This year AI became something new: as our team and much of our portfolio moved further into living inside AI tools day to day, we noticed that our work lives are starting to feel… metered. We expect more of us will start to recognize this “taxi meter” sense about AI.

That meter runs in three currencies: the time it costs to stay fluent, the dollars that pay for all that AI (or, if you prefer, the tokens), and the feeling that AI may be changing us somehow. We love AI, and, like some loves, it’s complicated and intense. We believe that people who see these meters running may be more capable of harnessing a healthy relationship with AI.

We tried something different this year: we wrote a (Claude) skill that read our past letters and the work we’ve done over the year to write the initial draft of this letter. To share the fare on this letter’s meter, she ultimately spent $65.56 in tokens, with three hours of hands-on work, 23 hours and 21 minutes of computer time running the still-state-of-the-art-as-of-this-typing Fable 5 model, and then another dozen hours of the team typing, discussing, and editing that draft.

Here is our collective take.

The Personal Meter: The Tax of Keeping Up

Every few months, we relearn what tools to use. To stay current, we keep our hands on the keyboard.

AI seems to live rent-free in many of our heads. Roy Bahat, head of Bloomberg Beta, calls this feeling AI-DHD: the low-grade overwhelm of tracking it all; the “will it or won’t it work” slot-machine pull of prompting an agent and seeing what happens (including, for others, hacking out of a supposedly-secure environment); the fact that instead of doing two or three things at once, we can now do a dozen. Inside the AI companies themselves, teams ship so much, so fast, that people at those labs describe the same scramble, just to keep up with their own colleagues.

James Cham, a Beta founding partner, thinks this, like everything, is a “diffusion of knowledge” problem: the models move faster than expertise in how to use them. This is why people say the same words (“agents”, “software factories”, “harnesses”) while meaning very different things. Real understanding requires longer conversations.

Karin Klein, also a Beta founding partner, sees this moment as an opportunity to rethink learning: “Education isn’t K-12 anymore. It’s K-100.” Her advice for keeping up is experimentation over perfection: everyone should try vibe coding, not necessarily to become a programmer, but to understand what AI now makes possible.

The Dollar Meter: Getting Better is Getting Expensive

Software was supposed to cost nothing to reproduce. It was the one industry with no raw materials. Except now that lines of code are abundant, the raw materials (chips, memory, and the electricity to run them) are scarce. Some CFOs now sign off on compute the way they once approved headcount.

The price of the memory chips that feed AI data centers roughly quadrupled in a few months at the end of last year, and some analysts price memory like they price oil. We are adjusting to a world where the bottleneck on knowledge work is electricity and GPUs rather than people.

The first “dark” software factory (where no humans participate in the work, so the lights are “off”) came from a Bloomberg Beta-backed founder. In Justin McCarthy’s software factory, which he now calls Diffusion, humans don’t write or read any code. Instead, they “manage” by writing the specifications and judging the results, spending on tokens rather than salary. And Replit, the vibe coding startup we backed, now believes we may be going from self-driving software to a self-driving company.

We also think about AI-made software differently depending on whether it is OK for it to fail. If we’re preparing for a meeting, it’s OK if the AI “can make mistakes.” On the other hand, if AI, for example deployed by our portfolio company Air Space Intelligence, rewrites our nation’s air traffic control system, then the stakes are much higher.

Who will actually pay for all of this, who will benefit, and will the rising bill slow the spread of AI into ordinary work? Will AI’s boon become concentrated among the people and companies who can afford the meter? Without a more thoughtful system, will AI exacerbate the distortions of power that already drive our society, government, and economy?

The Feelings Meter: the Intelligence Might Be Artificial, But the “Ick” is Real

As much as many AI builders are euphoric, others who use AI at work, often because their leaders mandated it, say that reading and generating AI slop makes them act like middle managers of a middling workforce.

AI needs more than “better messaging” so people understand the benefits—it has to actually deliver. Adoption has to reach far beyond the people already excited about it. Karin has been convening AI ”talking circles” of women leaders with Gloria Steinem to explore how AI can be adopted in more human, useful ways. We’ve seen how AI is as much about leadership and culture as it is about technology, and the people affected by AI need a greater voice in shaping it.

The companies that we want to see win the next decade will build AI nobody feels gross about.

Founders we backed are already building that kind of AI: Intuition Robotics puts a small robot named ElliQ into the homes of older adults living alone; The Daily spent an episode with an 85-year-old widow on a remote Washington peninsula whose memory-test scores improved and who calls their robot “the best roommate I could have ever asked for.” Vals AI built the first benchmark testing whether AI can help the 40 million Americans on food assistance navigate SNAP benefits.

A Few Victories

Thirteen years in, we’re most proud of what’s stayed the same: we continue to see founders as our customers, and recognize their work is so much more difficult than ours. We invest from a fund with the same size, same strategy, same three equal partners, and same backer in Bloomberg LP.

Since we started:

Founders we’ve backed have gone on to raise more than $14.1 billion

We’ve backed 10 unicorns (Replit, Lambda Labs, Shield AI, and MasterClass, to name a few).

In the first major round after we invest, a Bloomberg Beta-backed startup is worth $24 million more than the average startup.

We welcome any founders you send our way, especially the ones building the redeemed version of AI: products people trust enough to pay for because they genuinely make work better, not just busier.

— The Bloomberg Beta Team