AI in the News Makes Money by Cutting Jobs, While AI in Industry Starts Solving Real Problems

Frontier models with tens of trillions of parameters keep reshaping the world in the news, causing global stock markets to swing by hundreds of billions of dollars in a single day. But the technology a company uses usually depends on its budget. Outside the tech industry, which hires large numbers of high‑salary programmers, the most advanced large models are mostly talk fodder.

A cutting‑edge technology truly penetrates every industry and changes everyone’s life only after it ceases to be sexy and becomes “boring”ly mature. The latest to reach this stage are the pre‑Transformer stars—models based on CNN and RNN architectures.

Ten years ago, the tech media’s most watched innovative companies were the “AI Four Little Dragons” building these models. When Nvidia held an event in China, Jensen Huang listed Hikvision as an AI partner as important as Alibaba Cloud.

Today, those companies are no longer discussed, and the once‑cutting‑edge technology has become cheap enough to enter production lines.

On the eve of Singles’ Day, several Jitu logistics centers installed new AI servers. Each server monitors the real‑time footage from thousands of cameras in a logistics center, refreshing eight times per second. When a package gets stuck or a sorter throws a parcel forcefully, the server will issue a notification within a minute, dispatching someone to handle it.

On the production line of an Apple supplier in China, more than a hundred laser‑printed titanium alloy parts are assembled into a foldable‑screen hinge. Each part and the assembled hinge are finally placed on a fixture for inspection. There are no workers hunched over magnifying glasses as in documentaries. Instead, a row of a dozen cameras does the work, capturing ten frames per second and adjusting positions on the fly. All photos are inspected on an AI server that looks like a PC on the assembly line. Humans only need to carry away the qualified parts.

After hours, some judicial departments keep small servers in their office buildings processing historical archives. This data cannot—and does not need—to be uploaded to external large‑model services; the 27B Qwen running on the building’s server can organize it into a database. The next morning, staff can ask questions just like using Doubao and instantly retrieve how related historical cases were handled, complete with sources.

These scenarios don’t require the largest models. Visual models that were popular a decade ago, combined with smaller large language models—typically only a fraction or even a few tenths of the size of Kimi K3—can already accomplish a lot of work.

The GPUs in these scenarios are all provided by Denglin, priced from a few thousand to tens of thousands of yuan each. In the first four months of this year, Denglin sold over 10,000 units. In price‑sensitive industries, each unit has replaced Nvidia’s inference chips, and even gaming graphics cards.

Today’s courier industry looks more like an automated factory. Jitu’s massive 300,000‑square‑meter sorting hub has a total conveyor belt length of 40 km. Packages delivered by trucks are dumped onto sorting machines, scanned, then routed onto different belts before being loaded onto trucks for nationwide delivery.

The conveyor belt here experiences over 1,000 failures per day on average. If a package gets caught on a guardrail, it can block the packages behind it and jam an entire belt; sometimes a torn package gets pulled into the machine, stopping the whole line; and if the waybill is damaged, you can’t trace where the package came from or where it’s headed, creating the “headless item” that the courier industry talks about, waiting for a customer compensation claim.

A transit hub like this has nearly 3,000 cameras dedicated solely to monitoring the production line and sorting operations.

“People simply can’t keep up,” said an engineer. The monitoring room’s large screen usually shows 16 feeds at once. Each feed is observed for one minute to check for package anomalies, taking about three hours to complete a full cycle, with more than 170 minutes of blind spots in between.

Managers schedule patrol routes based on experience. The most accident‑prone route, a one‑kilometer stretch, has three people patrolling continuously. Even so, incidents take an average of 30 minutes to be discovered, and handling them takes a few more minutes.

SF Express was the first to deploy AI at scale to monitor every production step in its transit centers, and J&T followed closely behind. This year, several J&T transit centers received AI servers, each equipped with eight KS38 GPUs.

If you treat the AI on this server as a living being: it sits in total darkness, facing a screen larger than IMAX, composed of 1,000 1920×1080 frames. The massive screen refreshes 8 times per second, and it has to watch it 8 times. Running in its mind is an “image segmentation model”: first it recognizes the package in the frame, outlines its contour, then determines whether it’s stuck.

Now the package is stuck, the server notifies a person within 1 minute, resolves the fault within 10 minutes, and resumes operation. Humans no longer need intensive patrols or staring at screens; they just follow the AI’s prompts to handle the situation on site.

The detection uses open‑source vision models such as YOLO, some only dozens of MB in size, not large, but the challenge is processing 8 times per second. Every 0.125 seconds, the camera sends back a compressed video stream; the hardware decoder on the GPU first restores the video into individual frames, then extracts the portions to be inspected, adjusts the size and data format, feeds them to the corresponding model, and completes the detection.

Developers tried the earlier Nvidia V100, needing four servers and 32 cards to have enough compute—GPT‑3 was trained on more than ten thousand V100s. They tested another domestic GPU that had sufficient compute, but it couldn’t decode 1,000 video streams simultaneously, so they had to buy two machines, driving up the cost. The physical‑goods industry has to squeeze every penny. In the end, Jitu bought a server from Denglin, costing several hundred thousand yuan each.

The servers have spare compute, so the engineering team uses them to run large models that detect violent sorting. From several courier companies we learned that during sorting, packages get dropped, and couriers, pressed for time, upload photos early and falsely report delivery. These irregular practices bring platform fines and compensation, averaging more than 2 cents per order, amounting to hundreds of millions of yuan a year for a single company.

Packages vary in weight, and sorters have different strength, so just watching the motion makes it hard to tell if someone is throwing something too heavy. The engineering team first uses a small model to flag packages that move too fast during sorting, then compares body movements to judge possible violations, achieving only about 70‑80% accuracy. But when the suspicious cases filtered by the small model are handed to a large model, the accuracy can rise to over 90%, “and that’s usable.”

AI stealing people’s jobs, is the most talked‑about risk today. But in the low‑margin industrial sector, a more common situation is: companies simply cannot afford, nor can they find enough people willing to do the tedious work, so they have to let machines do it. Quality inspection in factories is exactly that kind of work. In the past, young workers would spend ten‑plus hours a day staring through a magnifying glass, looking for tiny scratches and cold‑solder joints on screens and circuit boards, often ending in tears.

At the entrance of an industrial testing company in Dongguan, our phone camera was covered with tape—this is an Apple requirement. Starting two or three years ago, Apple began mandating that manufacturers use AI optical inspection equipment, first purchasing it for them, debugging it to prove it works, and then having the manufacturers buy it themselves. The company we visited is one of the makers of those inspection devices for Apple.

Its product sells for over 1 million yuan per unit and sits at the end of the customer’s production line. A robotic arm picks up the finished parts onto the inspection station and secures them; up to 20 cameras start shooting, ten frames per second. The equipment continuously changes lighting colors, adjusts part positions, and determines whether dimensions are within spec and whether the surface has scratches.

Last month, we visited several testing companies in Shenzhen and Suzhou. Some check whether hundreds of components on a circuit board are all soldered, costing tens of thousands of yuan per unit; some check whether chips and display panels are damaged, with a single unit fetching up to tens of millions of yuan. There are both cheaper and more expensive options as well.

Visual inspection originally relied on engineers writing rules, and a CPU was sufficient. Shortly before the pandemic, inspection equipment began using AI to recognize defects and added GPUs.

These companies initially came standard with Nvidia RTX 3060, 4070 and other gaming GPUs—Nvidia’s lowest‑margin, best‑value products. But in the past two years, Nvidia’s gaming GPUs have risen in price along with AI chips, adjusting several times a year, and sometimes being out of stock for a month or two. Now these firms have all switched to Denglin’s KS19 GPU.

This industry has hundreds of companies, each specializing in a niche area, from picking fruit to inspecting wafer damage. Annual revenues of a few tens of millions are considered low; 200‑300 million is already medium‑scale. Even a hundred of them combined might not exceed the financing amount before SenseTime went public. They can’t afford AI PhDs with an annual salary of 2 million yuan, but that has led them to create sufficiently cheap equipment, bringing AI into equally low‑margin factory production lines and improving product quality control.

Like Jitu, inspection companies are also starting to pair large models with small models.

When Denglin was founded in 2017, chips could still be freely bought and sold between China and the US, and the story of domestic GPU replacements hadn’t emerged yet. At that time, any company making AI compute chips had to answer one question: why would customers use them instead of Nvidia?

This issue still holds today. You can’t legally buy NVIDIA’s most powerful chips in China, but many of its inference chips are still sold and used normally. Major internet companies and industrial enterprises have deployed hundreds of thousands of NVIDIA A10, A30, L40, L2, and L4 inference chips. Even the decade‑old inference chip Tesla T4 is still used by Alibaba Cloud for its GPU instances. For some industrial customers, domestic manufacturers still have to compete with NVIDIA gaming graphics cards that cost several thousand yuan to run vision‑detection models.

NVIDIA’s advantage is not just single‑card performance. TSMC’s capacity, a mature supply chain, and massive shipment volumes help lower production costs and amortize R&D investment. Decades of accumulated chip‑related patents and designs can be reused across multiple products, further gaining cost advantages.

CUDA is another major strength of NVIDIA; after 20 years on the market, more than 6 million developers use it to write programs, making it the preferred development platform for AI applications. Switching to other chips, including Google TPU and other ASICs, requires code changes and re‑testing.

Chip startups lack these advantages; they are smaller in scale and face higher production costs. If they try to build a product like NVIDIA’s and chase the most advanced components, they cannot be cheaper than NVIDIA.

“If a company’s competitiveness mainly relies on buying memory, buying IP, and then spending money to lock in capacity, in the end it’s just a race of who has more money.” Co‑founder Wang Ping of Denglin told us that after early research they decided to invest the money first in their own core compute design and software, get customers’ workloads running, then scale up and build more complex products.

Denglin ultimately chose AI as the breakthrough: strengthen the matrix computations commonly used in AI when designing the chip, while retaining the general‑purpose compute capability customers need, so that mature models for detection, recognition and other scenarios run more cost‑effectively—“be a specialist in AI.”

Wang Ping said Denglin defines its products “basically based on the metrics of the customer’s application.” The team first estimates how much compute can be packed within a given chip area budget, then uses power‑consumption models to predict performance, and only proceeds with design once the customer is willing to pay.

Video‑analysis customers want a single card to handle more cameras, so Denglin must also boost decoding capability—if compressed video can’t be restored to frames in time, the downstream model has nothing to recognize. Therefore, when defining the product, Denglin treats support for more decoding streams as a key metric to compete with NVIDIA’s comparable offerings.

The compute portion also allocates resources according to these requirements. Nvidia’s GPUs integrate matrix compute units (Tensor Cores, primarily responsible for AI computation) and general‑purpose compute cores in a fixed ratio. The performance gap between products of the same generation depends on how many Tensor Cores are integrated, but the ratio remains unchanged.

Denlin separates the matrix compute units that determine AI acceleration performance. The general‑purpose compute cores of multiple GPUs stay the same, but the matrix compute units are configurable. Thus, within the same chip area—and therefore the same production cost—it can allocate a larger proportion to AI acceleration, delivering higher performance for AI‑related workloads.

When handling parallel‑heavy tasks such as image preprocessing and result ranking, Denlin no longer has an advantage over Nvidia because the current architecture invests limited resources in general‑purpose computing. “This is the choice we made after weighing the trade‑offs,” says Wang Ping.

In memory design, Denlin makes similar compromises. Around the same time it was founded, Nvidia launched the T4 inference chip using GDDR6 memory, and in recent years its more mainstream A10, L2, L4 products—and recent reports that several internet companies purchased the RTX PRO 5500—also use GDDR memory, with only a few GPUs aimed at public‑cloud large‑model deployments using the expensive HBM high‑bandwidth memory. The first three generations of Denlin chips all use LPDDR memory like that in smartphones, which is cheaper and can be licensed as needed.

LPDDR memory consumes less power than contemporary GDDR, but its bandwidth is only one‑third. To address this, Denglin equips the chip with a relatively large SRAM to temporarily store computation results, and works with software so that the next layer of the model can directly use the data left by the previous layer, thereby reducing data shuttling between the chip and external memory. The trade‑off is that SRAM itself also occupies chip area.

The cost of a chip is determined by more than just design and manufacturing. The number of people the company needs to serve customers, and how much capital it spent on its startup, are all amortized into each chip.

Denglin controls these expenses by first limiting itself. After nine years, the company still has only 280 employees.

“We only sell cards; we don’t even sell servers,” said a Denglin representative. A small team does less, selling GPUs directly to actual enterprise customers; when serving large government or corporate projects, they only sell cards to system integrators and algorithm companies, leaving the complex requirements to partners.Chasing large government or corporate projects can increase revenue scale, but the profit margin is low, requiring a lot of manpower support, which ultimately is reflected in the cost of each card.

On November 11, 2017, Wang Ping boarded a flight from the United States, arrived in Suzhou the next day, and started working on the yet‑unregistered Denglin on the third day. The day before, he was busy buying locks, trash cans, and a printer—“really had nothing.” By the end of the year, the team had about 10 people.

In 2007, Li Jianwen, Wang Chenhui, and Wang Ping sequentially joined the chip design licensing company Vivante. Li Jianwen was responsible for GPU and general‑purpose computing licensing products, Wang Chenhui handled driver software, system integration, and technical support, and Wang Ping focused on performance analysis and feature definition. Vivante handed over the entire GPU design to the Chinese team, so the few of them participated in core compute design.

About half a year of customer and supply‑chain visits before the founding set Denglin’s direction: retain general‑purpose compute capability and add AI compute capability. Customers were worried that switching cards would require costly code changes, so Denglin also needed to support the NVIDIA CUDA programming model they were familiar with.

Technical breakthroughs may be scientific issues, but technology adoption is always a business issue

Denglin is beginning to prepare products it previously couldn’t afford, planning to add high‑bandwidth HBM memory to the next‑generation products and an interconnect design that enables multiple chips to compute collaboratively, gradually entering the public‑cloud market. According to the draft prospectus, these products will be launched in stages: first increase the compute power available to existing enterprise customers, then develop products for larger models.

“We’ve accumulated enough IP,” Wang Ping said. He was referring to design modules across different chips. Every time Denglin ships a new generation, it validates these designs: what hardware capabilities are still missing, which common modules are insufficiently supported, and which software‑hardware features have been built but no one needs. Real customers are the ones who can force chip companies to create more usable products that more companies require.

This accumulation has already changed the speed of product delivery. When Zheng Tao first arrived, the first‑generation chip delivered to Hikvision still required two months of software and driver adjustments. By the third generation, some small‑to‑medium customers with limited development capabilities received GPUs and could deliver them to their own customers for trial within two weeks.