Foreword:

In a single week, NVIDIA, Samsung, and Microsoft have strategically positioned themselves at the three most pivotal junctures within the same industrial landscape: computing power, cloud and quantum computing, and storage.

Author | Fang Wensan

Image Source | Internet

NVIDIA: AI Computing Power Expands into the PC Realm

NVIDIA's vision is to enable AI to permeate from data centers to every work interface, redefining the landscape of computing.

The newly unveiled RTX Spark propels NVIDIA's AI computing capabilities further into the PC domain, empowering local devices to run AI agents, process local models, and manage complex creative and development tasks.

The RTX Spark superchip, lauded by Jensen Huang as the cornerstone hardware to "reshape Windows PCs for the era of personal AI agents," signifies NVIDIA's formal foray into the PC processor market, which has been dominated by Intel for over four decades.

Built on TSMC's cutting-edge 3nm process, it integrates a 20-core Grace ARM v9.2 CPU with a Blackwell architecture GPU, enabling seamless CPU-GPU collaboration through NVLink-C2C interconnect technology and sharing up to 128GB of LPDDR5X unified memory.

At FP4 precision, its AI computing power reaches an astonishing 1 PFLOPS, capable of effortlessly running large models with 120 billion parameters on local devices, with a context length extending up to 1 million tokens.

The first batch of devices powered by RTX Spark will be manufactured by six major OEMs—ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI—and will officially hit the market in autumn 2026.

Additionally, the newly released Vera CPU epitomizes NVIDIA's strategic transformation. As the world's first general-purpose processor tailored specifically for AI agents, Vera boasts a task completion speed 1.8 times faster than traditional x86 CPUs.

Leading global AI labs, including Anthropic, OpenAI, and SpaceXAI, along with hyperscale cloud providers like ByteDance, CoreWeave, and Oracle Cloud Infrastructure, have all earmarked plans to adopt Vera to comprehensively upgrade their AI infrastructure.

Samsung: Dual-Wheel Drive Under Vertical Integration

At the recent Computex 2026 event, Samsung unveiled its eighth-generation high-bandwidth memory (HBM5) prototype, garnering significant attention. The underlying base die of HBM5 is entirely crafted using Samsung's proprietary 2nm advanced process technology, a departure from the 4nm process employed in previous generations, HBM4 and HBM4E.

Samsung also introduced a groundbreaking heat dissipation architecture known as the Heat Path Block (HPB). By incorporating dedicated independent thermal pillar structures within the memory stack, it efficiently exports heat from the core stacking area, effectively addressing the high-heat challenges associated with premium AI memory.

Samsung's resurgence in the HBM market is already yielding tangible results. On June 5, Jensen Huang officially announced that all three major memory chip giants—Samsung Electronics, SK Hynix, and Micron Technology—have obtained certification and will supply HBM4 high-bandwidth memory chips for NVIDIA's next-generation AI platform, Vera Rubin.

As an industry leader with comprehensive memory R&D and manufacturing capabilities, coupled with logic wafer foundry services, Samsung can achieve full autonomous production of HBM chip stacking, packaging, and underlying logic dies. This vertical integration advantage remains unparalleled among competitors.

Samsung's uniqueness lies in its simultaneous possession of memory, wafer foundry, and advanced packaging capabilities, which is why it is eager to showcase its prowess with HBM5.

HBM5 serves as a ticket back to the heart of the AI supply chain. Samsung aims to seamlessly integrate memory, foundry, packaging, and end-user devices, leveraging the SAFE ecosystem's EDA, IP, OSAT, and MDI alliance to bind customers' development processes, IP, and packaging solutions within the Samsung ecosystem from the outset.

Microsoft: Building a Closed-Loop AI Ecosystem with Integrated Hardware and Software

On June 2, at Microsoft Build 2026, Microsoft unveiled a series of groundbreaking chip products, underscoring its unwavering commitment to constructing a closed-loop AI ecosystem that seamlessly integrates hardware and software.

Azure Cobalt 200, Microsoft's second-generation Arm-based cloud chip, is optimized for agent AI workloads, delivering a 50% performance boost over its predecessor, Cobalt 100.

In its official blog, Microsoft explicitly stated that agents fundamentally differ from traditional workloads, necessitating reasoning, sequential decision-making, and sustained large-scale operation, which demands a novel computing architecture.

Cobalt 200 has undergone full-stack optimization from silicon to servers to services, integrating Microsoft's latest security, networking, storage, and offloading technologies, excelling in AI inference, data pipelines, and web services.

Maia 200, Microsoft's second-generation AI chip, has been gradually deployed in Microsoft's data centers in Iowa and Phoenix since early 2026.

It represents Microsoft's most energy-efficient inference system to date, outperforming similar chips from Google and Amazon in specific AI tasks.

Microsoft also unveiled its second-generation topological quantum chip, Majorana 2, which tackles long-standing core challenges in the quantum computing field.

The topological qubits on Majorana 2 achieve an average coherence time of 20 seconds, lasting up to 1 minute in specific ideal test environments, with operational reliability improved by 1000 times compared to the previous generation.

Microsoft also outlined an ambitious scientific commercialization roadmap, planning to officially launch the world's first "commercially viable, fully scalable, and fully fault-tolerant" quantum computer before 2029.

Its strategic intent remains measured and pragmatic. Instead of attempting to directly supplant NVIDIA in the high-end training market, Microsoft integrates its self-developed chips into Azure's "heterogeneous foundation."

The most expensive cutting-edge training and large-scale inference continue to rely on NVIDIA and AMD, while cost-sensitive, scalable inference and general cloud workloads gradually transition to Maia and Cobalt.

Thus, Microsoft Foundry connects with multi-model ecosystems like OpenAI, Anthropic, Meta, xAI, and NVIDIA, while deploying Maia 200 for GPT5.2 and Copilot, and utilizing Cobalt 100/200 to support a broader range of cloud workloads like Teams, Azure SQL, Elastic, Rescale, Databricks, and Snowflake.

What Microsoft truly seeks to control is the closed loop among cloud services, models, office software, operating systems, and developer platforms, focusing on cloud service gross margins and unit token costs.

Why Competition and Collaboration Coexist

While all three companies are launching chips, their strategic positions do not overlap.

NVIDIA safeguards the performance ceiling and software standards of AI infrastructure, Microsoft protects the cost curve and distribution entry points of cloud services, and Samsung secures the supply interfaces of HBM, foundry, and packaging—areas where competitors struggle to expand rapidly.

Thus, competition and collaboration can coexist harmoniously. Microsoft leverages NVIDIA Blackwell Ultra and future Rubin chips for high-end computing power in Azure while also utilizing its own Maia 200 to handle certain inference workloads.

Samsung aims to secure HBM and foundry opportunities with NVIDIA and Microsoft while also incorporating Exynos 2500 in end-user devices to utilize its own advanced processes.

They represent distinct control points along the same chain rather than simple one-to-one replacements.

This stratification becomes even more apparent when viewed through the lens of profit logic. NVIDIA's latest quarterly GAAP gross margin reached 74.9%, with data center revenue hitting $75.2 billion in a single quarter, indicating it still captures the most profitable segment.

Microsoft acknowledges in its official earnings report that AI infrastructure investments from Azure expansion are exerting pressure on cloud business gross margins.

Samsung's Foundry and System LSI businesses faced prolonged pressure until signs of improvement emerged in late 2025.

NVIDIA prioritizes profit growth, Microsoft deepens capital expenditures, and Samsung focuses on securing customer trust and validation.

Market share changes and substitutions are indeed occurring but in different regions and tiers.

Conclusion: Future Windows and Observation Metrics

Over the next one to two years, the industry tends to evaluate the three companies along a "system mass production capability curve."

NVIDIA's upward trajectory hinges on its ability to consistently deliver on its annual rhythm from Blackwell Ultra to Rubin.

Microsoft's value realization depends on whether its self-developed chips can genuinely enhance the unit economics of Azure and Foundry.

Samsung's turnaround potential rests on whether HBM4/4E and 2nm can surpass validation thresholds, transforming the "foundry + memory + packaging" vision into tangible orders.

Variables include TSMC CoWoS expansion and advanced capacity utilization, as well as whether Samsung's pure foundry share can continue to rise.

U.S. export controls and the speed of China's domestic alternatives will influence NVIDIA's revenue ceiling and alter Microsoft's regional computing power deployment and Samsung's order geography.

As AI penetrates deeper, the industry will revert to the fundamentals of hard technology: data must flow seamlessly, chips must dissipate heat efficiently, systems must collaborate effectively, costs must decrease continuously, and customers must receive stable, usable intelligence.

The three companies are positioned at the most congested and valuable bottlenecks in the AI industry chain. Whoever can hold these bottlenecks will transform growth from mere prosperity into structural dividends in the next technological cycle.

Partial References: Bloomberg Technology: "Jensen Huang Discusses the Global AI Chip Market Landscape," Guancha.cn: "NVIDIA Aims to Replicate CUDA's Success in PCs and Robots," Microsoft Official: "Full Text of Microsoft Build 2026 Developer Conference Keynote Speech," Samsung Electronics: "Samsung Showcases the World's First HBM5 High-Bandwidth Memory Prototype at Computex 2026."