DeepSeek is trying to do something most frontier AI labs have walked away from in the past two years. It wants its most advanced models to stay genuinely open and broadly accessible while still building a serious business and a path toward artificial general intelligence. That stance puts founder Liang Wenfeng right at the center of today’s argument about whether powerful AI should be locked behind corporate walls or treated as shared infrastructure.
The timing is not accidental. OpenAI and other Western leaders have steadily moved from publishing papers and model weights to tightly controlled APIs, arguing that safety, monetization, and competitive pressure make openness too risky at the frontier. In contrast, DeepSeek has risen to prominence by pushing low-cost reasoning models such as V3 and R1 that directly challenge assumptions about both China’s hardware limits and the viability of open source at scale. Additionally, the formation of WAICO reflects a broader push for inclusive governance in AI, emphasizing the voices of developing nations.
This is not an abstract philosophical debate. It affects who gets to experiment with cutting-edge AI, which countries can build credible domestic ecosystems, and how much of the future of intelligence is governed by a handful of firms versus a broader technical community.
From Quant Finance To An Open AI Mission
Liang’s path into frontier AI is unusual and important for understanding his choices. He first built a reputation in quantitative finance as cofounder of the hedge fund High Flyer, where large-scale data analysis and model-driven trading were the core of the business.
In 2023, he redirected that engine toward AI research, launching DeepSeek with a stated goal of pursuing artificial general intelligence rather than quick vertical applications.
Accounts of his early interviews with Chinese tech media and international outlets are consistent on one point. Liang describes the move into AGI as driven by scientific curiosity and a desire to push the technology forward, not primarily by short-term profit. That orientation matters because it shapes how much commercial pressure DeepSeek faces.
The firm is funded by substantial resources from High Flyer, which allows it to operate without the rigid revenue targets that often push startups toward closed models and enterprise contracts.
This background gives Liang a quant trader’s instinct for cost and risk but also a researcher’s tolerance for long-horizon bets. It explains why he talks about open source in terms of ecosystems and culture rather than as a narrow marketing tactic.
Why Liang Is All In On Open Source
In multiple interviews and investor meetings, Liang has been unusually explicit about his refusal to pivot to closed source. He has told both journalists and backers that DeepSeek “will not go closed source” and that this is a matter of principle, not a phase on the way to a proprietary business.
Rather than building a moat through secrecy, he argues that any advantage based on keeping models closed is temporary in a fast-moving field, even for companies as powerful as OpenAI.
His reasoning runs on two tracks.
First, he sees open source as the foundation of a durable technical ecosystem. Liang repeatedly says that the real moat is the team and its accumulated know-how, not hidden weights. Open models and published papers do not, in his view, cause meaningful loss. Instead, they attract ambitious engineers who want their work to be studied, reused, and cited, creating what he calls a developer tribe aligned with the company’s vision.
Second, he frames open source as a cultural act. In one widely cited interview, he calls open source “more of a cultural approach than a business strategy” and describes giving back to the community as an honor that builds soft power for the company.
This is an unusually clear statement among AI founders. It signals that DeepSeek is trying to earn influence through transparency and contribution rather than control.
There is also a geopolitical angle. Liang has suggested that open strategies give Chinese AI developers a structural way to compete with closed Western labs because a sufficiently large global market can resist permanent monopolies built on secrecy.
If developers worldwide can inspect and adapt high-performance Chinese models, he believes that talent and innovation will flow toward that open ecosystem rather than staying locked inside a few American firms.
Open Weights And The DeepSeek R1 Model
DeepSeek’s commitment is not just rhetorical. It is encoded in how it ships its flagship models. The company describes its systems as open weight. That means the parameter values for models such as R1 can be downloaded and run directly, even though the proprietary training datasets are not released.
DeepSeek R1 is a reasoning-focused model with 671 billion parameters that was trained on 2,048 Nvidia H800 GPUs at an estimated compute cost of 5.6 million dollars. Those numbers matter because they show that R1 reached what independent commentators describe as frontier-level performance in math, coding, and complex reasoning using resources far below the billion-dollar budgets associated with Western leaders. The team also highlights that DeepSeek-R1 uses roughly one-tenth the computing power of Meta’s Llama 3.1, underscoring how aggressively it optimizes hardware constraints.
Crucially, R1 and its related variants R1 Zero and R1 Distill were released with open weights under the permissive MIT license. That license allows anyone to use, modify, and integrate the models into commercial products as long as they retain the copyright notice, making DeepSeek’s systems truly deployable by startups, enterprises, and researchers without bespoke licensing deals.
DeepSeek paired these weights with a detailed technical paper describing architecture choices and reinforcement learning-based reasoning techniques. That combination of code, weights, and documentation gives external teams the ability to replicate and extend key methods rather than guessing at hidden training tricks.
It also sets DeepSeek apart from the common industry practice of releasing weakened variants while keeping more capable versions fully internal.
The Economics Behind Open Frontier Models
Open source at this scale only works if there is a credible business model. Liang has spent considerable time explaining how he thinks the economics can balance.
In a recent long-form discussion, he outlined a pricing philosophy that targets recovering hardware costs in roughly ten months while earning only modest profit.
He argues that as long as profit does not exceed something like six times the cost base, there is no fundamental conflict between open source and paid commercial services. In this view, the key is to avoid both extremes.
DeepSeek does not rely on loss-leading subsidies that might be unsustainable, but it also rejects pricing that extracts heavy margins simply because models are closed and scarce.
Open access does not mean free everything. DeepSeek still charges for hosted APIs and higher throughput services, but it positions those products as low cost compared to typical proprietary leaders in math, coding, and reasoning workloads.
That pricing, Liang suggests, is designed to make it rational for partners to build on top of DeepSeek’s open models, driving ecosystem growth and indirect revenue rather than locking users in through scarcity.
There are obvious risks. Hardware prices, regulatory compliance, and safety tooling all affect the real cost of running frontier models at scale.
DeepSeek’s ability to maintain generous openness depends on continued access to compute and on investor patience with comparatively modest margins. However, the company’s funding structure and quant heritage give it a more flexible runway than many pure-play AI startups.
Implications For Chinese AI And Global Competition
DeepSeek’s strategy is already shaping perceptions of China’s role in AI. For years, analysts assumed that export controls on advanced chips and a more fragmented software ecosystem would keep Chinese labs a step behind US leaders.
DeepSeek’s low-cost V3 and R1 releases have challenged that narrative by demonstrating competitive reasoning performance at dramatically lower training budgets.
Observers such as Marina Zhang at the University of Technology Sydney point out that DeepSeek is distinctive among Chinese AI firms in how it leans on software optimization and collective innovation rather than on access to the very latest hardware.
By squeezing more throughput out of domestic accelerators and opening its models, the company reduces reliance on imported chips and accelerates knowledge sharing that can benefit other actors in the ecosystem.
Open source also changes the global power dynamics around AI tools. Unlike closed systems such as GPT-4, open weights allow developers in emerging markets, small firms, and research labs to run frontier-level models locally, audit their behavior, and adapt them to local languages and regulations.
That can reduce the risk of a single company or country defining the default behavior of AI across sectors from education to finance.
At the same time, Western tech insiders and Chinese regulators are watching the safety and security implications closely. Open distribution of powerful models makes it easier for malicious actors to fine-tune systems for targeted misinformation, cyber intrusion, or biological misuse, and DeepSeek’s public materials do not yet fully resolve where it would draw the line if risk assessments change.
There is a live policy debate around whether national or international rules should constrain how far open weights can go.
Safety, AGI And The Case For Broad Access
Liang ties his open source stance directly to his long-term AGI ambitions. Both public interviews and investor discussions present frontier open models as essential for safety research, not as a side project.
In his view, many minds need to study and stress test the most capable systems if society is going to understand and control their behavior.
DeepSeek’s roadmap presentations describe broad access to frontier models as integral to the pursuit of AGI. The logic is straightforward. If only a handful of labs can see and experiment with the strongest models, safety breakthroughs and alignment methods may arrive later and be skewed toward those labs’ incentives.
Open weights plus detailed technical documentation widen the circle of researchers who can run systematic evaluations, probe failure modes, and propose mitigations.
Critics counter that openness can outpace our ability to secure downstream uses. They worry that releasing very strong models before robust monitoring and governance frameworks are in place could create irreversible harms in areas such as automated hacking or autonomous manipulation of financial markets.
DeepSeek’s current position assumes that the benefits of collective progress outweigh these risks, but the company may need to demonstrate more concrete safety tooling and red teaming practices if it wants to be seen as a leader in responsible open source.
What To Watch Next
DeepSeek’s bet is that high-performance open models can simultaneously accelerate global innovation, support a viable business, and strengthen China’s position in the AI landscape.
Liang’s refusal to go closed source, his emphasis on ecosystem building, and his willingness to encode openness in permissive licenses all make that bet unusually clear.
Several questions will determine how sustainable this strategy is.
- Can DeepSeek keep pushing frontier-level reasoning while maintaining a ten-month cost recovery model as compute becomes more expensive and safety demands grow?
- Will regulators in China or abroad decide that some classes of models are too risky to release as open weights, forcing compromises such as partially restricted variants?
- How many developers and companies choose to standardize on DeepSeek’s models instead of closed Western alternatives, and whether that developer tribe becomes the kind of durable moat Liang imagines.
If DeepSeek manages to grow a global ecosystem around truly open frontier models, it will not just be another AI startup. It will be a live test of whether openness can remain a core feature of advanced AI in an era defined by intense competition, geopolitical tension, and legitimate safety fears.
The outcome will influence how much of the intelligence infrastructure of the future is owned, and how much is shared.
Conclusion
DeepSeek founder Liang Wenfeng is making one of the clearest long term bets in frontier AI right now by promising that the company will keep its strongest models open source while pursuing artificial general intelligence rather than maximising short term profit. That stance puts DeepSeek directly against the prevailing closed model trend and turns its business strategy into a live test of whether openness can compete with proprietary systems at the very top of the market.
Background: How DeepSeek Got Here
Liang comes to AI from quantitative finance rather than consumer technology. He launched DeepSeek in 2023 as an offshoot of High Flyer, the hedge fund that has been building large GPU clusters and supercomputing infrastructure for years to support algorithmic trading and data processing. Under his guidance, DeepSeek chose not to chase flashy applications and instead focused its research talent on building foundation models that could rival or surpass leading systems from OpenAI and other US labs.
From the beginning, he framed DeepSeek as an AGI lab with a clear principle on money. In interviews he has said the company’s goal is to research large models and move toward artificial general intelligence while keeping prices at a modest margin above cost, avoiding both loss making and excessive profits. He has repeatedly argued that the real moat is not secrecy but compounding know how and an innovative culture, and that open sourcing models and publishing research do not significantly weaken that moat.
DeepSeek’s commitment to open models is not theoretical. The lab rose to global prominence when its R1 model matched top tier performance from OpenAI at roughly one thirtieth of the API cost while remaining fully open, which stunned many observers who assumed Chinese labs were mostly followers rather than leaders in foundational research. In the following year DeepSeek extended that strategy by releasing models such as V4 Pro and V4 Flash under permissive open source licenses, signalling that openness was central to its identity rather than a one off publicity move.
The Open Source Pledge, Financial Logic Included
Liang has distilled his open source stance into a remarkably blunt promise. He has said that DeepSeek will definitely open source its models and that even its strongest systems will be released openly because he does not see any inevitable advantage to keeping them closed. In private and public conversations he has added that the company will not go closed source, since it believes building a robust technology ecosystem matters more than defending secrets.
Unlike many founders who treat openness as a vague ideal, he has supplied explicit financial reasoning. Liang describes what he calls a disciplined strategy that aims for roughly six times profit on invested capital with payback in about ten months. At that level of return, he argues, open sourcing does not materially harm the business because third parties can deploy the models but the economics remain attractive for the original developer. The picture changes if a company chases one hundred times profit. In that regime open source becomes a constraint because others can run the same models at perhaps one twentieth of the cost, eroding pricing power for the originator.
By stating that DeepSeek is not aiming for that extreme profit multiple, he is effectively telling investors that they should expect sustainable returns built on continual technical progress rather than winner takes all extraction from a locked down platform. In fund raising meetings he has reportedly reiterated that DeepSeek prioritises the pursuit of AGI over profit and still intends to keep its most advanced models open source, stressing that open development and commercial monetisation are not mutually exclusive.
There is also an ideological dimension. Liang describes open source as a cultural approach more than a business tactic, and he believes a company that adopts it gains soft power by contributing to a shared ecosystem rather than guarding a fortress. He has emphasised that giving back through open models and papers is an honour that attracts talent and deepens the organisation’s capabilities over time.
Open Versus Closed: The Larger Model Ecosystem
DeepSeek’s position drops into a wider industry debate that has sharpened over the past few years. Many enterprises now run a hybrid stack in which they use flagship proprietary models such as GPT, Claude or Gemini for the most demanding workloads while deploying open source models such as Llama, Mistral, Qwen and DeepSeek for tasks where control, cost or customisation matter more. In this view open source is not a fringe movement but an essential part of the practical AI toolkit, especially for organisations that want to run models on their own infrastructure or embed them deeply into products.
Liang’s bet is that open models can occupy not only the cost sensitive segment but also the frontier tier. By releasing competitive systems at low prices and with permissive licenses, DeepSeek has already given developers in China and abroad a viable alternative to US based proprietary models for many use cases. Observers such as Marina Zhang have noted that DeepSeek’s open strategy helps it work around hardware constraints by focusing on software optimisation and collaborative innovation, turning what was once a structural disadvantage into a source of differentiation.
This is a departure from the classic Silicon Valley pattern in which a company begins in an open mode and then partially closes as its models become more powerful, a trajectory that many critics have pointed out in the history of OpenAI. In contrast, DeepSeek is effectively saying that the more powerful its models become, the more important it is to keep them open, because participation in global innovation and ecosystem building is the primary goal.
Technology Implications: Speed, Transparency and Talent
On the technical side, a sustained open source commitment at the frontier has several consequences.
First, it accelerates iteration. When model weights, training recipes and research insights are public, a global community can test, fine tune and extend them in ways that a single company never could on its own. DeepSeek is already seeing this effect as its models are integrated into a growing range of tools and applications that feed back empirical data and new ideas into the research loop.
Second, openness deepens transparency. External researchers can scrutinise model behaviour, training data approaches and safety mechanisms using more direct methods than are possible with closed APIs. DeepSeek’s founder and supporters have argued that this enables collective improvement of AI safety measures and supports more responsible development, rather than leaving the governance of powerful systems to a small set of internal teams.
Third, open models reinforce talent flywheels. Engineers and scientists who want to push the frontier often value the ability to publish, to be cited and to see their work used widely. By treating open sourcing and paper publication as central rather than peripheral, DeepSeek positions itself as a lab where contributing to a broad ecosystem is part of the job description rather than an exception. That matters in a world where top AI talent can choose among many labs and companies with different philosophies on secrecy, recognition and impact.
The flip side is that DeepSeek gives up some technical obscurity. Competitors can study and reimplement its architecture, data curation and optimisation strategies more easily than if they were hidden. Liang’s view is that the organisational learning curve and continuous innovation are harder to copy than any snapshot of a model, so the benefits of openness outweigh this risk over time.
Business and Market Consequences
From a business standpoint, DeepSeek’s stance challenges assumptions about how frontier AI should be monetised.
Most large model developers in the United States are under heavy pressure to prove recurring revenue through tightly controlled APIs, enterprise contracts and vertically integrated products. The standard story is that companies must own more of the application stack to justify the massive capital expenditure required to train ever larger models.
DeepSeek breaks that pattern. It has deliberately avoided building consumer facing products and instead focused on research, while allowing other firms to develop business to business and business to consumer services on top of its models. The lab has historically operated without external funding and has only recently moved toward larger capital raises, reportedly in the multibillion range, while still telling investors that AGI and openness remain the core priorities.
By articulating a disciplined profit target and refusing the lure of extreme multiples, Liang is signalling that DeepSeek aims to be a kind of infrastructure provider whose value comes from continual technical improvement and ecosystem dependence rather than exclusive access. That may prove attractive to companies that do not want to be locked into a single vendor, especially outside the United States, but it could also limit DeepSeek’s ability to capture the full upside of its own breakthroughs if competitors package and monetise those models more aggressively.
Investors will eventually judge whether this tradeoff is worth it. If DeepSeek continues to ship models that rival or beat closed systems at lower cost, market pressure could force proprietary players to reduce prices or offer more open variants, shifting the economic balance of the whole sector. If instead closed incumbents maintain a persistent capability lead or secure regulatory advantages, DeepSeek’s returns could look modest relative to the capital it must deploy.
Safety, Regulation and Economic Control
DeepSeek’s pledge also intensifies ongoing arguments about AI safety, regulation and economic control.
Regulators and policy makers worry that making cutting edge model weights widely available could lower the barrier to misuse, from automated cyber attacks to scaled disinformation and biologically relevant research. Those concerns are not unique to DeepSeek, but the company’s explicit commitment to open frontier models makes it an important test case for how open labs respond to regulatory pressure.
Supporters counter that transparency is a prerequisite for robust safety science. DeepSeek itself presents open source as part of responsible AI development, arguing that public models allow diverse actors to audit, stress test and improve safety techniques rather than trusting a single firm’s internal safeguards. There is a genuine tension here. Open weights empower both benign and harmful actors, and neither the market nor regulators have yet settled on a stable framework for frontier openness.
Economic control is another axis. Open models weaken the ability of any one company to dictate terms to downstream developers because those developers can host the systems themselves or switch among multiple open alternatives. That can redistribute value and bargaining power across the stack, potentially benefiting smaller firms and public institutions but reducing the rent extraction available to the original model creators.
Liang seems comfortable with that outcome. His focus on participation in global innovation rather than exclusive ownership implies a belief that long run influence comes from being indispensable infrastructure rather than from tight control over access. Whether that belief holds under future regulatory regimes and competitive pressure is one of the key uncertainties around DeepSeek’s strategy.
China, Soft Power and the Global AI Race
DeepSeek’s rise is especially significant in the context of China’s position in the global AI race. For years, US commentators have argued that open source practices were one reason Silicon Valley stayed ahead of Chinese competitors, since they encouraged rapid sharing of tools and research among labs and startups. DeepSeek is now using that same logic to flip the narrative, betting that a Chinese lab can gain influence by leaning further into openness than its US rivals.
By releasing high quality models at low cost and granting permissive rights to developers worldwide, DeepSeek projects technological soft power and invites integration into systems far beyond China. That is a different strategy from competing only through domestic applications or defensive data advantages. It positions DeepSeek as a contributor to global infrastructure, with the potential to shape norms and techniques across borders.
At the same time, DeepSeek’s dependence on open collaboration and transparent publication could make it more resilient to export controls and supply chain shocks than purely hardware intensive players, because it emphasises algorithmic efficiency and shared innovation over brute force scaling alone. In a world where access to top tier chips is increasingly politicised, that may prove to be a strategic hedge.
Key Takeaways and What To Watch Next
- DeepSeek is making an unusually clear commitment to keep its frontier models open source while prioritising AGI research over maximal profits, backed by explicit financial targets and a cultural argument for openness.
- The company’s success with models like R1 and V4 suggests that open systems can compete with proprietary leaders on capability and cost, strengthening the case for hybrid ecosystems where open and closed models coexist rather than a purely closed future.
- Liang’s strategy redistributes power across the AI value chain by favouring ecosystem growth over strict control, which could benefit developers and smaller firms but may limit DeepSeek’s ability to capture outsized economic returns and could create friction with safety focused regulators.
- Over the next few years, the critical signals to watch will be whether DeepSeek maintains a genuine frontier capability lead while staying open, how regulators respond to open access to advanced models, and whether other major labs adjust their own openness in reaction to DeepSeek’s example.
If DeepSeek can keep delivering state of the art models under open terms while building a sustainable business around them, it will not just validate one company’s philosophy. It will reshape expectations for what responsible stewardship of powerful AI can look like and force the industry to reconsider the assumption that secrecy is the only path to leadership.
