Big Tech Is Building Its Own AI Chips
Every major technology company is now designing chips specifically for AI. That shift — from general-purpose to purpose-built — is the most consequential hardware story of the decade, and it is only just beginning.
For thirty years, the semiconductor industry operated on a simple premise: build faster general-purpose processors and software will find a use for them. AI ended that bargain. The workloads that define modern computing — training and running large language models — are so specific, so demanding, and so economically significant that every company with the resources to do so has concluded it cannot afford to let someone else's chip define what its AI can do.
By Aaron Rose · Tech Reader Magazine · September 5, 2026
Why Big Tech Is Building Their Own AI Chips
The name Vera Rubin belongs to one of the 20th century's most important astronomers, whose patient observations of galactic rotation curves provided the strongest early evidence for dark matter — mass that shapes the universe without being directly visible. It is, accidentally or not, a fitting name for a chip platform whose deeper significance is not what it does but what it reveals about the structure of an industry reorganizing itself around a single workload.
Vera Rubin is not simply a faster GPU. Nvidia has been building faster GPUs for thirty years. What makes this platform architecturally distinct is that it treats the data center — not the chip, not the server — as the fundamental unit of compute. The GPU, the CPU, the networking fabric, the power delivery, the cooling — all of it designed together, from the ground up, for one purpose: running AI as efficiently as physics and engineering currently allow.
The numbers are striking on their own terms. Five times the inference performance of Blackwell. Ten times lower cost per token. Ten times greater throughput per watt. At rack scale, 3.6 exaflops of inference capacity from 72 GPUs in a single cabinet. At POD scale — 40 racks, 1,152 GPUs — 60 exaflops, enough compute to train the largest frontier models currently in existence.
But the numbers are not the story. The story is the design philosophy that produced them — and what it means that every other major technology company has independently arrived at the same conclusion.
$725BCombined capital expenditure planned by Amazon, Microsoft, Alphabet, and Meta in 2026 — up 77% from 2025.The overwhelming majority goes to AI infrastructure: GPU clusters, custom silicon, and data center construction.
The Economics That Changed Everything
To understand why purpose-built AI silicon is now a strategic imperative rather than an interesting experiment, start with the math that governs running a frontier AI model at scale.
Training a model is a cost you pay once. Inference — generating every response, every token, every output for every user — is a cost you pay continuously, at a rate that scales directly with how many people are using the system. For a model serving tens of millions of users, inference costs dwarf training costs within months of launch. Anthropic ran over budget on inference costs in 2025. OpenAI's inference bill is a matter of ongoing strategic concern. Google runs Gemini on custom TPUs specifically because the economics of serving billions of queries a day on general-purpose GPUs don't work at an acceptable margin.
A chip designed specifically for transformer-based inference — the mathematical operation at the heart of every major language model — can strip away the overhead that general-purpose processors carry for workloads they will never run in this context. Custom silicon doesn't need to be good at gaming, scientific simulation, video rendering, or the thousand other things an Nvidia GPU is engineered to handle. It needs to do one kind of tensor math, extremely well, at the lowest possible cost per operation. That specificity is where the 10x cost improvements come from. It is also why the economics of AI at scale have made chip design a core competency for companies that, five years ago, had no semiconductor engineering teams whatsoever.
Training a model is a cost you pay once.Inference is a cost you pay every day, at a rate that scales with every user. That math is what turned chip design into a strategic imperative for companies that never made hardware before.
Every Giant Has a Silicon Lab Now
Google has been designing Tensor Processing Units for over a decade, and Gemini runs almost entirely on TPU v6 and v7 — the company has, in effect, opted out of the Nvidia dependency for its own internal AI workloads. Amazon's Trainium and Inferentia chips have moved from interesting experiments to core infrastructure: Amazon's custom silicon business reached a $20 billion annual revenue run rate in the first quarter of 2026, with CEO Andy Jassy suggesting it could generate $50 billion if treated as a standalone entity. Microsoft's Maia 200 inference accelerator went live in Azure data centers in January 2026. Meta ships a new generation of its MTIA inference chip every six months.
OpenAI, which built its entire early infrastructure on Nvidia hardware, is collaborating with Broadcom and TSMC to bring its own custom chip to production by late 2026. Anthropic is in discussions with Samsung about chip production. Apple, which has led the industry in custom silicon for consumer devices with its M-series processors, is extending that philosophy into its AI server infrastructure.
The combined capital expenditure of Amazon, Microsoft, Alphabet, and Meta in 2026 is approximately $725 billion — up 77 percent from 2025, the largest coordinated technology buildout in history. Analysts project the number will cross $1 trillion in 2027. The overwhelming majority of that capital goes to AI infrastructure, and a growing share of it goes to silicon these companies design themselves rather than purchase from Nvidia.
Nvidia's market share in AI chips has declined from over 80 percent in 2023 to below 70 percent in 2026. The company lost roughly $1 trillion in market capitalization through the first half of the year as that dynamic became apparent to investors. It is still the dominant player in merchant AI silicon by a wide margin. But its largest customers are now also its most serious long-term competitors — a structural tension that will define the semiconductor industry for the rest of the decade.
What Vera Rubin Is Nvidia's Answer To
Vera Rubin is Nvidia's response to that pressure, and it is a sophisticated one. Rather than competing on the single-chip benchmark that custom ASICs can win by design — a chip built for one workload will always beat a general-purpose chip at that workload on some metric — Nvidia is competing on system-level performance and on the software ecosystem that no custom chip program has yet been able to replicate.
CUDA, Nvidia's programming platform, remains the foundational layer on which virtually all AI software is built. A developer building on CUDA can run their code on any Nvidia GPU, in any data center, across any cloud provider. A developer building on Google's TPU software stack, or Amazon's Neuron SDK, or Microsoft's Maia toolchain, is building for one platform. The fragmentation of the hardware layer is creating a software portability problem that the industry has not yet solved — and Nvidia, which caused that fragmentation by dominating so completely for so long, benefits from the resulting stickiness.
The Vera Rubin platform extends that advantage to the rack level. When the entire system — GPU, CPU, networking, interconnect — is co-designed by Nvidia, the integration benefits compound in ways that are difficult to replicate by assembling components from different vendors. The 3.6 terabytes per second of GPU-to-GPU bandwidth delivered by NVLink 6 is not achievable with standard networking infrastructure. It requires the switch, the interconnect protocol, and the GPU memory architecture to be designed together. That is what "extreme co-design" means in practice, and it is a moat that custom chip programs will take years to cross.
The Question Nobody Has Answered Yet
The deeper question raised by the silicon divergence is not who wins the chip race. It is what kind of AI ecosystem emerges from an industry where the hardware layer is owned by the same companies that own the models, the clouds, and the distribution.
When Google runs Gemini on TPUs, it controls the entire stack from silicon to user interface. When Amazon runs Anthropic's models on Trainium, it controls the infrastructure on which a nominally independent AI company depends. When OpenAI ships its own chip, it is not just reducing its hardware costs — it is eliminating a dependency that any competitor could theoretically exploit.
The open-source AI movement, represented most visibly by Hugging Face — now owned by Nvidia — exists in part as a counterweight to this vertical integration. Open models running on commodity infrastructure are the alternative to a world where the AI you can access is defined by whose silicon it runs on. But commodity infrastructure is increasingly custom silicon sold by the same companies building the closed models. The circle closes quickly when you follow it far enough.
An Extraordinary Piece of Engineering
Vera Rubin is an extraordinary piece of engineering. Sixty exaflops in a forty-rack configuration. Ten times lower cost per inference token than the generation before it. A CPU designed not for general computation but for keeping a 1,152-GPU cluster fed with data at the throughput AI factories require. It will make the AI systems it powers meaningfully better and meaningfully cheaper to run.
It will also make the companies that deploy it more dependent on Nvidia — or more motivated to build something that makes them less so. That dynamic, playing out simultaneously across every major technology company on earth, is the hardware story of the AI era. Vera Rubin is its latest chapter, not its conclusion.
The Inference Race
Vera Rubin. Fractile. Cerebras. Groq. Etched. The companies competing to define the economics of running AI are building different answers to the same question. Tech Reader Magazine is covering them all.