Why Every Tech Giant Is Suddenly Obsessed With Custom Silicon

Technology

Why Every Tech Giant Is Suddenly Obsessed With Custom Silicon

By reviewsideal Staff  |  September 27, 2026  |  7 min read

Google, Amazon, Microsoft, Meta, Apple, and OpenAI all now design their own chips. It looks like a sudden craze, but the logic is old and simple: when a computer is your product, the processor inside it is too important to leave to someone else's roadmap. AI's enormous appetite for compute has just made that logic impossible to ignore.

It Started Before AI

The playbook was proven years before the current boom. Google began building Tensor Processing Units for its own machine-learning workloads back in the mid-2010s. Amazon started designing Graviton server processors in the late 2010s to cut the cost of its cloud. Apple moved the Mac from Intel to its own M-series chips in 2020, gaining performance and battery life by designing hardware and software together. Each case showed the same payoff: a chip tailored to a company's specific workloads can beat a general-purpose part on cost and efficiency, and it gives the company control it can't get by shopping off the shelf.

Why It's Accelerating Now

1. The Economics of Scale

Hyperscalers are guiding to roughly $700 billion of combined capital spending in 2026, most of it for AI infrastructure. At that volume, shaving even a modest percentage off the cost of each accelerator is worth billions, and buying merchant chips means paying for a supplier's profit margin on top. Designing your own chip has a huge upfront cost, but the math flips once you deploy it by the hundreds of thousands.

2. Inference Is the New Workload

Training a frontier model happens occasionally; serving it to millions of users happens constantly. One industry analysis estimates inference now makes up about two-thirds of AI compute, and it rewards efficiency: chips built for one job, running it cheaply, are more valuable than flexible hardware doing everything. That is exactly what a custom accelerator offers. The same analysis projects custom AI chips growing at about 45% a year, though forecasts like this are estimates, not certainties.

3. Supply and Bargaining Power

Nvidia's GPUs have been the scarce, expensive ingredient in the AI boom, and being dependent on one supplier is uncomfortable for any company spending this much. A credible in-house alternative diversifies supply and improves negotiating leverage, even when the company keeps buying Nvidia as well.

4. Power Is the Real Limit

Data centers are increasingly constrained by electricity, not just chips. Performance per watt becomes the metric that matters, and designing hardware around a specific model architecture and network layout is one of the few ways to squeeze more useful work out of a fixed power budget.

5. Co-Design Beats Off-the-Shelf

When the same company writes the model, the software stack, and the chip, it can tune all three together, choosing number formats, memory layouts, and interconnects to match how its own models behave. That vertical integration is Apple's old advantage now applied to the data center.

Who Is Building What

  • Google: the TPU line, now in its seventh generation (Ironwood), sold to outside customers through Google Cloud.
  • Amazon: Graviton processors plus the Trainium and Inferentia AI chips, central to its partnership with Anthropic.
  • Microsoft: the Maia AI accelerators and Cobalt CPUs, developed for Azure.
  • Meta: its MTIA accelerators for recommendation and AI workloads.
  • OpenAI: reportedly working on its own chip with partners.
  • Apple and Tesla: proof that the strategy works outside the cloud, in phones, laptops, and cars.

The Catches

Custom silicon is not a guaranteed win. Nvidia's real moat is software: its CUDA ecosystem has years of tools, libraries, and developer habits behind it, and moving workloads to a new chip takes engineering effort. Designing chips is expensive, slow, and failure-prone, and a misjudged architecture can lock a company into hardware that doesn't fit where models go next. Nearly all of these chips are also manufactured by the same foundry, TSMC, on its leading-edge processes, where demand has reportedly outstripped supply, so custom design doesn't escape the manufacturing bottleneck. Analyst predictions that Nvidia's inference share will drop sharply by 2028 are forecasts, and Nvidia continues to advance its own roadmap. Most companies will use both for the foreseeable future.

What It Means for Everyone Else

For businesses buying cloud AI, the shift should mean more choice and pressure on prices, since cloud providers can offer their own chips as cheaper alternatives to GPU instances. For developers, it means paying attention to portability: frameworks that run across different accelerators reduce the risk of being tied to a single vendor. For investors and observers, the interesting question is where value ends up: with the chip designers, the foundries, the cloud owners, or the model builders.

The Bottom Line

The obsession with custom silicon isn't a fad. It is the rational response to spending hundreds of billions on computing while depending on a single supplier and a power grid that can't keep up. The chips won't replace GPUs overnight, but they are becoming the default way the biggest technology companies control their costs, their supply, and their pace of innovation.

Follow reviewsideal for ongoing coverage of the chips behind the AI boom.

Post a Comment

Previous Post Next Post