On November 15, Science and Technology Daily reported that analog computing chip company AnnaCore had completed the tape-out of its self-developed AX100 in October. According to the company, it is the world's first fully integrated analog computing chip capable of solving a matrix equation in a single step. Each solve takes less than 100 nanoseconds and consumes about 80 milliwatts of power.

Of those three figures, the most striking is not the speed but the 80 milliwatts, roughly the power draw of an indicator light.

500 Million Multiply-Accumulate Operations Are the Starting Point

First, it helps to clarify what is being solved.

Parameter optimization in large-model training, motion control and spatial planning for robots, and signal processing at communications base stations all repeatedly involve solving matrix equations. That is not an analogy; it is the mathematical form shared by all three tasks.

Digital chips translate the solve into vast numbers of iterative multiply-accumulate operations. As the matrix grows, the computational load rises with the cube of its dimension. Solving a 512-by-512 matrix equation, for example, requires about 500 million multiply-accumulate operations and takes tens to hundreds of microseconds. The cube of 512 is 134 million; once the constant factors involved in matrix inversion are included, 500 million is a plausible order of magnitude, indicating that the figure was not chosen casually.

Analog chips take a different approach. AnnaCore CEO Xiang Rui told Science and Technology Daily: “The matrix equation is mapped directly onto a circuit. Once the circuit reaches equilibrium, the result is read out in a single step, with no iteration required.”

That statement is worth unpacking. A circuit can solve equations not because it contains a hidden algorithm, but because Kirchhoff's laws are themselves a system of linear equations. The node voltages in a resistor network naturally satisfy current conservation. Apply voltages at the inputs corresponding to the equation's coefficients, and once the circuit reaches a steady state, the voltage at each node represents the solution. The solving process is simply the circuit settling into equilibrium after it is powered on, with no clock advancing the computation step by step.

This is what it means to replace algorithms with physics. The trade-off follows directly: the matrix is encoded in the devices' conductance, so a new problem requires the hardware to be reconfigured. The architecture is inherently not general-purpose.

Precision Was the Problem for Half a Century

Analog computing is not new. In the 1940s and 1950s, analog computers represented values using voltages and gears, and were used to solve differential equations and calculate ballistic trajectories. Once digital computers matured, the industry shifted decisively toward them for three reasons: reproducibility, error correction and storage. Analog computing then largely disappeared for nearly half a century.

The real constraint was never speed, but precision. Error rates in analog circuits long remained around 1%. If a system computes quickly but produces results that are off by 1%, any time saved is lost, making the approach pointless.

The turning point came in October 2025. A team led by Sun Zhong, a researcher at Peking University's Institute for Artificial Intelligence, published a paper in Nature Electronics describing an analog matrix-computing chip built with resistive memory. The team improved precision from roughly one part in 100 to one part in 10 million. According to Science and Technology Daily, the work reported single matrix-equation solves that were three orders of magnitude faster than advanced GPUs, with two orders of magnitude better energy efficiency.

One part in 10 million is roughly equivalent to 24-bit fixed-point resolution. That figure matters because it crosses a critical threshold: neural-network training typically requires numerical precision of 16 to 32 bits, a bar previous analog computing systems could not reach.

Only 12 months separated the paper from tape-out. That pace relied on China's mature manufacturing and packaging supply chain. The architecture does not require an advanced process node, so it does not have to queue for the industry's most constrained capacity.

A One-Year-Old Company Took a Paper to the Production Line

AnnaCore runs a lean operation.

Corporate records show that Beijing AnnaCore Technology Co., Ltd. was established on November 14, 2025, and is registered in the Yanqi Economic Development Zone in Beijing's Huairou district. Its legal representative, Xiang Rui, directly owns a 75% stake. The core team came from Peking University's Institute for Artificial Intelligence and has long worked on analog computing, compute-in-memory and AI chip architectures.

Capital arrived quickly as well. In June 2026, six months after its founding, the company raised several hundred million yuan in a round led by Matrix Partners China, with participation from FreeS Fund and Alpha Startup Fund. Existing investors iFlytek Venture Capital and Zhongying Venture Capital also increased their backing.

With fewer than 50 employees, the company reached tape-out in 11 months, a pace that would be almost impossible for a large digital chip. Backend design, verification and the software stack for digital chips can require hundreds of people working for years. Analog computing dispenses with precisely those layers: there is no instruction set, no compiler to write and no CUDA-like ecosystem to maintain.

Everything omitted today will become work that needs to be done later.

The 92% Figure Is the Most Valuable of the Three Sets of Results

The team released test results across three types of workloads, all based on its own testing.

For large-model training, tests on an autoencoder with 2.93 million parameters and a Transformer with 2.72 million parameters showed that a second-order optimization method reduced training-update costs by 92% and 56%, respectively. In embodied AI, end-to-end motion-control and planning algorithms ran 150% to 500% faster, while core components of 3D reconstruction and SLAM accelerated by 200% to 500%. In wireless communications, processing a 128-by-128 channel Gram matrix delivered 20 to 207 times the effective throughput of existing digital ASICs and GPUs, along with 27.6 to 61.3 times the effective energy efficiency.

Most reports have focused on claims of speedups of several hundred times. The 92% figure deserves closer attention because it reflects a change in the choice of algorithm, not simply an increase in speed.

The standard approach to training neural networks uses first-order optimization, including methods based on gradient descent. That is because every step of second-order optimization requires solving a curvature-related system of linear equations. At scale, solving that system becomes prohibitively expensive. As a result, second-order methods look elegant in textbooks but see little use in industry.

The analog chip cuts the cost of solving equations to almost zero, putting previously unaffordable second-order optimization methods back on the table. The 92% and 56% figures refer to reductions in “training-update costs,” not training time. Put simply, reaching the same target requires fewer steps, or each step delivers more value.

That is fundamentally different from making an existing algorithm run 1,000 times faster. The former changes the algorithm; the latter changes the hardware. Hardware upgrades are constrained by process technology, while algorithmic changes are not.

The scale, however, needs to be kept in perspective. The tests used models with 2.93 million and 2.72 million parameters, while today's large models have hundreds of billions of parameters, a gap of four to five orders of magnitude. Experiments on models with two or three million parameters can prove that the mechanism works, but not that it works at scale.

Not Requiring Advanced Process Nodes Does Not Mean Any Process Will Do

The claim most vulnerable to misinterpretation is that the architecture “does not rely on advanced process nodes.”

It does offer supply-chain autonomy: China has ample capacity at mature process nodes, so the chip does not have to wait for the most constrained production lines or get caught by equipment restrictions. From that perspective, it is accurate to say the company has taken a different path.

But there is another side to the argument. Digital chips absorb process variation through redundant design, error-correcting codes and timing margins; even if a device underperforms slightly, the logic remains correct. Analog chips have no such latitude. Their computational precision depends directly on the devices' actual conductance, meaning device mismatch, temperature drift and noise all flow directly into the result. Analog circuits therefore often demand greater process consistency than digital circuits. What they require is not narrower line widths, but more stable parameters.

In other words, compatibility with mature processes does not mean compatibility with just any production line. Whether this chip can be mass-produced depends on whether a foundry is willing to tune its process parameters and maintain tight consistency for a nonstandard chip.