Andrew Ng on AI Infrastructure, Open Models and Building Useful Applications

Harry Stebbings interviews Andrew Ng, founder of DeepLearning.AI and a general partner at AI Fund, about infrastructure, AI adoption, investment economics and geopolitics. Ng works on and invests in AI applications, so his views about demand and business opportunities have a commercial context.

Ng identifies electricity and semiconductors as important near-term constraints. He sees data centers as critical digital infrastructure and worries about delays in power supply and permitting, while acknowledging that local communities have interests in proposed projects. His comparison with Chinese infrastructure development is his assessment.

He says that throughout his AI career researchers have wanted more computing capacity. Inference for valuable tasks such as coding is generating strong demand even as models become cheaper and more efficient to run. He describes a supply-demand imbalance but does not quantify a market-wide shortage.

Falling token costs and growing use can coexist, Ng argues. He sees consumer assistants as horizontal tools and coding as one valuable vertical, with potential applications in recruiting, finance and marketing. His experience with coding tools and expectations for other sectors do not establish a universal productivity rate.

Ng says AI Fund engineers and he personally find coding assistants useful, while acknowledging substantial room for improvement. He considers them further along in practical value than image generation was in the period suggested by Stebbings. This is his comparison based on his team's use.

Asked about US policy, Ng welcomes what he sees as removal of unnecessary restrictions but warns against weakening the country's ability to attract talent or fund education and science. His view that some AI-safety narratives are overstated is a contested policy position, not a finding that AI risks are absent.

With a policy 'magic wand,' he would prioritize attracting talented people and securing semiconductor supply chains. He says dependence on a concentrated manufacturing base is a strategic concern. These are recommendations and risk judgments, not an assessment of any particular export-control rule's legal status.

Ng describes using AI-assisted coding to make multiplication flashcards and says a colleague built a small marketing-feedback application. His examples show uses within his organization; they do not prove that every nontechnical employee can build or maintain production software safely.

Stebbings asks whether productivity gains mean fewer entry-level employees. Ng says some jobs face risk but believes many workers will retain tasks that AI cannot perform. He offers hypothetical shares of a recruiter's work and a distant view of general AI, explicitly as estimates rather than measured employment effects.

On the junior talent pipeline, Ng says experienced engineers who understand AI tools are especially productive in his experience, while new graduates familiar with those tools can also contribute quickly. He worries about computer-science graduates who have not learned to use modern systems and urges universities to update curricula.

Asked about very large compensation packages for AI researchers, Ng says he does not know whether particular packages are justified. He has seen wealthy colleagues continue working hard, but his personal experience cannot establish how compensation affects motivation across the industry.

Ng hopes AI could contribute to substantially higher economic growth by making some forms of expertise less expensive. His discussion of productivity, medical advice and tutoring is an optimistic scenario, not evidence of a specific GDP growth rate or validated clinical outcomes.

On open and closed models, he welcomes the release of open-weight systems by US and Chinese teams. He argues that openness can accelerate the circulation of knowledge within local communities as well as internationally; the relative innovation effects he describes are a hypothesis, not a quantified study.

Ng also sees open-weight models as a potential source of geopolitical influence. A model's answers about history, borders or political events may reflect choices in its data and design. His point is about possible influence and requires scrutiny of actual systems, not a claim that every model from a country gives the same answer.

He resists treating AI as a single race with one finish line. Countries may cooperate in some areas and compete in others, and capabilities vary across tasks. He nevertheless expects stronger AI capacity to affect economic and strategic power, which is a forecast rather than a predetermined result.

Ng says China's coordinated industrial investment, training, infrastructure and access to critical materials should not be underestimated, while acknowledging problems in all regions. His assessment describes potential advantages and does not independently evaluate each program's effectiveness.

On semiconductor export controls, he argues that restrictions encouraged China to accelerate domestic chip development and may have backfired for US interests. That is his causal and policy judgment; the interview does not provide a counterfactual or establish the controls' net effect.

Asked about Europe, Ng wants regulators to focus more on investment and building. His opposition to what he calls overregulation is a policy preference; the conversation does not compare the benefits and costs of particular European rules.

Ng says infrastructure requires investment but sees many opportunities at the application layer, where others have already paid to train foundation models. A challenge for large funds, he suggests, is that individual application experiments can require much less capital than building data centers.

Stebbings counters that some AI applications face heavy model costs and engineering expenses. Ng agrees economics vary and notes that some usage is effectively subsidized by investors. He hopes falling inference costs and efficiency improvements will support durable businesses, while acknowledging that subsidies cannot continue forever.

He expects a range of large, medium and small models because tasks differ in difficulty and cost sensitivity. A simple spelling task may not need the same system as a complex coding problem; the best choice remains application-specific.

Ng disagrees with the idea that useful agents are necessarily a decade away. He describes an AI Fund portfolio company building workflows for tariff compliance and mentions other legal and medical-assistant projects. These are company examples offered by an investor, not independently verified measures of accuracy or medical effectiveness.

On margins, Ng says they matter eventually. His team may initially prioritize whether users value a prototype, then work on the cost of running it. He reports that his teams have often reduced costs after usage grew; that experience does not guarantee that every application can reach the same economics.

Ng argues that defensibility is often specific to an industry. Software that once took years to reproduce may be easier to copy, while distribution, brand, reputation and network effects can still matter. He does not claim those advantages are automatic or permanent.

For enterprise adoption, he sees people and change management as a larger barrier than data alone. Ng says teams can often start with available internal or public data, though sensitive information and permissions still need appropriate handling. His assessment differs from Stebbings's examples of organizations blocked by security and purchasing processes.

Ng expects meaningful progress in the next few years but says enterprise adoption will still be underway a decade later. He rejects very short timelines for general AI under his preferred definitions. These are forecasts; the interview does not settle the timeline or define a universally accepted threshold.

He regards telling students not to learn coding as poor advice. As tools become easier to use, he wants more people to learn to express problems precisely and use AI to help write programs. He distinguishes learning to direct and understand code from writing every line by hand.

Stebbings returns to whether AI value depends on shifting spending from labor to software. Ng says firms can also use AI to grow: redesign a workflow to deliver a result faster or serve more customers, rather than only reduce headcount. His loan-underwriting example is hypothetical and does not establish a safe or lawful lending process.

On vertical integration, Ng compares AI with early computing, when firms often had to build many components together because standards were immature. As interfaces mature, he expects specialist suppliers to interoperate more easily. The analogy is a framework, not proof of the industry's eventual structure.

He says it is possible to overinvest in infrastructure even while more power, data centers and chips are needed. Complex financing instruments and circular transactions are signs to watch, in his view, but he is not currently alarmed by every deal. He does not identify a precise threshold for a bubble.

Ng says he is less annoyed by questions about a bubble than by what he regards as exaggerated AI claims. He favors discussing workforce training and useful investment. His concerns about hype are his perspective; communities may also have legitimate reasons to examine data-center impacts and AI risks.

He urges educational institutions to update courses and teach students how to work with AI, including coding. His favorite development tools change frequently, he says, and he finds it hard to predict which provider will lead the coding market.

Reflecting on Baidu and China, Ng values speed and intensity but says people have different circumstances and should be respected when they cannot work long hours. He describes his own enthusiasm for coding on weekends as a personal choice, not a productivity standard for every worker.

Ng describes AI Fund as a venture studio that develops ideas, works with customers and recruits founders before or while financing companies. He says its initial investments often combine a check with common equity for work building a company. The approximate terms he mentions are a description of that firm's model, not a universal venture standard.

His greatest concern is helping people adapt to economic change within their own working lives. In previous transitions, he argues, children could enter new occupations while parents continued existing work; faster change may require adults to learn new skills too. This is his interpretation of the transition, not a quantified displacement forecast.

He says journalists' questions about AI have improved but believes financial and other incentives still encourage hype. He does not identify the companies he criticizes, and the discussion does not establish the motives of individual organizations.

Ng closes with a hope that more people can turn ideas into software, moving from asking whether an application exists to creating one. That is an educational and product aspiration, not a guarantee that access, safety or economic gains will be equally distributed.