Foreword:

Over the past few years, competition and isolation have been the defining features of the China-U.S. AI industry. Today, however, a different picture is emerging: ByteDance is collaborating with Microsoft on AI, while IBM is integrating the DeepSeek model into its enterprise AI ecosystem.

Technology, capital, and commercial demands are reshaping the global AI landscape. Competition persists, but connections are being restored.

Author | Fang Wensan

Image Source | Internet

Microsoft Stays in China: The Answer Lies in AI Demand

Microsoft's China business accounts for a relatively small portion of its overall revenue, and Azure faces local competition in China's public cloud market share.

However, AI is changing this trajectory.

Chinese companies like ByteDance need more than just a model—they require a comprehensive AI infrastructure to support global business expansion.

As a long-term partner of OpenAI, Microsoft can provide advanced model capabilities, while Azure's global infrastructure helps enterprises quickly enter overseas markets.

This has fostered a new business relationship: Chinese companies possess vast application scenarios, engineering capabilities, and globalization needs; Microsoft offers model ecosystems, cloud infrastructure, and compliance systems.

The two sides are not merely supplier and customer but have formed new complementarities within the AI industrial chain.

According to reports, ByteDance has become one of Microsoft's key AI customers in recent years, with annual spending on Microsoft's AI and cloud services nearing $1 billion.

ByteDance boasts a massive content ecosystem and user scenarios, while Microsoft provides global cloud infrastructure, enterprise service capabilities, and a long-accumulated AI technology system. Their collaboration essentially explores how large model capabilities can be integrated into more complex real-world applications.

The boundaries between model companies, internet platforms, and cloud computing enterprises are blurring. Past relatively independent technology supply chains are now forming tighter collaborative networks.

Enterprises with user access need underlying technical support, while those controlling infrastructure require real-world scenario validation. Only through this synergy can AI transform from laboratory capability into scalable productivity.

DeepSeek Enters IBM: Open Weights Enable Reverse Global Outreach

On August 11, IBM announced a multi-year, $240 million agreement with Together AI to deploy NVIDIA HGX B300 clusters on IBM Cloud, expected to go live in Q1 2027.

Reuters further revealed that the initial cluster will feature approximately 2,000 Blackwell 300 chips, with Together AI providing open model services including Chinese models like DeepSeek, MiniMax, and Kimi.

Here, DeepSeek functions as a workload—a key commodity on the open model shelf.

This path is enabled by open licensing. DeepSeek-R1's code and weights are released under the MIT license, allowing commercial use, modification, and derivative development.

DeepSeek's official repository shows that U.S. cloud providers can deploy the model on local clusters without special authorization, adding inference optimization, risk control, billing, and enterprise services.

Thus, a remarkably timely pipeline emerges: Chinese teams contribute model weights, U.S. companies add NVIDIA chips, IBM Cloud, and Together AI's inference engine, then sell the services to global enterprises.

IBM aims to create a "multi-model supermarket," where GPT handles high-value scenarios while DeepSeek, Kimi, and MiniMax populate the open inference pool. IBM profits from compute, integration, governance, and consulting revenues.

In the AI Era, Re-division of Labor Matters More Than Competition

Technological progress has never been a simple zero-sum game.

This held true in the semiconductor era, the internet era, and remains true in the AI era.

Creating a model requires algorithmic, computational, data, engineering, and commercial ecosystem support. For technology to generate real value, it must enter authentic industrial scenarios.

Microsoft-ByteDance and IBM-DeepSeek may appear as mere business collaborations, but they reflect broader shifts in the global tech industrial structure.

AI is driving tech companies to redefine their roles: some focus on foundational models, others provide cloud infrastructure, connect enterprise clients, or handle scenario implementation.

Future competition may no longer hinge on "who has the strongest model" but on who can transform models into capabilities truly needed by global industries.

Previously, China-U.S. tech discussions centered on "who will win." In the AI era, a new question arises: how to recombine different strengths within the global industrial chain.

Upstream Gatekeeping, Midstream Model Exchange

The rise in cross-procurement stems from the AI industrial chain being fragmented into independently tradable modules: chips, training compute, model weights, inference engines, cloud platforms, data governance, and application distribution.

Each layer faces varying policy intensities, cost structures, and cross-border difficulties.

The performance gap between Chinese and U.S. models is narrowing. Stanford's *2026 AI Index Report* notes that the performance gap has largely closed, with models from both countries Alternating lead (taking turns leading) since early 2025. By March 2026, Anthropic's top model led by just 2.7%.

With performance converging, enterprises prioritize price, latency, context length, tool integration, data boundaries, and governance capabilities over model origin.

While model provenance remains a risk factor, it rarely solely determines commercial orders.

Microsoft added DeepSeek-R1 to Azure AI Foundry and GitHub's model catalog in January 2025, stating the model underwent red-teaming, security assessments, and content filtering for serverless deployment.

The same Microsoft provides ByteDance with GPT and global cloud capabilities while making DeepSeek available to Western developers. Platforms increasingly act as customs, malls, and power utilities for the model world.

Thus, China-U.S. AI relations have adopted a "layered decoupling, modular reconnection" structure.

Upstream chips and large-scale training face barriers, midstream models, inference, and development tools seek compliant channels, while downstream enterprises pay based on results.

Winners of the New Narrative Hold the Switching Power

DeepSeek rapidly gains global distribution through open weights, allowing its technical influence to transcend borders. However, cloud revenues may remain with IBM, Together AI, and chip suppliers.

Chinese model companies face a tougher commercial challenge: benchmark leadership merely grants entry. Developer ecosystems, stable APIs, enterprise support, compliance adaptation, and overseas compute networks determine where value ultimately lands.

Delivering only weights often means models attract traffic while platforms collect rents.

ByteDance faces different risks: large-scale Azure purchases buy speed and global reach but accumulate supplier dependency.

Model interface abstraction, multi-cloud deployment, regional data governance, and dynamic switching between self-developed and outsourced solutions will shift from technical options to operational capabilities.

"Binding" must allow room for loosening. When policies, prices, or model rankings change, workloads must relocate smoothly.

Scarce resources are shifting from singular model capabilities to the ability to legally, stably, and cost-effectively run multiple models and switch between them.

Laying Motivations Bare Reveals the New Narrative's Framework

Microsoft needs China: It hosts the most willing AI buyers, provides samples for validating global revenues, and offers insights into local technological progress.

ByteDance needs Microsoft: Access to cutting-edge models and global compute pipelines is faster than building in-house.

IBM needs open-source inference: Inference has become a major driver of compute demand growth, and idle data centers represent the greatest losses.

DeepSeek needs IBM: Without trusted distribution channels, even superior technology cannot enter enterprise procurement lists.

Each party finds what it lacks in the other—this is the foundation of new mutual dependence.

A division of labor is taking shape: The U.S. guards closed-source premium layers (GPT and Claude charge per token), while China dominates open-source scale layers (Qwen and DeepSeek build ecosystems at ultra-low prices).

According to Artificial Analysis, Chinese open-source models like DeepSeek and MiniMax significantly outperform closed-source counterparts in "model intelligence per dollar."

Both sides profit within their domains while encroaching on each other's territories: Microsoft charges Chinese companies for models, while IBM helps Chinese models enter U.S. enterprise data centers. Political narratives remain zero-sum, but commercial ones have become bidirectional arbitrage.

IBM and Together AI's clusters deploy domestically in the U.S., while Microsoft keeps its China business at a "Congress-acceptable" scale.

Both have mastered the same art: confining conflicts to the chip layer while conducting transactions at the model layer. Firewalls rise higher, and API calls multiply.

Conclusion:

AI is redrawing boundaries. While technological competition remains fierce, the business world is forging new connections.

Signals from ByteDance-Microsoft and DeepSeek-IBM suggest the AI industry is transitioning from pure capability competition to ecosystem competition.

A decade ago, China-U.S. tech ties were bound by trade viscosity; today, token viscosity reigns. What truly cannot be severed is cost-effectiveness—the most resolute diplomat of this era.