AI Fund’s GP, Andrew Ng: LLMs as the Next Geopolitical Weapon & Do Margins Still Matter in AI?
Summary
Andrew Ng sees electricity and semiconductors—not a sudden end to scaling—as AI’s binding near-term constraints. Compute demand has felt insatiable throughout his roughly 20-year career, and cheaper inference is being met by still more usage, particularly in coding. Infrastructure investment therefore needs to remain “a lot,” though complex financing and circular deals make the exact amount harder to calibrate.
AI coding is the clearest evidence that application-layer ROI already exists. Work that once took six engineers half a year can sometimes be completed by one person over a weekend, while Claude Code, OpenAI Codex and Gemini CLI compete for developers who can switch tools quickly. Ng calls coding a “harbinger” for recruiting, marketing and finance—but the weak tool-level moat means today’s leader can change quickly.
In Ng’s software-engineering examples, the labor divide is between AI-enabled and AI-unprepared workers, not simply humans and machines. He describes experienced engineers fluent in AI as the most productive, followed by AI-fluent graduates, experienced developers still working “like it’s 2022,” and graduates without AI skills. Because AI might perform 30%—perhaps 50%—of a recruiter’s work but not the remainder, he expects augmentation and partial role redesign rather than wholesale replacement for many jobs; AGI is “decades away, maybe even longer.”
Ng hopes AI can push GDP growth toward 5%-6% or more by making intelligence cheap, but that requires redesigning products rather than merely trimming labor. Automating one of five equal workflow steps yields a useful 20% saving; turning a two-week loan process into a 10-minute initial answer changes the product. The larger opportunity is to do work faster or “a thousand times more,” extending services such as high-touch financial advice to customers who could not previously receive them.
Open-weight models are becoming innovation infrastructure and geopolitical soft power. China’s releases accelerate domestic knowledge circulation while influencing the answers users receive about borders, history and values; Ng compares that influence with Hollywood and K-pop. He believes US chip controls “largely backfired” by incentivizing China to accelerate semiconductor development, while Europe’s ambition to lead in regulation is “not a competitive advantage.”
Application margins matter, but Ng builds against the expected cost curve rather than today’s token bill. Token prices may be falling around 80% annually “depending on who you believe,” and his teams have repeatedly reduced costs faster still after first proving user demand. Harry counters with roughly 80% of Replit or Lovable’s pass-through going to Anthropic, supporting Ng’s warning that today’s “VC-subsidized AI computing” cannot persist indefinitely.
Enterprise adoption is constrained primarily by people and change management, not by an absolute shortage of data. Private transaction, sales, manufacturing and logistics data can already support scrappy projects; security, permissions and workflow redesign slow deployment. Ng expects major progress within one or two years but says enterprises will still be discovering applications a decade from now, making workforce reskilling a central unresolved problem.
Deep dive
1. Power and chips set AI’s near-term ceiling
Ng identifies electricity and semiconductors as the two immediate bottlenecks. US permitting and community resistance are constraining data centers just as China builds power plants “left and right,” including nuclear power.
The demand signal is unusually clean: “I have yet to meet a single AI person that ever felt like they had enough compute.” After roughly 20 years of that constraint, valuable generative-AI workloads now leave users rate-limited and providers unable to supply enough inference.
Harry’s GPT-5 pushback—perhaps scaling laws have hit limits and efficiency now matters—does not change Ng’s conclusion. Inference is becoming cheaper, including through what Ng recalls as an OpenAI open-weight model with roughly 120 billion parameters and, he thinks, 5.7 billion active parameters, but consumption keeps outrunning efficiency.
More infrastructure is clearly necessary; the unresolved investment question is how much. Ng is not alarmed by circular deals, yet calls complex risk-shifting instruments and the cited “$600 billion problem” signs that financing feels “a little bit more bubble-ish.”
2. Coding provides the strongest proof of application-layer value
Ng maps the market onto the internet era: ChatGPT may dominate horizontal information discovery, with Gemini’s Android and Chrome distribution making it a serious player, while large vertical markets remain open. Coding assistance is the clearest valuable vertical so far.
He rejects the comparison with image generation in 2016-2017: coding tools are already working extremely well. His engineering leader’s attitude was that the tools would have to be pried from his “cold dead hands”; Ng similarly never wants to return to unaided coding.
Rather than choose between replacing the bottom 5% of work and making people 10x better, Ng points to observed compression: projects once requiring six engineers for six months can sometimes be built by one engineer over a weekend. He even generated multiplication flashcards for his daughter rather than driving to a store.
Tool loyalty remains fragile. Ng’s favorite would have changed every three months; he loves Claude Code, has recently used OpenAI Codex much more, and thinks Gemini CLI may be improving faster than recognized. A consumer brand is more defensible than a developer tool users can switch “on a dime.”
3. AI fluency is reorganizing the talent hierarchy
Harry says a small subset of jobs may be in trouble, but Ng rejects generalized displacement narratives. If AI handled 30% of recruiting work—perhaps 50%, though “that feels a little bit high”—humans would remain necessary for everything outside that envelope.
His software-engineering hierarchy is explicit but not universal: experienced, AI-fluent developers usually move fastest; AI-fluent graduates follow; people with roughly 10 years of experience still coding “like it’s 2022 before ChatGPT” fall below them; and graduates without AI knowledge are the cohort genuinely struggling.
Harry’s talent-pipeline objection is worth keeping: eliminating junior work could leave no future seniors. Ng sees a different failure—slow university curricula producing computer-science graduates who have never called an internet API or used AI building blocks—while businesses cannot find enough graduates who have learned them independently.
Asked whether $100 million can dull an engineer’s motivation, Ng offers an honest non-answer: “I don’t know.” His experience is that wealthy Silicon Valley peers often keep working because building is fun; wealth makes people lazy “much less than one might guess.”
4. Cheap intelligence creates growth only when workflows change
Against Andrej Karpathy’s suggestion that AGI could blend into 2% GDP growth, Ng hopes for something “much closer to five, six or more percent.” His mechanism is cheaper intelligence: doctors, tutors and advisers are costly to train, but AI could give everyone “an army of smart, well-informed staff.”
The obstacle is treating AI exclusively as cost reduction. In a five-step process where each step consumes 20% of effort, automating one produces a worthwhile 20% saving—but “it doesn’t feel like a game changer” and leaves the underlying product intact.
Ng sees two larger patterns: faster and more. A lender that replaces a two-week wait with an initial answer in 10 minutes changes the product and enables growth; a business that extends high-touch service or financial advice beyond wealthy clients expands its market. AI creates the most value when it makes an activity feasible “a thousand times more.”
5. Open models are both innovation infrastructure and soft power
American labs often keep the frontier model closed and release the next tier down; China has moved toward releasing many strong open-weight models. Ng did not expect “China’s AI industry would end up being more open than America’s AI industry,” though he remains grateful for every open release.
Openness disproportionately benefits the nearby ecosystem. Published weights let teams use the work, contact one another and resolve implementation problems; closed development and $100 million talent compensation slow that circulation across American and European communities.
Open weights can influence answers about national borders, politically sensitive history and values. Ng calls that “a tremendous source of geopolitical influence,” comparing the resulting soft power with Hollywood’s American dream and South Korea’s disproportionate reach through K-pop.
He rejects a single US-China finish line—AI contains many capabilities and will improve for decades—but warns against underestimating China’s whole-country effort across semiconductors, K-12 education, business adoption and rare-earth elements. US chip controls “largely backfired,” incentivizing Chinese alternatives using more, individually weaker chips.
6. Application ROI is real, but capital does not fit neatly
Foundation-model spending lets application teams access capabilities costing billions to train for tens, hundreds or thousands of dollars. The VC problem is almost comic: the industry knows how to deploy $10 billion into data centers, but an application hypothesis may require only $1 million to test.
Harry presses on ugly economics, citing roughly 80% of Replit or Lovable’s pass-through going to Anthropic. Ng compares the moment with VC-subsidized food delivery: “VC-subsidized AI computing” cannot continue forever, though smaller applications already producing millions or tens of millions in revenue may be inexpensive to build and operate.
Ng directly disputes the claim that useful agents are a decade away. AI Fund built an agentic tariff-compliance workflow after a Biden–Trump debate: it reads regulations and product specifications—his bicycle example includes price and wheel size—then makes suggestions. The resulting portfolio company, Giga Dynamics, benefited as tariff compliance grew more complex.
AI Fund’s answer to cheap experimentation is operating rather than merely allocating capital. It develops ideas, validates them with customers, recruits founders and argues over product and pricing; its initial check is usually about $1 million at a $4 million cap, producing roughly 20% ownership on a SAFE plus common stock for sweat equity.
7. Future cost curves and industry structure determine margins and moats
Large, medium and tiny models will coexist because intelligence spans spelling “butterfly” through hours of technical reasoning. A tiny model, perhaps running locally, can handle grammar and spelling; more powerful models earn their cost on complex reasoning and code.
Margins ultimately obey “the laws of physics” or “the laws of finance,” but Ng first builds something users love. API bills may then exceed one engineer, two engineers or “a whole bunch of engineers”; almost every time so far, his teams have bent that cost curve down faster than market token prices.
AI weakens software itself as a moat: ten years of software development is much harder to defend than before. Defensibility instead remains industry-specific—two-sided marketplaces, consumer or enterprise relationships, brand, reputation and other structural advantages—not an automatic property of using AI.
Vertical integration is valuable in immature markets because component boundaries remain unclear. Ng’s computing analogy is that integrated players such as IBM could solve the interoperability problems and build working products; standards such as USB eventually enabled horizontal specialists. He says OpenAI’s infrastructure spending has paid off to date, while conceding that overinvestment remains possible as financial engineering grows more elaborate.
8. Enterprise AI is a change-management problem, not a data famine
Asked for the largest adoption barrier, Ng answers “people and change management.” Data matters, but it is “definitely not the bottleneck”; the industry has inflated a genuine need into a reason not to begin.
Data is vertical and more available than executives assume. Financial institutions can convert complex PDFs or SEC filing tables into analysis-ready Markdown or Excel, while transaction, sales, product, manufacturing and logistics records provide valuable private starting points for a scrappy team.
Harry’s pushback is operational: major institutions struggle with permissions, security and custom systems, sometimes refusing ChatGPT and building internally. Ng expects adoption anyway, comparing it with cloud computing—years into that transition, many workloads still remain on-premises.
The timeline is therefore long. Progress and returns will be substantial over the next one or two years, but companies will still be identifying applications ten years from now. Claims of AGI in two years are, for most reasonable definitions, “just ridiculous.”
9. Policy and education determine whether the gains diffuse
Ng credits the US federal government with resisting extinction-driven restrictions and clearing unnecessary regulation, but says failing to attract talent and invest in scientific institutions would sacrifice core advantages. His regulatory wish is to keep attracting ambitious talent while securing a semiconductor supply chain overly dependent on TSMC.
Europe receives a blunter prescription: aspiring to lead in AI regulation is “not a competitive advantage.” It still has smart people and time, but should “stop regulating so much and just focus on investing and building,” including letting willing people work hard.
Hype has tangible costs. Ng recounts a high-school student rejecting an AI career because she associated it with human extinction; the same narrative can weaken community support for data centers. Schools should instead update curricula, embrace AI and teach every student to code—even if writing code manually is becoming obsolete.
His deepest concern is the speed of reskilling: agricultural transitions allowed farmers to keep farming until retirement while their children changed trades, but today’s workers themselves must adapt. His optimistic endpoint reverses the default question from “Is there an app for that?” to “I built an app for that,” turning software users everywhere into creators.