Why China Is Betting on Analog Chips for the Next AI Generation

01The Claim, and How to Read It

The news, as reported by the South China Morning Post and circulating widely since, is straightforward: a Chinese research team has demonstrated an analog AI chip that, on certain workloads, is claimed to run up to 1,000 times faster than a flagship Nvidia GPU while consuming a fraction of the power. The chip is described as a proof of concept built around in-memory analog computation rather than the digital arithmetic that dominates machine learning hardware today.

The right way to read the 1,000x figure is as a per-workload result, not a universal one. Such claims are typically measured on a narrow class of operations, often matrix-heavy inference, where the analog approach is strongest. Nobody is claiming an analog chip will render a game engine or run a database 1,000 times faster. But the claim matters less for its precision than for its direction: it signals that serious research money, in a country with strong incentives to look past digital silicon, is moving toward compute-in-memory architectures.

02Physics as a Co-Processor: How Analog Actually Computes

Digital chips compute by shuffling bits between memory and arithmetic units, paying energy at every step. Analog in-memory computing takes a different route: it uses the physics of the circuit itself as the arithmetic engine. Conductance values stored in a crossbar of memory devices represent the weights of a matrix. Input voltages are applied along the rows, and by Ohm's law each device produces a current proportional to voltage times conductance. By Kirchhoff's current law, the currents sum naturally along each column.

The result is that a full multiply-accumulate operation, the inner loop of every neural network, happens in a single step, in the physical domain, effectively for free. Where a digital GPU schedules thousands of multiply and add instructions, the analog crossbar simply lets current do the math. This is why the efficiency numbers from analog demonstrations look so extreme: the operation is not being executed faster, it is being executed in a different physical regime where movement of data has been engineered out of the loop.

03Why Neural Networks Tolerate This, and the Wall It Sidesteps

Modern AI is dominated by multiply-accumulate (MAC) operations, trillions of them per inference. On digital hardware these are bounded by the memory wall: fetching weights from DRAM or HBM costs orders of magnitude more energy than the arithmetic itself. Processors spend most of their power and their idle cycles moving data, not computing on it.

Analog in-memory computing dissolves that wall by placing the computation where the data already lives. And crucially, neural network inference happens to be unusually tolerant of the approach. Deep models are statistical machines trained amid noise and regularization; they do not need 32-bit precision in every weight to classify an image or generate text. Quantization to 8 bits or 4 bits already proved that digital AI can shed enormous precision with little quality loss. That tolerance is precisely what analog exploits, converting a precision budget into an energy budget.

04The Precision Tax: Noise, Drift, and the Converter Overhead

Analog computing is not free lunch. Signals are continuous, so noise accumulates, device characteristics drift with temperature and aging, and the achievable effective precision in practice has hovered around 4 to 8 bits per operation. Digital designs can chase exactness to 64 bits; analog must live with the physics of the substrate.

The other tax is conversion. Real systems still live in a digital world, so inputs must be converted to voltages by digital-to-analog converters and results read back through analog-to-digital converters. These ADCs and DACs are often the dominant cost, in area and in power, of the whole chip, and they scale badly as precision increases. This is the honest engineering reason analog AI has spent a decade in ISSCC papers and Nature Electronics demonstrations without displacing GPUs: the core operation is nearly free, but everything around it is expensive. Any 1,000x claim survives only on workloads where the analog core dominates the converter overhead.

05Export Controls and the Post-Digital Bet

The geopolitical context is impossible to separate from the technical one. Successive rounds of United States export controls have restricted China's access to the most advanced AI GPUs and to the leading-edge fabrication processes needed to compete on digital silicon directly. Faced with a road closed at the frontier, the rational strategy is to change the road.

That is what the analog pivot is: a workaround bet. If the digital GPU race at 3-nanometer and 2-nanometer nodes is unwinnable given tooling and export restrictions, then architectures whose economics do not depend on the most advanced digital process become attractive. Analog crossbars can be built on mature, domestically available process nodes, because their advantage comes from physics and architecture rather than transistor density. For Beijing, funding compute-in-memory research is not just science policy, it is a hedge against being permanently locked out of the dominant computing paradigm. The wager is that the next generation of AI hardware might reward a different kind of physics.

06Analog Is Old: The Machines That Came Before

It is worth remembering that analog computing is not a new idea disrupted by digital; it is the elder technology that digital displaced. Differential analyzers solved differential equations with rotating shafts and wheel integrators from the 1930s onward, tracing the trajectory of artillery shells before digital machines existed. Naval gunnery computers aboard warships, electromechanical analog devices, kept range and lead angles on moving targets for decades.

Even electronic analog computing, the operational-amplifier machines of the 1950s and 1960s, was a mainstream engineering tool for simulation and integration before the digital revolution made it obsolete. The pattern is consistent: analog dominates wherever the problem is continuous and the physics can be recruited to solve it, then loses to digital's exactness, reproducibility, and generality. What has changed in 2026 is that machine learning, for the first time, is a mass-market continuous-ish problem that fits the analog sweet spot.

07What Would Have to Be True for Analog AI to Scale

For the demonstrations to become products, several things must hold. Analog arrays need to be programmable and retrainable at scale, not just hardwired for a single demonstration model. Drift and noise must be compensated, either through on-chip calibration or training methods robust to device variability. Converter overhead must shrink so it stops eating the energy budget. And the software stack, compilers, and model formats, must map real workloads onto crossbars as smoothly as CUDA maps onto tensor cores today.

None of that is impossible; most of it is under active research in laboratories worldwide, and the published efficiency literature has been improving steadily. But it is the difference between a striking paper and an industry. What the Chinese result guarantees is not victory but attention: the assumption that AI compute will remain digital, von Neumann, and GPU-shaped is now being challenged at the architectural level, and the challenge is being funded like it matters. That alone makes the 1,000x claim worth watching, whatever its precise magnitude.