The Puzzle of Compute Without Progress
The starting point is an uncomfortable fact. Since GPT-4 arrived in 2022, the money flowing into AI infrastructure has grown enormously. Wang points to Nvidia's data center revenue, which rose from roughly $5 billion a quarter to more than $20 billion, with well over $100 billion spent on high-end chips in that window. Yet no lab has yet shipped a model that is dramatically better than GPT-4.
Wang frames progress as resting on three pillars: compute, algorithms and data. Algorithmic advances such as the transformer and reinforcement learning from human feedback have historically moved the field, and compute keeps scaling. The weak link, in his view, is the third pillar. The industry has scaled compute hard while the other two have not kept pace in step.
He suggests the recent plateau is best explained as a data wall. Building more chips does not help if the training material has run out.
The Data Wall and the Limits of Internet Text
Wang describes GPT-4 as a model trained on nearly all of the easy internet: open web pages, social media, anything not behind a paywall, and some material that is torrented. That easy supply has been used up. The first stage of AI improvement, pre-training, taught models to imitate that text, and on that task they now outperform any human.
But imitation is not the goal. Wang wants systems that can solve hard problems, carry out tasks and collaborate with people on daily work. Those abilities cannot be learned from text that already exists in quantity, because the text is missing the process.
He gives a concrete example. A fraud analyst at a large bank reviews a set of suspicious transactions, weighs many pieces of evidence, and reaches a judgment. The reasoning moves through that person's head, and almost none of it is written down step by step. A model trained on the internet has no way to learn from it. The thinking that powers the economy is largely unrecorded.
βData is one of the few areas where you can actually produce a long-term sustainable competitive advantage.β
What Frontier Data Actually Is
To name what is missing, Wang introduces the term Frontier data. In his definition it includes complex reasoning chains, expert discussion, and agent traces in which a model looks up one piece of information, reasons over it, looks up another, corrects an error, and uses tools. Each of those components is what makes an agent able to do work rather than just describe it.
Wang's thesis is that the next stage of progress needs an abundance of this material, not the scarcity the field is in now. The question then becomes practical: where does it come from? He answers with a three-part picture, starting with the data that companies already hold.
The scale he cites is striking. JP Morgan's internal data set is reportedly around 150 petabytes, while the internet data used to train GPT-4 was under one petabyte. Large enterprises sit on enormous reserves of process knowledge that has never been published.
Mining Existing Enterprise Data
The first route is mining. Each large organization can go through its own records to find the reasoning embedded in them, then refine that material so AI systems can learn from it. Wang notes this is essentially a one-time benefit, but it can be very meaningful. Every company would have to do this for its own problems, because the data does not travel.
He expects that in five years only the most sophisticated companies will have fully mined their largest sources. Even so, models will still need to improve, which pushes the question toward ongoing production of new data.
Wang also distinguishes this kind of mining from broader collection. Process mining, which records which applications people use and how they move information between them, and consumer devices that log a person's day, will create useful datasets. Neither, he argues, produces the complex reasoning that moves models forward.
Producing New Data Like Safety Drivers
The second route is manufacturing new data, and Wang uses a familiar comparison. Chip makers keep building bigger fabs and finer processes, so the field instinctively thinks in terms of expanding production capacity for compute. He wants the same instinct applied to data.
His proposed method is a hybrid of human and synthetic work. Algorithms do much of the heavy lifting in generating material, while human experts step in when the system gets stuck, makes a factual error, or meets a situation it has never seen. He compares this to autonomous vehicle development, where safety drivers sit in the car, let the software run, and take over when it starts to fail. The same arrangement, he argues, is needed for AI systems.
The role has a name: AI trainer or contributor. Wang calls it one of the highest-leverage jobs a skilled person can hold. A mathematician who improves a model slightly on math reasoning contributes to every student, company and developer who later relies on it, which is a far wider reach than a single research paper.
Why Data Is the Real Moat
Wang's most pointed claim concerns competition. Algorithms are intellectual property, but they diffuse, because the rest of the industry eventually learns them. Compute can be matched by anyone willing to spend more. Data, he argues, is among the few areas that can produce a lasting advantage.
He points to existing deals as early signals. OpenAI's agreements with the Financial Times and Axel Springer give it access to historical archives that rival models may lack, which he expects will shape the quality of its answers. He predicts the field will move from bragging about how many GPUs a lab owns to bragging about which data it can access and produce.
Wang doubts that data strategy will stay differentiated for long. Content producers are unlikely to sign exclusive deals with a single model maker, so each lab will need its own route to unique data. Anthropic might build around enterprise use cases, while OpenAI might leverage its large user base. Either way, labs will have to find where they can get proprietary material.
Enterprises Will Guard Their Data
A related shift concerns where companies run their AI. Wang reports that enterprise leaders are quickly realizing their data may be the only competitive edge they have in an AI world. Handing it to a model developer risks mortgaging their future, so they are cautious about any arrangement that exposes it.
That caution creates an opening for models that run on premises. Enterprises can take an open or semi-open model, customize it on their own data, and never send that data back to the developer or a cloud provider. Wang believes this need is largely unmet and that most serious enterprises will move toward it. The guarantee they want is simple: their data will not be used to improve their competitors.
He also predicts that AI services will generate more revenue over the next five years than AI models themselves. The value, he argues, sits both below the model, in infrastructure such as Nvidia, and above it, in the applications built on top. Treat that forecast as his opinion, since the interview offers no hard evidence for it.
Policy: Pooling Data Without Strangling Progress
On policy, Wang argues that some data should be shared broadly. Safety data in aerospace, for example, would advance the entire industry if pooled, and fraud and compliance data in financial services could do the same. Some sectors should deliberately pool data rather than keep it locked inside individual firms.
He also points to consumer-facing rules that block progress. Healthcare laws such as HIPAA and personal-information regulations largely prevent patient records from training AI models. Wang suggests a civilization that wants to cure disease should find clear, narrow provisions that let existing medical data improve future health outcomes.
He criticizes the EU's restrictive approach to data and argues that more permissive rules are compatible with a liberal democracy. He says the United States has put real effort into keeping chip production strong and should apply the same lens to data. Whether his regulatory prescriptions would work in practice is a matter for debate, and the interview does not test them against counterexamples.
The China Question and Closed Systems
Wang rejects the idea that China is comfortably behind. He cites the Chinese company 01.AI, whose Yi-Large model he says has become one of the best in the world, sitting just behind GPT-4o, Gemini and Claude 3 Opus on leaderboards. In his view Chinese AI capability is now close to neck and neck with the United States.
He credits China's centralized industrial policy, which has built world leadership in solar and made cheap electric vehicles. The system is less inventive at the frontier, he says, but very good at turning the crank once an industry is established.
The most unsettling part of his argument is military. He describes AI as potentially one of the greatest military assets humanity has seen, possibly more decisive than nuclear weapons, and worries about outcomes if a hostile power held it alone. That leads him to a split view of openness. The most advanced systems, he argues, should stay closed for geopolitical and military reasons. Less advanced open models can still deliver large economic value, and the hard question is where that line sits and how close the field is to crossing it.
Company Building and Public Narrative
Wang also discusses how he builds Scale AI. He favors direct channels over traditional press, arguing that media built around clicks tends to inflate companies on the way up and tear them down on the way down. Podcasts and owned channels let a founder explain what they are working on without distortion. He says he has received fairer treatment testifying before Congress than from many press outlets.
He is candid that the media tone toward his company shifted around 2022, after the market crashed and the narrative turned against tech. He says the company's decision to work with the US Department of Defense, which predates the current defense-tech wave, drew criticism rather than a balanced hearing.
His hiring philosophy is demanding. He wants people who care intensely about the quality of their work and will push through obstacles, and he compares the goal to a small elite unit rather than a large organization. He still approves every hire and says he overrides the team's recommendation in roughly 25 to 30 percent of cases, usually to calibrate a new hiring manager or handle an edge case.
He admits a mistake. Between 2020 and 2022, Scale grew from about 150 people to more than 700 as it chased hypergrowth. The bar slipped, challenges the organization once handled became harder to solve, and the company has since held headcount near 800 while growing revenue, aiming to keep its standards high.