Essay No. 078  ·  Semiconductors / Arm Servers / Cloud-Native CPUs
Ampere Computing Arm Servers Cloud Native CPUs SoftBank AI Infrastructure AmpereOne Oracle Cloud Graviton Cobalt Axion Semiconductors

Ampere's Cloud-Native Pitch Wasn't Fluff. The IPO Story Became a SoftBank AI Infrastructure Bet. Ampere AmpereOne AmpereOne M Cloud-native CPUs Arm Servers Oracle A4 Qualcomm AI 100 SoftBank Graviton Cobalt Axion AI Inference

Ampere was right that cloud fleets reward predictable throughput per watt per dollar. The problem is that AWS, Microsoft, and Google learned the same lesson and built their own silicon.

PM
PUGALENTHI MAGENDRAN
May 27, 2026  ·  Research memo  ·  Updating a 2022 cloud-native Arm CPU thesis
16 MIN
Thesis
The 2022 article was right that Ampere's cloud-native strategy had real technical substance: many small cores, one thread per core, predictable performance, low cost per core, and high throughput density. But by 2026, Ampere's thesis became too correct. AWS, Microsoft, and Google built their own Arm CPUs, while AMD and Intel copied the dense-core direction. Ampere still has real proof points through AmpereOne, Oracle A4, the Qualcomm inference partnership, and the Aurora roadmap, but the independent IPO story gave way to SoftBank's US$6.5 billion acquisition. Ampere's future is no longer just merchant cloud CPUs. It is whether SoftBank can turn Ampere into a power-efficient CPU and AI infrastructure layer inside the Arm ecosystem.
Executive summary
  • Ampere's 2022 "cloud native" pitch was not pure marketing. It reflected real architecture choices: many smaller cores, one thread per core, no SMT, predictable clocks, private L2, smaller shared L3, and low cost per core.
  • The 2022 article's cost table showed Ampere's cost-per-core advantage under its assumptions, while clearly warning it was not a full TCO model.
  • AmpereOne moved Ampere beyond stock Arm Neoverse cores into custom Arm-compatible server CPU design.
  • The business problem is that hyperscalers copied the thesis: AWS Graviton, Microsoft Cobalt, and Google Axion internalized cloud-native Arm CPUs.
  • The IPO story did not happen. SoftBank acquired Ampere for US$6.5 billion, turning it into a strategic AI infrastructure asset rather than a public merchant CPU company.

Section 1  ·  Historical frameWhat the 2022 article got right

The 2022 SemiAnalysis piece asked whether Ampere's "cloud native" phrase was real or marketing.[1] The conclusion was that the architecture had substance. Ampere targeted cloud workloads, focused on high core counts using smaller cores, avoided SMT and boost-clock unpredictability, used one thread per core to improve predictability and security posture, and chose private L2 with smaller shared L3 to reduce noisy-neighbor behavior. The article also argued that Ampere's cost-per-core advantage came from smaller cores, smaller L3, and fabric choices, while being careful that Ampere lagged Intel and AMD in single-threaded and floating-point-heavy workloads.[1]

2022 PDF page references used in this essay
  • Server-chip manufacturing cost estimate table comparing Intel Ice Lake, AMD Milan, Ampere Altra, and Ampere Altra Max under explicit assumptions.
  • Neoverse N1, V1, and N2 architecture diagrams, useful for explaining the Arm server-core trade-off between small efficient cores and larger high-performance cores.
  • Inferred AmpereOne custom-core block diagram.
  • AmpereOne framing as Armv8.6 with some Armv9.1 features, width similar to N1, doubled L2 cache to 2 MB, two branch units, no SVE.
  • Comparison table contrasting AmpereOne with Neoverse N1, N2, and V1.

Ampere's insight was not that Arm is magically better. It was that cloud fleets reward predictable throughput per watt per dollar.

Section 2  ·  Cloud nativeWhat it meant in silicon

"Cloud native" should not be treated as a slogan. In Ampere's case, it meant a specific set of design choices that show up in the silicon, not the marketing deck.[1]

Ampere's cloud-native design choices
Many cores
More independent workers for parallel cloud services.
One thread per core
Less SMT interference and more predictable tenant performance.
No SMT
Lower side-channel and noisy-neighbor concerns, with possible utilization trade-offs.
Predictable clocks
Less variance across core counts and workloads.
Larger private L2
More per-core isolation and less dependence on shared cache.
Smaller shared L3
Less area spent on a cache structure that does not help every cloud workload.
Smaller cores
Lower cost per core and more throughput density per die.
Lower single-thread focus
Acceptable for horizontally scaling workloads; a weakness for some databases, HPC, and FP-heavy code.

Cloud-native CPUs are not designed to win every benchmark. They are designed to make a cloud fleet easier to pack, price, and predict.

Section 3  ·  CostThe cost-per-core argument was real but incomplete

The 2022 article's page 6 table estimated Ampere Altra and Altra Max as having lower cost per core than AMD Milan and Intel Ice Lake under its modeling assumptions, with inputs covering die size, good dies per wafer, wafer cost, packaging cost, and cores per package.[1] The article was explicit that the model excluded many real factors. Die harvesting, parametric yield, binning, fixed design costs, royalties, and full server TCO were left out by design. The cost table was directional, not final truth.[1]

Important caveat from the 2022 article
The page 6 cost-per-core estimates intentionally excluded die harvesting, parametric yield, binning, fixed design costs, royalties, and full server TCO. Use the table as directional support for Ampere's cost-density argument, not as a comprehensive TCO conclusion.

Ampere had a cost-density advantage, not a universal workload advantage.

Section 4  ·  Trade-offPredictable throughput vs peak per-thread performance

Ampere traded peak single-thread performance for throughput density and predictability. Intel and AMD big cores offer stronger single-thread and floating-point performance. SMT can improve utilization when execution units are waiting on memory. Boost clocks let Intel and AMD push performance under power and thermal budgets. Ampere avoided those features because the cloud-native pitch valued predictability over peaks.[1] That made Ampere strong for many small independent tasks and weaker for workloads that need maximum per-thread performance.

Ampere's design was not better everywhere. It was better where predictability and density mattered more than peak thread speed.

Section 5  ·  AmpereOneBeyond borrowed Neoverse cores

The 2022 article used LLVM and GitHub clues to infer that AmpereOne would carry a custom core, with the page 18 microarchitecture block diagram showing approximately N1-class width, a 2 MB L2, two branch units, and no SVE on the inferred design.[1] Ampere later productized AmpereOne with up to 192 custom Arm ISA-compliant cores and 2 MB L2 cache per core, alongside DDR5, PCIe Gen5, and Armv8.6+ ISA support.[2] AmpereOne moved Ampere from "merchant Neoverse CPU vendor" to "custom Arm server CPU company."

Cores per AmpereOne SoC
96 to 192
Custom Arm ISA-compliant cores per AmpereOne SoC, per Ampere's product brief.
L2 cache per core
2 MB
Private L2 cache per core in AmpereOne family.
ISA
Armv8.6+
Armv8.6+ baseline per Ampere product material.
Memory + IO
DDR5 + PCIe Gen5
DDR5 memory and PCIe Gen5 connectivity in AmpereOne family.

AmpereOne was the moment Ampere stopped being just a Neoverse implementation and became a custom cloud CPU company.

Section 6  ·  Business problemHyperscalers copied the thesis

Ampere validated the cloud-native Arm thesis, but the biggest cloud companies had the scale and incentive to internalize it. AWS Graviton4 is positioned as a custom AWS Arm-based processor for EC2 instances, with cloud-native Arm now mainstream inside AWS.[11] Microsoft Cobalt 100 is described as Microsoft's first fully custom Arm-based Azure CPU.[12] Google Axion powers C4A instances, with Google citing price-performance and energy-efficiency benefits for cloud-native workloads.[13] These chips are not only CPUs. They are part of vertically integrated cloud platforms.

AWS
Graviton4
Custom AWS Arm-based processor for EC2 instances; cloud-native Arm at AWS-internal scale.
Microsoft
Cobalt 100
Microsoft's first fully custom Arm-based Azure CPU, generally available for Azure workloads.
Google
Axion (C4A)
Google's first custom Arm CPU; powers C4A instances with stated price-performance and energy-efficiency claims.

The structural problem for Ampere is straightforward. A hyperscaler controls the chip, host, hypervisor, networking, storage path, pricing, fleet deployment, and workload placement. An independent merchant CPU vendor cannot always match that level of platform control. AWS missing from Ampere's customer list back in 2022 was not a small detail; it was the preview of the strategic problem.

Ampere validated the market, but hyperscalers captured the highest-value version of the market for themselves.

Section 7  ·  OracleThe clearest Ampere proof point

Oracle introduced OCI A4 Standard instances powered by AmpereOne M, positioning A4 for AI inference and cloud-native applications, with A4 Standard shapes that include 96 AmpereOne M cores.[4] Ampere highlights AmpereOne M's 192-core configurations, 12-channel DDR5 memory, 2 MB L2 cache per core, 100G networking on A4, and expanded memory bandwidth aimed at demanding AI inference workloads and LLM inference framing.[3][5]

A4 Standard shape
96 cores
96 AmpereOne M cores per A4 Standard shape, per Oracle.
AmpereOne M peak
192 cores
Up to 192 cores in AmpereOne M family.
Memory channels
12-channel DDR5
12-channel DDR5 memory subsystem on AmpereOne M.
Networking
100G
100G networking on A4 framing per Ampere.

Oracle is the proof that Ampere's merchant CPU story did not disappear. It narrowed.

Section 8  ·  AI inferenceA real pivot, framed carefully

Ampere is not an Nvidia replacement. Its AI relevance is around inference infrastructure and the CPU-heavy work around AI systems. Ampere and Qualcomm announced a joint AI inference solution using Qualcomm Cloud AI 100 inference accelerators and Ampere CPUs, with Ampere also announcing a 256-core CPU roadmap.[6] Reuters framed the Ampere-Qualcomm pairing around reducing the power cost of operating AI chips.[7] The realistic AI workload set for Ampere is the infrastructure layer around AI, not the model itself.

CPU inference for small and medium models
RAG pipelines (retrieval, ranking, orchestration)
Vector databases and embedding stores
Data preprocessing for AI training and inference
Orchestration and control-plane services
Networking and storage path workloads
Multi-tenant inference serving
Cloud-native applications around AI systems

Ampere's AI role is not "beat Nvidia." It is "make inference infrastructure more power-efficient."

Section 9  ·  AuroraThe big roadmap bet

Ampere announced AmpereOne Aurora as a future product roadmap. Ampere says Aurora scales to 512 Ampere cores, integrates Ampere AI IP, supports high-bandwidth memory, and is framed around AI inference and training use cases including RAG and vector databases.[8] Treat Aurora as a roadmap and company claim, not independent validation. The strategic intent is clear regardless: Aurora is Ampere's attempt to move from cloud-native CPU to AI-native infrastructure CPU.

Aurora core target
up to 512
Up to 512 Ampere cores in the announced Aurora roadmap.
AI IP
Integrated
Integrated Ampere AI IP framed for inference workloads.
Memory
HBM
High-bandwidth memory support per Ampere's Aurora framing.
Workload framing
RAG + vectors
RAG and vector database use cases highlighted by Ampere.

Aurora is Ampere's attempt to turn the cloud-native CPU thesis into an AI infrastructure CPU thesis.

Section 10  ·  Dense-core x86AMD and Intel copied the lesson too

The dense-core idea did not remain unique to Ampere. AMD's EPYC portfolio includes dense-core Zen 4c configurations with up to 128 cores in the 4th-generation product family, framed at cloud-native and scale-out workloads.[14] Intel's Xeon 6 family includes efficiency-core server platforms aimed at cloud-native and scale-out workloads alongside performance-core configurations.[15] The cloud-native CPU idea won at the industry level. That made Ampere's differentiation harder, not easier.

The cloud-native CPU idea won. That made Ampere's differentiation harder.

Section 11  ·  SoftBankThe IPO story ended differently

The 2022 article discussed Ampere's IPO prospects. That story did not happen. SoftBank announced a US$6.5 billion all-cash acquisition of Ampere.[9] SoftBank completed the transaction on November 25, 2025 (U.S. time), with Ampere becoming a wholly owned subsidiary of SoftBank inside the broader SoftBank AI infrastructure strategy.[10] The key question is no longer "Can Ampere IPO?" The question is whether SoftBank can use Ampere as part of a broader Arm-centered AI compute strategy.

Deal value
~ US$6.5B
All-cash acquisition value announced by SoftBank.
Structure
All cash
All-cash transaction per Ampere's announcement.
Completion
Nov 25, 2025
Acquisition completed on November 25, 2025 (U.S. time), per SoftBank.
Status
Wholly owned
Ampere is now a wholly owned subsidiary of SoftBank.

Ampere did not get the IPO ending. It got the strategic-infrastructure ending.

Section 12  ·  What SoftBank getsAnd what it has to manage

SoftBank gets more than a CPU startup. It gets an experienced Arm server CPU team, custom cloud CPU IP, a working relationship with Oracle and cloud / OEM customers, a power-efficient AI infrastructure CPU roadmap, a possible complement to Arm and to broader AI data-center ambitions, and a path into server CPU infrastructure without building the team from scratch. It also inherits a set of risks the IPO market would have been pricing.

What SoftBank gets
  • Experienced Arm server CPU team.
  • Custom cloud CPU IP and roadmap.
  • Oracle relationship and cloud / OEM ecosystem.
  • Power-efficient AI infrastructure CPU direction.
  • Possible complement to Arm and SoftBank's AI bets.
  • A path into server CPU infrastructure without building from scratch.
Risks SoftBank has to manage
  • Customer neutrality questions inside Arm and AI ecosystems.
  • Merchant CPU economics against captive hyperscaler silicon.
  • Competition with Graviton, Cobalt, Axion, AMD Zen 4c / Zen 5c, and Intel E-core Xeons.
  • Execution risk on Aurora and AI infrastructure positioning.
  • Sustaining roadmap investment across cycles.
  • Building an AI infrastructure identity without overclaiming as a GPU rival.

SoftBank bought Ampere because AI infrastructure is not only GPUs. It is the CPU layer, memory layer, orchestration layer, and power-efficiency layer around GPUs.

Section 13  ·  Proof pointsWhat Ampere must prove now

Ampere's open proof points after the SoftBank acquisition
  1. Can Ampere keep winning customers that do not want to design their own CPUs?
  2. Can Oracle A4 become a meaningful proof point beyond one cloud partner?
  3. Can AmpereOne M show real inference and memory-bandwidth advantages?
  4. Can the 256-core roadmap ship competitively?
  5. Can Aurora become a real product, not only a roadmap slide?
  6. Can Ampere remain a neutral merchant supplier under SoftBank?
  7. Can Ampere compete against Graviton, Cobalt, Axion, AMD Zen 4c / Zen 5c, and Intel E-core Xeons?
  8. Can Ampere build an AI infrastructure identity without pretending to be a GPU company?
  9. Can SoftBank turn Ampere into a strategic asset rather than another expensive AI bet?

Ampere's next test is not whether cloud-native was real. It is whether cloud-native is enough.

Section 14  ·  Then and now2022 thesis vs 2026 reality

2022 thesis 2026 reality Lesson
Cloud-native was not just marketing High-core-count, no-SMT, predictable-performance strategy became mainstream Ampere's technical thesis was real
Cost per core was Ampere's advantage x86 and hyperscaler Arm competitors also moved toward dense-core designs Advantage narrowed
Ampere could become a public merchant Arm CPU company SoftBank acquired Ampere for US$6.5 billion IPO story became strategic-infrastructure story
AmpereOne would move beyond Neoverse cores AmpereOne shipped as a custom-core product family Architecture roadmap matured
Public clouds were key customers Hyperscalers built internal CPUs, while Oracle remains the clearest proof point Merchant opportunity narrowed
Cloud-native CPU was the core pitch AI inference, memory bandwidth, RAG and vector workloads now shape the pitch Cloud-native evolved into AI infrastructure CPU

Section 15  ·  EvidenceEvidence ledger

Claim
Evidence
Interpretation
Ampere's cloud-native pitch had technical substance
The uploaded PDF explains high core counts, one thread per core, no SMT, predictable clocks, private L2, smaller shared L3, and low cost per core.
Cloud-native was not pure marketing.
Ampere had a cost-per-core argument
Page 6 of the PDF estimates lower cost per core for Altra / Altra Max than AMD Milan and Intel Ice Lake under the article's assumptions.
Ampere had a density / cost argument, not a universal workload argument.
AmpereOne moved Ampere to custom cores
Pages 18 to 20 infer AmpereOne custom-core details; Ampere later described AmpereOne with 192 custom Arm ISA-compliant cores and 2 MB L2 per core.
Ampere evolved beyond stock Neoverse.
Hyperscalers copied the thesis
AWS Graviton4, Microsoft Cobalt 100, and Google Axion are custom Arm CPUs for cloud workloads.
Ampere validated the market, but hyperscalers internalized the best version.
Oracle remains a proof point
Oracle A4 Standard uses AmpereOne M for AI inference and cloud-native workloads, with 96-core A4 shapes.
Ampere still has a merchant cloud CPU path.
Ampere is pivoting toward AI inference infrastructure
Ampere and Qualcomm announced AI inference work using Qualcomm Cloud AI 100 and Ampere CPUs; Reuters reporting frames it around power-cost reduction.
Ampere's realistic AI role is efficient inference infrastructure.
Aurora is the roadmap bet
Ampere says Aurora scales to 512 cores, integrates AI IP, and supports HBM.
Ampere wants to move toward AI-native infrastructure CPUs.
The IPO path was replaced by acquisition
SoftBank acquired Ampere for approximately US$6.5B and completed the transaction on November 25, 2025 (U.S. time).
Ampere became a SoftBank strategic asset, not a public merchant CPU story.
Dense-core competition became mainstream
AMD and Intel both moved toward dense-core / efficiency-core server platforms.
The cloud-native CPU idea won, making Ampere's differentiation harder.

Section 16  ·  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.

The 2022 cost-per-core table excluded die harvesting, parametric yield, binning, fixed design costs, royalties, and server-level TCO. Treat it as directional.
Ampere's performance and Aurora claims (256-core, 512-core, HBM, integrated AI IP) are company claims pending independent validation in shipping silicon.
Hyperscaler captive silicon is a structural headwind for merchant CPU vendors. AWS, Microsoft, and Google will use their own Arm CPUs where it suits them.
Customer neutrality questions arise once Ampere is inside SoftBank, especially given SoftBank's other AI and Arm-ecosystem positions.
Aurora may slip or change scope. Future-node roadmap claims should not be treated as guaranteed delivery.
Dense-core x86 (Zen 4c, Zen 5c, Intel E-core Xeons) can erode Ampere's differentiation in cloud-native workloads even where Arm is acceptable.
Oracle dependence is real. If Oracle's AI infrastructure strategy shifts or its A4 ramp underperforms, Ampere's merchant-CPU proof point narrows further.
Software-stack maturity for cloud-native Arm has improved but still varies by workload. Some applications need x86 binaries and engineering effort to migrate.
AI workloads continue to evolve. CPU-side AI inference relevance depends on model size, memory bandwidth, and accelerator economics that can shift quickly.
Power, regulatory, and trade-policy dynamics can change merchant CPU economics independently of Ampere's technical roadmap.

Section 17  ·  Bottom lineBottom line

Bottom line

The 2022 article was right that Ampere's cloud-native strategy had real technical substance: many small cores, one thread per core, predictable performance, low cost per core, and high throughput density.

But the 2026 update is that Ampere's thesis became too correct. AWS, Microsoft, and Google built their own Arm CPUs, while AMD and Intel copied the dense-core direction. Ampere still has real proof points through AmpereOne, Oracle A4, Qualcomm inference work, and the Aurora roadmap, but the independent IPO story gave way to SoftBank's US$6.5 billion acquisition.

Ampere's future is no longer just merchant cloud CPUs. It is whether SoftBank can turn Ampere into a power-efficient CPU and AI infrastructure layer inside the Arm ecosystem.

Ampere proved cloud-native CPUs mattered. Then the clouds built their own. Now SoftBank has to prove Ampere still matters in the AI infrastructure stack.

Section 18  ·  DefinitionsGlossary

Cloud-native CPU
A server CPU designed around predictable throughput per watt per dollar, with many small cores, often one thread per core, optimised for scale-out cloud workloads.
Arm server CPU
A server processor that implements the Arm instruction set rather than x86. Used by hyperscalers (Graviton, Cobalt, Axion) and merchant vendors (Ampere).
Ampere Altra
Ampere's first family of cloud-native Arm server CPUs, based on Arm Neoverse N1 cores.
Ampere Altra Max
A higher-core-count variant of Ampere Altra, extending the cloud-native CPU strategy.
AmpereOne
Ampere's custom Arm ISA-compliant server CPU family with up to 192 cores and 2 MB L2 per core.
AmpereOne M
An AmpereOne family product positioned around AI inference and memory-bandwidth-heavy workloads, used in Oracle A4.
AmpereOne Aurora
Ampere's announced roadmap product, claimed to scale to 512 cores with integrated AI IP and high-bandwidth memory.
SMT
Simultaneous multithreading. Allows one CPU core to run multiple threads. Ampere chose to avoid SMT for predictability.
One thread per core
A design choice where each core runs a single hardware thread, reducing SMT-related variability and side-channel concerns.
Predictable performance
A property of a CPU where performance does not vary widely across tenants, workloads, or load levels.
Noisy neighbor
A workload that interferes with other tenants on shared resources (cache, memory bandwidth, threads). Reducing noisy-neighbor impact is a cloud-native CPU goal.
Private L2 cache
A second-level cache assigned exclusively to one core, improving per-tenant isolation.
Shared L3 cache
A last-level cache shared across cores in a CPU. Larger shared L3 helps some workloads but costs area and may not help every cloud workload.
Neoverse N1
Arm's efficiency-oriented server core used in Ampere Altra and AWS Graviton2.
Neoverse N2
Arm's successor efficiency-oriented server core, supporting Armv9 features.
Neoverse V1
Arm's higher-performance server core used in HPC-oriented designs such as early generations of AWS Graviton3.
Graviton
AWS's family of custom Arm-based server CPUs, including Graviton2, Graviton3, and Graviton4.
Cobalt
Microsoft's custom Arm-based Azure CPU family. Cobalt 100 is Microsoft's first fully custom Arm CPU for Azure.
Axion
Google's first custom Arm CPU. Powers Google Cloud C4A instances.
AI inference
Running a trained AI model to generate outputs. CPU inference is increasingly used for smaller models and infrastructure-adjacent AI work.
RAG
Retrieval-augmented generation. An AI pattern that retrieves relevant context from external sources and feeds it into an LLM at inference time.
Vector database
A database optimized for similarity search over high-dimensional embedding vectors. Often part of RAG and AI inference pipelines.
Merchant silicon
Silicon sold to multiple external customers by an independent semiconductor vendor.
Captive hyperscaler silicon
Silicon designed by a hyperscale cloud provider for its own internal use rather than sold as a merchant product.
Throughput per watt per dollar
A composite cloud fleet metric Ampere optimised for. Combines aggregate work done, power consumed, and cost paid.

Section 19  ·  MethodSources and method notes

How this essay reads sources

The 2022 SemiAnalysis Ampere piece is treated as historical context for the cloud-native CPU thesis, the page 6 cost-per-core table (with its explicit caveats), the pages 12 to 20 Arm Neoverse and AmpereOne architecture analysis, and the broader argument that cloud-native was substance rather than marketing. Ampere's product claims (96 to 192 cores, 2 MB L2 per core, 12-channel DDR5, 100G networking, 256-core roadmap, Aurora's 512-core / HBM / integrated AI IP framing) are treated as Ampere's claims rather than as independently verified figures.

The 2026 read is built primarily from Ampere's AmpereOne and AmpereOne M product briefs, Oracle's OCI A4 Standard announcement and Ampere's own A4 / AmpereOne M blog, Ampere's 256-core and Qualcomm AI inference announcement and Reuters coverage of the pairing, Ampere's AmpereOne Aurora announcement, Ampere's SoftBank acquisition announcement, SoftBank's acquisition-completion press release, and the AWS Graviton4, Microsoft Cobalt 100, Google Axion, AMD EPYC Zen 4 / 4c, and Intel Xeon 6 reference pages used to anchor the hyperscaler and dense-core x86 competitive context. The structural argument that Ampere's thesis was correct but its differentiation narrowed as hyperscalers and x86 vendors copied the pattern, and that SoftBank's acquisition reframes Ampere as a strategic AI infrastructure asset rather than a public merchant CPU story, is independent analysis.

Footnotes  ·  primary sources

  1. SemiAnalysis, “Is Ampere Computing's Cloud Native Marketing Fluff? — Siryn AmpereOne 5nm Architecture, Cost Analysis, and IPO Analysis,” 2022 (PDF supplied by author). Historical anchor used in this essay for the cloud-native CPU thesis, the page 6 server-chip manufacturing cost estimate table comparing Intel Ice Lake, AMD Milan, Ampere Altra, and Ampere Altra Max (with explicit caveats about die harvesting, parametric yield, binning, fixed design costs, royalties, and TCO), the pages 12 to 17 Neoverse N1 / V1 / N2 architecture diagrams, the pages 18 to 20 inferred AmpereOne (Siryn) custom-core analysis, and the broader public-cloud customer-list framing.
  2. Ampere Computing, “AmpereOne Family Product Brief,” amperecomputing.com/…/ampereone-family. Source for the AmpereOne family framing including 96 to 192 custom Ampere Arm ISA-compliant cores, 2 MB private L2 cache per core, Armv8.6+ ISA support, DDR5 memory, and PCIe Gen5 connectivity.
  3. Ampere Computing, “AmpereOne M Product Brief” and “Processors Page,” amperecomputing.com/…/ampereone-m; amperecomputing.com/…/processors. Source for AmpereOne M positioning at 96 to 192 cores with 12-channel DDR5, 2 MB L2 per core, and AI inference framing.
  4. Oracle, “Introducing OCI Ampere A4 Standard,” blogs.oracle.com/…/oci-ampere-a4-standard. Source for OCI A4 Standard instances powered by AmpereOne M, the AI inference and cloud-native workload framing, 96-core A4 shapes, and Oracle's positioning of A4 inside its compute portfolio.
  5. Ampere Computing, “AmpereOne M-Powered A4 Instances Coming to Oracle Cloud,” amperecomputing.com/…/a4-oracle. Source for 100G networking on A4, 12-channel DDR5 memory bandwidth framing, demanding AI inference workloads, and Ampere's own LLM inference positioning around A4.
  6. Ampere Computing, “Ampere Scales AmpereOne to 256 Cores and Announces AI Inference Partnership with Qualcomm,” amperecomputing.com/…/256-core-qualcomm. Source for the 256-core AmpereOne roadmap, the Qualcomm Cloud AI 100 joint inference solution, the power-efficient AI inference framing, and Ampere's claims around the partnership (treated as Ampere claims).
  7. Reuters, “Ampere Computing pairs with Qualcomm AI, unveils new chip,” reuters.com/…/ampere-qualcomm-2024-05-16. Source for external reporting on the Ampere-Qualcomm pairing, the AI inference power-cost framing, the 256-core CPU context, and the TSMC 3nm manufacturing context where stated.
  8. Ampere Computing, “Introducing AmpereOne Aurora,” amperecomputing.com/…/aurora. Source for the AmpereOne Aurora roadmap framing including up to 512 Ampere cores, integrated Ampere AI IP, high-bandwidth memory support, and RAG and vector database use-case framing. Treated as a roadmap and company claim, not independent validation.
  9. Ampere Computing, “SoftBank Group to Acquire Ampere Computing,” amperecomputing.com/…/softbank-acquires-ampere. Source for the approximately US$6.5B all-cash acquisition, SoftBank's AI infrastructure strategy framing, the wholly owned subsidiary outcome, and the explicit end of the IPO story.
  10. SoftBank Group, “Completion of the Acquisition of Ampere Computing,” group.softbank/…/20251126. Source for the completion of the Ampere acquisition on November 25, 2025 (U.S. time) and the post-acquisition status of Ampere as a wholly owned subsidiary of SoftBank.
  11. AWS, “Amazon EC2 R8g Instances (Graviton4),” aws.amazon.com/…/r8g. Source for AWS Graviton4 framing as a custom AWS Arm-based processor for EC2 instances and the broader cloud-native Arm context inside AWS.
  12. Microsoft Azure, “Azure Cobalt 100-Based Virtual Machines Are Now Generally Available,” azure.microsoft.com/…/cobalt-100-ga. Source for Microsoft Cobalt 100 framing as Microsoft's first fully custom Arm-based Azure CPU and the hyperscaler internal CPU strategy used in this essay.
  13. Google Cloud, “First Google Axion Processor C4A Now GA with Titanium SSD,” cloud.google.com/…/axion-c4a-ga. Source for the Google Axion framing as Google's first custom Arm CPU, the C4A instance launch, and Google's price-performance and energy-efficiency claims for cloud-native workloads.
  14. AMD, “4th Generation EPYC Server Processors,” amd.com/…/epyc-4th-gen. Source for AMD Zen 4 / Zen 4c dense-core EPYC configurations and the x86 industry shift toward dense-core cloud / scale-out server platforms.
  15. Intel, “Intel Xeon 6,” intel.com/…/xeon6. Source for Intel Xeon 6 family framing including efficiency-core server platforms aimed at cloud-native and scale-out workloads, used in this essay as evidence that the dense-core / efficiency-core idea became mainstream across x86 as well.
← All essays