Marc Andreessen's 2026 Outlook: AI Timelines, US vs. China, and The Price of AI

Andreessen traces AI's origins to the 1930s debate between building computers as 'adding machines' versus modeling human cognition. He explains how neural network theory existed since 1943 but remained impractical for 80 years until the ChatGPT moment in late 2022. The technology is now ultra-democratized, with the best AI available to anyone, and Silicon Valley is reallocating talent and capital at unprecedented speed.

• •AI represents a fundamental architectural shift comparable to the microprocessor, steam engine, and electricity - not just another internet wave
• •The computer industry took the 'adding machine' path for 80 years; AI finally delivers on the alternate 'human cognition' model from the 1940s
• •ChatGPT's December 2022 launch crystallized 80 years of research - we're only 3 years into an 80-year revolution
• •Silicon Valley's unique ability to recycle talent across technology waves is accelerating AI adoption faster than any previous cycle
• •Daily breakthroughs in research papers and new startups continue to exceed expectations

The cost of AI tokens is collapsing faster than Moore's Law due to improvements across the entire stack - chips, algorithms, and infrastructure. This drives elastic demand growth as prices fall. AI companies are growing revenue faster than any previous technology wave, with consumer products successfully charging $200-300/month premium tiers. The infrastructure build-out will cause per-unit costs to 'drop like a rock' over the next decade.

• •AI token costs are falling much faster than Moore's Law across all inputs - creating hyper-deflation in intelligence pricing
• •Leading AI companies show unprecedented revenue growth rates - actual customer dollars in bank accounts growing faster than any prior wave
• •Consumer AI products successfully charging $200-300/month, demonstrating willingness to pay for value vs. traditional SaaS pricing caps
• •Shortages create gluts: GPU and data center constraints will resolve as hundreds of billions flow into infrastructure build-out
• •Enterprise value proposition is straightforward: intelligence injection directly improves customer service, reduces churn, increases upsells

Andreessen explains the dynamic between frontier models and small models, where capabilities demonstrated by large models get replicated in smaller, cheaper versions within 6-12 months. Recent example: China's Kimi model replicates GPT-5 reasoning capabilities on 1-2 MacBooks. He predicts an industry structure similar to computing - a few 'god models' at the top, with massive volume in smaller models proliferating to embedded systems everywhere.

• •Capabilities of leading-edge models consistently shrink to small model size within 6-12 months - creating a 'chase function'
• •Kimi (Chinese open source) recently replicated GPT-5 reasoning on 1-2 MacBooks - demonstrating rapid capability compression
• •Most business tasks don't require 'Einstein-level' intelligence - 120 IQ models sufficient for vast majority of applications
• •Predicted structure: Small number of supercomputer-scale 'god models' plus cascade of smaller models down to embedded chips
• •Volume will be in smaller models proliferating everywhere, similar to how microchips evolved from mainframes to embedded systems

NVIDIA's dominance stems from historical accident - GPUs designed for graphics happened to work well for AI. But massive profits signal the entire chip industry to compete. AMD, hyperscalers building custom chips, and Chinese companies are all entering. AI chips designed from scratch would be more economically efficient than repurposed GPUs. In 5 years, AI chips will likely be cheap and plentiful compared to today.

• •NVIDIA's GPU architecture was designed for graphics/gaming, not AI - it's a twist of fate that GPUs work well for parallel AI computation
• •Hyperscalers (AWS, Google, Microsoft) now building custom AI chips - backward integration to control costs
• •Chinese government reportedly requiring DeepSeek to build on Chinese chips (Huawei) to accelerate domestic chip ecosystem
• •Startups building AI-specific chip architectures from scratch - more economically efficient than repurposed GPUs
• •Historical pattern: chip shortages become gluts as competition enters high-profit markets

China has 3-6 major AI companies (DeepSeek, Qwen/Alibaba, Kimi/Moonshot, Tencent, Baidu, ByteDance) plus numerous startups. DeepSeek's release was surprising - came from a hedge fund, not a national champion, and was released as open source. This kicked off a trend of Chinese open source releases. The reality of US-China competition has dramatically improved the DC policy landscape by making restrictions politically untenable.

• •DeepSeek came from a quant hedge fund, not government lab - suggesting smart teams can build state-of-art AI without massive resources
• •China releasing multiple competitive open source models: DeepSeek, Qwen (Alibaba), Kimi (Moonshot)
• •Chinese open source strategy may be 'dumping' to commoditize Western AI advantage, but also demonstrates genuine technical capability
• •US-China AI competition has improved DC policy environment - hard to justify restrictions when in a two-horse race
• •China ahead on robotics supply chain but working to catch up on chips - government pushing domestic chip development

Federal AI regulation risk has decreased dramatically, but now 1,200+ bills are being tracked across 50 states in both red and blue states. California's SB-1047 (modeled on EU AI Act) would have killed AI development by assigning downstream liability to open source developers. Colorado passed draconian rules and is now trying to reverse them. Federal government needs to assert interstate commerce authority to prevent state-level fragmentation.

• •Tracking 1,200+ AI bills across 50 states - both Republican and Democratic legislators proposing problematic regulations
• •California's SB-1047 would have assigned downstream liability to open source developers for any future misuse - 'completely insane'
• •Example: If open source model used in nuclear plant that melts down 5 years later, original developer liable
• •Colorado passed restrictive AI law, now trying to reverse it one year later after seeing economic impact
• •Federal moratorium on state AI regulation attempted but failed - active discussions on next approach to assert federal authority

The EU AI Act has effectively killed AI development in Europe - even American companies like Apple and Meta won't launch leading-edge AI features there. European entrepreneurs who moved to the US are the 'darkest' about Europe's situation. The Draghi Report on European competitiveness now acknowledges overregulation is holding Europe back, and there are efforts to unwind both the AI Act and GDPR.

• •EU AI Act so draconian that Apple and Meta won't launch advanced AI features in Europe
• •European philosophy: 'If we can't be leaders in innovation, at least we can be leaders in regulation' - self-harm strategy
• •Mario Draghi's competitiveness report acknowledges overregulation problem - EU now trying to reverse course
• •European entrepreneurs who relocated to US are most critical of Europe's approach
• •California's SB-1047 was essentially a copycat of the failed EU AI Act

AI companies are more creative on pricing than previous SaaS/consumer internet companies. While infrastructure providers offer tokens-by-the-drink (usage-based), application companies are exploring value-based pricing - charging based on the value of a replaced worker or productivity uplift. Consumer AI products successfully charging $200-300/month premium tiers. High prices benefit customers by enabling better R&D investment.

• •Big tech companies commoditized their AI magic through cloud services - tokens-by-the-drink model enables easy startup adoption
• •Core pricing principle: price by value, not cost - charge percentage of business value delivered, not compute consumed
• •AI companies exploring pricing based on replaced worker value (doctor, lawyer, coder) or productivity multiplier
• •Consumer AI routinely charging $200-300/month - higher than traditional consumer internet, demonstrating value perception
• •High prices are underappreciated: they enable vendors to invest more in R&D, ultimately benefiting customers with better products

Both proprietary and open source models continue advancing rapidly. Proprietary labs report 800+ new ideas and ongoing rapid progress. Open source enables education and knowledge proliferation - critical for training new AI researchers. Leading application companies are backward integrating, building their own models rather than remaining 'ChatGPT wrappers.' The answer may be 'both' - god models for highest intelligence, open source for volume applications.

• •Proprietary model labs report continuous rapid progress with hundreds of new improvement ideas in pipeline
• •Open source models catch up to proprietary capabilities within 6-12 months, then become learning tools for next generation
• •Open source critical for education: enables CS professors, students, and basement entrepreneurs to learn and experiment
• •AI application companies (like Cursor) backward integrating - building own models rather than staying dependent on third-party APIs
• •Best application companies now use dozens of models for different tasks, including custom-built and open source options

Multiple companies have caught up to OpenAI's capabilities in under 12 months - XAI (Grok) reached state-of-art in less than a year from standing start. Four Chinese companies caught up in under 12 months. This suggests no permanent moat for any single incumbent. Application companies are becoming full-fledged technology companies building their own AI. Venture's advantage: can bet on multiple contradictory strategies simultaneously through portfolio approach.

• •XAI reached OpenAI/Anthropic capability levels in under 12 months from standing start - suggests catch-up is achievable
• •Four Chinese companies (DeepSeek, Qwen, Kimi, others) caught up to US state-of-art in under 12 months
• •Leading AI application companies building own models, using dozens of different models, backward integrating into infrastructure
• •Venture portfolio advantage: can simultaneously bet on big/small models, open/closed source, consumer/enterprise - multiple paths to win
• •Companies need single coherent strategy; venture can invest behind every plausible strategy even when contradictory

The biggest ongoing discussion between Marc and Ben is managing the firm's public footprint - how outspoken and controversial to be. Being publicly visible and taking clear stands has proven to be an incredible competitive advantage with founders, who want to work with people demonstrating courage. It also educates policymakers in DC who otherwise only get anti-tech information from East Coast media. However, there are externalities to being controversial that require careful navigation.

• •Outspoken, controversial public presence is massive competitive advantage - founders know who a16z is before first meeting
• •Active marketing strategy from day one: broadcast clear beliefs so best founders understand the firm before walking in door
• •Public content educates DC policymakers who would otherwise only consume anti-Silicon Valley East Coast media
• •Tripling down on being thought leaders on tech and business issues through expanded communications team
• •Ongoing tension: how many 'third rail' topics to touch and how frequently - trying to moderate while staying on front foot

Andreessen distinguishes between what people say (stated preferences) and what they do (revealed preferences). Surveys show Americans panicked about AI killing jobs, but revealed preferences show everyone using AI tools. Historical pattern: every technology wave from printing press to automation sparked panic, but technology creates more jobs than it destroys. AD sectors (energy, materials, manufacturing) seeing unprecedented job demand due to AI infrastructure needs.

• •Stated preferences (surveys): Americans in total panic about AI destroying jobs and ruining everything
• •Revealed preferences (behavior): Everyone actively using AI tools - adoption happening at unprecedented speed
• •Historical precedent: 300+ years of technology panic from printing press to Marxism to 1960s automation fears - all wrong
• •AD sectors experiencing massive job growth due to AI infrastructure demands - physical world jobs booming
• •Technology historically creates more jobs than destroys by making economy larger and creating new categories of work