[ Thought Leaders
](https://www.unite.ai/series/thought-leaders/)
[Add Unite.AI to your preferred sources on Google](https://www.google.com/preferences/source?q=unite.ai)
At DVC, we’ve spent a decade watching the AI market up close — backing companies and working closely with the portco founders. This year, we decided to share the framework we’ve built with the world. This is how the State of AI Report was born: a living, continuously updated analysis of the AI economy across every layer, from silicon and energy to foundation models and applications. We update it with AI and review it personally — because in a market that moves this fast, a static annual snapshot is outdated the moment it’s published.
Here are 5 findings both founders and investors should pay attention to.
1. The Rules of Scale Have Been Rewritten #
Never in the history of technology have five-person teams routinely challenged $100 billion incumbents and won, disrupting their business models and offering the market expansion to $500B. Perplexity added 50% to $300M ARR in a month. Higgsfield went from zero to $300M in a year. Cursor hit $500M ARR in June 2025, $1B by fall, $2B by February 2026. None of them built their own factories. They competed at the application layer, expanding their markets tenfold while infrastructure scaled beneath them.
2. The SaaS Playbook Doesn’t Apply Here #
Traditional SaaS valuation rests on one premise: margins of 70–80% that compound as the business scales. AI companies don’t work like this. Even the best ones pay what we call the “inference tax” — 40 to 60 cents of every dollar goes to compute, foundational models, cloud, silicon, and energy. OpenAI’s gross margin in H1 2025 was ~42%. Anthropic’s ~40%. Cursor ~35%. GitHub Copilot is currently 0–15%, subsidized for strategic lock-in.
AI companies are closer to industrial manufacturers than to software vendors. Jensen Huang said it best: think of a datacenter as a factory — electrons go in, intelligence comes out. Your “AI agent for banking AML and compliance” isn’t SaaS. It’s a silverware customization service. The spoons are mass-produced in a factory, from sheet metal, sourced from mills that process ore from mines.
Does this mean paying SaaS multiples is wrong? No. But let’s be honest: we’re paying for extraordinary revenue growth potential, not long-term terminal value. That’s a different bet. We should make it knowingly.
3. Foundation models are becoming a commodity — faster than anyone expected #
Sixteen companies now have frontier-capable models. GPT-4-class performance cost $37.50 per million tokens in 2023. By 2025: $0.14. DeepSeek V3 brought it to $0.006 — a 99.6% decline in two years, and 6,000× cheaper than where this started.
When the model layer becomes a commodity, the advantage shifts to whoever controls access to the user. Distribution, workflow depth, and interface ownership — not model quality — will determine who defines the next decade. The model quality won’t be that important.
4. The Business Model Is Still Being Invented #
There are currently three ways to make money in AI, and none of them is obviously right. You can charge per token — simple, scalable, and racing toward zero; prices dropped 10× in 18 months. You can charge per outcome (per resolved ticket, per completed task), which sounds elegant until you spend a week trying to define what “resolved” actually means. Or you can go the subscription and advertising route. OpenAI launched ads in February 2026 and hit $100M annualized in six weeks, which sounds impressive until you do the math: that’s $0.12 per user per year, while Google makes $60. Anthropic responded by running a Super Bowl ad that mocked the whole idea — Claude jumped to #7 on the App Store the same day.
Nobody has figured this out yet. The company that cracks sustainable monetization for AI-mediated workflows may matter more than the company with the best model. We give it two to three more years before this is resolved.
5. Energy, not chips, is the constraint that defines the decade #
Data center power demand is set to increase 160% by 2030. Google, Amazon, Microsoft, and Meta collectively contracted 10+ gigawatts of new US nuclear capacity in 2025 alone. They’re not doing this for fun. They have no other choice.
The asymmetry: the US has frontier GPUs but is heading for a 44-gigawatt power shortfall by 2028 (Morgan Stanley). China generates twice the electricity of the US, added 543 gigawatts of new capacity in 2024, and is projected to have 400 gigawatts of spare capacity by 2030. The US has chips. China has electrons. Whoever resolves their binding constraint first shapes the decade.
Energy infrastructure doesn’t get built in a quarter. The decisions being made right now — nuclear PPAs, data center siting, grid investments — will determine who’s winning in 2030. We’re not paying nearly enough attention to this.
The full DVC State of AI Report is available at state-of-ai.dvc.ai — and unlike most reports, it’ll still be relevant next month.