Essay No. 073  ·  AI Infrastructure / Platform Strategy
Nvidia Intel AI Infrastructure Edge AI OpenVINO vRAN CUDA Data Centers Semiconductors Software Moats

Nvidia's Empire Was Not Eroded. But Intel's Network and Edge Playbook Still Matters. CUDA OpenVINO Open Edge Platform FlexRAN OpenRAN IPDK P4 BlueField DOCA AI-RAN Vera Rubin

Nvidia won the AI training empire. But the edge, telecom, infrastructure processing, private inference, and mixed-hardware world still rewards open abstraction, standard hardware, and deployment flexibility.

PM
PUGALENTHI MAGENDRAN
May 27, 2026  ·  Research memo  ·  Updating a 2022 Nvidia vs Intel Network and Edge thesis
16 MIN
Thesis
The 2022 article was right about the playbook, but wrong if read as a direct prediction that Intel would erode Nvidia's core AI empire. Nvidia became much stronger in data-center AI. But Intel's Network and Edge strategy still matters because not every workload wants Nvidia's full-stack platform. Edge AI, vRAN, infrastructure processing, private inference, AI PCs, and mixed-hardware deployments reward open abstraction, standard hardware, integrated accelerators, and low-TCO deployment. The anti-Nvidia playbook is not a better CUDA clone. It is open infrastructure where customers keep hardware choice.
In this essay
  1. Historical context: what the 2022 article got right
  2. Nvidia's empire was not eroded
  3. Intel did not win frontier AI training
  4. The real playbook: open abstraction
  5. OpenVINO is not a CUDA killer
  6. Intel's modern edge strategy is Open Edge Platform
  7. Telecom is still Intel's strongest battlefield
  8. IPDK and P4 still explain the infrastructure-processing playbook
  9. But Nvidia learned the playbook too
  10. Vera Rubin shows Nvidia's full-stack direction
  11. The Nvidia-Intel partnership changes the story
  12. Where Nvidia wins vs where Intel still matters
  13. The honest verdict
  14. Evidence ledger
  15. Risks and limitations
  16. Bottom line
  17. Glossary
  18. Sources and method notes
Executive summary
  • The 2022 article argued that Nvidia's hardware-software moat could be challenged by Intel's Network and Edge playbook.
  • That playbook was based on open standards, abstraction layers, compilers, containers, standard hardware, and workload-specific accelerators.
  • The direct "Nvidia will be eroded" prediction did not happen. Nvidia's data-center AI business became vastly larger.
  • But the playbook still matters outside frontier AI training: edge AI, vRAN, private inference, AI PCs, infrastructure processing, and mixed-hardware deployments.
  • The real anti-Nvidia strategy is not building a better CUDA clone. It is building open infrastructure where customers keep hardware choice.

Section 1  ·  Historical frameWhat the 2022 article got right

The 2022 SemiAnalysis piece, How Nvidia's Empire Could Be Eroded — Intel Network And Edge Has The Playbook, opened with two observations.[1] Nvidia's power came from software-hardware co-optimization, with CUDA making Nvidia hardware easier to program and harder to leave. And Nvidia was expanding far beyond GPUs into automotive, enterprise AI software, Omniverse, and full-stack platforms. Page 2 of the public excerpt framed Nvidia as a full-stack platform company across gaming, chips and systems, AI enterprise software, Omniverse, and automotive.[1]

Against that, the article positioned Intel's Network and Edge organization as having a different playbook. Page 4 walked through Intel's network end-to-end model from device to cloud. Page 5 mapped OpenRAN deployment architectures across distributed, centralized, and integrated RAN. Page 10 separated datacenter AI inference, datacenter AI training, edge inference, and edge training. Page 15 detailed Mount Evans and the IPU architectural breakdown. Page 18 illustrated the AI and HPC scaling-up communication problem with all-to-all and all-reduce patterns. Page 21 showed an Intel edge stack with applications, reference code, messaging, Docker, orchestration, an operating system, and Intel hardware blocks underneath.[1] Across those visuals, the argument was consistent. The counter to Nvidia's closed stack was not another closed stack. It was open abstraction: OpenRAN, FlexRAN, OpenVINO, IPDK, P4, Mount Evans IPU, Open Visual Cloud, SmartEdge, and the edge software stack.[1]

Four years later, the playbook description still reads accurately. The prediction is what needs updating.

Section 2  ·  EmpireNvidia's empire was not eroded

Nvidia's reported FY2026 revenue was approximately US$215.9 billion, with data-center revenue at approximately US$193.7 billion.[2] Q1 FY2027 revenue was approximately US$81.6 billion, with data-center revenue at approximately US$75.2 billion.[3] Those are not erosion numbers. They are the financial signature of platform dominance at industrial scale.

Nvidia FY2026 revenue
~ US$215.9B
Fiscal-year revenue, per Nvidia.
FY2026 data center
~ US$193.7B
Data center segment FY2026.
Q1 FY2027 revenue
~ US$81.6B
Quarterly Nvidia revenue.
Q1 FY2027 data center
~ US$75.2B
Q1 FY2027 data center compute and networking.

The 2022 article was right about the playbook. It was wrong if read as a direct prediction that Intel would weaken Nvidia's core AI training empire.

Section 3  ·  Frontier AIIntel did not win the frontier AI training battle

Intel's own 2025 Form 10-K does not hide the picture. Intel disclosed that customers prioritized GPU systems from competitors for compute-heavy generative AI workloads, often at the expense of CPU investment, and that Intel's data-center business faced pressure from GPU-led AI spending.[4] Intel did not become a meaningful high-end AI GPU competitor over this window. The frontier AI training battlefield is not where the 2026 version of this argument should be fought.

Intel 2025 Form 10-K admission
Intel disclosed that customers prioritized GPU systems from competitors for compute-heavy generative AI workloads, often at the expense of CPU investment. That is Intel telling investors the same thing the Nvidia numbers tell the market: CUDA-driven GPU systems captured the high-end AI training spend.

CUDA won the cloud training war. That does not mean CUDA wins every edge, network, and private inference deployment.

Section 4  ·  The playbookThe real playbook: open abstraction

Intel's Network and Edge strategy is not replace Nvidia with another closed Intel stack. It is abstraction. The pattern is consistent across the portfolio. Take fragmented specialized workloads. Move them onto standard hardware. Add just enough acceleration. Use open software and compilers to hide hardware complexity. Let customers keep deployment flexibility. The names change across markets but the shape does not: FlexRAN for virtualized radio workloads, OpenVINO for inference deployment, IPDK for infrastructure offload, P4 for programmable packet processing, Open Edge Platform for edge orchestration and lifecycle management.

The anti-Nvidia playbook is not a better CUDA clone. It is open infrastructure.

Section 5  ·  OpenVINOAn edge inference abstraction layer, not a CUDA killer

OpenVINO's job is not to beat Blackwell in frontier training. Its job is to make inference easier across Intel CPUs, GPUs, NPUs, and edge systems. The 2025.4 release notes describe scheduling optimizations for Intel Core Ultra Series 3, simplified NPU deployment, Qwen3-MoE support, BitNet optimizations on Xeon and client processors, and Triton backend deployment with Intel GPUs and NPUs.[6] Those are not headline AI lab claims. They are the work of an inference runtime designed for factories, retail, cameras, smart cities, local LLMs, AI PCs, industrial edge, private inference, and mixed CPU / GPU / NPU systems.

OpenVINO 2025.4 highlights  ·  what an edge-first runtime ships
Core Ultra S3
Scheduling optimizations for Intel Core Ultra Series 3 client and edge platforms.
NPU deployment
Simplified deployment direction on Intel NPUs for on-device inference.
Qwen3-MoE
Support added for Qwen3-MoE-class models in the runtime.
BitNet
Optimizations on Xeon and client processors for BitNet-class quantized models.
Triton backend
OpenVINO usable as a Triton backend for Intel GPU and NPU serving.
Mixed hardware
A single inference stack across CPUs, GPUs, NPUs, and FPGAs.

CUDA wins when the workload wants a giant accelerator cluster. OpenVINO matters when the workload wants to run wherever the customer already has compute.

Section 6  ·  Open Edge PlatformEdge AI as managed infrastructure

Intel's Open Edge Platform supports secure onboarding and management of edge-node fleets, dynamic application deployment, near-zero-touch provisioning, orchestration and lifecycle management, and OpenVINO runtime optimizations for edge-to-cloud inference.[5] Edge AI is not only a model-serving problem. It is a deployment, fleet management, update, security, and lifecycle problem. The hard parts at the edge are not training a model. They are securely bringing thousands of devices online, updating them safely, and keeping them current without site visits.

Intel Open Edge Platform  ·  what it provides
Secure onboarding
Authenticated bring-up and identity for edge-node fleets.
Dynamic apps
Dynamic application deployment across edge nodes.
Zero-touch
Near-zero-touch provisioning to reduce site visits.
Orchestration
Orchestration and lifecycle management for fleets at scale.
OpenVINO runtime
OpenVINO runtime optimizations for edge-to-cloud inference.
Fleet security
Security and updates designed for distributed deployments.

Nvidia wants to define the AI factory. Intel wants to make edge AI look like open, manageable infrastructure.

Section 7  ·  TelecomTelecom is still Intel's strongest battlefield

The 2022 article's telecom section was strong because it showed how Intel used standard servers, virtualization, containers, and FlexRAN to move telecom workloads away from closed appliance-style systems. The OpenRAN diagrams and the Layer 1 acceleration trade-offs between performance, energy efficiency, and programmability still describe the strategic shape of the market.[1] Intel's MWC Barcelona 2026 press kit extends that argument with Xeon 6 SoC positioned for network and edge solutions in an AI-driven world, with Intel vRAN Boost and Intel AMX as the acceleration story across vRAN, media, AI, and network security, alongside a Xeon 6+ and Clearwater Forest-AP direction aimed at high-density network and edge workloads.[7]

Xeon 6 for network and edge  ·  what Intel is selling
Xeon 6 SoC
Designed for network and edge solutions in an AI-driven world.
vRAN Boost
Integrated vRAN acceleration on the SoC to consolidate workloads.
AMX
Advanced Matrix Extensions for inference and signal processing.
vRAN coverage
Targeting vRAN, media, AI, and network security workloads.
Roadmap
Xeon 6+ and Clearwater Forest-AP direction for high-density edge workloads.
Consolidation
Standard servers replacing closed telecom appliances.

Intel's strongest anti-Nvidia argument is not peak AI performance. It is consolidation.

Section 8  ·  InfrastructureIPDK and P4 still explain the infrastructure-processing playbook

IPDK is described by its project page as an open-source, vendor-agnostic framework of drivers and APIs for infrastructure offload and management, capable of running on CPU, IPU, DPU, or switch silicon, and built on top of tools such as SPDK, DPDK, and P4. It targets network virtualization, storage virtualization, workload provisioning, root-of-trust, and offload capabilities.[8] That is the same anti-lock-in logic the 2022 article identified: make infrastructure programmable across hardware, so customers do not get locked into one vendor's smart-NIC, DPU, or IPU silicon.

IPDK  ·  what it covers
Open and vendor-agnostic
Open-source framework of drivers and APIs, not tied to one silicon vendor.
Hardware targets
Runs on CPU, IPU, DPU, or switch silicon as the deployment target.
Building blocks
Uses SPDK, DPDK, and P4 to compose infrastructure offload pipelines.
Network virt
Network virtualization across the deployment target.
Storage virt
Storage virtualization with consistent APIs.
Workload + trust
Workload provisioning, root-of-trust, and offload capabilities.

The best way to fight a proprietary hardware moat is to move the abstraction boundary above the hardware.

Section 9  ·  Counter-adaptationBut Nvidia learned the playbook too

The 2022 contrast between Intel open abstraction and Nvidia closed stack is still useful, but it is less clean now. Nvidia has adapted. BlueField-4 combines a Grace CPU and ConnectX-9 networking, supporting AI factories described by Nvidia as up to 4x larger than BlueField-3, with DOCA microservices for networking, storage, security, cloud elasticity, service chaining, and multi-tenant infrastructure.[9] Nvidia's Aerial CUDA-Accelerated RAN runs on Red Hat OpenShift for AI-native 5G and 6G RAN development.[10] Read together, those are Nvidia adopting the language of Kubernetes, cloud-native infrastructure, AI-RAN, storage, security, and service chaining when it suits the platform.

Intel's playbook was right. Nvidia noticed.

Section 10  ·  Vera RubinNvidia's full-stack direction

Vera Rubin makes the direction explicit. Vera Rubin NVL72 combines 72 Rubin GPUs and 36 Vera CPUs, with NVLink 6, ConnectX-9 SuperNICs, and BlueField-4 DPUs, scaling up inside a rack-scale platform and scaling out through Nvidia networking.[11][12] The product page anchors the rack-scale AI infrastructure framing with Quantum-X800 InfiniBand and Spectrum-X Ethernet.[12] Nvidia is integrating compute, networking, security, communication, and system architecture into one platform. The trajectory is from GPU vendor to AI factory platform owner.

Nvidia wins by making the system one platform.

Section 11  ·  PartnershipThe Nvidia-Intel partnership changes the story

In September 2025, Nvidia and Intel announced a partnership in which Intel will build Nvidia-custom x86 CPUs for Nvidia AI infrastructure platforms, Intel will build x86 SoCs integrating Nvidia RTX GPU chiplets for PCs, and Nvidia agreed to invest US$5 billion in Intel common stock, subject to closing conditions.[13] That is not the relationship the 2022 article described. By 2026, the story is more complex: Intel may be both a counterweight to Nvidia in fragmented edge and network markets, and a component supplier inside Nvidia's data center and PC platforms.

Custom x86 for Nvidia AI
Intel-built
Intel will build Nvidia-custom x86 CPUs for Nvidia AI infrastructure platforms.
x86 + RTX chiplets
PC SoCs
Intel will build x86 SoCs integrating Nvidia RTX GPU chiplets for PCs.
Nvidia investment
US$5B
Nvidia agreed to invest US$5B in Intel common stock, subject to closing conditions.
Linkage
NVLink
Announcement framed around NVLink integration between Nvidia and Intel architectures.

Intel did not erode Nvidia's core platform. Nvidia found a way to pull Intel's x86 ecosystem into its own roadmap.

Section 12  ·  MatrixWhere Nvidia wins vs where Intel still matters

Market Nvidia advantage Intel advantage Likely outcome
Frontier AI training CUDA, HBM, NVLink, Blackwell/Rubin, cloud availability Weak high-end GPU position Nvidia dominates
Large AI factories Full-stack racks, networking, BlueField, NVLink x86 partnership, CPUs, edge-to-cloud integration Nvidia leads, Intel may supply pieces
Edge inference Strong GPUs, Jetson, CUDA ecosystem OpenVINO, x86 installed base, NPUs, standard hardware Contested
Industrial AI GPU acceleration Power, manageability, Open Edge Platform, existing x86 Intel can win selective deployments
vRAN Aerial, GPU acceleration FlexRAN, vRAN Boost, Xeon, telecom history Contested
Infrastructure processing BlueField + DOCA IPDK, P4, IPUs, open abstraction Contested
AI PCs RTX ecosystem Core Ultra, NPUs, x86, Windows PC footprint Mixed, partnership changes the story
Private inference Nvidia Enterprise stack x86 servers, OpenVINO, cost and data control Fragmented

Section 13  ·  Honest verdictThe verdict

The 2022 article was too optimistic if read as Intel will erode Nvidia's core AI empire. That did not happen. Nvidia became much stronger. But the article was directionally right about where Nvidia is vulnerable, even now.

Where the 2022 playbook still applies in 2026
  1. Fragmented workloads that do not fit one giant accelerator cluster.
  2. Local deployment, including private inference inside enterprise data centers.
  3. Power-sensitive environments where every watt of inference matters.
  4. Latency-sensitive workloads that cannot tolerate a round-trip to the cloud.
  5. Telecom infrastructure, where vRAN and AI-RAN converge on standard servers.
  6. Private inference, where customers prefer to keep models and data in house.
  7. Mixed hardware environments combining CPUs, GPUs, NPUs, FPGAs, and DPUs.
  8. Cost-sensitive edge deployments where TCO beats peak performance.
  9. Customers that explicitly refuse full-stack lock-in to any single vendor.

Nvidia wins by making everything one platform. Intel can still win where customers refuse to make everything one platform.

Section 14  ·  EvidenceEvidence ledger

Claim
Evidence
Interpretation
The 2022 article identified Intel's open playbook
The uploaded SemiAnalysis piece focused on FlexRAN, OpenVINO, IPDK, P4, OpenRAN, Mount Evans IPU, edge software, and hardware abstraction.
Intel's counter to Nvidia was abstraction, not CUDA cloning.
Nvidia's empire was not eroded
Nvidia FY2026 revenue was approximately US$215.9B with data-center revenue at approximately US$193.7B.
Nvidia's core AI platform became much stronger.
Nvidia's current scale is even larger
Q1 FY2027 revenue was approximately US$81.6B with data-center revenue at approximately US$75.2B.
Nvidia dominates the AI infrastructure profit pool.
Intel's own filing admits the GPU shift
Intel's 2025 Form 10-K says customers prioritized GPU systems from competitors for generative AI, often at the expense of CPU investment.
Intel did not win the core AI training battle.
OpenVINO is still evolving
OpenVINO 2025.4 adds Core Ultra Series 3 scheduling, Qwen3-MoE, BitNet, NPU and Triton backend improvements.
Intel's edge inference abstraction layer remains alive.
Intel is pushing Open Edge
Intel Open Edge Platform includes provisioning, orchestration, lifecycle management, and OpenVINO edge-to-cloud optimization.
Intel's modern playbook is edge infrastructure management.
Xeon 6 keeps the vRAN strategy alive
Intel's MWC 2026 materials highlight Xeon 6 SoC, vRAN Boost, AMX, and network/edge workloads.
Intel still has a credible telecom and network-edge wedge.
Nvidia learned the playbook too
BlueField-4 and Aerial on Red Hat OpenShift show Nvidia moving into DPUs, software-defined infrastructure, AI-RAN, and cloud-native ecosystem language.
Nvidia is attacking Intel's edge and network territory.
Intel and Nvidia now collaborate
The Nvidia-Intel partnership announcement describes Intel-built Nvidia-custom x86 CPUs and x86 SoCs integrating Nvidia RTX chiplets, plus a US$5B Nvidia investment in Intel common stock.
Intel may become both counterweight and component inside Nvidia's platform.

Section 15  ·  Risk registerRisks and limitations

This essay is an analysis of public disclosures and historical context. It is not investment advice. The honest risks against the read above run in several directions, and they are listed here so the argument can be stress-tested.

CUDA's software moat is wider than any single open-abstraction story can dislodge in the short term. The "open infrastructure wins fragmented markets" argument should not be read as "CUDA loses cloud training."
Nvidia adopts open standards and cloud-native ecosystems when it suits the platform (DPUs, OpenShift, Kubernetes). "Closed" vs "open" is a useful frame, not a clean binary.
Intel still has to execute on Xeon 6, Xeon 6+, vRAN Boost, NPUs, and Open Edge Platform at scale. Strategy framing does not substitute for product delivery.
AMD, custom hyperscaler silicon, and ASIC platforms also pressure Nvidia in some segments, so "Intel vs Nvidia" oversimplifies the competitive picture.
vRAN and AI-RAN economics depend on telco capex cycles and regulatory decisions that neither company fully controls.
The Nvidia-Intel partnership is subject to closing conditions and may evolve. Treat it as a directional signal, not a finalized roadmap.
Open-source projects (IPDK, OpenVINO) depend on continued contributor investment. Roadmaps can slow if industry attention shifts.
Hyperscalers may consolidate around fewer abstractions over time, narrowing the surface area where open-infrastructure strategies pay off.
An AI demand normalisation could compress both Nvidia and Intel opportunities asymmetrically. Concentration-driven moats and fragmentation-driven counter-strategies do not move together.
Performance and capability claims from vendor materials are vendor claims. Field results across diverse workloads may differ from benchmark headlines.

Section 16  ·  Bottom lineBottom line

Bottom line

The 2022 article was right about the playbook, but wrong if read as a direct prediction that Intel would erode Nvidia's core AI empire. Nvidia became much stronger in data-center AI.

Intel's Network and Edge strategy still matters because not every workload wants Nvidia's full-stack platform. Edge AI, vRAN, infrastructure processing, private inference, AI PCs, and mixed-hardware deployments reward open abstraction, standard hardware, integrated accelerators, and low-TCO deployment.

The anti-Nvidia playbook is not a better CUDA clone. It is open infrastructure where customers keep hardware choice.

Section 17  ·  DefinitionsGlossary

CUDA
Nvidia's parallel computing platform and programming model that turns GPUs into a general-purpose accelerated computing platform.
OpenVINO
Intel's open-source inference runtime designed to run efficiently across Intel CPUs, GPUs, NPUs, and FPGAs, with bindings into ecosystem servers like Triton.
Open Edge Platform
Intel's platform for secure onboarding, dynamic deployment, near-zero-touch provisioning, orchestration, and lifecycle management of edge-node fleets.
oneAPI
Intel's unified programming model spanning CPUs, GPUs, FPGAs, and other accelerators using a common API and toolchain.
FlexRAN
Intel's reference architecture and software for running radio-access-network workloads on Intel servers, supporting vRAN and OpenRAN deployments.
OpenRAN
An open radio access network architecture allowing multi-vendor, software-defined RAN deployments instead of closed proprietary appliances.
vRAN
Virtualized radio access network. Running RAN functions in software on standard servers rather than dedicated hardware.
AI-RAN
A direction that integrates AI workloads into the RAN, often using accelerators alongside traditional baseband processing.
IPDK
Infrastructure Programmer Development Kit. An open-source framework of drivers and APIs for infrastructure offload across CPU, IPU, DPU, or switch silicon.
P4
A programming language for programmable packet processing. Used in IPDK pipelines and in modern data-plane silicon.
DPU
Data Processing Unit. A class of silicon that offloads networking, storage, and security from CPUs in modern data center servers.
IPU
Infrastructure Processing Unit. Intel's term for a class of silicon similar to DPUs, used for infrastructure offload at the server boundary.
DOCA
Nvidia's software framework and microservices stack for BlueField DPUs, covering networking, storage, security, and cloud elasticity.
BlueField
Nvidia's family of DPUs. BlueField-4 combines a Grace CPU with ConnectX-9 networking and DOCA microservices for AI factories.
AMX
Advanced Matrix Extensions. Intel's matrix-acceleration instructions used to accelerate inference and signal processing on Xeon.
vRAN Boost
Intel's integrated vRAN acceleration on Xeon-class SoCs, designed to consolidate radio workloads without external accelerator cards.
Edge inference
Running AI inference at or near the source of data (factories, cameras, retail, vehicles), as opposed to in a central cloud.
Private inference
Inference run on infrastructure controlled by the customer, often inside an enterprise data center, for data-control or cost reasons.
Hardware abstraction
A software layer that hides differences between hardware platforms so applications can run across multiple kinds of silicon with less rework.
Platform lock-in
A state in which switching cost from one vendor's platform to another is high enough that customers stay even when alternatives improve.
Open infrastructure
An approach that uses open standards, open-source software, and abstraction layers to keep deployment flexibility across multiple hardware vendors.

Section 18  ·  MethodSources and method notes

How this essay reads sources

The 2022 SemiAnalysis piece is treated as historical context for the Nvidia full-stack platform framing, Intel's Network and Edge open-abstraction playbook, OpenRAN, FlexRAN, OpenVINO, IPDK, P4, Mount Evans IPU, edge software, the AI hardware market separation across training and inference, and the strategic argument that open abstraction is the way to limit Nvidia-style lock-in.

The 2026 read is built from primary corporate disclosures: Nvidia FY2026 and Q1 FY2027 results, Intel's 2025 Form 10-K, Intel's Open Edge Platform pages, OpenVINO 2025.4 release notes, Intel's MWC Barcelona 2026 press kit, the IPDK project page, Nvidia's BlueField-4 announcement, Red Hat's Aerial on OpenShift blog, the Vera Rubin platform announcement, the Vera Rubin NVL72 product page, and the Nvidia-Intel partnership announcement. Company claims about revenue, performance, capability, and roadmap are treated as company claims, not as endorsed forecasts. The structural arguments that Nvidia won the core AI training battle, that the open-abstraction playbook still matters at the edge and in fragmented markets, and that the Nvidia-Intel relationship is now both competitive and collaborative are independent analysis.

Footnotes  ·  primary sources

  1. SemiAnalysis, “How Nvidia's Empire Could Be Eroded — Intel Network And Edge Has The Playbook,” 2022 (PDF supplied by author). Historical anchor used in this essay for the page 2 Nvidia full-stack platform framing (gaming, chips and systems, AI enterprise software, Omniverse, automotive), the page 4 Intel network end-to-end visual, the page 5 OpenRAN deployment architectures across distributed, centralized and integrated RAN, the page 10 AI hardware market separation into datacenter training, datacenter inference, edge training, and edge inference, the page 15 Mount Evans IPU architectural breakdown, the page 18 scaling AI / HPC communication framing with all-to-all and all-reduce patterns, and the page 21 Intel edge software stack visual covering applications, reference code, messaging, Docker, orchestration, OS, and Intel hardware blocks.
  2. Nvidia, “Nvidia Announces Financial Results for Fourth Quarter and Fiscal 2026,” nvidianews.nvidia.com/…/fy2026. Source for Nvidia FY2026 revenue of approximately US$215.9B and FY2026 data-center revenue of approximately US$193.7B used in this essay.
  3. Nvidia Investor Relations, “Financial Reports,” investor.nvidia.com/…/financial-reports. Source for Nvidia Q1 FY2027 revenue of approximately US$81.6B and data-center revenue of approximately US$75.2B used in this essay.
  4. Intel Corporation, 2025 Form 10-K sec.gov/…/intc-20251227. Source for Intel's disclosure that customers prioritized GPU systems from competitors for compute-heavy generative AI workloads, often at the expense of CPU investment, and the broader Intel data-center and AI context cited in this essay.
  5. Intel, “Open Edge Platform,” intel.com/…/edge-platform. Source for the Intel Open Edge Platform capabilities used in this essay, including edge-node fleet management, dynamic application deployment, near-zero-touch provisioning, orchestration, lifecycle management, and OpenVINO edge-to-cloud inference optimization.
  6. Intel, “OpenVINO 2025.4 Release Notes,” intel.com/…/openvino-2025-4-release-notes. Source for the OpenVINO 2025.4 highlights used in this essay including Core Ultra Series 3 scheduling, NPU deployment improvements, Qwen3-MoE support, BitNet optimizations on Xeon and client processors, and Triton backend deployment with Intel GPUs and NPUs.
  7. Intel, “Intel at MWC Barcelona 2026 Press Kit,” newsroom.intel.com/…/mwc-2026-press-kit. Source for the Xeon 6 SoC positioning for network and edge solutions in an AI-driven world, Intel vRAN Boost and Intel AMX in vRAN, media, AI, and network security workloads, and the Xeon 6+ / Clearwater Forest-AP direction for high-density network and edge workloads used in this essay.
  8. IPDK Project, ipdk.io. Source for the description of IPDK as an open-source, vendor-agnostic framework of drivers and APIs for infrastructure offload and management, running on CPU, IPU, DPU, or switch, and using SPDK, DPDK, and P4 to compose network virtualization, storage virtualization, workload provisioning, root-of-trust, and offload pipelines.
  9. Nvidia, “BlueField-4 AI Factory,” blogs.nvidia.com/…/bluefield-4-ai-factory. Source for BlueField-4 combining a Grace CPU and ConnectX-9 networking, supporting AI factories up to 4x larger than BlueField-3, and the DOCA microservices framing for networking, storage, security, cloud elasticity, service chaining, and multi-tenant infrastructure.
  10. Red Hat, “Using Nvidia Aerial CUDA-Accelerated RAN on Red Hat OpenShift to Accelerate Development of AI-Native 5G and 6G RAN Solutions,” redhat.com/…/nvidia-aerial-openshift. Source for Nvidia Aerial CUDA-Accelerated RAN running on Red Hat OpenShift for AI-native 5G and 6G RAN development, used in this essay to show Nvidia adopting cloud-native ecosystem positioning where useful.
  11. Nvidia, “Nvidia Vera Rubin Platform,” nvidianews.nvidia.com/…/vera-rubin-platform. Source for Vera Rubin NVL72 combining 72 Rubin GPUs, 36 Vera CPUs, NVLink 6, ConnectX-9 SuperNICs, and BlueField-4 DPUs in a rack-scale AI supercomputer platform.
  12. Nvidia, “Vera Rubin NVL72,” nvidia.com/…/vera-rubin-nvl72. Source for the Vera Rubin NVL72 product framing including Rubin GPUs, Vera CPUs, ConnectX-9, BlueField-4, NVLink 6, Quantum-X800 InfiniBand, and Spectrum-X Ethernet in the rack-scale AI infrastructure positioning used in this essay.
  13. Nvidia, “Nvidia and Intel to Develop AI Infrastructure and Personal Computing Products,” nvidianews.nvidia.com/…/nvidia-intel-partnership. Source for Intel building Nvidia-custom x86 CPUs for Nvidia AI infrastructure platforms, Intel building x86 SoCs integrating Nvidia RTX GPU chiplets for PCs, Nvidia's US$5B agreed investment in Intel common stock subject to closing conditions, and the NVLink integration framing between Nvidia and Intel architectures.
Read next  ·  related essays
Essay 072
Nvidia's Hopper Supplier Cut Was a Warning: AI Hardware Design Wins Are Not Moats
Essay 069
Nvidia Was Hacked in 2022. The Real Lesson Is AI Infrastructure Security.
Essay 058
Intel Bet the Farm. Now It Has to Prove the Foundry
Essay 060
AWS Turned the CPU Into Cloud Infrastructure
Essay 061
The Package Became the Computer
Essay 056
The Node Race Became a System Race
← All essays