DeepSeek's Hundred-Billion Financing and Zero AI Returns: Two Sides of Capital Frenzy

DeepSeek has raised another 100 billion yuan. This round, which set a record for a single domestic AI funding round, was led by Tencent and CATL—one representing the gateway to China's mobile internet, the other representing the throne of EV batteries. As soon as the news broke, voices saying "industrial capital is going all in on AI" flooded the screens.

But on almost the same day, Apollo chief economist Torsten Slok offered another set of figures: setting aside the tech giants, AI has delivered zero financial return for American companies. Revenue per employee, profit margins, total factor productivity—all flat across the S&P 493 non-tech companies. The money has not produced results.

On one side is a hundred-billion-level arms race in computing power; on the other, a verdict that "AI's economic return is zero." Only when these two pictures hold true at the same time do they constitute the most contradictory moment of 2026 for capital and technology.

A hundred billion: whose money is it

In this funding round, DeepSeek went from a target of 50 billion yuan all the way to nearly 100 billion in oversubscription, with a pre-money valuation of about 500 billion, absorbing more than 150 billion in capital across two rounds within half a year.

Tencent took a heavy position in two consecutive rounds (10 billion in Series A, and adding more in this round), aiming for priority participation in the ecosystem gateway—when WeChat, gaming, and cloud businesses need to connect to model capabilities, it wants to be sitting in the front row. CATL raised its contribution to about 10 billion, but what it wants is a different calculation: data center electricity costs, energy storage peak-valley arbitrage, direct green power supply—putting "electricity," the largest cost item in computing power, into its own business map.

One manages traffic, the other manages electricity. What they are betting on may not be the large models themselves, but the positioning of the "computing power foundation"—SMIC Juyuan (backed by SMIC), Shixi Capital (backed by GigaDevice), and Geely have also squeezed in; the demand for computing power from chip manufacturing and the automotive industry is forming a positive cycle.

Where this money flows writes the ambition on its face: DeepSeek plans to build a data center in Ulanqab, Inner Mongolia, deploying at least 160,000 Huawei Ascend chips, using domestic computing power for frontier model pre-training for the first time. Liang Wenfeng himself calls it "the biggest gamble."

"What we have never lacked is money" no longer holds. Back then he said the company's problem was the embargo on high-end chips; now the funding logic has become—since the most advanced chips cannot be bought, use massive capital to first smash out a domestic computing power ecosystem. Money has become a weapon against the ban.

The economy has not caught this money

Revenue per employee for the S&P 493 (companies excluding the Mag 7) has long been flat in the 450–550 range, while the Mag 7 has climbed toward 800; profit margins are flat for the former and rising for the latter; productivity growth—whether TFP or output per hour—shows no sign of being lifted by AI.

More striking is the distribution of spending. For model services and inference, the top 10% of enterprises account for 99.5% of spending; among OpenAI+Anthropic customers, the top 10% account for 95.6%; among companies in the top 1% of AI spending, 80.4% are VC-backed, while among customers with median spending, the VC share is only 13.1%.

A considerable portion of AI consumption is VC money circulating among itself—like the subsidy-burning shared bikes of the past, lively in the app, but no money on the books.

Torsten Slok nailed it in one sentence: mainstream enterprises are beginning to redefine token spending from a "source of growth" to a "cost center that erodes profit." They do not need the latest models, nor are they willing to pay for them.

Foreign media have erected a monument to this frenzy, calling it the AI Graveyard. According to S&P Global Market Intelligence, about 42% of enterprise AI projects are ultimately abandoned by their parent companies, for reasons no more than insufficient funding, technical difficulties, competition, difficulty scaling, and weak demand.

Even the giants themselves are cutting. OpenAI cut its video platform Sora—high operating costs plus low retention made it unsustainable; the browser Atlas lasted less than a year before being folded back into ChatGPT; Operator and DALL-E were likewise folded into the main app, and even the "super app" redesign was rolled back due to user complaints.

Counting down, there is a long list of names: Relay, which made a Zapier alternative, was crushed by platforms; the wrapper Phind shut down a month after raising money; Tome, which makes PPTs, had over 20 million users and annual revenue of $3 million; hardware pioneer Humane AI Pin raised $230 million and was ultimately bought by HP for $116 million; Yupp burned $33 million, Figgs AI was too expensive to survive even for free, and top companion app Character.ai went bankrupt—the "graves" in the AI Graveyard are still multiplying.

Domestically, the clearing is also underway: in 2025 alone, at least 25 domestic AI applications were shut down, delisted, or stopped updating for long periods; according to AI Graveyard statistics, 392 AI tools worldwide ceased service that year. Tencent Yuanqi, QClaw, and the self-built agents of Doubao and Qianwen were successively shut down and taken offline. They were unified and merged with enterprise services into the "office" agent.

Autonomous driving has already died once

This 2019–2026 list of autonomous driving companies that exited is in fact the first complete-cycle autopsy report of the AI industry chain. Drive.ai, Starsky Robotics, Uber ATG, Argo AI (Ford and Volkswagen pulled their funding, the capital chain broke), lidar leader Velodyne (product involution), TuSimple (difficulty commercializing technology + geopolitical pressure + internal management, delisted), Luminar, Zongmu Technology (three failed IPOs), HoloMatic, Qian'gwa Technology, Qingtian Smart Card... the causes of death are highly consistent: financing rupture, difficulty of commercial deployment, slow technology iteration, tight capital chains.

These were not small workshops. Argo AI was backed by Ford and Volkswagen, Velodyne was once the industry leader, and TuSimple was the "first autonomous driving truck stock." But the commercialization leg never landed, and as soon as the capital market tightened, it broke. In eight years, autonomous driving has proven one thing with a list of more than forty exited companies: technology working does not equal business working.

Today's general-purpose large models are reenacting the frenzy of autonomous driving in 2019–2021, and the ebb that began in 2022—only with a bigger stage, faster cash burn, and more frightening scale.

Embodied intelligence, the next bubble?

The same script is already being rushed onto the stage. In August, DeepSeek took a 141 million yuan stake in Unitree Robotics, and Liang Wenfeng is again set to lead investment in the humanoid robot company Winverse AI. Embodied intelligence is standing exactly where autonomous driving once stood: hot financing, valuations running ahead, and commercialization nowhere in sight.

Robots burn more money than autonomous driving—hardware, supply chain, mass production, scenarios, every link is a cash-burning black hole, and currently not a single company has a sustainable business model that works. What's worse, it shares the same lifeline as large models: capital. Once this round of the computing power arms race cools first, embodied intelligence will be the first track to feel the chill of a supply cutoff. It doesn't even need to become a second autonomous driving bubble on its own—it just needs to walk the same path once more, and that's enough.

My judgment is, no. They are the peak-seers of the dividend—mobile internet growth has peaked, new energy vehicles have been involutioned into overcapacity—and all are looking for the next certain growth engine. And AI, especially computing power, happens to be the certainty closest to "infrastructure" in the eyes of capital right now. So what they are going all in on is not AI applications, but the AI foundation: one wants a traffic gateway, one wants electricity, and in essence both are staking out positions.

Risk and opportunity coexisting is a correct but useless platitude, but concretely it comes down to two sides: the opportunity is that if DeepSeek can truly use hundreds of billions in capital to blast open a domestic computing power ecosystem, with chips—energy—models—applications as one integrated whole, this will be an infrastructure transformation more foundational than the mobile internet; the risk is that if this round of investment, like autonomous driving, still only blooms and bears no fruit, the shock of a hundred-billion-level bubble bursting will far exceed the sum of those years in autonomous driving.

The endgame of the AI story is not written in the Mag 7's financial reports, nor in financing news, but in the profit margin tables of the S&P 493. Only when the economy truly catches the AI money does it count. Before that, this hundred-billion all-in is more of a gamble on the future—win the bet and it's a foundation, lose and it's a tombstone. The list in the AI graveyard is still getting longer.