Domestic HBM Breakthrough + DeepSeek Multimodal Open Source: The AI Hardware Wars Escalate

The tech news on September 1st was dense with signal: domestic memory chips made a breakthrough in a core AI component, DeepSeek open-sourced its first multimodal model, and AI giants are locked in a battle over who gets to define the hardware layer. These seemingly scattered events collectively sketch out the full competitive landscape of the AI supply chain — from silicon to application.

Domestic HBM Enters Small-Scale Production, Breaking Samsung-SK Hynix-Micron Stranglehold

Changxin Memory Technologies (CXMT) has begun small-scale production of HBM3 high-bandwidth memory — news that deserves serious attention from anyone tracking the AI chip industry.

HBM (High Bandwidth Memory) can be thought of as the "short-term memory" of an AI chip: it determines how efficiently data flows during training and inference. Until now, the high-end HBM market has been virtually monopolized by Samsung, SK Hynix, and Micron — a persistent bottleneck in the AI compute supply chain.

CXMT's entry into small-scale production means domestic AI chips now have a homegrown memory option, with commercial availability expected as early as 2027. While full-scale deployment is still some ways off, from a supply-chain independence standpoint, this is a critical step toward filling a longstanding gap in AI hardware. Memory and compute are the two legs of an AI chip — you can't walk without both.

DeepSeek Open-Sources Multimodal Vision Model, Engineering Chops Surpass Rivals

On the last day of August, DeepSeek open-sourced its first multimodal vision model — V4 FlashVision EXP.

The model can process both text and images simultaneously, excelling at tasks like screenshot recognition and chart analysis. Most notably, it topped the leaderboard on real-world software engineering benchmarks, outperforming rival models, while its tool-calling capabilities nearly matched the best in class.

Also worth highlighting is the release cadence: the model went open-source just 10 days after its API launch. This "rapid iteration + open sharing" approach is precisely the playbook that has helped DeepSeek stand out among domestic large model developers. Open-sourcing multimodal capabilities will further lower the barrier for developers building vision-intelligence applications, and inject new energy into the domestic model ecosystem.

The AI Hardware Shadow War: Anthropic vs. OpenAI in the Battle for Definition

Beyond the application layer, AI giants are waging a quieter war over who gets to define the hardware layer — and two starkly different strategies have emerged.

The Protocol Route: Anthropic Lets AI Control Physical Devices Directly

Anthropic released an open protocol called MHS, enabling AI agents to directly control lab and factory equipment. The protocol compresses what used to be months-long device integration cycles down to minutes.

In a real-world test on a quantum computer, four Claude agents iterated hundreds of rounds overnight — cutting laser lock recovery time from 2.5 minutes down to 6 seconds. This demonstrates the enormous potential of AI agents in scientific research and industrial automation: no longer just a chatbot, but an executor capable of directly intervening in the physical world.

The Terminal Route: OpenAI's Closed Ecosystem Ambitions

Unlike Anthropic's open protocol approach, OpenAI is reportedly taking a different path. Its rumored first hardware product is a home smart speaker — an attempt to embed its models directly into endpoints, following a closed-ecosystem playbook reminiscent of Apple.

One side is setting standards and gatekeeping platforms; the other is building devices and chasing "the next iPhone." This battle has only just begun. At its core, it's a contest over who controls the entry point — and the narrative — of the AI era.

NVIDIA Invests $3.5 Billion in MediaTek to Build AI Factories Together

On the hardware ecosystem integration front, NVIDIA has made another strategic move.

NVIDIA is reportedly planning to invest $3.5 billion in MediaTek, with the two companies set to jointly build "AI factories" and MediaTek integrating NVIDIA's interconnect technology.

The implications are significant: for MediaTek, this is a major opportunity to pivot from mobile chips toward data centers; for NVIDIA, it further cements its authority in AI infrastructure. Through this dual binding of capital and technology, NVIDIA is constructing an increasingly robust AI hardware alliance.

Other Notable AI Industry Developments

Beyond these headline events, several other stories reflect the current state and direction of the AI industry:

• Diverging AI monetization paths: OpenAI's ChatGPT advertising business has crossed $1 billion in annualized revenue, while Anthropic holds firm on its no-ads strategy. Meanwhile, OpenAI is piloting "outcome-based pricing" — charging only when AI actually delivers results — a genuinely imaginative model innovation.

• Technology outrunning management: Reports from Deloitte, KPMG, and Accenture all point to the same uncomfortable reality — companies are deploying AI agents quickly, but very few have actually redesigned their human-AI collaboration workflows.

This exposes the central tension in current AI adoption: technical capabilities are ready, but organizational and process adaptation is severely lagging. The real challenge for enterprises may not be "whether to use AI" but "how to restructure workflows."

• A new approach to bypassing HBM bandwidth bottlenecks: Hanxu Technology unveiled a new architecture that keeps model weights resident in memory chips, attempting to route around HBM bandwidth constraints — an interesting contrast to the domestic HBM breakthrough discussed earlier, reflecting the industry's diverse approaches to the memory bottleneck problem.

• Policy support for AI deployment: China's Ministry of Industry and Information Technology (MIIT) has launched an initiative to cultivate AI application service providers, targeting the creation of a resource pool of 3,000 service providers within two years — a policy-level push for scaled AI deployment.

Conclusion

From the foundational HBM memory breakthrough, to the middle-layer open-sourcing of multimodal models, to the top-layer battle over hardware ecosystems — this early-September cluster of developments clearly illustrates the full-stack nature of today's AI competition. Domestic players are closing critical gaps; open-source models continue to challenge closed ones; and the strategic divergence between giants will ultimately determine who controls the entry points of the AI era.

The consensus from three major consulting firms is worth taking seriously: technology has already outrun management. For any organization hoping to embrace AI, the real barrier is shifting from technical capability to the ability to reconstruct processes.

(Note: Nothing above constitutes investment advice.)