Nvidia Didn’t Buy Arm. It Built the AI Factory Anyway.Original analysisNot investment advice
In 2020, the fear was that Nvidia would buy Arm and control the future of computing. The deal failed. But the strategy survived. Nvidia built the CPU-GPU-DPU-networking-software stack anyway, and the data center became its product.
In 2020, the Nvidia-Arm story looked like a battle over ownership. Nvidia wanted to acquire Arm. Critics worried that a vertically integrated AI company buying the world’s most important CPU IP licensor would distort the entire semiconductor ecosystem. Arm was supposed to be neutral. Nvidia was not.
If Nvidia owned Arm, it could theoretically pressure Qualcomm, Samsung, MediaTek, hyperscalers, embedded vendors, and future Arm server CPU players. The fear was simple. Jensen Huang wanted to control the data center, and Arm was the missing CPU piece.
That deal failed. Nvidia and SoftBank terminated the acquisition in February 2022 because of significant regulatory challenges.23
The acquisition failed. The strategy did not.
Nvidia did not need to own Arm to build the data-center computer. It needed an Arm license, a CPU roadmap, a GPU monopoly, a networking stack, a DPU, a software moat, and the ability to package everything into a system customers could buy as one machine. That is exactly what happened.
The correct 2026 claim is not that Nvidia destroyed Arm. The correct claim is that Nvidia failed to buy Arm but still achieved the strategic goal: control more of the data-center computer. Arm survived and grew. The ecosystem is now split between Nvidia’s vertically integrated AI factory and hyperscalers building their own Arm-based compute stacks.
I. The 2020 thesis
In August 2020, Dylan Patel published a SemiAnalysis piece arguing that Jensen Huang’s real goal in trying to acquire Arm was not Arm’s licensing revenue. The deeper play was data-center vertical integration.1 Nvidia already had the GPU and CUDA. Mellanox had given Nvidia the networking base and a credible DPU path. The missing leg was CPU. Owning Arm would have given Nvidia direct leverage across the data-center compute stack, with implications for Arm’s neutrality, hyperscaler in-house silicon, and the slower rise of RISC-V.
I revisited that piece because its framing aged unusually well. The acquisition path was wrong. The strategic destination was right. Six years later, Nvidia owns the data-center computer in every practical sense that matters, even without the Arm purchase.
If the data center becomes one computer, the company that controls CPU, GPU, DPU, networking, and software controls the most valuable layer of computing.
That sentence, written from the 2020 framing, is still the cleanest summary of how Nvidia operates today. The interesting question is how Nvidia got there without the Arm acquisition.
II. The deal failed, but the strategy survived
In February 2022, Nvidia and SoftBank announced the termination of the proposed Arm acquisition, citing significant regulatory challenges that made completion impossible.2 The US FTC had already filed a complaint to block the transaction over competition concerns, and dismissed the matter after termination.3 Arm then prepared for its own public offering and continued to operate as a neutral CPU IP licensor.
Nvidia lost the ownership path. It kept the architecture path. The Grace and Vera CPU programs continued through Arm architectural licensing, and Nvidia integrated those CPUs into its own systems alongside the GPUs, DPUs, NICs, and switches it already controlled.
Nvidia did not buy Arm. It rented enough of Arm to build the machine it wanted.
III. The data center is the computer
Jensen’s working argument is that the unit of computing has changed. The old data center was a collection of servers stitched together with commodity networking. The new AI data center is a single machine.
Why that shift matters in practice: AI training needs thousands of accelerators behaving like one computer. Inference needs high utilization, memory bandwidth, scheduling, networking, and low latency. GPUs alone are not enough. CPUs feed, schedule, and orchestrate. DPUs handle networking, storage, security, and infrastructure offload. Networking determines cluster scale. Software determines whether the hardware is usable at all.
A pile of servers
- Commodity servers · x86, generic NICs
- Independent networking · switch + cable layer
- Optional accelerators · added per workload
- Software · mostly customer responsibility
One vertically integrated machine
- CPU + GPU + DPU · tightly co-designed
- NIC + NVLink + InfiniBand or Ethernet · rack-scale
- Software stack · CUDA, NIM, NeMo, AI Enterprise
- Cooling and power · designed in, not added
IV. Nvidia now has the CPU-GPU-DPU trifecta
The 2020 piece described a three-legged stool. CPU, GPU, DPU. The implicit fear was that Nvidia could only complete it by acquiring Arm. In 2026, Nvidia completed the stool without owning Arm.
The Vera Rubin NVL72 platform is the clearest single artifact. Nvidia says it unifies 72 Rubin GPUs and 36 Vera CPUs into a rack-scale system with ConnectX-9 SuperNICs, BlueField-4 DPUs, NVLink 6, and scale-out networking, designed as one machine rather than as a tray of cards.4 Grace already uses high-performance Arm-based CPU cores. Vera moves further toward custom Olympus Arm-compatible cores, paired with the Nvidia Scalable Coherency Fabric, and Nvidia frames Vera as its first fully custom data-center CPU core targeted at AI factory workloads.56
Grace / Vera
Arm-based CPUs via architectural licensing, paired with Nvidia’s coherency fabric.
Blackwell / Rubin
Accelerator monopoly that scales from one die to a 72-GPU rack.
BlueField + Mellanox
DPUs, SuperNICs, and switches as one networking and offload stack.
V. Nvidia’s business became the data center
The financial picture tracks the strategic one. Nvidia’s Q4 FY2026 revenue was $68.1 billion, full-year FY2026 revenue was $215.9 billion, and Q4 Data Center revenue was $62.3 billion, up 75 percent year over year.7
Read that ratio again. Data Center revenue alone was roughly 91 percent of total revenue in Q4 FY2026. This is no longer a graphics company with a data-center business. It is a data-center company with a graphics legacy.
VI. Arm was not destroyed. It became the battlefield.
The 2020 framing feared that Nvidia ownership would destroy Arm’s neutrality. What happened instead is more interesting. Nvidia did not buy Arm. Arm remained independent. Arm became more important in data centers, not less. Arm reported $1.49 billion in Q4 FYE26 revenue and $4.92 billion for the full fiscal year, with cloud and infrastructure adoption as a structural driver.8
The data-center CPU map filled in around Arm in three directions. Google introduced Axion as its first custom Arm-based CPU for the data center, designed on Arm Neoverse.9 Microsoft launched the Cobalt 100 family for Azure, built on Arm Neoverse N2.10 AWS Graviton continued to power a growing share of new AWS CPU capacity year over year, anchored on Arm Neoverse generations.11 Arm itself positioned Neoverse and Compute Subsystems (CSS) as a way for hyperscalers and chipmakers to build their own custom CPUs faster, without each one designing from scratch.12
That is the unexpected outcome. Arm was not silenced. Arm became the language many companies now use to build alternatives to x86 and Nvidia.
Nvidia did not kill Arm. Nvidia helped prove why Arm mattered.
VII. The new war is stack versus stack
The competition is no longer just GPU versus GPU, or CPU versus CPU, or ISA versus ISA. It is stack versus stack. The question for any AI buyer is no longer “which chip is fastest” but “who controls the full cost per token across training and inference, including software, networking, and rack-scale design.”
Vertically integrated AI factory
- Grace / Vera CPU
- Blackwell / Rubin GPU
- BlueField DPU
- ConnectX / Spectrum-X / Quantum InfiniBand
- NVLink interconnect
- CUDA, CUDA-X libraries
- DGX / NVL rack systems
- NIM, NeMo, AI Enterprise
Cloud-native AI factories
- AWS Graviton + Trainium + Inferentia
- Google Axion + TPU
- Microsoft Cobalt + Maia
- Meta MTIA
- Custom networking fabrics
- Internal software layers
- Workload-specific optimisation
- In-house data center design
Neutral CPU IP layer
- Neutral IP licensor
- Neoverse data center platforms
- CSS pre-validated subsystems
- Wide ecosystem reach
- Multiple foundry partners
- Same architecture across both stacks
- Royalty + licensing model
- Beneficiary of both sides
VIII. Why Nvidia’s lock-in is stronger than ISA ownership
The 2020 fear assumed that owning Arm was the strongest lever in computing. The 2026 reality is that the strongest lever turned out to be CUDA plus rack-scale integration.
The Nvidia moat is not one product. It is the combination of CUDA, the developer ecosystem around it, optimised libraries, model frameworks, the networking stack, full rack-scale systems, supply allocation, customer trust, faster deployment, performance predictability, and the long tail of software support that comes from running the largest installed base of AI accelerators on Earth.
Owning Arm would have given Nvidia leverage. Owning the AI factory stack gave Nvidia dependency.
ISA ownership would have given Nvidia leverage. CUDA plus rack-scale integration gave Nvidia dependency.
IX. Hyperscalers are fighting back
Hyperscalers are not passive customers. They are building internal silicon because they want control over capex, opex, power efficiency, inference cost, workload specialisation, supply chain risk, software integration, and the architecture of every data center they construct.
X. The physical stack underneath
The AI factory is not only a software architecture. It is a manufacturing and physics problem. TSMC’s 2025 annual report frames AI/HPC as a structural driver of advanced process demand and continued packaging investment.13 ASML’s 2025 strategic report frames AI as requiring leading-edge processors and a meaningful step up in DRAM, with advanced Logic and DRAM driving further EUV lithography exposures and spending.14
Read those reports together and the AI factory looks less like a Nvidia product and more like a stack with many interlocked layers. Software sits on top. Lithography, packaging, memory, power, and cooling sit underneath. Nvidia controls more of that stack than any other company, but it does not control the bottom.
XI. What 2020 got right and wrong
Six years is long enough to grade the 2020 framing honestly. The strategic insight aged well. The mechanism by which it would play out did not.
The strategic destination
- Jensen’s endgame was data-center control.
- CPU + GPU + DPU was the correct framework.
- Mellanox was strategically important.
- CUDA and software lock-in mattered.
- Arm would become central in cloud and server.
- Customers would worry about Nvidia dependence.
The mechanism
- The Arm acquisition was blocked.
- Arm did not collapse. It grew.
- RISC-V did not rush in to replace Arm at the top.
- Hyperscalers accelerated Arm-based silicon.
- Nvidia built Grace and Vera through licensing.
- Lock-in moved from ISA ownership to full stack.
Quick terms
- CPU
- General-purpose processor that runs operating systems and orchestration logic.
- GPU
- Accelerator optimised for parallel compute, central to AI training and inference.
- DPU
- Data processing unit for networking, storage, security, and infrastructure offload.
- NIC
- Network interface card. SuperNICs add programmable offload on top.
- NVLink
- Nvidia’s high-speed interconnect that ties GPUs and CPUs into one system.
- CUDA
- Nvidia’s GPU programming platform and software ecosystem.
- Arm
- CPU instruction set architecture and IP licensing ecosystem.
- ISA
- Instruction set architecture. The abstract contract between software and CPU.
- Neoverse
- Arm’s data-center CPU platform family.
- Hyperscaler
- Major cloud or data-center operator.
- AI factory
- Data-center system optimised to produce training and inference compute.
- HBM
- High-bandwidth memory. Stacked DRAM next to the accelerator on advanced packaging.
- CoWoS
- TSMC’s advanced packaging family used to integrate logic and HBM.
- TCO
- Total cost of ownership across compute, networking, power, and operations.
XII. What could break the thesis
A serious piece needs counterarguments. The case for Nvidia as the AI factory company has more than one honest failure mode.
- Hyperscaler in-house silicon. AWS, Google, Microsoft, and Meta are explicitly trying to reduce Nvidia dependence for inference and selected training workloads.
- AMD catches up at system level. If AMD’s software and rack-scale designs close enough of the gap, the Nvidia premium compresses.
- Ethernet at scale. If standards-based Ethernet alternatives become credible at AI fabric scale, Nvidia’s networking leverage softens.
- CUDA alternatives improve. ROCm, Triton, SYCL, OpenXLA, and custom compiler stacks all chip at the edges of the CUDA moat.
- Export controls. Restrictions limit Nvidia’s addressable market in some regions and pull demand toward local alternatives.
- Supply constraints. TSMC, HBM, and CoWoS limits cap how fast Nvidia can ship, regardless of demand.
- Lock-in fatigue. Customers grow uneasy about full-stack lock-in and price aggressively against it.
- Power and cooling. Grid capacity and liquid cooling could become the binding constraints, not silicon.
- Arm ecosystem competition. A thriving Arm ecosystem ironically gives Nvidia’s rivals a stronger CPU foundation.
- Inference efficiency wins. Quantization, sparsity, and better KV management reduce demand for the highest-end systems per token.
The correct claim is not that Nvidia has permanent control. The correct claim is that Nvidia currently sells the most complete AI factory stack.
XIII. The AI factory endgame
In 2020, the question was whether Nvidia would buy Arm and control the future of computing. In 2026, the answer is stranger. Nvidia did not buy Arm, but it still built the future it wanted.
It built CPUs without owning Arm. It built DPUs through BlueField. It built networking through Mellanox. It built the accelerator layer through GPUs. It built the programming model through CUDA. It built the system layer through NVLink and NVL racks. It built the deployment layer through DGX and enterprise software.
Arm survived. Hyperscalers fought back. The ecosystem did not collapse. But Nvidia still changed the shape of the data center.
That is the real 2026 update.
1 Patel, D. (Aug 2020). Jensen Huang’s Vision For Data Center Dominance Will Destroy The Arm Ecosystem. SemiAnalysis. Historical anchor for the CPU + GPU + DPU framework and the data-center-as-computer thesis. Used as inspiration only. No content, structure, or charts reproduced.
2 Nvidia (Feb 2022). Nvidia and SoftBank Group Announce Termination of Nvidia’s Acquisition of Arm Limited. Acquisition terminated in February 2022 due to significant regulatory challenges, with Arm preparing for a public offering.
3 US Federal Trade Commission. Nvidia / Arm matter. FTC challenge over competition concerns and dismissal of the matter after the deal’s termination.
4 Nvidia. Vera Rubin NVL72. 72 Rubin GPUs, 36 Vera CPUs, ConnectX-9 SuperNICs, BlueField-4 DPUs, NVLink 6, and scale-out networking presented as a single rack-scale system.
5 Nvidia. Grace CPU. Grace using high-performance Arm-based CPU cores, Vera using custom Olympus Arm-compatible cores, and the Nvidia Scalable Coherency Fabric.
6 Nvidia Developer Blog. Vera CPU for AI factories. Olympus framed as Nvidia’s first fully custom data-center CPU core, with detail on bandwidth, efficiency, and the AI factory framing.
7 Nvidia (FY2026). Q4 and FY2026 financial results. Q4 FY2026 revenue $68.1B, full-year FY2026 revenue $215.9B, Q4 Data Center revenue $62.3B up 75% year over year.
8 Arm Holdings (FYE26). Arm Q4 FYE26 results. Q4 FYE26 revenue $1.49B and full-year FYE26 revenue $4.92B, alongside cloud and data-center adoption commentary.
9 Google Cloud (2024). Introducing Google’s new Arm-based CPU (Axion). Google’s first custom Arm-based CPU for the data center, with performance and energy efficiency claims for Google Cloud.
10 Microsoft Azure. Cobalt 100 overview. Microsoft Azure’s Cobalt 100 CPU built on Arm Neoverse N2 for general-purpose cloud workloads.
11 AWS. AWS Graviton and AWS Graviton5 announcement. Graviton family on EC2 with successive generations and AWS’s public statements that Graviton powers a growing share of new AWS CPU capacity.
12 Arm. Neoverse CSS V3 and Cloud and AI data center. Neoverse and CSS as the neutral CPU IP foundation across hyperscaler and merchant silicon designs.
13 TSMC. 2025 Annual Report. AI/HPC as the structural driver of advanced process demand, advanced packaging investment, and roadmap context for N2, N2P, A16, and A14 nodes.
14 ASML (2025). 2025 Annual Report, strategic report section. AI requiring leading-edge processor chips and significant additional DRAM, with advanced Logic and DRAM driving further EUV lithography exposures and spending.
- The AI Memory Wall. Companion essay on DRAM, HBM, packaging, and semicap as the new center of computing.
- The Dry Resist War. Companion essay on Lam’s Aether dry resist and why patterning is central to AI-era chipmaking.
- MediaTek and the Fragmented Compute War. Companion essay on a Taiwan fabless platform between Android, edge AI, automotive, and hyperscaler ASICs.
- Nvidia Investor Relations. Quarterly results, segment breakdown, and rack-scale system disclosures.
- Arm Newsroom. Quarterly results and Neoverse / CSS adoption updates.
- The AI Field Manual. Reference layer for the AI stack: hardware, memory, models, agents, safety, economics.
This is Essay No. 015. The topics: intelligence, AI, systems, knowledge, and the questions underneath the questions everyone else is asking. If you read this far and disagreed with any part of it, write to me. I read everything.