The State of Startups, Consumer AI and Mental State for Founders
Answers Are Easy, Questions Are Hard.
Good morning,
Happy New Year, everyone!
In today’s edition, among other things:
The State of Startups
The State of Consumer AI
The Next Mental State for Founders
2026 Predictions, Outlooks & Forecasts
Onwards!
The State of Startups
Two founders could run the same fundraising process in 2025 and come away with opposite conclusions: “fundraising is open” vs. “fundraising is shut.” Both were right. They are just living in different percentiles. 2025 was defined by dispersion: checks concentrate, valuations stretch, and the gap between the median and the top decile becomes the actual constraint.
The market is paying now for right-tail outcomes, and it is pricing everything else as optional.
Here’s Carta with The State of Startups 2025:
“Today’s seed stage: more $, higher valuations, fewer rounds”
That sentence is doing three jobs at once. Here’s how to unpack it so a normal reader can follow:
“More $” means the system is willing to deploy capital.
In 2022, seed cash raised peaks around $10.4B.
In 2024–2025 it sits around $8.5B and $8.1B.
“Fewer rounds” means access is narrowing.
Primary seed rounds fall from roughly 2,438 (2022) to about 1,494 (2025, FY est.).
Put together, you get the what is saw many times over in 2025:
If you are fundable, you may raise faster and larger.
If you are not, you get “soft no’s” indefinitely because the market is conserving tickets.
This is why two founders can look at 2025 and report opposite realities. They are sampling different parts of the distribution
The next step is to think about what’s going on with seed rounds.
The report shows the gap between the top of the market and the middle widening.
The median is moving up, but the ceiling is moving faster.
The market is paying up for “obvious winners,” not “promising companies.”
If you plan your company around median fundraising outcomes, you are planning around the part of the market that is losing relevance.
Now let’s talk about terms. SAFE specifically.
Many founders assume the world got easier because the instrument got standardized.
“Everybody has chosen SAFEs as the default instrument”
Yes. SAFEs dominate. But standardization mainly removed paperwork friction. It did not remove economic variance.
The instrument is standardized, but your outcome is still negotiated through caps, tranches, and timing. If you can’t model dilution ranges quickly, you are operating blind in the one place compounding never forgives.
And next up - time and ESOP.
“Startups are taking longer to go from Seed to Series A”
Median time from seed to Series A rises to about 2.1 years in 2024–2025 (up from roughly 1.4 years in 2017).
Then look at hiring behavior:
Median time to first hire rises to 284 days (with the average up to 476 days).
Early engineer equity remains chunky:
First engineer median grant about 1.54% (range roughly 0.61%–4.13%).
Option pools start around 12% at priced seed and grow with later rounds.
If fundraising takes longer, every “we’ll hire after the round” plan becomes a risk.
The market is pushing you to prove more with fewer people for longer.
The mistake is building a decent company and hoping the market notices. The market is not short on companies. It is short on conviction.
Consumer AI
Consumer AI (report from a16z) looks like a settled market. ChatGPT has 800-900 million weekly active users. Gemini sits at 35% of that on web, 40% on mobile.
But absolute user counts are the wrong metric. The economics of AI aren't driven by users in total. but they're driven by power users.
Look at the growth rates, not the totals. ChatGPT is growing 23% year-over-year on desktop. Gemini is growing 155%. And accelerating.
The report frames this as “potential shift in market dynamics.” That’s underselling it. This is the pattern that precedes market flips:
Incumbent has massive base, slowing growth
Challenger has smaller base, compounding growth
Lines cross faster than anyone expects (usually 18-24 months after the pattern emerges)
Google’s distribution advantage is finally activating. Android users see Gemini at 50% of ChatGPT’s scale versus 35% elsewhere. That’s not product superiority. That’s placement. And placement compounds, hence “code red” at OpenAI,
From the report:
“ChatGPT is the Kleenex of AI... it is the brand.”
I don’t buy the Kleenex framing. Kleenex won a commodity category where the product is identical. AI assistants are not identical. They have different personalities, different strengths, different failure modes. Brand matters less when the product experience diverges. And it’s diverging fast.
The stickiest number in the report: fewer than 10% of ChatGPT users even visited a rival LLM provider in 2025. Only 9% of consumers pay for more than one service.
This looks like lock-in. It’s not. It’s inertia.
Lock-in requires switching costs. What are ChatGPT’s switching costs?
Data? Your conversation history isn’t portable, but it’s also not that valuable. Most people don’t revisit old chats.
Workflow integration? Minimal for consumer use cases. You’re not rebuilding infrastructure.
Learning curve? Near zero. Every chat interface works the same way.
The “loyalty” is just default behavior. Defaults are powerful but fragile. They break when:
Something visibly better appears (viral creative moment)
External force changes the default (enterprise mandate)
Distribution shifts (Android/Chrome integration)
All three are happening simultaneously for Google.
Creative tools are the only proven mechanism for breaking default behavior at scale.
“Best-in-class creative models generate ‘nearly infinite demand,’ creating powerful viral loops that can pull users from an entrenched leader, even for a single-purpose task.”
The pattern is specific:
Launch state-of-the-art creative capability (Nano Banana Pro, VEO, Sora)
Output spreads on social platforms (the content markets itself)
Users download app or visit site they’d never have sought
Even if they return to ChatGPT for general tasks, you’ve established a beachhead
Google’s Nano Banana Pro went “insanely viral,” comparable to OpenAI’s Ghibli moment. This matters because creative virality doesn’t require users to decide to switch. They just want to make the thing they saw. The switching happens as a side effect.
The vulnerability in ChatGPT’s position isn’t competitors building better chatbots but competitors building better creative tools that happen to come with chatbots.
Here's where the report gets interesting. The incumbent labs face constraints that aren't temporary or fixable but architectural.
“The ‘promo committee’ culture incentivizes PMs to build safe, incremental features that extend core metrics, not risky, opinionated new products that could fail.”
Three constraints bind the labs:
Organizational inertia. Big company PMs get promoted for shipping features that move existing metrics. They don’t get promoted for launching new products that might fail. This selects for incrementalism. OpenAI’s “Everything App” strategy (Pulse, group chats, shopping, research tasks all inside ChatGPT) is a symptom: it’s easier to add features to the core product than to launch risky standalone bets.
Compute scarcity. Every GPU running inference for viral consumer features is a GPU not training next-generation models. Labs face this tradeoff constantly. Startups at the application layer don’t. They can consume inference from whoever offers the best price/performance and spend zero cycles on training.
First-party model lock. Labs will always prioritize their own models. Startups can be multi-model, choosing the best engine for each task. When Claude is better at writing and GPT-4 is better at code and Gemini is better at multimodal, the multi-model startup wins on user experience. The lab is stuck with their own model’s weaknesses.
This is why Perplexity’s Comet browser achieved higher sustained traffic than ChatGPT’s Atlas browser despite massive distribution disadvantage. Opinionated product beats default product when the builder isn’t constrained.
The most important slide in the deck is page 12. It describes something that wasn't supposed to be possible in consumer software.
“For the first time ever, consumer software products are achieving over 100% net revenue retention.”
Traditional consumer SaaS struggles to hit 70-80% annual retention. Users churn. The ones who stay don’t expand. You’re always backfilling.
AI products are different. Usage-based pricing (pay-per-token, credits on top of subscription) means power users can 10x their spending over time:
Casual users churn at normal rates
Power users don’t just retain, they expand
Expansion from power users exceeds revenue lost to casual churn
Net retention crosses 100%
You don’t need to win the mass market. You need to win the users who will spend $50/month, then $100, then $200.
This is why Claude’s position (3x fewer US teen users than Character.ai, but beloved by technical professionals) might be stronger than it looks. The report calls it “confinement to the tech bubble.” I’d call it “concentration on high-LTV users.”
The power user thesis makes niche strategy not just viable but potentially superior to broad reach.
The report claims the enterprise flywheel will lock in consumer behavior:
“As users are required to use ChatGPT for work, it may solidify its position as their default personal AI, locking in consumer behavior.”
I agree, but it’s still unsupported with data. Google is a major player in enterprise, so the most probable winner (?).
The report offers no data on:
What percentage of ChatGPT consumer users also have enterprise access
Whether enterprise-to-consumer spillover actually happens
How this compares to Google’s workspace integration (which could work the same way for Gemini)
The 8-9x YoY growth in enterprise usage is real. The claim that it creates consumer lock-in is speculation. Worth watching, but not proven.
Nine percent of consumers pay for multiple AI services. The conventional read is “winner-take-all.” The contrarian read is “specialization.”
Winner-take-all assumes AI is one market. It’s not. It’s dozens of markets wearing similar interfaces:
Writing assistance (Claude’s strength)
Research and citations (Perplexity’s lane)
Creative generation (Gemini’s viral wedge)
Code assistance (Cursor’s position)
Workflow automation (still open)
The power users in each vertical don’t need an “Everything App.” They need the best tool for their specific task, and they’ll pay for it separately.
ChatGPT is the default. Defaults matter. But the economics of AI reward depth over breadth, obsession over adoption, power users over everyone else. The 9% who multi-home today are the leading indicator, not the laggards.
The Next Mental State for Founders
From NFX:
People like to say that AI is making it possible for everyone to be a founder. That is somewhat true – it’s never been technically easier to build a product than it is today. But psychologically, not everyone has what it takes to succeed in this agentic era of 1,000 experiments.
The few people you do hire at your company matter more than ever. They must be entrepreneurs within your organization. They need to be thinking like someone who has the resources of a 1,000 person team at their disposal.
These are the traits that lead to success in that arena:
High agency: They don’t wait for instructions. They take responsibility and act.
Multi-domain fluency: They’re able to think marketing, operations, and engineering all at once. They refuse to stay in one lane.
Builder’s instinct: They make things with their own hands. If something breaks, they fix it. If something’s missing, they prototype it. They have a bias toward creation. It doesn’t mean they all need to write code – a lot of the agent creation and implementation will be done through prompting – but they need to know how to create and manage agents.
Comfort with chaos: Agents fail, models drift, things break. They stay calm, and adapt.
Truth-seeking over status-seeking: They share information, including failures. Everything is public and out in the open.
Low ego: experimentation is failure. They are able to separate themselves psychologically from the product.
..but not low self-worth: if you believe you are destined to fail you will handicap yourself before you even start. They have confidence to move forward in the first place, without fear of failure.
Internal locus of control: They operate as if everything is their responsibility, even things they don’t formally own. If something is broken, if someone is stuck – they step in.
Systems intuition: They sense when a workflow, team, or agent network feels “off.” They can diagnose bottlenecks instinctively.
Willingness to unlearn: Old intuitions about “what takes time” or “what requires a team” are now liabilities. They can abandon outdated mental models instantly and adopt new ones when the world shifts.
Interesting Analysis and Trends
2026 Predictions, Outlooks & Forecasts
2026 Investor Predictions (Insight Partners) LINK
2026 Predictions (7wire Ventures) LINK
2026 Predictions from the M13 Constellation LINK
26 Predictions for 2026 (Equal VC) LINK
12 Gaming Predictions for 2026 LINK
The 26 Most Important Ideas for 2026 LINK
Stanford AI Experts Predict What Will Happen in 2026 LINK
AI in 2026: A Tale of Two AIs LINK
12 Predictions for 2026 (Aashay) LINK
2026 Predictions (Tunguz) LINK
Our Predictions for Security in 2026 (BCV) LINK
12 Outlooks for the Future (2026) LINK
AI, SaaS & Platform Shifts
AI Agents Are Starting to Eat SaaS LINK
AI Agents Are Starting to Eat SaaS (alt link) LINK
State of LLMs 2025 LINK
GTM, Sales & Growth
Post-Sales GTM: Following Customers LINK
An Outbound Playbook for 2025 LINK
AEO: Golden Age or Blood Bath? LINK
Markets, Strategy & Decision-Making
Don’t Confuse Market Share with Winning LINK
Being Right vs Getting It Right LINK
The Two Kinds of Pivots LINK
Energy, Infrastructure & Healthcare
Year-in-Review
Year in Review 2025 (Karpathy) LINK
Meditations
Life, February 1966:
I don’t explain, I explore.
Thank you for your time,
Bartek