TSMC reports 1 billion processor units backlog; DeepSeek raises API prices citing compute scarcity; Anthropic plans custom AI chip design

On August 6, Huawei Fellow and semiconductor chief scientist Liao Heng stated in an interview: "Chips are now almost everywhere—they have become as important as water and air." The significance of the domestic chip industry chain targeted by the Science and Technology Innovation Chip ETF Huifutian (588750) is evident from this alone.

The AI wave has created unprecedented massive demand for chips. Chip shortages, scarcity of processing capacity, and chip inflation have become industry norms. DeepSeek's rare price increase announcement and Anthropic's move into self-designed chips both underscore this trend.

Looking forward, accelerating domestic large model iteration and the trend of "domestic models paired with domestic chips" suggest that domestic chip innovation breakthroughs—driven by 3D stacking and advanced node technologies—position domestic chips to leapfrog competitors, entering a golden development period of rising volume and pricing.

**TSMC's Supply Challenge**

According to industry sources, TSMC is ramping up high-end capacity to meet demand for Apple's A20 Pro processors. However, due to DRAM shortages, the company has accumulated approximately $1 billion worth of Apple processors awaiting final packaging, pending DRAM arrivals.

**DeepSeek Raises Prices**

On August 6, DeepSeek announced on its official website that it plans to substantially raise pricing for its API services in the near term. Multiple factors likely drive this decision:

First, peak-time call volumes and concurrent usage have dramatically spiked, creating acute compute shortages. OpenRouter's weekly rankings show V4-Flash calls leading at 7.22 trillion tokens; the open-source project OpenCode disclosed that August 1 alone reached 8 trillion tokens daily; on August 4, the service temporarily experienced capacity constraints due to "unprecedented access levels."

Second, inference cost structures—encompassing GPU, HBM, electricity, and operations—are rising during demand surges. The previous strategy of "ultra-low pricing plus high usage" eroded margins; pricing must return to "sustainable cash flow and value-based pricing."

Third, industry-wide coordination is raising the price floor. Domestic cloud providers and multiple model vendors have already increased prices multiple times this year. DeepSeek's price adjustment will shift the "low-price anchor" upward, moving the industry from "competing on price" to "paying for performance, stability, and ROI."

Near-term, this supports load-balancing and business stability; medium-term, the rising price ceiling will repair model-side cash flow, reinvesting in upstream compute and infrastructure, driving a positive flywheel of "compute rental—domestic compute—applications."

**AI Companies Turn to Custom Chips**

AI large models drive explosive growth in training and inference compute demand. To meet massive chip requirements and ensure supply stability, leading AI companies are developing proprietary chips.

According to media reports, AI startup Anthropic has publicly confirmed for the first time that it is assembling an internal team to design custom AI chips tailored to its Claude model. This move leverages self-developed hardware to close compute gaps, fully aligning with Claude's escalating large-scale training and inference compute requirements, and establishing the hardware foundation for scaled model deployment. Similarly, OpenAI previously launched its first self-developed AI inference chip, Jalapeño.

**Demand Outlook**

Just how large is chip demand under the AI wave? Goldman Sachs projects that global AI accelerator chip demand for 2026–2028 will reach 19 million, 27 million, and 32 million units respectively—14% to 22% higher than previous forecasts.

**Policy Support and Infrastructure**

On the policy front, computing power network construction is accelerating, enabling cross-regional and cross-industry efficient resource scheduling. As one of "six networks" (water systems, new-generation power grids, computing power networks, next-generation telecommunications networks, urban underground pipelines, and logistics networks), the computing power network aims to coordinate national computing resources through unified dispatch, reducing per-token costs and functioning like a "national power grid." This directly spurs investment while driving multiplier effects across AI, big data, and industrial internet ecosystems. In concrete implementation, China's AI self-innovation pace accelerated significantly in the first half of 2026: the nation's first fully domestic 100,000-GPU cluster went operational; intelligent computing capacity roughly doubled year-over-year; domestic large model downloads exceeded 10 billion; and domestic computing and chip architectures are accelerating alignment.

**Domestic Model Momentum**

From the model perspective, domestic large models are concentrated in rapid iteration, strengthening the domestic compute logic. CITIC Securities notes that DeepSeek, ByteDance's Seedance, MiniMax, Moonshot, and other domestic large models have undergone dense upgrades recently. DeepSeek-V4-Flash formally launched API public testing with Codex compatibility; ByteDance Seedance 2.5, SiYu MiniMax H3, Kimi K3, and others simultaneously breakthrough in video generation, multimodality, and parameter scale. Model evolution is shifting from "stacking parameters and Q&A" toward agentic capabilities, lower inference costs, and multi-model coordination—potentially transitioning AI adoption from low-frequency question-answering to high-frequency, complex, multi-step tasks. Price decreases need not suppress total compute demand. If call volume growth outpaces per-token cost declines, compounded by agents' multi-turn planning and tool-use incurring higher token consumption, inference compute demand could still rise.

As domestic model competitiveness increases and applications penetrate enterprise production workflows, beneficiary chains are expanding from GPUs to CPUs, AI servers, storage, switching, optical modules, PCBs, liquid cooling, data centers, and power infrastructure. The logic strengthening the domestic compute supply-chain ecosystem remains robust.

**Technical Advances**

Super-node technology is driving compute upgrades. According to IT Home, Tencent Cloud plans large-scale deployment of domestically developed compute capacity and intends to deploy NPO super-nodes by Q4 2026. Tencent Cloud believes optical interconnection has become critical to breaking bandwidth and power bottlenecks as super-node scale expands. NPO—placing photo-to-electrical conversion chips on the same board as compute chips—represents a more pragmatic route for domestic GPUs to build super-nodes than CPO, shortening the distance to the primary chip and eliminating DSP chips from optical modules.

WAIC 2026 demonstrated that domestic GPU competition is shifting from single-card performance to system-level competition spanning super-nodes, clusters, software ecosystems, and application deployment. Vendors like Biren, Moore Threads, Tengsun, Muxi, and Graphcore collectively showcased large-scale interconnects, backplane-free designs, general-purpose GPUs, open-source ecosystems, and commercialization results. Domestic compute has transitioned from the "whether we have it" phase to "how well we use it," and industry focus has shifted from "designing chips" to "deploying chips effectively."

**Investment Implications**

An AI infrastructure super-cycle remains within reach, with storage, CPU, wafer, and advanced packaging all experiencing supply gaps exceeding expectations. The Science and Technology Innovation Chip ETF Huifutian (588750) target index carries high concentration in compute (CPU/GPU/ASIC) plus storage and advanced manufacturing at 86%, delivering pure-play chip-sector exposure. A 20-day volatility range provides convenient access to high-growth-momentum opportunities. Off-exchange investors can track linked funds (A: 020628; C: 020629), available for 7×24 redemption.

For more targeted positioning in domestic compute and storage, the Science and Technology Innovation Chip Design ETF Huifutian (589290) offers compute (CPU/GPU/ASIC) plus storage at 72%, positioned to fully benefit from the three-pronged tailwind of AI compute demand, storage chip price supercycle, and domestic substitution—capturing a historic development opportunity in domestic compute and storage supply chains amid the AI wave and self-reliance trend. Its 20-day volatility characteristic helps capture this super-cycle.