, Meta and now OpenAI all design their own AI accelerators. The reasons are cost, supply and control—but the trade-offs are real, and most performance claims still come from the companies themselves.

By PrimetimeGeek AI Desk

Original ink illustration of several custom silicon chips on a workbench with drafting tools, a rising cost curve and cloud server racks in the background.
Every major cloud company now designs its own AI silicon. Original conceptual illustration, not a depiction of any specific chip.

For a decade, buying AI compute meant buying Nvidia. That is still largely true—Nvidia holds roughly 70% of the AI chip market—but the company's biggest customers have quietly become its competitors. Google, Amazon, Microsoft and Meta all have custom AI accelerators in production, and OpenAI is developing its own chip with Broadcom. Microsoft said this month that its next chip, Maia 300, will become the primary inference system for its own products when deployment begins in 2027.

This is not a hobby. Designing a leading-edge chip costs hundreds of millions of dollars before the first unit ships, and each of these companies continues to buy enormous quantities of Nvidia hardware at the same time. So why do it?

The three reasons: cost, supply and fit

Cost. Merchant GPUs carry a large margin, and at hyperscaler volumes that margin compounds into billions. A chip designed for one company's own workloads can skip the generality a merchant GPU has to carry. Microsoft claims Maia 200 delivers 30% better performance per dollar than the latest hardware in its fleet, and that Maia 300 will deliver 40% more AI tokens per megawatt than leading GPU-based systems. Treat these as vendor claims: they are measured on the company's own workloads, under conditions it chose, and no independent benchmark has confirmed them.

Supply. After years of GPU shortages and allocation queues, depending on a single supplier for your core input is a strategic risk. Custom silicon does not remove that risk—nearly all of these chips are fabricated by TSMC, and advanced packaging and high-bandwidth memory remain constrained—but it moves the bottleneck from a merchant's product roadmap to a foundry slot the buyer negotiates directly.

Fit. Google, Amazon, Microsoft and Meta run enormous, well-understood internal workloads. A chip co-designed with the serving stack and compilers that will actually run on it can win on performance per watt and cost per token for those specific jobs. Google's TPU program is the proof of concept: eight generations in, TrendForce projects TPUs will account for nearly 78% of AI servers shipped to Google in 2026—the only major cloud whose build-out is more custom silicon than GPU.

Who is doing what

Google is the earliest and most mature player. Its latest TPU generation splits into separate training and inference variants for the first time, and Google now sells TPU capacity externally—Anthropic has committed to up to one million TPUs.

Amazon shipped Trainium3 at re:Invent in December 2025, its first 3-nanometer AI chip. The deployment that matters is Anthropic's Project Rainier, which AWS says grew past 500,000 Trainium2 chips and now exceeds one million. GPUs are still expected to represent roughly 60% of AWS's AI server build-out in 2026, per TrendForce—custom silicon is a complement, not a replacement.

Microsoft introduced Maia 200 in January 2026 as an inference accelerator and announced Maia 300 this month, with deployment starting in 2027. It is the most recent major entrant and the clearest statement that inference economics—tokens per dollar and per watt—are now the design target.

Meta deploys its MTIA line for ranking and recommendation workloads and has committed to a Broadcom-backed custom-chip roadmap, while also being an early gigawatt-scale AMD customer. OpenAI is developing its first chip with Broadcom, initially for inference, and says it is already working on subsequent generations.

Broadcom is the quiet enabler behind much of this: it co-designs chips for Google, Meta and OpenAI, reported $8.4 billion in AI semiconductor revenue for its first fiscal quarter of 2026 (up 106% year over year), and says it has line of sight to more than $100 billion in annual AI chip revenue in 2027. Together with Marvell, it controls roughly 95% of the custom AI chip co-design market.

The honest trade-offs

Custom silicon is not free dominance. The developer ecosystems around these chips are far smaller than Nvidia's CUDA, which remains the default training environment. A chip tuned to today's model architectures is a bet that those architectures will not shift underneath it. Custom chips also lock workloads to one cloud, and most performance-per-dollar figures are self-reported on internal workloads. TrendForce projects custom ASIC shipments will grow 44.6% in 2026—nearly triple the growth rate of merchant GPUs—but from a much smaller base, reaching about 28% of AI server shipments.

What this means

For buyers of cloud AI services, the practical effect is downward pressure on inference prices and more choice among platforms—with a new form of lock-in attached. For Nvidia, the risk is real but gradual: its biggest customers are building alternatives for the workloads they understand best, while still buying its GPUs for everything else. The question for the next two years is whether these programs can move from vendor claims and anchor-customer deployments to reliable multi-gigawatt production on the announced schedules, given that foundry capacity, substrates and high-bandwidth memory are shared constraints for everyone.

We covered the market side of this buildout in Why America's AI Chip Boom Is Facing a New Market Test and the recent sell-off in AI stocks slide as investors reassess infrastructure costs.

Sources