Inside Liang Wenfeng’s four-hour DeepSeek meeting—and the collision between open intelligence, Wall Street, Silicon Valley and the Chinese party-state

By Jonathan Brown

At an ordinary fundraising meeting, the founder explains how the investors will make money.

Liang Wenfeng appears to have spent nearly four hours explaining what DeepSeek would refuse to do for money.

It would not abandon its pursuit of artificial general intelligence to chase fashionable products. It would not build a super-app merely to capture users. It would not reserve its best models for private use while distributing inferior versions to the public. It would not impose the highest price the market could bear. It would not sacrifice the stability of its research team. It would not fill every employee’s calendar with assigned work, performance indicators and managerial supervision.

Investors were not being offered control of DeepSeek’s mission. They were being asked to finance it.

That inversion makes the reported meeting one of the most remarkable documents to emerge from the artificial-intelligence boom. If the transcript is substantially authentic—and that qualification matters—it reveals a founder whose central proposition is not that profit is unnecessary, but that profit must be denied sovereignty over the enterprise.

Liang is no anti-capitalist innocent. He made his fortune in quantitative finance. His hedge fund turned mathematics, computing power and market inefficiency into an enormous private pool of capital. DeepSeek’s independence was initially possible because that machinery worked so well. He understands accumulation, pricing and competitive advantage from inside the engine room.

What makes him unusual is that he appears to have concluded that the highest purpose of capital is to purchase freedom from capital’s normal demands.

THE MANUSCRIPT AND ITS SHADOW

The document circulating online is described as a transcript of a May 20, 2026 investor meeting lasting three hours and 44 minutes. The closest thing to a complete manuscript is a roughly 42-page Chinese text derived from an audio file identified as “deepseek 0520.m4a.” Its own cover warns that automatic speech recognition was used, speakers were not consistently separated, and names, numbers and technical terms may contain errors. The original audio has not been made publicly available.

DeepSeek and Liang have not officially authenticated the transcript. A detailed Chinese examination of its provenance reported that an institution whose representatives attended the meeting considered its substance credible, but also emphasized the missing audio, uncertain speaker attribution and possibility of machine-generated smoothing. An English translation of 118 selected answers appropriately calls the material “alleged.”

The right response is neither credulous worship nor automatic dismissal.

The transcript’s ideas closely match Liang’s authenticated interviews, DeepSeek’s actual pricing, its open-weight releases, its unusual organization and its published research priorities. Contemporary reporting also confirms that management was telling prospective investors that breakthrough research would take precedence over rapid commercialization. Yet that consistency cannot itself prove authenticity; it would also make a plausible imitation easier to construct.

Accordingly, the manuscript should be read as a reportedly credible but unverified record. Exact figures and stray technical phrases deserve particular caution. Its repeated strategic arguments are more persuasive because they correspond with observable behavior.

The aftermath adds another layer. DeepSeek reportedly paused its second fundraising round after Liang became frustrated by accounts of investor conversations appearing online. Reuters reported that the company might resume the process later, while noting that neither Reuters nor Bloomberg had verified every circulating post.

If true, Liang’s reaction is entirely consistent with the man depicted in the transcript: intensely private, suspicious of financial theater and willing to leave capital waiting outside the laboratory.

THE QUANT WHO ESCAPED THE MARKET

Liang was born in Guangdong in the mid-1980s and entered Zhejiang University at 17, studying electronics and communications before completing a master’s degree. He later co-founded the quantitative hedge fund High-Flyer, which used artificial intelligence and large-scale computing to trade Chinese markets.

By the end of 2021, High-Flyer reportedly managed more than 100 billion yuan. It had also assembled the computing infrastructure and engineering culture from which DeepSeek would emerge. DeepSeek was formally created in 2023, initially financed by High-Flyer rather than outside venture capital.

A Reuters profile of Liang describes a founder more interested in research than public celebrity. Earlier interviews present him as a technical participant who reads papers, writes code and joins engineering discussions rather than behaving like a conventional executive. He has repeatedly criticized the tendency of Chinese technology companies to prioritize immediately profitable applications over original research.

The irony is productive rather than embarrassing. Liang did not discover some pure realm beyond commerce. He first became exceptionally good at commerce.

High-Flyer gave him what most frontier laboratories lack: a source of money that did not initially require him to satisfy a venture-capital board, prepare a public listing or sell a story about quarterly adoption. Finance built the protected enclosure in which DeepSeek could investigate intelligence for its own sake.

This helps explain why Liang’s outlook is more sophisticated than a simple denunciation of greed. He knows that a laboratory consumes chips, electricity, salaries, data, buildings and time. He wants DeepSeek to become financially durable. In the transcript, he even presents the API business as capable, in a pessimistic scenario, of supporting a substantial public company.

But commercial survival is the floor, not the destination.

Liang’s objection is to the conversion of every technical choice into an extraction problem: How much can we charge? How completely can we bind users to our platform? Which research project produces the fastest revenue? Which feature increases switching costs? How can access to intelligence be divided into increasingly expensive tiers?

He is not proposing the extinction of profit. He is proposing its demotion.

REASONABLE PROFIT

The phrase at the center of the meeting is “reasonable profit.”

Liang reportedly argued that DeepSeek should charge enough to recover its costs and continue its work, but not the maximum price permitted by temporary scarcity. He discussed computing equipment being paid back in roughly ten months, far faster than its normal accounting depreciation. That is not charity. A ten-month capital recovery period can support an extremely healthy business.

What is different is the refusal to treat customers’ dependence as an invitation to extract every available dollar.

At one point, according to the manuscript, DeepSeek had priced a model high to suppress demand while capacity was constrained. The team disliked the decision. When the price was cut to roughly one-quarter of its previous level, employees celebrated.

That small episode says more about organizational character than a hundred mission statements. In most technology companies, a successful price increase is cause for celebration because it demonstrates pricing power. At DeepSeek, the team reportedly celebrated when a capability became dramatically cheaper.

Liang’s formula—“restraint is a strategy”—should not be mistaken for sentimental generosity. Lower prices widen adoption, attract developers and damage competitors whose economics depend on expensive inference. Open models create an ecosystem around DeepSeek’s architecture. Sharing capability can increase its technical influence, recruiting power and global prestige.

Restraint can be both principled and strategically devastating.

DeepSeek is therefore not relinquishing competition. It is changing what competition measures. Instead of asking who can construct the largest tollbooth, it asks who can make intelligence sufficiently efficient that the toll collapses.

In the manuscript, Liang describes fashionable products and minor revenue opportunities as “sesame seeds” beside the “watermelon” of AGI. The metaphor is revealing. He is not indifferent to commercial products because they are worthless. He fears they are valuable enough to distract a company from something vastly larger.

The contemporary corporation normally treats distraction as discipline. A research program must continually justify itself through milestones, addressable markets and near-term deliverables. Liang reverses the burden: commercialization must justify any attention it takes away from the central scientific problem.

That is why the meeting sounds less like a pitch than a qualification examination for capital. Investors must decide whether they can tolerate a founder who regards their preferred metrics as side effects.

THE STAIRCASE TO AGI

Liang’s technical vision is ambitious but more structured than mystical proclamations about an imminent machine god.

He begins with current language models, which can perform remarkable tasks when given the right instructions and context. Reasoning models add longer chains of inference. Agents then use models to operate tools, write and execute code, search, test hypotheses and complete multi-stage work.

Coding agents occupy a privileged position because software development is recursively useful. A model that can improve the tools used to build models can accelerate the next generation of research. It is both a product and an instrument of its own evolution.

The next step, in Liang’s proposed sequence, is continual learning.

A human employee joins an organization with incomplete knowledge but gradually learns its systems, people, exceptions, history and unwritten practices. Today’s models generally do not accumulate experience in the same durable way. They can receive a long context or retrieve stored documents, but they do not reliably transform each interaction into stable, generalized competence without additional training.

Liang treats that limitation as one of the principal barriers between impressive models and more general intelligence. If a system could learn continuously from work, preserve useful discoveries, correct its habits and incorporate new knowledge without destroying older abilities, it could become a persistent participant rather than a repeatedly reset visitor.

Beyond that lies self-improvement: systems contributing meaningfully to their own research and gradually accelerating scientific progress. Embodied intelligence—machines operating in the physical world—comes later in the sequence.

This is a coherent research wager, not an established map of the future. Continual learning remains difficult. New training can overwrite old knowledge. Persistent memory creates privacy and security problems. A model may learn errors, manipulation or institutional prejudice as easily as expertise. Measuring whether it has genuinely improved is harder than measuring whether it has merely adapted to a benchmark.

Nor does continual learning automatically produce judgment, consciousness, moral agency or safe self-improvement. The transition from a more useful coding agent to AGI is not a law of nature.

Liang appears to understand the uncertainty. His roadmap is presented as a hypothesis rather than a timetable guaranteed by destiny. What is extraordinary is his willingness to organize a company around testing that hypothesis while much of the industry is organizing around monetizing what already exists.

That choice also explains some of DeepSeek’s apparent deficiencies. The transcript reportedly treats polished consumer products, exhaustive support and even some forms of hallucination reduction as downstream engineering matters rather than the laboratory’s overriding concern. Partners can build applications around the models; DeepSeek wants to work on the next underlying capability.

That attitude can be bracing, but it can also become dangerous. Hallucination is not a cosmetic defect when a model is used in medicine, law, infrastructure or scientific research. A laboratory pursuing general intelligence cannot permanently leave reliability to someone else.

The visionary’s defense—that applications are byproducts—does not relieve the model maker of responsibility for the systems it releases.

A COMPANY ORGANIZED AROUND CURIOSITY

DeepSeek’s internal culture may be the most radical portion of Liang’s design.

The manuscript says team stability is the one interest he will not compromise. People are selected for curiosity, ability and identification with the problem. Large portions of their time are left unassigned. Formal work is supposed to occupy less than half of a researcher’s capacity, leaving room for bottom-up experiments.

Liang also rejects excessive overtime. Research, he argues, needs concentration, intellectual freedom and a relaxed environment rather than permanent exhaustion.

Within Chinese technology culture, this is almost subversive. The notorious “996” schedule—9 a.m. to 9 p.m., six days a week—was declared illegal by Chinese authorities, yet long hours and hypercompetitive “involution” remain embedded in parts of the economy. DeepSeek’s answer is not slower ambition. It is the belief that frontier discovery cannot be beaten out of people through time sheets.

This resembles the best traditions of scientific laboratories more than the standard mythology of the heroic startup. It also resembles the early periods of some American technology companies, when engineers could explore without demonstrating an immediate revenue contribution.

But informality has limits. An organization with unwritten values and few formal controls can feel emancipated while everyone agrees with the founder. As it grows, ambiguity can conceal authority rather than eliminate it. Liang acknowledges in the transcript that some departments already require hierarchy and structure.

“Bottom-up” does not mean leaderless when one founder defines the north star, controls the capital and decides which questions belong to the mission.

OPENNESS AS BOTH ETHIC AND WEAPON

Liang’s commitment to open models is not incidental. It is connected to his theory of progress.

The manuscript says DeepSeek intends to release its strongest models rather than maintaining an inferior public version beside a secret internal system. The company has already made model weights, technical papers and significant code available. Its DeepSeek-R1 repository permits commercial use, modification and distillation under permissive terms.

This matters. Open weights allow researchers and companies to inspect, adapt and run models without sending every request to a centralized American or Chinese provider. They reduce dependence on a single company’s pricing, policy changes and continued existence. They give smaller countries and institutions a measure of technological sovereignty.

Open release is also a competitive weapon. If highly capable intelligence becomes a downloadable commodity, companies whose valuations rely upon proprietary access face shrinking margins. DeepSeek can surrender some direct rent while establishing its architecture, methods and cost structure as industry reference points. Its remaining moat is the organization capable of producing the next model more efficiently.

Liang has previously described open source as a cultural act. It is also industrial strategy and national soft power.

The term must nevertheless be used precisely. DeepSeek provides substantial access to weights and code, but it does not publish every training example, filtering decision, data contract, experiment and operational system required to reproduce a frontier model from the ground up. “Open weights” and “fully reproducible open source” are not identical.

Nor does an open model make every DeepSeek service open. The hosted platform remains a centralized service subject to Chinese law. DeepSeek’s privacy policy says relevant personal information may be processed and stored in the People’s Republic of China. The public service applies content controls associated with the country’s political and regulatory environment.

DeepSeek therefore embodies a striking modern contradiction: technically open artifacts emerging from a politically closed system.

The reverse contradiction also exists. In the United States, a politically open society increasingly depends on privately governed AI systems whose most capable models, training data and internal decisions remain closed.

Neither contradiction should be concealed by national mythology.

WHEN ABUNDANCE BECOMES BAD NEWS

The stock market’s reaction to DeepSeek supplied the clearest demonstration of the problem Liang is addressing.

When DeepSeek’s efficiency became globally visible in January 2025, Nvidia shares fell almost 17 percent in a day, erasing approximately $593 billion in market value. The Nasdaq declined, along with companies tied to data centers, energy consumption and the assumption that frontier AI would require ever-expanding amounts of expensive infrastructure. Reuters described one of the largest technology selloffs in modern market history.

Consider what had happened in material terms. A company claimed that advanced intelligence could be produced and delivered with fewer resources. If true, that was an increase in social capability. It implied that students, researchers, businesses and governments might obtain more intelligence for less money and energy.

Wall Street interpreted this improvement as destruction.

The market was not irrational according to its own logic. Shares represent claims on future profits, not votes on whether humanity is receiving a useful gift. If intelligence becomes cheaper, some expected rents disappear. If less hardware can perform the same work, projections based on unlimited hardware demand must be revised.

That is precisely the problem.

An economic system whose immediate response to abundance is panic has confused value with scarcity. It can celebrate a cure only after determining who will own it, how strongly it can be patented, and what recurring payment can be imposed upon those who need it.

Closed AI platforms fit this logic naturally. Users send information into systems they cannot inspect, receive outputs under changing terms, and remain dependent upon providers that control access, prices, permitted uses and model retirement. The system is not merely a product. It is a privately governed environment.

Closed development is not inherently malicious. Training frontier models is enormously expensive. Safety testing can require controlled deployment. Open release creates real dual-use risks. American laboratories have produced foundational discoveries, and Silicon Valley also gave the world extraordinary open-source institutions.

OpenAI, moreover, officially remains controlled by a nonprofit foundation even as its commercial public-benefit corporation raises vast sums. Anthropic has embedded a public-benefit mission in its corporate structure. These are meaningful distinctions.

They do not abolish financial gravity.

By mid-2026, OpenAI and Anthropic had raised sums measured in tens of billions of dollars and were moving toward public markets at valuations approaching the scale of the world’s largest companies. Reuters reported confidential IPO preparations alongside enormous continuing capital requirements.

Money at that scale arrives with an implied future. Revenues must become immense. Products must capture users. Enterprise contracts must deepen. Platforms must become difficult to leave. Intelligence must generate returns capable of validating valuations already assigned to it.

The critique of “Silicon Valley” is therefore not that every American founder is venal or every Chinese founder virtuous. It is that an incentive architecture can turn even mission-driven organizations toward enclosure. The more capital a laboratory requires, the more difficult it becomes to tell capital that its claims are secondary.

Liang’s achievement was to reach the frontier by another road. Quantitative finance financed the laboratory, technical efficiency reduced its dependence, and open release converted openness itself into competitive power.

CAPITAL WITH CHINESE CHARACTERISTICS

It would be a serious mistake to convert Liang into evidence that China has transcended capitalism.

Modern China contains some of the fiercest profit competition in the world: price wars, property speculation, platform monopolies, financial engineering, punishing work schedules and founders who became vastly wealthy. Chinese companies can be every bit as aggressive about market share and labor costs as their American counterparts.

The difference is not the absence of capital. It is that capital does not possess the final formal authority.

The Chinese Communist Party defines the political direction of the state and expects private enterprise to contribute to national objectives. Company law provides for Party organizations inside firms. Regulation can rapidly alter entire industries. Access to financing, procurement, licenses and political favor is inseparable from broader state priorities.

China does not abolish capital. It denies capital the last word.

The Party claims that word.

Liang’s position inside this system became unmistakable in February 2025, when Xi Jinping invited him to a high-profile symposium with leading private entrepreneurs including Jack Ma, Ren Zhengfei and executives from Tencent, BYD and CATL. Xi urged private companies to pursue high-quality development while practicing patriotism and contributing to Chinese modernization, according to the Chinese government’s account.

The meeting also signaled a partial rehabilitation of private technology after years of regulatory crackdowns. Reuters noted that the presence of Jack Ma carried particular symbolism after the cancellation of Ant Group’s enormous public offering and Beijing’s campaign against what it considered disorderly expansion of capital.

DeepSeek is useful to the Chinese state. Its models demonstrate indigenous scientific competence, weaken American platform dominance, advance technological self-reliance and offer other countries an alternative AI ecosystem. Its cost efficiency is especially valuable under American restrictions on advanced semiconductor exports.

Liang’s reported assertion that the gap between China and the United States is primarily computing power rather than talent fits both his experience and the Party’s strategic narrative. Scarcity forced DeepSeek to pursue architectural and engineering efficiency. What might have remained a commercial disadvantage became a culture of frugality and optimization.

The company is private, but it is also becoming a national champion. Those categories are not mutually exclusive in China.

Its reported financing structure captures the relationship with almost diagrammatic clarity. According to Reuters, DeepSeek’s first external round raised more than 50 billion yuan through a limited partnership managed by Liang. Most investors reportedly accepted a five-year lockup and no voting rights. The National AI Industry Investment Fund was the exception: it invested directly, obtained voting rights and avoided the same lockup.

The reporting was not officially confirmed by DeepSeek, but if accurate, it reveals two different disciplines imposed upon capital.

Private investors were required to be patient and silent. State capital retained a voice.

That does not mean the CCP dictates DeepSeek’s daily research agenda or that every open-model decision originated in Beijing. Liang’s commitment to open research predates his elevation as a national figure. The transcript presents a genuine scientific mission, not a government slogan pasted over a business plan.

But founder autonomy and state strategy currently point in the same direction. Both favor Chinese technical capability, reduced dependence on American platforms, domestic semiconductor development and models that spread Chinese influence internationally.

The unresolved question is what happens when they no longer align.

A party-state can restrain private monopolies and support patient investment in projects markets neglect. It can also impose political obedience, censorship and opaque strategic priorities without independent judicial or democratic accountability. Capital disciplined by the state is not necessarily capital disciplined by the public.

Patient capital can become obedient capital.

This is why DeepSeek cannot be interpreted through the lazy binary of capitalist America versus communist China. The real contrast is among competing systems of authority: shareholders, founders, states, workers, users and the public.

In Silicon Valley, founders often invoke missions until investors acquire sufficient economic gravity to shape the outcome. At DeepSeek, private investors may have remarkably little formal influence, while the founder and party-state occupy the commanding positions.

Liang has created distance from the stock market. He has not created democratic governance.

THE DANGERS INSIDE THE VISION

Every compelling vision contains its characteristic danger.

DeepSeek’s danger is founder dependence. Its restraint, openness and research culture appear to rest heavily upon Liang’s personal convictions. An unwritten mission can be powerful when transmitted through a small, trusted team. It is less reliable when thousands of employees, state institutions, commercial partners and future executives become involved.

There is no public constitutional mechanism guaranteeing that DeepSeek’s strongest future model will always remain open, that prices will remain restrained or that AGI research will continue to outrank political and commercial demands.

Its open releases also create legitimate security questions. Powerful models can lower barriers for scientific discovery and local innovation; they can also assist cyber operations, surveillance, manipulation and weapons development. Openness distributes power, but it does not choose only benevolent recipients.

Continual learning presents its own governance problem. A model that remembers an organization over years could be extraordinarily useful. It could also become an unprecedented repository of private behavior, institutional secrets and accumulated human vulnerability. The closer an agent comes to being a permanent colleague, the more urgent questions of consent, correction, deletion and accountability become.

Liang’s economic language also deserves realism. “Reasonable profit” does not mean modest wealth. DeepSeek may become one of the world’s most valuable companies. Liang’s existing fortune and control already place him far from ordinary economic life. A company recovering expensive computing hardware in ten months is not operating at subsistence level.

His position is not saintly renunciation. It is closer to enlightened abundance: accept extraordinary success, but do not destroy the mission by squeezing every possible rent from it.

That distinction is still profound.

The choice facing AI is not profit or no profit. It is whether profit remains the oxygen an organization needs to live or becomes the god it is organized to serve.

THE FOUR-HOUR REFUSAL

If the manuscript is authentic in substance, Liang Wenfeng’s meeting matters because he used a fundraising room to establish limits on money.

He told capital that DeepSeek’s products were byproducts of research. He told it that the best models would remain available outside the company. He told it that maximum pricing was strategically and morally inferior to sufficient pricing. He told it that researchers needed unassigned time, organizational trust and relief from performative overwork. He told it that user growth, fashionable applications and the construction of a super-platform could all become distractions.

He did not reject wealth. He rejected its right to decide what the company was for.

That posture may prove impossible to sustain. AGI may remain distant. Continual learning may not provide the staircase Liang imagines. Open models may be overtaken by closed systems financed at vastly greater scale. Political pressure may narrow the company’s freedom. Growth may force DeepSeek into the same hierarchy, product accumulation and institutional caution it currently resists.

But even failure would not make the experiment meaningless.

Liang has demonstrated that many supposedly inevitable choices in artificial intelligence are choices after all. A frontier company can release weights. It can lower prices when customers would pay more. It can treat efficiency as liberation rather than merely margin expansion. It can ask investors to accept a mission they do not control. It can recruit talented young researchers by offering difficult questions instead of only compensation and prestige.

Most importantly, DeepSeek exposes the poverty of an economic imagination that recognizes intelligence primarily as an asset to be enclosed.

When cheaper intelligence erased hundreds of billions of dollars in market value, the event revealed the divergence between human benefit and financial expectation. Wall Street saw threatened rents. Liang saw confirmation that intelligence could become more abundant.

That does not make China innocent, the CCP benevolent, open models automatically safe or Liang infallible. DeepSeek exists inside a political economy capable of immense coordination and immense coercion. Its founder is a billionaire produced by the very financial machinery he now seeks to subordinate. Its openness is technically consequential but politically incomplete.

The contradictions are not blemishes around the story. They are the story.

Liang Wenfeng is extraordinary because he stands inside those contradictions without resolving them through slogans. He is a capitalist trying to civilize capital, a Chinese national champion distributing models across borders, a private founder operating beneath a Leninist party-state, and a technologist using one of the world’s most profitable industries to finance work he insists should not be governed by profit maximization.

Perhaps his greatest innovation is not a model architecture.

It is the proposition that intelligence becomes more valuable when no single company can fully possess it; that a company can become powerful by refusing some forms of power; and that the appropriate return on scientific discovery is not always the largest toll society can be forced to pay.

For nearly four hours, in a room assembled to discuss money, Liang appears to have talked about the things money must not be allowed to change.

Then he asked the investors whether they still wanted in.

Jonathan Brown is a cybersecurity researcher and investigative journalist at bordercybergroup.com.

If you would like to support our work — useful, well-researched, ad-free cybersecurity intelligence — subscribe, comment, or buy us a coffee! Thanks.