Mary Meeker published 340 AI slides. The models were not the point.

The cost of intelligence was falling. So was the value of simply having access to it.

In May 2025, Mary Meeker and the team at BOND published 340 slides on artificial intelligence. I read all of them. My eyes still hurt. Most summaries focused on the size of the numbers. ChatGPT had reached 800 million weekly users. The Big Six technology companies had reported $212 billion in capital expenditure during 2024. AI job postings were climbing while other technology roles were falling.

The useful story was quieter. AI capability was becoming cheaper, easier to access and harder for any one company to own. That created an opening for startups. It also removed one of their favourite excuses.

The cost collapse

Training frontier models remained brutally expensive. Using them was moving in the opposite direction. The report showed that the price of serving comparable AI models had fallen 99.7 per cent in two years. The energy required to generate a token on NVIDIA hardware had also collapsed over the previous decade. A small company no longer needed to build a model to use powerful AI. It could rent the capability through an API, test an idea quickly and switch providers when the economics changed.

This was good news for builders. It was less comforting for anyone who thought access to a model counted as differentiation.

The report did not show that every model was the same. It showed that performance was converging. Different models still had strengths in reasoning, code, speed, privacy and cost. But for tasks such as summarisation, extraction and classification, a cheaper model was becoming good enough. That changes where the value sits. The model matters. The customer rarely buys the model.

The vertical AI lesson

The most interesting part of the report covered specialised products. Cursor was growing in software development. Harvey was doing the same in legal work. Abridge was spreading through healthcare. Their revenue curves made vertical AI look like the next easy startup formula.

Pick an industry. Add AI. Wait for the money.

Sadly, industries continue to contain customers.

Those companies were not growing because they had put a profession after the letters AI. They were entering expensive workflows where the problem happened frequently and the result could be measured. Developers wanted to produce code faster. Lawyers wanted to reduce repetitive review. Doctors wanted to spend less time writing notes after seeing patients.

The AI was important. The work it removed was the product.

A vertical label does not create demand. You still need access to the customer, enough trust to enter the workflow and a problem painful enough to survive procurement, implementation and somebody asking why the current process is not good enough.

The moat moved somewhere less exciting

Falling costs and converging performance make building easier. They also make copying easier. The report showed open models gaining momentum and Chinese models narrowing parts of the performance gap with their American competitors. The model underneath your product may change several times. A feature that looked rare in January can become a standard API call by June.

The model provider built the road. You rented a faster car. Useful, certainly. A moat, probably not.

The harder advantages sit closer to the customer. Proprietary data you have permission to use. Integrations buried inside the workflow. Distribution competitors cannot buy overnight. Regulatory approval. A reputation for answering when the system produces something strange at 2 am.

This matters most in industries such as finance, healthcare and legal services, where better output must arrive with privacy, accountability and someone willing to put their name behind the decision.

The report also showed AI skills moving rapidly into the labour market. That did not mean every company needed an AI department. It meant every function would need people capable of deciding where AI belonged, checking its work and understanding what should never be handed over.

The numbers in Meeker’s report describe a moment in 2025. Many will date quickly. The commercial consequence will not. As intelligence becomes cheaper, the expensive part moves back to understanding the customer, fitting into the company, earning trust and taking responsibility when the technology fails.

The winners will not be the startups that found a model first. They will be the ones that found a problem customers already pay to make disappear.

See you out there.

Martin