The Twenty Minute VCAI Fund’s GP, Andrew Ng: LLMs as the Next Geopolitical Weapon & Do Margins Still Matter in AI?
CHAPTERS
Compute bottlenecks: electricity, semiconductors, and the data-center buildout
Andrew Ng argues that the biggest near-term constraints on AI progress aren’t algorithms or data first, but physical infrastructure—especially electricity availability and semiconductors. He frames data centers as foundational infrastructure for the digital economy and contrasts Western permitting friction with China’s rapid power-plant expansion.
- •Electricity capacity and permitting delays are slowing data-center expansion in the US and parts of the West
- •Semiconductors remain a hard constraint as demand for inference and training continues to surge
- •Data centers should be treated like roads/railways—critical national infrastructure
- •China’s aggressive energy buildout (including nuclear) may shift competitive dynamics
Why compute demand stays ‘insatiable’ even as token costs fall
Despite efficiency gains and cheaper token generation, Ng says demand keeps outpacing supply because new high-value workloads expand to consume available compute. AI-assisted coding is highlighted as a major value bucket that drives massive inference demand, similar to how the early internet created many vertical winners beyond a single horizontal platform.
- •AI practitioners rarely feel they have ‘enough compute’; demand expands to fill supply
- •Rate limits and inference constraints show real excess demand for token generation
- •Efficiency improvements don’t reduce demand; they unlock more use cases
- •AI coding assistants are a leading driver of value and consumption
- •Market shape: dominant horizontal assistants plus many valuable vertical applications
AI coding assistants as a preview of AI-augmented work across functions
Ng pushes back on comparisons that place coding assistants at an early ‘2016 image generation’ stage, arguing they already deliver strong value while still having headroom. He broadens the lens: coding tools foreshadow productivity changes coming to marketing, recruiting, finance, and other knowledge-work roles.
- •Coding assistants are already indispensable for many developers
- •Productivity gains can compress months of engineering work into days or weekends
- •These tools serve as a template for how other job functions will be augmented
- •There is substantial remaining headroom for improvement
- •Adoption intensity is visible even inside Ng’s own teams
‘Vibe coding,’ learn-to-code advice, and how jobs change (without AGI hype)
Ng argues that making coding easier means more people should learn to code, not fewer—coding remains the best way to precisely instruct computers. He downplays near-term AGI narratives, describing AI as able to automate meaningful fractions of workflows while leaving substantial work for humans, and warns that workers who don’t adapt to AI tools will be disadvantaged.
- •‘Vibe coding’ aside, everyone benefits from being able to code with AI assistance
- •Examples: marketers building quick apps; recruiters using prompts to scale screening
- •AI won’t replace most knowledge workers end-to-end; it automates portions of work
- •AGI that does everything humans do is likely decades away (in Ng’s view)
- •Big risk cohort: experienced professionals and new grads who fail to adopt AI tooling
Infrastructure policy, regulation, and talent: what actually moves the needle
Ng credits regulatory rollback and bipartisan efforts for resisting overly restrictive AI rules, especially those targeting open source/open weights under extreme safety narratives. He emphasizes that America’s enduring advantage is attracting talent and funding research institutions, while also calling out public distrust of AI as a societal friction.
- •Unnecessary regulations can slow infra buildout and AI innovation
- •Overblown ‘extinction’ narratives can be used to justify anti-competitive rules
- •Attracting immigrant talent and supporting universities are key advantages
- •Semiconductor supply chain resilience is strategically important
- •Public skepticism and distrust can indirectly slow deployment and investment
Open vs closed models—and why open weights can become geopolitical soft power
Ng describes a shifting landscape where frontier US models are often closed while a tier below is open, and notes China’s surprising leadership in releasing strong open-weight models. He argues openness accelerates local knowledge circulation and innovation, and that model origin can shape values and narratives worldwide—making open models a lever of geopolitical influence.
- •US pattern: keep frontier closed, open the next tier down; China releases many strong open models
- •Open models speed innovation via fast knowledge sharing and iteration within ecosystems
- •Model origin can influence answers on sensitive topics—values, borders, history
- •Open-weight models function as a critical part of the AI supply chain
- •Soft power parallels: Hollywood/K-pop as influence; LLMs as the next communications frontier
Rethinking the ‘AI race’: multiple finish lines, cooperation, and China’s velocity
Ng challenges the idea of a single AI finish line, describing AI as a general-purpose technology with many distinct capability races. He acknowledges areas for cooperation alongside competition, and warns that China’s whole-of-nation mobilization—education, semiconductors, industrial coordination, and resource control—should not be underestimated.
- •No single ‘finish line’ (including AGI); capabilities will keep improving for decades
- •AI strength translates into national power, prosperity, and economic growth
- •China’s speed and coordinated state/industry effort is a major competitive force
- •Whole-economy investments include semiconductors, education, and international distribution
- •Control over rare earth elements adds strategic leverage
Do chip export controls work? Ng’s case that restrictions can backfire
Ng argues US chip export controls have largely incentivized China to accelerate domestic semiconductor development. He suggests China is building competitive systems by scaling out with more (potentially less powerful) chips, and questions whether the policy aligns with long-term US self-interest.
- •Export controls spurred faster Chinese investment and urgency in semiconductors
- •China can compete via scale-out architectures even with weaker individual chips
- •US restrictions may reduce US leverage while accelerating a competitor’s capability build
- •Policy outcomes should be evaluated by strategic self-interest, not intent
- •Semiconductors remain central to AI competitiveness and national security
Europe’s position: stop over-regulating and invest in builders
Speaking to a European audience, Ng argues that leading in regulation is not a competitive advantage. He advocates for reducing regulatory friction, enabling ambitious work, and investing in building across the stack while it’s still early in the AI cycle.
- •Europe has ample talent but risks slowing itself through heavy regulation
- •‘Being leaders in regulation’ doesn’t create durable competitiveness
- •Enable people who want to work hard to build and ship faster
- •AI is still early; catch-up is possible with the right policy and investment posture
- •Competitiveness requires building products and companies, not just rules
Why the application layer is the biggest opportunity—and why deploying capital is weird
Ng argues that while massive capital can be absorbed by data centers, many valuable application-layer bets are highly capital-efficient, creating a mismatch for investors trying to deploy huge funds. He notes a common flow where application funding largely passes through to model providers and then to NVIDIA, yet he remains bullish on many smaller, high-ROI application businesses.
- •Infra can absorb billions easily; applications can often be tested with comparatively little capital
- •Some VCs struggle to deploy large sums at the app layer because experimentation is cheap
- •Observed money flow: app startups pay model providers who pay GPU vendors
- •Despite pass-through dynamics, many smaller app businesses can be efficient and profitable
- •The key is building products users love, then optimizing economics as tech improves
Margins in AI: build for user love now, bend the cost curve later
Ng agrees margins matter, but argues AI businesses shouldn’t assume today’s cost structure is static. He describes a build-first approach—prove demand, then apply cost-reduction techniques and benefit from falling token prices to improve unit economics faster than the market’s cost declines.
- •Margins ultimately matter, but forecasting future tech/cost curves is essential
- •Prototype phases often ignore token costs to validate product value
- •Real risk appears when API bills scale; then teams must optimize aggressively
- •Techniques can reduce costs faster than token-price declines alone
- •Rejects both ‘ignore margins forever’ utopianism and rigid present-day cost assumptions
Defensibility after AI: software moats weaken, industry moats and distribution matter more
Ng argues AI changes the nature of moats: building software is faster to replicate, so software-as-a-moat is weaker. Defensibility shifts to industry-structured advantages like marketplaces, brand, reputation, and other domain-specific barriers rather than AI itself.
- •Moats depend more on industry structure than on AI as a technology
- •Traditional software moat (time to build) is less defensible in an AI era
- •Marketplaces, brand, reputation, and distribution can still create strong moats
- •Enterprise vs consumer dynamics change the defensibility landscape
- •AI accelerates competition but doesn’t eliminate the need for strategic moat design
Enterprise adoption: change management over data, plus the long slog of implementation
Ng says the main blocker in large enterprises isn’t data availability but people, organizational change, and workflow redesign. He predicts meaningful progress in the next couple of years, but emphasizes enterprise adoption will take a decade-plus—similar to the long transition from on-prem to cloud—despite AGI-speed hype.
- •Biggest enterprise bottleneck is people and change management, not data
- •Data is often sufficient to start; scrappy teams can unlock value from private/internal datasets
- •Security, permissioning, and legacy systems slow adoption but won’t stop it long term
- •AI adoption will take longer than hype suggests; even 10 years from now the work continues
- •Bad advice to ‘not learn to code’ will age poorly; coding literacy becomes more valuable
Growth vs cost savings: redesign workflows to go faster or serve many more people
Ng reframes AI’s value creation: automating a single step in a workflow yields limited savings, while real upside comes from rethinking the entire process. The two scalable growth patterns he highlights are doing the same work dramatically faster and expanding capacity to serve far more customers at high-touch quality.
- •Cost-savings automation often feels incremental (e.g., 20% of a five-step task)
- •Major value requires workflow redesign, not isolated step automation
- •Growth pattern #1: dramatically faster turnaround times (e.g., loan decisions in minutes)
- •Growth pattern #2: serve 10x–1000x more users with previously high-touch services
- •AI can change the product itself, not just reduce headcount or expense
Vertical integration, bubbles, and hype: calibrating infra spend and keeping society onboard
Ng compares today’s stack to early computing: vertical integration can win before standards and APIs mature, but boundaries will clarify over time. He’s cautiously attentive to frothy financial engineering and circular deals in data-center investment, while criticizing hype (especially doomer narratives) for distorting policy, talent pipelines, and public support.
- •Early industries favor integrated players until standards and interfaces mature
- •OpenAI’s big infra bets have paid off so far; over-investment remains a risk to monitor
- •Complex risk-shifting financial instruments can be a sign of bubble dynamics
- •Application layer shows clearer ROI; infra investment level is harder to calibrate
- •Hype harms adoption and talent: fear narratives can deter students and communities
