AI Trends

Computing power surges while net profit margin falls below 5%: Why can't industrial AI calculate the true profits of real-economy enterprises?

TMTPost · AGIUpdated Nov 15 16:18

AI summary

The root cause of industrial AI's inability to calculate the true profits of real-economy enterprises lies in treating sales on-time delivery, finance's low inventory, the workshop's high utilization rate, and procurement's low cost as a zero-sum game; single-point optimization will offset each other under the coupling of shared resources. The author uses a manufacturing enterprise with annual revenue of 10 billion yuan and a net profit margin of 5% for calculation: global decoupling can recover 170 million to 300 million yuan in pre-tax profit each year, equivalent to selling 5 billion yuan more in goods, and pull ROIC back from 3%–6% to above 15%. Lenovo's global supply chain IPS system has been running continuously for 18 years, with on-time delivery rates increasing by more than 30% and inventory turnover nearly doubling.

Body · AI translation

By GritMeng

Over the past two years, a visible chasm has opened up between the tech world and the real economy.

On one side, compute infrastructure and large language models are advancing at breakneck speed, with parameter counts and inference speeds doubling every six months, and the concept of Agents sweeping every tech summit; on the other side, the financial reports of the vast majority of real manufacturing enterprises are growing heavier and heavier—revenue surging from a few billion to tens of billions, software and digitalization investment rising year after year, yet net profit margins have generally fallen to the knife's edge of 4%–5%, and return on invested capital (ROIC) has dropped into the thin-margin range of 3%–6%, unable even to cover the cost of capital.

Why is it that silicon-based compute is already so cheap and powerful, yet cannot save the profit statements of real enterprises?

If you walk into the boardroom of any ten-billion-level manufacturing enterprise and sit in on a budget approval meeting, the answer will surface on the spot.

The R&D head applies to invest 300 million to develop next-generation hardware, the sales head applies to spend 200 million to expand overseas channels, and the manufacturing head applies to spend 500 million to build a new automated production line—the top boss usually decides extremely quickly, because everyone takes it for granted that R&D, channels, and production lines are the engines of profit. Only when it is the turn of the digitalization and supply chain head to apply to build a core planning and decision algorithm system, the expected benefits written on the project proposal are almost invariably the same: estimated inventory reduction of 5%–8%, on-time delivery rate improvement of 2%–3%, and 30% savings in manual scheduling hours.

Before a single line of code has been written, this project proposal has already strategically sentenced industrial decision intelligence to death: it is not the operating mother machine that determines the life or death of the enterprise, but merely a back-office auxiliary tool for "wringing the towel to save some pocket money." What is crueler still is that two years after the system goes live, when you open the company's consolidated income statement, even the promised 5%–8% book gains have vanished without a trace—no sooner has the finished-goods warehouse emptied a little than the semi-finished-goods warehouse swells up; no sooner has the on-time delivery rate been pulled up two points than the month-end expedited air freight charges and year-end write-offs of dead stock swallow it up clean.

Having battled for twenty-two years on the algorithmic foundation of hundred-billion-level manufacturing networks and complex systems, I see with increasing clarity a truth that the entire industry has overlooked: the reason industrial AI and decision systems cannot calculate big money is by no means that the return rate of the physical world is low, but that the "ruler" we have used to measure value over the past forty years has been engraved wrong from its physical foundations.

I. After the High-Margin Tide Recedes, the Submerged Wounds of Internal Friction Are Laid Bare

Why was it that over the past twenty or thirty years, real enterprises could still live well by treating their core scheduling systems as a "logistics supporting role"? Because the era of extensive incremental growth provided a sufficiently thick margin cushion.

When the industry's overall gross margin was as high as 40%–50%, enterprises were in fact doing only one thing in operational scheduling: guaranteeing only the quantity of front-end deliveries, without asking about the physical cost behind that quantity.

As long as the goods could be shipped out, the massive internal friction occurring underwater every day went entirely unexamined: shortage orders at the front of the queue locked up in advance the shared chips needed by complete-set orders behind them; the workshop, in order to make up equipment utilization rates, put incomplete semi-finished goods into production first, turning them into stalled work-in-process piled up in the aisles; procurement, in firefighting mode, scrambled everywhere to buy spot goods at high prices and placed duplicate orders, which at quarter-end turned into massive write-offs of dead stock; logistics chartered planes for air freight all year round; and hundreds of planners stayed up late manually scrambling for materials against offline spreadsheets.

In the greenhouse of 40%–50% gross margins, all this dead stock, redundant inventory, air freight premiums, and idle capacity were silently absorbed by the thick gross margin, and no wound could be seen on the financial statements. So from business schools to the software industry, everyone comfortably defined this central system as a low-value tool for "saving a few heads and helping pull some reports," and turned around to spend all the latest large-model compute on meeting-minute summaries, writing emails, and chat Q&A.

But today the tide has fully receded. When manufacturing net profit margins are squeezed down to 4%–5%, the gross margin cushion is pulled away, yet the extensive scheduling logic that has persisted for thirty years underwater has not changed for a single second. The several percentage points of internal friction waste heat that were originally hidden in the gross margin now pierce straight through the income statement.

In the incremental era, market dividends masked internal friction; in the stock era, silicon-based compute must eliminate internal friction. Unfortunately, the vast majority of so-called "industrial AI" today is still using chat dialog boxes to scratch the surface of the enterprise's skin, and has not reached into the arteries that are truly bleeding.

II. A Ruler Engraved Wrong for Forty Years: Mistaking "Coupled Common Origin" for "Zero-Sum Conflict"

Since there are bleeding holes of several hundred million buried in the arteries, why is it that for decades, whether top consulting firms or industrial software giants, when evaluating the value of core scheduling systems, have dared to count only those meager marginal gains?

The root of the disease lies in an implicit assumption that has ruled management science and traditional operations research for forty years: the default assumption that the "high on-time delivery" wanted by sales, the "low inventory" wanted by finance, the "high utilization rate" wanted by the workshop, and the "low cost" wanted by procurement are inherently zero-sum trade-offs that rise and fall against one another.

As long as these four goals are assumed in the underlying algorithms and business cases to be zero-sum contradictions, two disasters are mathematically foreordained.

The first disaster is that the ceiling on returns is welded shut on the very day the algorithm project is initiated. Since the four metrics are like a seesaw—press down one end and the other necessarily rises—the total pie is fixed. Whether you use a traditional linear programming solver to set weights and compromise, or deploy a few agents to game out a balance, at best you can shift a few crumbs at the edge of the zero-sum balance—either pad a few tens of millions in safety stock to trade for two points of on-time delivery rate, or sacrifice a bit of delivery speed to press down month-end inventory. A referee who can only mediate on a zero-sum balance cannot, mathematically, create net increments.

English text is being prepared.

In a strongly coupled network, when traditional systems ignore kit-completeness bottlenecks and lock resources exclusively based only on static priorities or isolated batches, a single microscopic mismatch will smash through four sets of books in the same second: orders that could have been completed and shipped have their shared components pulled away, causing delivery breaches; orders missing a small short-board item cling tightly to 99 high-priced long-board items sleeping in the warehouse, causing inventory buildup; the workshop cannot wait for truly kit-complete work orders and is forced to stop the line and change molds, causing effective capacity to sit idle; and procurement urgently buys high-priced spot goods and charters air freight, causing costs to soar.

Conversely, as long as the cross-locking of shared underwater resources is not unlocked, any single-point AI optimization will trigger violent cross-departmental vector cancellation: a local algorithm forcibly pushes down the inventory of one warehouse by 10%, and the production line dragged by the same coupling rope underwater runs out of material, while breach penalties and air-freight expediting fees immediately double and erupt in other accounts. Looked at locally, every project closure report is stunning; but when the year-end consolidated income statement is summed up, the forces in the four directions cancel each other out two by two inside the enterprise, and net effective work done approaches zero.

III. The Real Gold Mine of Physical AI: The "ROIC Squared Double Hit" of Numerator and Denominator Resonance

Once we throw away that zero-sum balance that has been drawn wrong for forty years and use a deterministic global physical AI engine to do something truly hardcore—penetrate millions of coupled nodes and eliminate cross-locking and cross-departmental vector cancellation at the shared-resource layer—what will happen to the financial books of a real enterprise?

Taking a typical manufacturing enterprise with annual revenue of 10 billion yuan, a net margin of 5% (annual net profit of 500 million yuan), and ROIC stuck at 3%–6% as the baseline, there actually lie four pure-profit and cash-flow gold mines underwater whose boundaries are strictly mutually exclusive and involve no double counting:

First is the marginal gross profit released by effective production-line working hours. Traditional factories permanently have 10%–20% of effective capacity that is not due to lack of orders, but is wasted on idle waiting for work stoppages caused by "goods in the warehouse but falsely missing materials due to mismatch" and on frequent line teardown and changeover; after unlocking the kit-completeness deadlock, the working hours left idle for nothing are directly converted into effective shipment gross profit.

Second is the direct reduction of slow-moving inventory write-downs and engineering-change scrapping. Eliminating "high-priced long-board main components waiting alongside short-board small materials and turning into dead stock," hard-cut scrapping during old-to-new material switching, and tail-end overbuying caused by isolated batch amplification keeps the real money that would otherwise be written off every year-end in the income statement.

Third is stopping the bleeding from abnormal firefighting costs and hidden gross-margin leakage. Plug the black hole that eats gross margin when frontline firefighting uses "high-grade materials to substitute for low-grade materials" without guardrails, and cut charter air freight, unpacking and rework, and high-priced spot premiums.

Fourth is the massive release of funds trapped by mismatch. Eliminate the false kit-complete work-in-process and redundant buffer inventory in the workshop where "ninety-nine wait for one," release 250 million to 400 million yuan of free cash flow in one go, and save the cost of capital tied up in inventory year after year.

Adding these four mutually exclusive hard accounts together, a manufacturing enterprise with 10 billion yuan in revenue can, without building a single new production line or adding a single machine, genuinely recover 170 million to 300 million yuan of pre-tax profit and comprehensive income every year.

Measure this number in the real business world, and you will see two highly striking facts:

First is the revenue equivalent. In a red-ocean market with a net margin of only 5%, recovering 250 million yuan of pure profit through internal orthogonal decoupling is equivalent to the sales team selling an extra 5 billion yuan of goods through a price war. Selling an extra 5 billion yuan of goods requires building factories, hiring workers, financing receivables, and bearing bad-debt risk; by contrast, every cent recovered from the waste heat of internal vector cancellation is pure profit with zero market risk, directly lifting the enterprise's net margin by 2 to 3 percentage points.

Second is the "squared double hit" to return on invested capital. The ultimate yardstick for measuring the quality of an enterprise is return on invested capital (ROIC = after-tax net operating profit ÷ invested capital). Traditional single-point management actions can never attend to more than one end; global orthogonal decoupling rewrites both the top and bottom of the fraction at the same time—on the numerator side, operating profit jumps by 40% to 60%; on the denominator side, materials engage precisely onto the line only at the moment they are truly kit-complete, inventory turnover nearly doubles, and net working capital contracts sharply. The nonlinear amplification produced by one multiplication and one division between the jump in the numerator and the contraction in the denominator is enough to pull a manufacturing enterprise's ROIC directly out of the 3%–6% quagmire back above 15%.

IV. Conclusion: From a Chat Toy in the Bit World to a Work-Doing Engine in the Atom World

Some may ask: is it really engineering-feasible to make delivery rates soar and inventory turnover double at the same time across millions of strongly coupled nodes?

Not only is it feasible, but fully fourteen years before generative large models entered corporate meeting rooms, it had already been proven at industrial scale in the atom world: in Lenovo's global supply chain integrated planning system (IPS) at the hundred-billion scale, an autonomous closed loop based on an orthogonal decoupling structure ran continuously and stably for 18 years, with more than 95% of daily planning decisions completed automatically "outside the human loop," producing a real record of on-time delivery rate up more than 30% while inventory turnover nearly doubled. Moreover, after this underlying logic was thoroughly abstracted and reconstructed into a general in-memory computing architecture, without any need for expensive GPU compute clusters, a single ordinary personal laptop (16 cores, 32GB memory) can, in 296 seconds, compute through 500,000 concurrent demands, 20 levels of BOM, 2 million material nodes, and 150,000 constraints for lossless network-wide substitution.

What determines whether a system can create real money in the atom world has never been how much it chats like a human on the screen, but whether its underwater data model and algorithm structure have unlocked the real physical couplings.

As this wave of artificial intelligence reaches its second half, both the tech world and the industrial world should make a sober valuation switch: AI that writes poetry, paints pictures, and summarizes meeting minutes in the bit world is certainly lively, but the super gold mine that can truly rewrite the balance sheets of major-power manufacturing and conjure several percentage points of net margin out of thin air in a red ocean will always belong to the "physical AI" that can compute through millions of atomic nodes and make the entire giant industrial system do work in the same direction with zero internal friction.

Source

TMTPost · AGI