NVIDIA Prepares a Special AI Chip for China Based on Groq Technologies: A Bet on LPU and SRAM to Bypass Export Restrictions

Against the backdrop of a protracted standoff over the supply of American accelerators to China, NVIDIA, according to available reports, is developing a specialized solution that should allow the company to maintain the competitiveness of its products in the Chinese market while staying within the sweeping export restrictions imposed by the United States. The key feature of the new chip, reportedly, will be Language Processing Units (LPU) — technologies belonging to the startup Groq, whose rights and a portion of whose assets NVIDIA acquired back in December 2025. In doing so, the company is attempting not merely to circumvent regulatory barriers but to adapt its architecture to the new realities of the Chinese market before it loses its foothold there entirely.

Groq and LPU: A Bet on SRAM and Pre-Planned Computations

Unlike traditional GPU accelerators, Groq's architecture places its main bet on a massive amount of fast SRAM memory and computations that are planned in advance by the compiler. By design, this approach allows developers to minimize latency and deliver very high speeds in processing queries to AI models. For the Chinese market, where demand for inference and generative models is growing, this looks like an attempt to offer a product with predictable performance in a situation where conventional accelerators with cutting-edge HBM memory are subject to restrictions.

The Export Labyrinth: H200, HBM Requirements, and CoWoS

The development is particularly timely given the ambiguous situation surrounding the supply of American accelerators to China. The Trump administration, according to reports, has authorized the export of the NVIDIA H200 model, but Beijing has responded by urging local companies to forgo purchasing them. Amid this dual uncertainty, the new chip must account for the current US restrictions, including requirements on the volume of HBM memory and the availability of advanced packaging technologies such as CoWoS. In other words, NVIDIA is searching for a "golden middle": a product powerful enough for Chinese clients, yet not exceeding what Washington is willing to allow for export.

Chinese Alternatives and the Growing Role of SMIC

Meanwhile, China is actively developing its own alternatives, which increases pressure on NVIDIA. Local manufacturers are experimenting with RISC-V processors, three-dimensional memory architectures, and compute-near-memory. Against this backdrop, SMIC is already reaping significant benefits from growing domestic demand for local chips. For NVIDIA, this means that the window of opportunity in the Chinese market is narrowing: the longer the company waits, the more market share shifts to domestic vendors supported by a state policy of self-sufficiency.

Contradictory Data

Sources differ on how deeply Groq is integrated into NVIDIA. One description speaks of acquiring "rights to technologies and part of the assets" of Groq, while other publications indicate that around 90% of the startup's employees will move to work at NVIDIA and receive generous payouts, which in effect means the acquisition of the team, not just the technologies. Moreover, some materials portray Groq as a potential "hidden threat" to NVIDIA's own dominance, whereas in the context of the Chinese chip, Groq appears as a tool that NVIDIA is using to retain its position. These discrepancies in emphasis and in the assessment of the scale of integration should be taken into account when interpreting the news.

Strategic Significance: Adaptation, Not Circumvention

Taken together, these steps point to a shift in tactics: NVIDIA appears to be trying not simply to bypass restrictions but to restructure its products to fit the specific regulatory and market conditions of China. The use of LPU with an SRAM architecture allows the company to reduce its dependence on the components (HBM, advanced packaging) that are under the greatest pressure from export controls. If this approach works, the company will retain access to one of the largest AI markets; if it does not — Chinese alternatives based on RISC-V and SMIC's local manufacturing capacity may fill the vacated niche before NVIDIA offers a competitive product.