---
title: "AMD Bought Li Feifei, But Teaching Every Watt of Electricity to Think Is a Hard Problem"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-10-08T16:08:58+00:00"
canonical: "https://ffcap.cn/en/research/amd-bought-li-feifei-but-teaching"
source: "https://uniqueresearch.substack.com/p/amd-bought-li-feifei-but-teaching"
language: "en"
---

# AMD Bought Li Feifei, But Teaching Every Watt of Electricity to Think Is a Hard Problem

[![cover](https://substackcdn.com/image/fetch/$s_!XuSy!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa7ae24d4-1c53-432f-bfaf-4df90f429db1_2730x1536.jpeg)](https://substackcdn.com/image/fetch/$s_!XuSy!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fa7ae24d4-1c53-432f-bfaf-4df90f429db1_2730x1536.jpeg)

_Industrial Physical AI · Deep Observation_

* * *

The endpoint of AI is compute, and the endpoint of compute is electricity — and the protagonist of electricity is exactly new energy and energy storage.

As the “Chat moment” of large models swept across the world, a group of Tsinghua-trained people were thinking about another problem: a language model can write poetry and code, but it cannot read the current curve of a storage station at three in the morning.

**Zhisuan Intelligent**, a “physical AI” company focused on industrial time-domain multimodal large models, is giving its own answer. The company is currently raising its Series B, with Unique Capital serving as financial advisor.

At two in the morning in Foshan, Guangdong, tens of thousands of cells inside an industrial-and-commercial storage station are still silently cycling through charge and discharge.

No one is on duty, and no one notices that a battery cluster is quietly drifting away from its health baseline — until the system issues a level-three warning: this battery cluster is at risk of lithium plating, and the charge/discharge strategy should be adjusted.

A possible thermal runaway is extinguished inside a millisecond-level signal shift.

This is not imagination. It is exactly what Zhisuan Intelligent’s “battery health diagnosis platform” does every day at dozens of storage stations across the country.

There is an open secret in the storage industry: batteries are a “black box.”

The internal state of a single 314Ah LFP (lithium iron phosphate) cell cannot be measured directly; it can only be inferred through external signals such as current, voltage, and temperature. If state-of-charge (SOC) cannot be estimated accurately, the station dares not fully charge or discharge; if state-of-health (SOH) cannot be calculated clearly, asset valuation becomes a muddle; if thermal runaway cannot be prevented, one accident could set the whole industry back a year.

Zhisuan Intelligent’s approach is to treat the entire storage station as a “living being that can speak.”

Edge agents capture nine categories of multidimensional data — current, voltage, temperature, SOC, SOH, and PCS operating conditions — at millisecond-level latency, build precise physical models of 280Ah/314Ah LFP and even sodium-ion cells, and then use time-series large models to catch micro-deviations in the massive waveforms that the human eye cannot detect.

Thermal runaway, lithium plating, internal short circuits, and self-discharge — hazards that traditional BMS “cannot see” — are broken down into a quantifiable, warning-capable, traceable three-tier risk system, with daily, weekly, and monthly reports generated automatically.

Today, Zhisuan Intelligent performs real-time monitoring and data collection on more than **3 million storage cells**, ranking among the top in the country.

The other side of safety is revenue.

At a centralized photovoltaic station in Xinjiang, traditional power forecasting relied on a single weather model, producing large forecast errors, severe curtailment, and heavy grid penalty pressure.

Zhisuan Intelligent fuses remote-sensing cloud imagery from the Fengyun-4 and Sunflower satellites with numerical weather prediction (NWP), “tokenizes” the 4D satellite cloud images, and feeds them together with ground time-series data into a multimodal Transformer — letting the model “read the clouds and know the sky.”

The result: the number of days meeting the forecast standard rose, curtailed energy fell markedly, overall key-node forecast accuracy improved by more than 3%, and station revenue grew **9.2%**, with a significant lift in economic returns.

At an integrated photovoltaic-storage-charging station in Guangdong, the three sides — solar, storage, and chargers — were each scheduling on their own, so revenue never improved.

Using reinforcement learning to generate the optimal charge/discharge strategy, Zhisuan Intelligent achieved coordinated optimization across solar, storage, and load: daily revenue rose **24%**, curtailed solar volume fell **62%**, and management efficiency improved sharply.

From a benchmark “unmanned storage station” project for a large state energy enterprise, this technical path received an even more extreme validation: operating costs fell **8.5%** and revenue rose **10%**.

Another driving force behind this revenue revolution is electricity-market liberalization.

2026 is an acceleration period for China’s power-market reform. Electricity prices begin to fluctuate like stock prices, and storage assets must shift from the static “build it and earn” arbitrage of the past to extremely complex dynamic market operation — peak-valley spreads, spot-market arbitrage, demand management, and grid dispatch instructions, interwoven with physical constraints, weather, and user behavior across a sea of variables. The complexity is no less than that of financial quantitative trading.

And yet most stations lack exactly what an “intelligent dispatching soul.”

From “keeping safety in check,” to “calculating revenue well,” to providing decision support for power trading — Zhisuan Intelligent makes every cell and every watt of green power “know its own state.”

Behind this capability is a complete technical chain: edge gateways handle millisecond-level collection and data cleaning; physical modeling translates “cold equipment” into “understandable semantics”; time-series large models capture long-range dependencies and cross-correlations in waveforms; multimodal fusion puts sensor curves, on-site images, maintenance logs, and pricing policy into the same cognitive framework.

While most peers still use point algorithms to solve point problems, Zhisuan Intelligent has already closed the loop of “perception — cognition — decision — execution.”

What is even more interesting is how “small” these cases are.

They are not beautiful lab numbers from papers, but rooftops in Foshan industrial parks, integrated solar-storage-charging stations in Longhua, Shenzhen, and PV-plus-storage stations in Shule, Xinjiang. They are the real money saved by peak shaving and valley filling, and the plain ledger of earning one more cent by wasting one less watt.

Industrial-storage owners calculate payback period: one more point of safety, one more year of life, one more percentage point of revenue — and the asset valuation is completely different.

Zhisuan Intelligent’s software-subscription plus performance-share model ties its own returns directly to station returns. The company has already achieved a comprehensive service gross margin of over **90%**, and customer stickiness comes from the simplest logic: if it works well, they renew.

Small things are not small. These percentage improvements, station by station, accumulate into a revaluation of storage assets — and the first turn of the “data — model — revenue” flywheel.

Over the past three years, the landscape of general large language models (LLMs) has largely settled: OpenAI, Anthropic, Google, Meta, plus Chinese players such as DeepSeek, Zhipu, and Kimi. Hundreds of billions of dollars in capital, compute, and talent have already taken their positions.

The battle for the “intelligence foundation” of text, code, and images is, in effect, no longer a startup’s battlefield.

But the industrial world is another matter.

A chemical plant, a storage station, a gas turbine — what they produce every day is not text, but massive, high-frequency, continuous sensor time-series signals.

This data is strongly structured, requires extremely high fault tolerance (≥99.9%, zero-tolerance errors), and is closed and sensitive. It cannot be scraped from the internet, nor can it be effectively represented by a language model.

“General large models are learned generalists; what industrial sites need is a ‘master craftsman.’”

Training specialized intelligence that understands mechanisms, processes, and decision control with a small but extremely precise set of on-site data is a completely different technical route.

China’s industrial AI practice over the past decade has fallen into a “small-model trap”: heterogeneous equipment and countless protocols create data silos; “one machine, one model” means tens of millions of custom small AI algorithms that cannot be reused, with maintenance costs rising exponentially; each model only optimizes locally, and no one considers global optimum.

On-site industrial data is also generally low-quality, fragmented, and hidden: fault samples are extremely rare, the labeling threshold is very high, and small and medium-sized enterprises are deterred.

More critically, continuous-process industry runs 24/7, like the human circulatory system, with zero fault tolerance — the opposite of the internet culture of “test, iterate, ship fast.”

Industry has long suffered from “fragmented small models.” What it needs is a foundation model that unifies the industrial time-series world the way language models unified the text world.

The divide between the two paradigms can be laid out clearly.

General large models feed on public internet text: trillions of tokens, abundant and fault-tolerant, using volume to compensate for quality, training learned “generalists.” Industrial large models feed on sensor time series, equipment states, process parameters, and quality-inspection records: scarce samples, hard to obtain, closed and sensitive, extremely low fault tolerance, training “master craftsmen” who deeply know the site.

The former competes on the upper limit of intelligence and the open-source ecosystem; the latter competes on closed-loop commercial validation, industry know-how, and full-stack integration.

Completely different data endowments determine that these are two barely overlapping markets — and that a language-model victory does not automatically equal an industrial-intelligence victory.

This is Zhisuan Intelligent’s track: **industrial time-domain multimodal large models.**

It natively processes multi-source, heterogeneous data — sensor waveforms, visual images, text manuals — maps continuous physical waveforms into tokens in an abstract latent space, and predicts the evolution of the physical world. It provides a “general infrastructure layer” for understanding complex, heterogeneous industrial business on the industrial internet.

In the founder’s words: “All things exist in the time domain.”

This is not Zhisuan’s niche judgment alone.

Over the past decade, AI has mainly learned the language and symbols recorded by humans. The real industrial world, however, is governed by physical laws — atoms, molecules, mechanical structures, thermodynamic gradients, and material fatigue. AI is moving from “semantic understanding” to “physical reasoning,” from “information generation” to “entity creation,” from virtual digital space into factories, workshops, and laboratories.

There is a plain consensus in the industrial-AI circle: process industry is “multi-time-series, little text, scarce images.” Any event on a factory floor can be traced back to a time-series record. An ideal industrial large model should be like a “superhuman engineer who has read every equipment manual and historical record in the world” — an **Artificial General Engineer (AGE)** — discerning patterns and extrapolating the future from digital sequences.

There is also a saying in this circle: “Industry does not accept hallucinations.”

The track consensus around physical AI and industrial foundation models is accelerating globally.

Jensen Huang, founder of NVIDIA, compares the AI industry to a “five-layer cake”: at the very bottom is energy, and only above it sit chips, infrastructure, models, and applications.

He proposes that “the endpoint of compute is electricity,” that “electricity is the new currency,” and at GTC 2026 introduced a new metric — **“tokens per watt.”** In an AI factory, every token is a measurable unit of profit, and the core anchor of token price is electricity.

Musk likewise warns: “The fundamental constraint on AI deployment is electricity.”

The International Energy Agency estimates global data-center electricity consumption at about 415 TWh in 2024, rising to about 945 TWh by 2030 — more than doubling.

At the same time, compute is spreading to the edge. As AI applications move from the training side to the inference and execution side, edge compute has become the essential foundation for the “last mile” of AI deployment. China’s edge-computing market is expected to exceed 130 billion RMB in 2026.

The 2026 Politburo meeting, for the first time, listed the compute network alongside the water network, the new-type power grid, the new-generation communications network, and the urban underground pipe network as one of the “six networks” of national key infrastructure.

This is exactly the tailwind for Zhisuan’s “cloud-edge-device” collaborative architecture: its edge agents are deployed closest to the data, closest to the motors and valves.

The endpoint of AI is compute, the endpoint of compute is electricity — and the protagonist of electricity is new energy and storage.

In 2025, China’s new-type storage installations grew 84% year on year, reaching a cumulative 136 GW / 351 GWh — over 40 times the level at the end of the 13th Five-Year Plan.

New-type storage has become the largest regulating resource in the new power system, with nationwide equivalent utilization hours rising to 1,195. China is the world’s first “electricity kingdom” whose annual electricity consumption has surpassed 10 trillion kWh, with industrial output accounting for more than 30% of the global total.

The direction of the tide is clear: the storage safety-and-O&M AI market is estimated at 5–10 billion RMB per year, while the storage power-trading AI market is on the order of a hundred billion RMB.

Whoever can make these 351 GWh of storage assets safer, smarter, and more profitable stands at the intersection of the era.

This “compute-power coordination” wave is also written into China’s policy language: the 2026 Government Work Report explicitly calls for “implementing new infrastructure projects such as ultra-large intelligent-computing clusters and compute-power coordination”; the outline of the 15th Five-Year Plan proposes building a nationwide integrated compute network, and four ministries jointly issued an action plan to promote two-way empowerment of AI and energy.

In the west, green power that once sold for only 0.3–0.5 RMB per kWh at peak wind and solar is being turned into “tokens” through compute. The electricity never crosses a border, but its value has already been exported. In storage stations, the same green power is repriced through peak-valley shaving and spot-market arbitrage.

The value of a watt-hour is being rewritten by AI.

This judgment is being validated by global capital markets.

Palantir, with its Ontology, front-deployed engineers (FDEs), and AIP platform, has surpassed a market capitalization of 450 billion USD.

Bezos’s Prometheus is built around the core goal of creating an “Artificial General Engineer” (AGE); it completed a 12-billion-USD Series B in June 2026 at a valuation of 41 billion USD, while setting up a 100-billion industrial strategic fund with a blunt, aggressive strategy — directly acquiring manufacturing companies to obtain data.

Industrial AI has become the next main channel that global tech giants recognize.

Zhisuan Intelligent’s founding team comes from the Department of Computer Science at Tsinghua University.

Several core co-founders are serial entrepreneurs from different Tsinghua schools, with years of industrialization experience. Other core team members come from Peking University, Fudan University, the University of Chinese Academy of Sciences, and Shibaura Institute of Technology in Japan, while senior executives from JD Industrial, Yuntian Lifely, China Mobile, and China Unicom have joined. Research strength and industrial resources complement each other.

Backed by the Next-Generation Internet R&D Center of the Shenzhen Tsinghua Research Institute, the team has built joint laboratories with many central and state-owned enterprises. R&D staff account for more than 70% of the team, and the master’s/PhD ratio exceeds 50%.

They have participated in more than 10 national major science-and-technology projects and hold over 50 intellectual-property rights. The team has published 50+ high-level papers in journals and conferences such as IEEE TC, TMC, TMM, TVT, TNSM, TNSE, TII, _Computer Communications_, _Computer Networks_, as well as IEEE INFOCOM, ICDCS, IWQoS, and ICCCN.

Between research and production, the distance is only a corridor: the lab does source innovation in next-generation internet, edge computing, and new-energy-oriented large models, while the industrialization team turns it into products customers are willing to pay for.

On the industry side, Zhisuan Intelligent has formed stable cooperation with more than a dozen central and state-owned enterprises, including CNPC, CNOOC, China Huadian, PowerChina, Energy China, China Resources Power, China Merchants Group, Shenzhen Energy, and China Mobile.

Since its founding in 2020, the company has received investment from professional institutions such as Linear Capital and Inno Capital, and in 2023 won the second prize of the Shenzhen Science and Technology Award for scientific and technological progress.

Zhisuan Intelligent’s product system carries a complete “Dune universe” naming scheme:

-   At the top is the physical-AI industrial foundation large model **Melange** — from Melange 1.0, an industrial time-domain multimodal model, to Melange 2.0, an industrial causal time-series model, to Melange 3.0, an industrial physics model fusing causality, space, and physical laws.
    
-   In the middle is the **Dune** industrial decision platform / energy agent, replicated rapidly across industry versions such as energy storage, water, and gas.
    
-   At the bottom is the **Hulud** series of edge-end agents — industrial intelligent sensors, gateways, and collectors — responsible for collecting data and holding the data threshold.
    

This three-layer architecture of “foundation brain — industry agent — edge hardware” forms a complete loop from perception and decision to control, aiming to replace the experiential judgment of industrial master craftsmen and plant managers, pushing digitalization and intelligence into the hardest core of continuous-process industry.

The company chose to start from storage — where demand is strong and standardization is high, the most cost-effective starting point in its industry quadrant — then expanded to new-energy power forecasting, power trading, and charging stations, and has since validated scenarios such as gas turbines (1,967 concurrent feature measurement points, simulation accuracy above 95%, and major-fault warnings 48 hours in advance), logistics, and water.

The technical roadmap is equally clear: build a 1-billion-parameter industrial model in 2026, a 10-billion-parameter foundation in 2027, and a 100-billion-to-trillion-parameter industrial foundation model by 2028, ultimately becoming a cross-industry industrial-foundation-model provider.

Across the Pacific, the U.S. layout has already unfolded: first relying on bottom-layer chips and compute, then leading the world with general large models, and now doubling down on world models, physical AI, and AGE industrial-intelligence engineers, hoping to maintain global leadership.

Palantir positions itself as “the core operating system of global decision-making and integration,” as “the shield and weapon of the Western world,” serving national security and physical productivity. Bezos’s Prometheus uses a hundred-billion-dollar fund to directly acquire manufacturing companies, targeting “AI models that understand the physical world.”

In Silicon Valley’s narrative, industrial AI is no longer just a business — it is a contest of national strength.

After reform and opening up, we became the world’s factory through demographic dividends, and then led manufacturing and innovation through the engineer dividend. Today, we possess the most complex and largest energy and industrial system in the world, having accumulated massive industrial data that has not yet been effectively used.

Over 30% of global industrial output, the world’s first 10-trillion-kWh power system, and the structural advantage of “compute-power coordination” — while U.S. data-center electricity prices repeatedly hit new highs, China’s western green-power cost advantage is being converted into competitiveness for the AI industry.

The “narrow ditch” of chips may exist, but the “moat” of electricity and industry is deeper and wider.

“China’s industrial AI should not fall behind — and it cannot afford to lose.”

The decisive point of this contest is not in the office, but on the factory floor.

The core capability system of industrial large models is, in the end, three things: scenario acquisition that understands the site, the process, and the data; edge management that understands equipment, deployment, and compute; and data analysis and modeling that understands analysis, mechanisms, and models.

“Proprietary data × professional models × expert scenarios” forms the closed loop of industrial intelligence — and this is exactly the battlefield Chinese companies know best.

Zhisuan Intelligent’s route is precisely its answer to this contest: it does not fight general large models, but lets general models provide the “intelligence foundation” while industrial models inject the “industry soul,” building a cloud-edge-device collaborative industrial-intelligence system and a high-margin, high-retention, high-priced, deep-relationship industrial-intelligence platform.

The capability coordinates of this “industrial brain” are clear:

-   **Industrial ontology cognition** — deeply understanding product structure, material mechanisms, process constraints, and quality standards, reconstructing fragmented industrial experience into a computable engineering knowledge graph.
    
-   **Physical-law reasoning** — fusing multiphysics simulation with deep learning to make millisecond-level precise predictions of force, heat, sound, light, electricity, and fluid processes.
    
-   **Human-machine co-creation** — deeply embedding into R&D and operations to become the autonomous intelligent partner of human engineers.
    
-   **Enterprise collaborative connection** — breaking the technical barriers between giants and small and medium enterprises, letting engineering intelligence flow seamlessly on an open platform.
    

Changing watts with bits, rewriting industrial civilization with engineering intelligence — that is the weight of AGE.

Storage is the starting point, not the end.

The methodology can transfer because continuous-process industry shares the same “data genes.”

Whether it is an aluminum electrolysis workshop, a paper machine, a cement rotary kiln, a fine-chemical reactor, a bioreactor, or an urban water-supply network, gas network, or district heating station, their data endowments are highly isomorphic: multiple time series, little text, strong physical constraints, and continuous, non-stop production.

Temperature, pressure, flow, vibration, current — any anomaly can be traced back to a time series.

This is the true meaning of “all things exist in the time domain”: the skills Zhisuan Intelligent built in storage stations — millisecond-level collection, physical modeling, time-series tokenization, cross-modal semantic alignment, and physical-causal reasoning — transfer to another production line without starting from zero.

What changes is the industry; what does not change is the method.

The company’s name had the answer written into it all along.

**Zhi-Suan (Weave-Compute)** places “weaving” and “compute” together: the warp is time, the weft is modality (current waveforms, on-site images, O&M logs, weather cloud images, grid instructions), and the loom is compute.

Edge agents are the shuttles shuttling back and forth, weaving the compute and data scattered at equipment ends, valves, and motors back into one network. The industrial ontology is the unified “pattern blueprint,” ensuring that cloth woven in different workshops and industries can eventually be stitched into the same bolt of brocade.

The compute of a single device is like a fragile, isolated thread; only when thousands of threads interlace according to the same warp and weft can they become an industrial “fabric” that both bears load and breathes — this is the essence of cloud-edge-device collaboration, and another language for how industrial AI understands the world: not splitting the factory into countless isolated algorithms, but re-weaving industry into a whole.

The order of advance on the industry map is equally clear.

Zhisuan Intelligent starts from storage and new-energy power forecasting, where demand is strongest and standardization highest, forging and validating the methodology; then spreads horizontally into infrastructure such as water, heating, gas networks, and communications networks; then pushes into the more complex continuous-process depth of oil refining, fine chemicals, steel smelting, non-ferrous metal electrolysis, cement and glass furnaces, pulp and papermaking, pharmaceutical and food fermentation, polysilicon purification, and coal chemical industry.

This path relies not on “laying on more people” but on “weaving section by section”: first building trust where trust is hardest to earn (safety, zero fault tolerance), then extending along directions of similar mechanisms and isomorphic data.

Each additional station delivered brings an on-site time-series sample that no one else can obtain; the denser the data, the stronger the foundation; the stronger the foundation, the faster new scenarios go live and the lower the marginal cost.

This is the unique compound interest of industrial AI — it does not happen in internet public data, but between real pipes, valves, and production lines.

The goal is therefore clear: build a cross-industry industrial foundational model, moving from 1 billion and 10 billion parameters toward 100 billion and even trillion parameters, uniformly supporting language, time-series, image, and video input, and becoming the “general infrastructure layer” that continuous-process industries can call upon.

In the end, Zhisuan Intelligent is not making ten thousand custom small models; it is making one foundation model understand every factory.

Let every production line no longer depend on scarce master-craftsman experience, and let industrial intelligence become a public capability like water, electricity, and gas. That is the full meaning of “weave”: using compute to weave a network so that every node of industry can be understood, predicted, and optimized.

_Building a Smarter and Better World_ — what is being woven is an intelligent network covering the entire industrial world.

The endpoint of the story points to the company’s vision: building an **Artificial General Engineer (AGE)**, making industrial intelligence a public good shared by all humanity.

On that day, every chemical plant, every storage station, and every gas turbine will have its own “super brain”; the industrial machinery that has run continuously for centuries will be truly awakened.

_Building a Smarter and Better World._

This is both the mission and the meaning of this contest.

---

Original publication: https://uniqueresearch.substack.com/p/amd-bought-li-feifei-but-teaching
On-site reading page: https://ffcap.cn/en/research/amd-bought-li-feifei-but-teaching