Late last week, analyst and investor Mary Meeker – formerly of Morgan Stanley and Kleiner Perkins, now founder and general partner at venture capital firm BOND (Bond Capital) – dropped her first trends report since 2019. First released in 1995, Meeker’s lengthy fact-filled reports became bibles to industry watchers, and eventually a much-anticipated annual tradition in which reports as long as 350+ pages were delivered by Meeker as rapid-fire, on-stage presentations in roughly 30 min. The hiatus after 2019 came during the early days of Meeker’s departure from Kleiner Perkins and founding of BOND as a new VC firm (late 2018).

The just-released report is about 340 pages and focuses on artificial intelligence. The report highlights the growth of AI, particularly in the realms of global user/developer adoption, capex spend, and losses among AI players, despite falling inference costs. The growth has spurred momentum in adjacent arenas such as open models and physical AI (e.g. robotaxis, robots), as well as rivalry from China and tumult in the IT labor market.

Rather than recap the themes (many of which we’ve covered in prior briefs) in exhaustive detail, we highlight here what we consider to be the most interesting. (Note: We actually read the deck and wrote this brief ourselves, rather than running it through an AI model.)

The top 10 things that AI will likely do in 5 years (per ChatGPT): (1) Generate human-level text, code, and logic; (2) create full-length films and games; (3) understand and speak like a human; (4) power advanced personal assistants; (5) operate humanlike robots; (6) run autonomous customer service and sales; (7) personalize entire digital lives; (8) build and run autonomous businesses; (9) drive autonomous discovery in science; (10) collaborate creatively like a partner. While there’s certainly room here for caveats and qualms (e.g. human oversight, won’t outperform best-in-the-world humans), perhaps the most interesting aspect of these predictions is that they don’t seem very outlandish given where we are today.

Computing-related patents are exploding. There were 6,000 more computing-related patents granted in 2024 vs. 2023, in a marked near-vertical inflection point that mirrors the explosion in computing-related patents during the dot-com era (1995-2003). US patent applications related to generative AI also rose 56% in 2024. While AI still cannot hold a patent, humans using AI and developing AI systems can certainly do so. In particular, large AI players – led by Google – are investing in patent applications to protect their intellectual property (IP) and build defensive portfolios.

It’s easy to underestimate how much better AI has gotten, especially if your reference points are months old. Meeker’s report points out how much better AI models have gotten in simulating humans in conversation, in generating realistic images, in generating audio voices, and in translating voices into other languages.

In terms of consumer technology adoption, OpenAI’s ChatGPT dominates the field. A Jan 2025 survey found that 52% of US adults had used a large language model like ChatGPT, Gemini, Claude, and Copilot. ChatGPT was the most popular among the AI users, reaching 72% (followed by Gemini at 50%). As of Apr 2025, ChatGPT had an estimated 800M+ weekly active users (WAU), with 10%+ of the world using OpenAI's systems. ChatGPT’s 5.1B monthly website visits is far ahead of rivals such as DeepSeek, Grok, Perplexity, and Claude, which are all still below 500M monthly visits. ChatGPT continues to see rising engagement (about 19 min/day) and retention (75%+ of weekly users retained) over time. (Note this is well below the engagement of a social platform like TikTok, which sees an average of 108 min/day.)

Enterprises are more focused on AI for growth and revenue vs. reducing costs. In a survey, the top 4 improvements targeted were (1) production/output, (2) customer service, (3) sales productivity, and (4) revenues. Bringing AI into customer-facing interactions can have both revenue-enhancing and cost-reduction implications. Bank of America’s Erica concierge built into its mobile app, for example, has already handled 2.5B cumulative interactions for 40M+ clients.

A growing number of AI-enabled medical devices have been approved by the US Food & Drug Administration (FDA). As of Sep 2024, 1,016 AI-enabled medical devices had been approved in total, up from 624 at the end of 2022. Earlier this week, the FDA launched an AI tool called Elsa, which will be rolled out agency-wide by the end of Jun 2025 and is already being used to accelerate clinical protocol reviews, shorten the timeline for scientific evaluations, and identify high-priority inspection targets.

Meeker’s report points out the differences between capex and opex in AI: “CapEx is driven by land, power provisioning, chips, and cooling infrastructure – especially as AI workloads push thermal and power limits far beyond traditional enterprise compute. OpEx, by contrast, is dominated by energy costs and systems maintenance, particularly for high-density training clusters that operate near constant load…[D]ata centers aren’t just physical assets – they are strategic infrastructure nodes. They sit at the intersection of real estate, power, logistics, compute, and software monetization.”

Tracing the bottlenecks in the data-center buildout leads directly to electricity: “There is no AI without energy – specifically electricity…Globally, data centre electricity consumption has grown by around 12% per year since 2017, more than four times faster than the rate of total electricity consumption. As power demand rises, so too does its concentration: The United States accounted for…[45% of global data centre electricity consumption], followed by China (25%) and Europe (15%)…nearly half of data centre capacity in the United States is in five regional clusters.”

At the same time, inference costs have fallen shockingly fast. On a per-token basis, the cost to serve a model has fallen 99.7% in just two years (2022-2024). While technology costs often fall as user adoption grows, the pace of these twin dynamics has been much faster than prior technologies (e.g. electric power, computer memory).

Developers are rapidly pivoting towards AI. Surveys indicate that 3/4 of developers are now using AI in their work – perhaps not surprising given that AI job postings are up 448% over 7 years while non-AI IT jobs are down 9%. Nvidia’s computing ecosystem, as a proxy (Nvidia’s GPUs and CUDA programming model are still in wide use among AI developers), grew by 2.4x in developers and 3.9x in AI startups between 2021-2025. According to Meeker’s report, “What’s emerging is a flywheel of developer-led infrastructure growth. As more developers build AI-native apps, they also create tools, wrappers and libraries that make it easier for others to follow. New front-end frameworks, embedding pipelines, model routers, vector databases, and serving layers are multiplying at an accelerating rate. Each wave of developer activity reduces the friction for the next.”

Google’s AI ecosystem has been gaining traction among developers. It reported processing 480T+ tokens/month in May 2025 – up 50x from the year prior. For comparison, Microsoft reported processing 100T+ tokens over a full quarter (Q3 2025) – which was up “only” 5x from the year prior.

Big tech’s long-awaited AI chips are here and gaining adoption. While Nvidia (which recently reported 69% year-over-year growth for Q1 FY26) isn’t going anywhere, sales of Google’s TPUs (est. $8B+ in 2024) and Amazon’s Trainium chips (est. $3.6B in 2025) are ramping up.

China is and will be a major AI player on the world stage. Despite US efforts to slow its progress down with chip sanctions and visa revocations, China has been inexorably pushing forward in chips, energy, and AI models. (People seem to forget that China has been in the AI game for more than a decade at this point.)

AI is entering the physical world, and it’s not just robotaxis and humanoid robots. Examples in Meeker’s report also include autonomous defense systems, mining exploration, and agricultural equipment (e.g. weeders, livestock collars).

The urban-rural divide in connectivity – and more broadly, the divide between who is equipped to capitalize on AI vs. not – could mean that AI may end up increasing inequality. Certainly, as Meeker points out, AI will be an enormous wealth generator but who will that wealth accrue to? It’s more likely to be people who are investors or already wealthy, who own businesses or found startups, who are developers or knowledge workers, and who are early technology adopters (due to the pace of change).

If we had one critique of Meeker’s trends report (which will show our bias), it would be that the scale of the report and the multi-year nature of its charts sometimes neglect the granularity in what is changing in AI from month-to-month and week-to-week. For instance, who cares about the growth in the number of large models between 2020 and 2024? There were relatively few before and many more today, and what’s more interesting now are the changes over the past month (or since last week).

On the other hand, Meeker’s broader vantage does reinforce a few pertinent points – that we’re in the early stages of AI, that things are moving more quickly than we can easily grasp, and that many of these AI players already have big businesses (if not yet profitable). OpenAI had $3.7B in revenue in 2024, Anthropic recently reached $3B in annual recurring revenue (ARR), and Perplexity now has $120M in ARR, and there’s also the big tech firms bringing AI to already-large user bases.

If we were to look back at prior eras as analogs, we’d find that, while picking the exact winners would have been tough, drawing a circle that includes many of the eventual winners might not have been all that hard. Of course, you’d want to keep redrawing that circle in response to how the market is shifting, while having the grit to hang onto the winners – which might be the hardest part.

May 30 2025 (3 Shifts): “Vibe coding” and personal apps

Apr 18 2025 (3 Shifts): The growth of AI big tech and $200-per-month subscriptions