In The Loop Episode 19 | Mary Meeker AI Trends 2025: Three Reasons Why AI Is Different From Any Other Tech In History

After a six-year hiatus, the famous analyst Mary Meeker is back with her first trends report since 2019, and this time, it's entirely focused on AI.

For those unfamiliar with Mary Meeker, think of her as the oracle of tech trends. Starting as a Wall Street analyst in the mid-90s, her annual Internet trends reports became the must-read document for anyone trying to understand where technology was heading. These weren't just reports—they were cultural moments that shaped how investors, entrepreneurs, and tech leaders understood the digital revolution.

Now, after six years of silence, she's returned with a singular focus: artificial intelligence.

Let me walk you through the three key ways this report demonstrates that AI is distinct from any technology that came before it—and offer you a candid assessment of both the report and a significant issue that most people are not yet aware of.

Get Mary Meeker’s AI Trends report on Bondcap. (All images in this post are sourced from this report.)

This is In The Loop with Jack Houghton. I hope you enjoy the show.

‍

‍

Reason #1: Speed of growth

The first way AI differs from past tech trends is speed.

Meeker explicitly draws analogies to past platform shifts, citing Vint Cerf's 1999 observation that "a year in the Internet business is like a dog year," then noting that AI's pace is even more extreme.

Technology compounds and AI is building on everything that came before it to move faster. Since 2010, we've seen 260% annual growth in data to train AI models and 360% in compute power. This has led to a 167% growth in the number of powerful AI models over the last four years, which is now driving unprecedented uptake.

The speed of adoption across different technology eras tells the whole story. The PC era took 20 years to reach 50% adoption in US households. The desktop internet era took 12 years. Mobile internet took six years. The AI era? About three years. Each cycle is roughly half the time of the previous one.

And you can see this with ChatGPT, which reached 365 billion annual searches….5.5 times faster than Google.

The recurring theme throughout this report is that ChatGPT is utterly dominant. Its growth is unprecedented and it is currently the AI product to beat. The report shows ChatGPT's subscriber and revenue growth is exponential at this stage, and ChatGPT has hit this milestone in record time.

The numbers are staggering. ChatGPT now has over 800 million active users and 20 million paid subscribers, with revenue now approaching $4 billion.

The growth curve has gone exponential, and what's interesting is the demographics. While the majority of users are between 18 and 49, 20% of people aged 65 and older say they've used ChatGPT at least once. This isn't just a young person's technology. Older adults are using it as a Google replacement, while younger users treat it more like a life advisor.

We discussed the most popular use cases of AI in a recent episode: in the top spot is now Therapy & Companionship. People use AI as a therapist, life coach, or advisor.

And people aren't just trying it once and walking away. Daily usage has nearly doubled, growing from less than 10 minutes per day to close to 20 minutes daily.

This is fundamentally changing how we interact with technology. We're moving from clicking and typing to conversational interfaces, and this shift could drive people away from traditional platforms like social media, Netflix, and YouTube.

As a bit of a proxy to show how much developer involvement and engagement is going on in the developer ecosystem, Google saw their AI developer community grow from 1.4 million to 7 million in just one year between May 2024 and May 2025—that's a 5x increase in 12 months.

Performance is also accelerating at breakneck speed. AI models are becoming so sophisticated that as of February this year, using GPT-4.5, AI systems are consistently fooling humans into thinking they're chatting with another human. Previously, the human win rate was well over 70%. Now, AI has crossed the Turing Test threshold in practical terms.

Meeker equates this speed of adoption and scale of investment with profound impact, repeatedly using the term "unprecedented" to drive the point.

In parallel, the report shows how the speed of AI-related infrastructure investments: the "Big Six" tech companies' capital expenditures hit $212 billion in 2024, up 63% year-over-year, far above historical norms.

‍

This brings us to one of the most fascinating aspects of AI right now: the huge cost reduction that is happening in real time. AI inference costs have dropped 99% over the past two years. What used to cost pounds now costs pennies, and what costs pennies may soon cost fractions of a penny to run these models.

To understand the speed of this reduction, it took around 80 years for a light bulb to become very cheap to use versus about a year for some AI models to achieve similar cost accessibility.

This cost reduction is democratizing AI access—which will increase usage and adoption worldwide.

Another area of speed highlighted in the report is adoption at the Enterprise level. This isn't a consumer trend that slowly trickles up to enterprise—everyone jumped in at once.

The proportion of S&P 500 companies mentioning AI during earnings calls grew from around 10% at ChatGPT's launch to over 50% now. A 2024 Morgan Stanley survey found that 75% of global CMOs were already running AI tests or experiments, with virtually everyone else planning to start within 12 months.

‍

‍

Reason #2: Global from Day 1

Another big way AI differs is its immediate global reach. Previous tech trends started in the US and slowly spread worldwide. The internet took 23 years for 90% of users to be outside North America. ChatGPT? Just three years.

. The top country using ChatGPT isn't the United States - it's India, representing 13.5% of usage compared to the US's 8.9%. After that comes Indonesia and Brazil, both with over 5%, followed by Egypt with nearly 4%.

Geopolitical implications are baked in from the start. This isn't a technology that developed in a vacuum and later became geopolitically relevant===the US-China AI competition is defining the entire landscape.

Geopolitically, Meeker implies a model reminiscent of a tech "arms race" in its intensity. One of her key takeaways: "The reality is AI leadership could beget geopolitical leadership—and not vice-versa." Essentially, the nation or company that leads in AI could gain a major strategic edge.

The report backs this by comparing progress across countries. For example, China is portrayed as surging ahead in certain areas: by 2025, China had released three major large-scale open-source models, more than any other country. A Chinese model suite called DeepSeek went from 0% to 21% of global LLM user share in a matter of months, highlighting how quickly China is mobilizing its huge user base. She also notes that Alibaba's open-source model Qwen 2.5 outperformed some Western models on key benchmarks.

The data is stark: the total number of large-scale AI systems in the US and China absolutely dwarfs the rest of the world combined.

China has a significant advantage in embodied AI - AI embedded in robots. China has more industrial robots installed right now than the rest of the world combined, while the US lags embarrassingly far behind in this dimension.

‍

‍

Reason #3: Breadth of disruption

Perhaps the most significant way AI differs is the sheer scope of what's being disrupted. Mary Meeker isn't just saying things are changing faster—she's saying everything is changing.

The report predicts that by 2035, AI will be conducting scientific research, coordinating global logistics systems, and performing complex physical tasks like assembling components. This paints a picture of AI disrupting virtually every sector of the economy simultaneously.

We're already seeing new AI companies growing at unprecedented rates. Take Lovable, featured in the report—their annual recurring revenue went from essentially zero in December 2024 to $50 million by May 2025. These aren't incremental improvements; these are companies creating entirely new categories almost overnight.

The competitive dynamics are unlike anything we've seen before. As Mary Meeker puts it:

Throughout the report, she sprinkles examples of AI moving from research into the real world. She cites deployments in domains like transportation, industry, and healthcare.

For instance, one chart shows autonomous vehicles carving out a share of the ride-hailing market in San Francisco: an "Autonomous Taxi" service achieved a measurable ride-share market share in San Fransisco against traditional ride-hailing. This is presented not as a sci-fi experiment but as a competitive reality in a major city.

‍

‍

My take on the AI trends report

Looking at Meeker's comprehensive report, I think she nails the macro story but there's a significant gap between the 30,000-foot view and what's actually happening on the ground when companies try to extract value from AI. Let me walk through where I see the biggest disconnects.

The capital deployment story is, of course, fascinating. Meeker documents Big Tech spending $212 billion on AI in 2024 alone, which is staggering by any measure. But when you follow the money trail, so the ways people are making money, back to P&L impact, the results are very uneven. The wins are absolutely real, but they're specific rather than broad-based.

Take Stripe's fraud detection system—they have an assistant that analyses payment patterns and jumped their catch rate from 57% to 97% practically overnight. That's not incremental improvement, that's a big change that flows directly to the bottom line. Ive done episodes on these a couple of weeks ago.

But notice what these successful implementations have in common: they're B2B focused, deeply integrated into specific industry or customer workflows, and solving very particular problems with clear, measurable outcomes. These aren't general-purpose AI assistants trying to be the consumer of everything-to-everyone chatbot. They're purpose-built and designed to slot into existing processes where the value proposition is obvious and quantifiable.

The contrast with consumer AI is stark. Two years after ChatGPT's explosive debut, what do we have in the consumer space? We have ChatGPT and ChatGPT Plus. Despite thousands of startups and billions in venture funding chasing the next big consumer AI breakthrough, the pull of the general-purpose tools has proven enormous. Most specialized consumer AI applications have struggled to find sustainable footing.

This connects to a broader issue with how people interpret AI capabilities. Every quarter brings huge headlines about AI achieving "human-level performance" on some new benchmark—GPT-4o beats humans at task X. The benchmark improvements are genuine, and the models really are getting better rapidly. But there's often a massive gap between controlled benchmark performance and real-world business value.

I think of it as a translation problem. Benchmarks measure performance per token in carefully controlled conditions with clean data and clear objectives. But businesses have messy systems, legacy constraints, and operational friction.

For programmers, if you are using an AI tool on a new product or perfect code base, Cursor users genuinely experience 10x productivity gains. But on a multi-million line enterprise code base, it doesn’t work like that. The model hits a wall.

This pattern repeats across all work domains. Feed a frontier model a perfectly formatted PDF, and it will extract all the information easily. But hand it a folder of chaotic documents and random file types with poorly scanned images,  and everything breaks quickly.

The AI agent phenomenon illustrates this gap perfectly. Meeker notes that Google searches for "AI agent" climbed over 1,000% in early 2025, and the vision driving this interest is undeniably compelling—autonomous AI systems that can plan and execute complex multi-step workflows without constant human oversight. Point a frontier model at your business process, let it autonomously orchestrate everything, and return to find everything completed. In very specific, constrained scenarios, this vision actually works.  

But many, and I’d argue most, successful AI agent deployments share three critical characteristics that the broader hype tends to overlook:

1. The scope is extremely tight—draft one specific type of note, categorise one type of customer inquiry, etc.
2. The inputs are highly predictable—medical notes or customer support chats.
3. The risk of a mistake is well managed—there is some form of human oversight or ability to audit and a human to be involved.

The moment you step outside this carefully constructed comfort zone, AI agents will become exponentially more fragile. Complex memory across multiple interactions is not easy. Multi-tool orchestration is a huge development challenge to overcome that must be maintained. Despite cool-looking videos, a production agent system still requires human operators who can interpret errors and data and make improvements—sometimes early in the morning.

What I’m saying is that the real story isn't about the technology curves or benchmarks—it's about implementation specifics that make the difference between videos or demos and reliable business tools.

When AI capabilities become increasingly commoditized and competitive features get replicated within weeks rather than months, the durable competitive advantages start looking surprisingly traditional—direct relationships with budget holders and decision-makers, contractual renewal cycles that create switching costs, and brand trust that signals reliability and attention to detail rather than flashy innovation.

Now, that was a chunky episode, I hope you found it interesting. Please share this post with someone who might enjoy it. I will see you next week!