And then your cloud provider started making chips...

Yesterday I put “the provider started designing the hardware” on my list of the ten biggest changes to cloud capacity in the last decade, and I gave it two sentences. It deserves a post of its own, because it is the item that changes who you are buying from. Ten years ago the cloud resold other companies’ chips. Today the companies selling you capacity design a growing share of it themselves, and on the CPU side that share is the majority. But it did not take over everything: GPUs are still Nvidia’s market, and Google’s TPUs show where the road ends.

It started earlier than you think

The beginning is 2015, and it is one acquisition and one secret project. In January 2015, Amazon bought Annapurna Labs, a small Israeli chip design firm. It looked like a deal about network cards and storage controllers. Jeff Barr’s launch post for the first Graviton instances tells the real story: AWS acquired Annapurna “after working with them on the first version of the AWS Nitro System,” and a few years later “the team started to think about building an Amazon-built custom CPU designed for cost-sensitive scale-out workloads.”¹ AWS did not set out to become a chip company. It bought a chip team to offload hypervisor work, and the team was already in the building when someone asked the next question.

Google got to the same place from the other direction. The TPU was running inside Google’s data centers in 2015, and Google announced it in May 2016 with an unusual admission: the chips had been running for more than a year and delivered “an order of magnitude better-optimized performance per watt for machine learning,” and the team went from first tested silicon to “running applications at speed in our data centers within 22 days.”² Neither company announced a customer product. The first TPU ran Google Search and Street View; the first Graviton was an instance family for scale-out workloads. The transition started as a purchasing decision that turned into an engineering one.

ARM is the transition point

Why did the CPU side move first? Because Arm changed what it means to buy a CPU. Intel and AMD sell finished processors. Arm licenses an instruction set and core designs, which means a company with enough engineers can build its own processor on proven parts, with only the features it wants. That is the difference between buying a car and buying the right to build one.

So when the providers started designing hardware, they all walked through the same door. AWS launched Graviton on Arm cores in November 2018; Graviton2 arrived in December 2019 and went mainstream with the M6g, C6g, and R6g families at up to 40% better price performance than equivalent non-Graviton instances.³ Microsoft announced its Arm-based Cobalt CPU at Ignite in November 2023, alongside the Maia accelerator, with Scott Guthrie naming the reasons directly: “optimize and integrate every layer of the infrastructure stack to maximize performance, diversify our supply chain and give customers infrastructure choice.”⁴ Google announced Axion in April 2024, “our first custom Arm-based CPUs,” with claims of 50% better performance than comparable current-generation x86 VMs.⁵ Three competitors who agree on almost nothing picked the same architecture within six years of each other. Arm did not win by selling processors. It won by making processors something its customers could design.

How much capacity it drives now

On CPUs, this stopped being a projection a while ago. At re:Invent 2025, AWS stated it plainly: “For the third year in a row, more than half of new CPU capacity added to AWS is powered by Graviton, with 98% of the top 1,000 EC2 customers” already using it.⁶ AWS’s own services run on it too, over 90% of Redshift Serverless and over 60% of services like Aurora and RDS, per the Graviton session at the same conference,⁷ and the current Graviton5 packs 192 cores per chip with a five-times-larger cache.⁸ When AWS adds a rack of general-purpose compute today, the odds are better than even that the chip in it was designed by Amazon.

On accelerators, look at where the largest new blocks of capacity came from. Project Rainier, the cluster AWS built for Anthropic, came online in October 2025 with nearly half a million Trainium2 chips, and AWS’s own announcement says Claude will run on more than a million of them by the end of the year, on infrastructure “70% larger than any other AI computing platform in AWS history.”⁹ Two weeks later, Anthropic and Google announced that Anthropic would have access to up to one million TPUs, “well over a gigawatt of capacity coming online in 2026,” in a deal worth tens of billions of dollars,¹⁰ extended in April 2026 with a multi-gigawatt agreement that brings Broadcom in, with capacity online from 2027.¹¹ Google’s seventh-generation TPU, Ironwood, connects up to 9,216 chips in a single pod.¹² Read those numbers as capacity, not as chip news: the biggest single capacity commitments announced in the last twelve months are provider-designed chips.

The contrast: Nvidia

On GPUs, the providers did not take over. They became customers. Nvidia booked $89.0 billion of data center revenue in the quarter ended July 26, 2026, up 117% year over year, and its quarterly filing breaks out $48.7 billion of that from hyperscale customers, which means the same companies designing their own chips are buying more Nvidia silicon than anyone else on earth.¹³ Full-year fiscal 2026 revenue was $215.9 billion.¹⁴ Third-party estimates put Nvidia’s share of AI accelerator revenue somewhere between 75% and 90%, and the range itself is the point: the lower end counts custom chips like TPUs and Trainium as part of the market, the higher end does not.¹⁵

The same companies make both moves in the same season. On August 26, 2026, the day Nvidia reported that quarter, AWS and Nvidia announced plans to deploy 2 million additional Nvidia GPUs across AWS in 2027 and 2028, with AWS framing the expansion as something that “complements Amazon’s own custom silicon, giving customers the freedom to choose.”¹⁶ Vera CPUs are coming to AWS, and next-generation Trainium chips will support Nvidia’s NVLink Fusion interconnect.¹⁷ So the honest picture is not provider-designed versus Nvidia. The providers design the CPU layer, almost entirely on Arm, plus a growing slice of the accelerator layer, and they still buy most of their accelerator capacity from Nvidia. The reason is software: CUDA and a decade of tooling keep most training workloads on Nvidia hardware, and AMD has only reached the 5% to 7% range of that market.¹⁸

The outlier: Google’s TPUs

Google never went through the phase the others did. It did not buy accelerators for its main workloads and then decide to design around them. It started building its own accelerator years before it said anything in public, had the chip running in production by 2015, and showed it off in May 2016, two years before Graviton existed. When the AI boom arrived, Google was the one provider with a decade of production experience on silicon it designed itself.

Then it did something none of the others have done: it turned the internal chip into a product other companies buy at scale. By October 2025 the customer list included Anthropic, the company whose model competes directly with Google’s own Gemini, signing up for up to a million chips, with Google Cloud’s CEO describing the TPU line as “our already mature AI accelerator portfolio.”¹⁹ That is the endpoint of provider-designed hardware. Not a cost saving buried in the provider’s margin. A new product line, sold to customers the provider competes with.

What this means for a capacity plan

Three things follow, and none of them are speculative. The provider’s roadmap is now a hardware roadmap: when more than half of new CPU capacity at the biggest cloud runs on the provider’s own chip, the product list keeps being rewritten from inside the provider, not by its suppliers. The default fleet is no longer one architecture: a realistic plan today spans x86, Arm, GPUs, and provider accelerators, the choice among them is a price-performance decision you re-make every generation, and the evidence standard still applies, measure your workload on the new hardware before you believe the marketing number (chapter 11). And when a whole class of capacity depends on one company’s design decisions, architecture questions become capacity questions (chapter 17).

Watch the seams too. The newest announcements are all about the boundary between designed and bought: NVLink Fusion reaching into Trainium, Vera CPUs arriving on AWS, Arm CPUs pairing with Nvidia GPUs.²⁰²¹ Ten years after Amazon quietly bought a chip firm, the answer to “who designs the hardware” is: increasingly the provider. On CPUs, almost entirely. On accelerators, partly. On GPUs, not yet.

Sources

---

1. Jeff Barr, AWS News Blog, “New – EC2 Instances (A1) Powered by Arm-Based AWS Graviton Processors” (November 2018) — Annapurna acquired in 2015 after working on the Nitro System; the team “started to think about building an Amazon-built custom CPU designed for cost-sensitive scale-out workloads”; A1 instances launched on Arm cores. https://aws.amazon.com/blogs/aws/new-ec2-instances-a1-powered-by-arm-based-aws-graviton-processors/ (accessed October 17, 2026).↩︎

2. Norm Jouppi, Google Cloud Blog, “Google supercharges machine learning tasks with TPU custom chip” (May 19, 2016) — TPUs running “inside our data centers for more than a year”; “an order of magnitude better-optimized performance per watt”; from first tested silicon to running applications at speed “within 22 days”; RankBrain, Street View and AlphaGo named as TPU workloads. https://cloud.google.com/blog/products/ai-machine-learning/google-supercharges-machine-learning-tasks-with-custom-chip (accessed October 17, 2026).↩︎

3. AWS News Blog, “Announcing New Amazon EC2 M6g, C6g, and R6g Instances Powered by Next-Generation Arm-based AWS Graviton2 Processors” (December 3, 2019), https://aws.amazon.com/blogs/aws/coming-soon-graviton2-powered-general-purpose-compute-optimized-memory-optimized-ec2-instances/ ; AWS News Blog, Graviton4 R8g preview post — “December 2019 – The Graviton2 processor debuted with the announcement of M6g, M6gd, C6g, C6gd, R6g, and R6gd instances with up to 40% better price performance than equivalent non-Graviton instances.” https://aws.amazon.com/blogs/aws/join-the-preview-for-new-memory-optimized-aws-graviton4-powered-amazon-ec2-instances-r8g (both accessed October 17, 2026).↩︎

4. Microsoft Source, “With a systems approach to chips, Microsoft aims to tailor everything ‘from silicon to service’ to meet AI demand” (November 2023, Microsoft Ignite) — Azure Maia 100 AI accelerator and Azure Cobalt 100 Arm-based CPU announced; Guthrie: “At the scale we operate, it’s important for us to optimize and integrate every layer of the infrastructure stack to maximize performance, diversify our supply chain and give customers infrastructure choice.” https://news.microsoft.com/source/features/ai/in-house-chips-silicon-to-service-to-meet-ai-demand (accessed October 17, 2026).↩︎

5. Google Cloud Blog, “Introducing Google’s new Arm-based CPU” (April 2024, Cloud Next ’24) — “Google Axion Processors, our first custom Arm®-based CPUs”; built on Arm Neoverse V2; up to 50% better performance than comparable current-generation x86-based VMs; Arm CEO Rene Haas on custom silicon optimized for Google’s infrastructure. https://cloud.google.com/blog/products/compute/introducing-googles-new-arm-based-cpu ; Google Cloud Blog, “Compute and infrastructure enhancements at Next ’24” — Google services including Bigtable, Spanner, BigQuery and the YouTube Ads platform already deployed on Arm-based servers. https://cloud.google.com/blog/products/compute/compute-and-infrastructure-enhancements-at-next24 (both accessed October 17, 2026).↩︎

6. About Amazon, “AWS re:Invent 2025: Amazon announces Nova 2, Trainium3, frontier agents” (December 2025) — “For the third year in a row, more than half of new CPU capacity added to AWS is powered by Graviton, with 98% of the top 1,000 EC2 customers” using it; Graviton5 with 192 cores per chip, 5x larger cache, up to 25% higher performance than Graviton4; Trainium3 UltraServers with up to 144 chips. https://www.aboutamazon.com/news/aws/aws-re-invent-2025-ai-news-updates (accessed October 17, 2026).↩︎↩︎

7. AWS re:Invent 2025 session CMP360, “AWS Graviton: The best price performance for your AWS workloads” — “over 90% of Redshift Serverless runs on Graviton”; “Over 60% of all of these services run on Graviton” for MSK, Amazon Aurora and RDS. https://www.youtube.com/watch?v=haSHl2BLhm8 (accessed October 17, 2026).↩︎

8. About Amazon, “AWS activates Project Rainier: One of the world’s largest AI compute clusters comes online” (October 2025) — nearly half a million Trainium2 chips; “Claude is expected to be on more than 1 million Trainium2 chips” by the end of the year; the Trainium2 infrastructure is “70% larger than any other AI computing platform in AWS history.” https://www.aboutamazon.com/news/aws/aws-project-rainier-ai-trainium-chips-compute-cluster (accessed October 17, 2026).↩︎

9. Google Cloud Press Corner, “Anthropic to Expand Use of Google Cloud TPUs and Services” (October 23, 2025) — access to “up to one million TPU chips”; “well over a gigawatt of capacity coming online in 2026”; deal “worth tens of billions of dollars”; Thomas Kurian: “building on our already mature AI accelerator portfolio, including our seventh generation TPU, Ironwood.” https://www.googlecloudpresscorner.com/2025-10-23-Anthropic-to-Expand-Use-of-Google-Cloud-TPUs-and-Services (accessed October 17, 2026).↩︎↩︎

10. Silicon Republic, “Anthropic, Google, Broadcom announce 3.5GW TPU deal” (April 2026) — multi-gigawatt next-generation TPU capacity online from 2027; Broadcom regulatory filing notes Anthropic’s TPU consumption depends on its “continued commercial success.” https://www.siliconrepublic.com/machines/anthropic-google-broadcom-announce-3-5gw-tpu-deal (accessed October 17, 2026).↩︎

11. Google, “Ironwood: The first Google TPU for the age of inference” (2025) — scales up to 9,216 chips in a pod. https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/ironwood-tpu-age-of-inference ; The Register, “TPU v7, Google’s answer to Nvidia’s Blackwell is nearly here” (November 2025) — TPU v4 pods up to 4,096 chips, v5p up to 8,960. https://www.theregister.com/software/2025/11/06/tpu-v7-googles-answer-to-nvidias-blackwell-is-nearly-here/650535 (both accessed October 17, 2026).↩︎

12. NVIDIA Newsroom, “NVIDIA Announces Financial Results for Second Quarter Fiscal 2027” (August 26, 2026) — revenue $96.2 billion; Data Center revenue $89.0 billion, up 117% year over year. https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027 ; NVIDIA Form 10-Q for the quarter ended July 26, 2026 (SEC EDGAR) — Data Center revenue by market: Hyperscale $48,710 million. https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm (both accessed October 17, 2026).↩︎

13. NVIDIA Newsroom, “NVIDIA Announces Financial Results for Fourth Quarter and Fiscal 2026” (February 25, 2026) — Q4 revenue $68.1 billion; record quarterly Data Center revenue $62.3 billion; full-year fiscal 2026 revenue $215.9 billion. https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-fourth-quarter-and-fiscal-2026 (accessed October 17, 2026).↩︎

14. Silicon Analysts, “NVIDIA AI GPU Market Share 2024–2026” and “AMD vs NVIDIA AI GPU Market Share 2026” — third-party estimates: Nvidia peaked near 87% of AI accelerator revenue in 2024, roughly 75–80% in 2026; AMD estimated at 5–7%; ranges vary by whether hyperscaler custom silicon is counted in the market definition. https://siliconanalysts.com/analysis/nvidia-ai-accelerator-market-share-2024-2026 ; https://siliconanalysts.com/analysis/amd-vs-nvidia-ai-gpu-market-share-2026 (both accessed October 17, 2026).↩︎↩︎

15. About Amazon, “AWS and NVIDIA expand partnership for next-gen AI infrastructure” (August 26, 2026) — 2 million additional NVIDIA GPUs across AWS in 2027–2028; the expansion “complements Amazon’s own custom silicon, giving customers the freedom to choose”; NVIDIA Vera CPUs coming to AWS; NVLink Fusion support in next-generation Trainium plus NVIDIA custom NVHBM memory technology; AI factories for the U.S. government including 100,000 GPUs. https://www.aboutamazon.com/news/aws/aws-nvidia-2-million-gpus-ai ; NVIDIA Newsroom version: https://nvidianews.nvidia.com/news/aws-and-nvidia-to-deliver-2-million-additional-gpus-and-next-generation-infrastructure-for-agentic-and-physical-ai (both accessed October 17, 2026).↩︎↩︎↩︎

16. CNBC, “Arm custom chips get a boost with Nvidia partnership” (November 17, 2025) — Arm CPUs will integrate with Nvidia’s NVLink Fusion; aimed at customers “namely hyperscalers” pairing custom Neoverse CPUs with Nvidia GPUs; Arm “licenses its instruction set technology”; Nvidia’s $40 billion Arm acquisition attempt failed in 2022. https://www.cnbc.com/2025/11/17/arm-nvidia-nvlink-partnership-custom-cpu.html (accessed October 17, 2026).↩︎