[1/2] Marc Andreessen's 2026 Outlook: AI Timelines, US vs. China, and The Price of AI (AI column)
I watched with great interest. story Marc Andreessen from the company Andreessen Horowitz (a16z)It explains why we are only at the beginning of a major technological revolution. The cost of intelligence is collapsing faster than Moore’s Law, the revenue of AI companies is growing at an unprecedented rate, and products through the Internet are growing. 5 The years will be different.
The main points of this story are such
1The main thesis: we are at the beginning of the revolution Andreessen conducts a historical excursion: 1930-x had the choice to build computers like computers (von Neumann architecture) or (brain). Pick one, get one. 80 Classical Computers (from mainframes to smartphones). Neural networks remained academic exotic until ChatGPT (ending 2022). In the last three years after that, you see it's more of the internet, it's the microprocessor/steam engine level. What's interesting for engineers is that literally every month, breakthrough research comes out with new possibilities that seemed impossible. The products we build today will probably become obsolete. 5 In the years to come, it's gonna get better.)
2Economy: Revenue is explosive, but costs are rising - what is missing? AI companies are growing faster than anyone Andreessen has seen (Real money in banks, unprecedented takeoff). But the criticism is "burn catches up with revenue." His answer is in two parts:
- Consumer AI The Internet is already built. 5–6 Billions of people online, smartphones at $10. AI can be downloaded, unlike electricity or running water - there is no distribution problem. Monetization is strong, including high tariffs ($200–300/me standard - above SaaS).
- Enterprise AI The question is, "How much is intelligence?" If AI raises service levels, lowers churn, strengthens upsell products, that’s a direct business outcome. The model of tokens of intelligence per dollar, and the price of AI falls faster than Moore's Law - we get a deflation of costs. Demand rises, waste. (GPUs, chips, data centers) Now trillions of dollars go to infrastructure, and the per-unit cost will fall in the future due to the introduction of these capacities under AI.
3Big vs Small Models: Structure as in the Computer Industry Andreessen describes how top-end models (For example, GPT-5) through 6–12 Months are copied by small models with comparable capability. An example from recent weeks: Kimi (Chinese Open Source Model) It replicates reasoning GPT-5but working on 1–2 MacBook. He predicts that the industry is being structured like a computer - a handful of "god models" (Supercomputers in giant data centers) at the top, a cascade of small models down, up to embedded (in every physical object). Most tasks do not require breakthrough abilities – conventional ones are sufficient.120 IQ, no need for "PhD on string theory" If you’re building a product, you can choose between big/small, open/close. (a16z) Bet on all strategies at once, because it is unclear who will win. Companies like Cursor are using 10–100 parallel models.
4China, Open Source and the Global Race Chinese companies (DeepSeek by hedge fund, Kimi, Qwen by Alibaba, Tencent, Baidu, ByteDance) We have released open-source models at the frontier level, with much lower inputs. This confirms that there is no permanent leadership. xAI/Grok catches up with OpenAI12 months. In China. 3–6 Top AI Companies, Progress Rapidly (despite US sanctions on chips). For engineers, this shows that open-weights changes the rules - you can take Kimi/DeepSeek and deploy locally. (privacy, cost control).
5Regulating AI Mark described the problem that the US is trying to regulate AI at the state level, in his opinion, federal regulation is better, and laws at the state level will lead to fragmentation of the EU. (That'll slow down innovation.). China as a motivator - the U.S. cannot afford to fall behind.
In the sequel, I will finish with Mark’s theses about the future of AI.
#AI #ML #Trends #Software #Engineering #Future