a16zMarc Andreessen's 2026 Outlook: AI Timelines, US vs. China, and The Price of AI
CHAPTERS
Why this AI wave is historically different (and still early)
Marc frames AI as the biggest technological revolution of his lifetime—larger than the internet—with comparisons to electricity and the microprocessor. He argues we’re only a few years into an 80-year arc of ideas (neural nets) finally working at scale, so today’s products are likely primitive compared to what’s coming.
- •AI is positioned as a general-purpose breakthrough on par with major industrial-era inventions
- •The 'ChatGPT moment' crystallized decades of research into real-world capability
- •Democratized access: consumers can use state-of-the-art models instantly via apps
- •Silicon Valley’s core advantage is reallocating talent/capital into new tech waves
- •Current AI product forms are unlikely to be what users rely on in 5–10 years
Revenue vs. burn: how AI companies can scale profitably
The discussion tackles skepticism that AI revenue growth is offset by equally fast-growing costs. Marc splits the landscape into consumer and enterprise/infrastructure models, arguing both have unusually rapid adoption because distribution is piggybacking on the already-built internet.
- •Two primary business models: consumer AI and enterprise/infrastructure AI
- •Unlike the internet era, the 'carrier wave' (global broadband + smartphones) already exists
- •Consumer AI can spread to billions quickly, enabling faster adoption than prior tech revolutions
- •AI monetization is strong, including higher consumer price tiers ($200–$300/month)
- •Enterprise value proposition centers on what 'intelligence' is worth in operational uplift
“Tokens by the drink” and AI’s deflationary cost curve
Marc explains why usage-based pricing (tokens) has worked well for infrastructure providers and startups: low friction and scalable. He emphasizes that unit costs are collapsing faster than Moore’s Law, suggesting expanding demand and improving economics as compute becomes cheaper and more abundant.
- •Core infra model: usage-based 'tokens by the drink' for intelligence
- •AI input costs are falling rapidly, creating hyperdeflation in per-unit inference/training
- •Lower costs drive elasticity: cheaper tokens unlock more use cases and demand
- •Supply/demand dynamics: shortages motivate massive buildout that tends to flip into gluts
- •Over a decade, optimization and buildout should drop unit costs significantly
GPUs, chips, and the infrastructure bottleneck (and why it won’t last)
The conversation turns to GPU longevity, data center buildout, and the chip roadmap. Marc argues that current reliance on GPUs is partly historical accident; purpose-built AI chips, hyperscaler silicon, and new competition should reshape pricing and supply in coming years.
- •GPU shelf life can be extended (e.g., hyperscalers optimizing older hardware)
- •Big profit pools attract competition: AMD, hyperscalers, startups, and Chinese entrants
- •AI on GPUs is path-dependent; from-scratch AI chips could be more efficient
- •New chip startups may win or be acquired by larger scaling players
- •Expect many AI chip options globally (US, China, Japan, Korea), driving competition
Big models vs. small models: the cascading “pyramid” structure
Marc outlines a dual-track future: a few “God models” at the top, with a large volume of smaller models cascading down into devices and embedded systems. He highlights how small models often catch up to big-model capability with a time lag, changing deployment economics.
- •Data center build is oriented to large models, but small-model capability is rapidly improving
- •Smaller models often match prior big-model performance within 6–12 months
- •Example: open models approaching frontier reasoning, enabling local deployment
- •Economic argument: not every task needs the most intelligent/expensive model
- •Predicted industry shape: few supercomputer-like models + widespread smaller embedded models
US vs. China in AI: open source, chips, and the two-horse race
The episode frames AI as a geopolitical and economic contest primarily between the US and China, with global proliferation at stake. Marc discusses China’s model ecosystem (DeepSeek, Qwen, Kimi/Moonshot, others), chip catch-up efforts, and how this competition reshapes Washington’s policy posture.
- •US–China ties are economically intertwined, unlike US–USSR, complicating 'Cold War' framing
- •DC increasingly views AI as a national competition: whose AI proliferates globally?
- •Chinese ecosystem: DeepSeek, Alibaba’s Qwen, Moonshot’s Kimi, plus Tencent/Baidu/ByteDance
- •DeepSeek surprised markets (quality, open-source release, origin from a hedge fund)
- •China’s chip push (e.g., Huawei ecosystem) and robotics supply chain advantage matter long-term
Policy & regulation: federal vs. 50-state chaos
Marc argues the federal outlook has improved because policymakers don’t want to handicap the US against China, but states are introducing a flood of bills. He explains why fragmented state regulation is mismatched to interstate AI markets and discusses attempts (so far unsuccessful) to preempt state action.
- •Federal risk of 'ruinous' AI legislation has declined markedly over the last two years
- •States have introduced ~1,200 AI-related bills, across both red and blue states
- •AI is inherently interstate; state-by-state compliance could be crippling
- •A federal moratorium/preemption attempt was explored but didn’t pass
- •Some states (e.g., Colorado) are already trying to unwind overly restrictive laws
What “draconian” AI bills look like: EU AI Act and California SB 1047
Marc uses Europe’s AI Act as a cautionary tale, arguing it has slowed deployment and withheld features from European users. He then describes California’s SB 1047 (vetoed) and highlights provisions that would have imposed downstream liability on open-source developers, potentially freezing research and startup activity.
- •EU AI Act described as overreaching; even major US firms have delayed EU AI launches
- •Europe is reportedly reconsidering aspects of its regulatory approach (competitiveness focus)
- •California attempted an EU-style approach; SB 1047 passed legislature but was vetoed
- •Key concern: downstream liability assigned to open-source model creators for future misuse
- •Such liability would chill open source, academia, and startups; federal alignment is needed
AI pricing models: usage-based vs. value-based (the trillion-dollar question)
Marc explains why tokens-by-usage is powerful for infrastructure and early-stage builders, but may not be optimal for applications. He emphasizes pricing discipline: avoid cost-plus pricing when possible, and instead price to capture a portion of the customer’s realized value (labor replacement or productivity uplift).
- •Infra usage pricing enables fast startup iteration and low fixed costs
- •Application pricing can move beyond tokens to outcomes, productivity, or value delivered
- •Potential models: replacing a role (coder/doctor/lawyer) or augmenting a professional’s output
- •a16z emphasizes pricing as a core capability and encourages experimentation
- •Higher prices can benefit customers by funding faster product improvement and reliability
Open vs. closed models: why both may win
Marc argues the open/closed debate is still unresolved: proprietary labs keep advancing, while open source keeps rapidly matching capabilities and spreading know-how. He highlights education and skill diffusion as a strategic advantage of open models, accelerating the supply of AI talent and builders.
- •Closed models continue improving quickly; lab insiders report strong ongoing progress
- •Open models frequently 'catch up' and shrink frontier capabilities into cheaper form factors
- •Open source is a key educational substrate for students, engineers, and founders
- •Talent constraints (researcher scarcity) may ease as knowledge spreads and more people upskill
- •Likely end state: a mix—premium frontier models plus ubiquitous open/small models
Incumbents vs. startups: from “GPT wrappers” to full-stack AI companies
The conversation reframes the startup debate: application-layer companies aren’t just thin wrappers if they orchestrate many models, customize stacks, and sometimes build their own models. Marc notes fast catch-up dynamics (xAI, and multiple China players) as evidence that leads may not be durable.
- •Incumbents (Google, Meta, Microsoft, Amazon) are investing aggressively; new incumbents include OpenAI/Anthropic
- •Rapid catch-up examples suggest no permanent lock-in for a single frontier lab
- •Application startups increasingly use multiple models, not just one provider
- •Many app companies backward-integrate: fine-tune/build their own models and leverage open source
- •Portfolio advantage: venture can invest across contradictory strategies (big/small, open/closed, apps/foundations)
a16z AMA: strategy, org design, and the cost of being outspoken
Marc discusses how he and Ben operate—frequent debate but generally converging—and where tension does exist: the firm’s public footprint. He argues that clear, sometimes controversial public positioning attracts founders and educates policymakers, but creates real externalities that must be managed.
- •“Disagree and commit” is less common because internal debate usually converges
- •Biggest ongoing debate: how outspoken/controversial the firm should be publicly
- •Outspokenness helps founders self-select and builds trust before first meetings
- •Public content also targets Washington to counterbalance hostile/biased information channels
- •a16z sees 'little tech' policy advocacy as necessary given high stakes and collective action failures
Jobs, labor, and adoption: panic in polls vs. revealed preferences
Marc places current AI job fears in a long historical pattern of technology panics, arguing society ultimately adapts. He emphasizes the divergence between what people say (surveys) and what they do (rapid adoption), predicting widespread normalization as AI proves practically valuable.
- •Tech panic cycles recur: printing press, industrial automation, outsourcing, robots, now AI
- •Marc argues Silicon Valley must take social concerns seriously and explain technology clearly
- •Observed behavior shows mass adoption despite negative polling sentiment
- •Examples of daily AI use: workplace help, health queries, relationship advice, complaint letters
- •Expectation: turbulence near-term, but longer-term normalization and dependence on AI tools
Lightning round: beliefs, cryonics, reality distortion, and Mars
In closing rapid-fire questions, Marc reflects on continually updating beliefs, skepticism about current cryonics, and practical checks on ego via market feedback and public criticism. He also downplays personal interest in going to Mars while expressing confidence that routine trips may become plausible.
- •Changing mind frequently, often influenced by younger builders and new ideas
- •Cryonics: not with current technology, given poor track record
- •Reality distortion is real; mitigated by partners’ candor and markets delivering fast feedback
- •Missed investments and public reminders reinforce humility
- •Mars: unlikely personally, but he expects the possibility of routine travel within a decade
