Controlling the tokens, the programmable cancer vaccine, and who is doing the learning?

This week saw Stripe enter into an agreement to acquire OpenRouter for a reported $7.5 billion. The deal is a clear step by Stripe to bridge the worlds of payments and AI infrastructure. The figures accompanying the deal explain the scale of the bet. OpenRouter processed over 70 trillion tokens in the week beginning 10 August, and is reportedly reaching an annualised rate of more than 4.5 quadrillion tokens, after three years in which volume has doubled roughly every 11 weeks.

Stripe is also growing quickly. Its investor letter reported revenue growth of 41% during the first half of the year and free cash flow growth of 43%. The company says that 88% of the Forbes AI 50 use Stripe, while the contribution from AI and cryptocurrency companies has more than doubled over the past year. OpenRouter was valued at approximately $1.3 billion during a funding round in May, making the acquisition price a substantial strategic premium.

An investor letter obtained by Axios from Patrick and John Collison explains the excitement. It says Stripe treats 1 January as the start of the "singularity" and has operated on that basis since. The term is used less as a claim that AGI has arrived than as a description of an economic discontinuity. Intelligence is becoming cheaper, more capable and available through software. Agents are beginning to consume services and may eventually hold and exchange money. Stripe believes it should provide the infrastructure on which this new economy is built.

OpenRouter fits into a broader sequence of investments. Metronome provides usage metering and billing. Bridge and Privy provide stablecoin infrastructure and programmable wallets. Tempo is intended to support high-volume settlement. OpenRouter adds model discovery, inference routing and consolidated access. Together, these assets could allow Stripe to route intelligence, measure its consumption, bill for it and settle the resulting transactions.

Supporters see a clear "synergy" with payments. Stripe provides one interface across payment methods, banks, currencies and regulatory environments. OpenRouter provides one interface across hundreds of models and inference providers. It normalises APIs, manages credentials and billing, and can route requests according to price, latency, availability and data policy. These capabilities reduce engineering and procurement friction.

That matters increasingly for agentic systems. An application calling one model can integrate directly with its provider. An agentic system operating continuously across many tasks will need several models, different service tiers and automatic fallbacks. It may also need to change its model selection as prices and capabilities evolve. A "gateway" becomes more valuable as the harnesses and runtimes become more dynamic.

Stripe’s investor letter also paints a bullish picture on long-term global economic growth. It argues that we do not necessarily live in a zero-sum state. Rather, humanity has repeatedly increased the energy, information and productive capability at its disposal. This reminds us of the macrohistorical work of Ian Morris, who examines how energy capture, technology, social organisation and values have evolved together. His development index suggests that continuing historical trends will produce more change during this century than humanity experienced during much of its previous history.

But this perspective sets up some hard challenges ahead. If historical trends continue, then long-term social and economic development cannot follow a stable middle path in which growth simply stops for several generations while existing institutions continue unchanged. Systems that cannot adapt to new productive technologies are displaced by those that can. The trajectory either continues through another technological transition (Stripe’s discontinuity) or encounters constraints severe enough to produce instability and destructive decline. On this basis something resembling a singularity may be less an optional destination than the requirement to continue our long-term development trajectory.

The doubts about this specific OpenRouter acquisition remain. It does not control the production of tokens. Model laboratories control proprietary models, while cloud and inference providers control computing capacity. Customers can use direct provider contracts, supply their own keys or operate alternative gateways. The basic software can be reproduced through open-source and commercial products. OpenRouter’s normalisation makes models easier to mix and match, but it may also make OpenRouter easier to replace.

Its real assets are distribution, aggregated demand, provider relationships and operational data. At sufficient scale, OpenRouter can observe model adoption, provider reliability, price sensitivity and movements in demand. It can potentially negotiate capacity and use its telemetry to improve routing. That could create a marketplace effect, but it has not yet created strong lock-in. A gateway to payments and inference remains a convenience rather than a strategic necessity for many customers.

A more consequential gateway would provide controlled access to internal and external knowledge, data, applications, skills and other capabilities. It would propagate identity, enforce permissions, apply policy and maintain an audit trail across every model and tool invocation. It would solve a wider set of enterprise friction problems than model routing alone. (Such a gateway happens to be one of the current focus areas at ExoBrain.)

OpenRouter has some of the necessary components, including standardised tool calling and support for MCP-based tools. It does not yet provide the complete enterprise context and capability plane. There is no suggestion that it is a malicious intermediary, but it remains an intermediary whose influence depends on customers continuing to route traffic through it.

Takeaways: This acquisition looks less like the completion of Stripe’s strategy than its next stage. For now Stripe is purchasing a large and rapidly growing stream of inference demand, and the option to build more valuable services around it, rather than controlling the tokens themselves. The reported price assumes that this position can become a fixed control point. That is plausible, but it will require Stripe to extend beyond token transfer into identity, knowledge, capability and agent governance, areas where we all must start to focus if we are to provide economically critical agent infrastructure.

On Wednesday Moderna’s share price almost tripled after it announced the first successful Phase 3 trial of a personalised mRNA cancer treatment. Headlines quickly presented the result as an "AI-designed cure". The reality is more complex, but also perhaps just as consequential. Moderna and Merck have reported positive Phase 3 results for intismeran, which is intended to stop melanoma returning after surgery. The trial involved over 1,000 patients and compared the treatment, plus Keytruda, with Keytruda alone. Detailed results have not yet been published, but the combination improved both recurrence-free survival and the time before cancer spread elsewhere.

What is historic here is that each dose of intismeran is made for one patient. Doctors sequence the patient’s tumour and compare it with healthy cells to identify mutations unique to the cancer. This can produce a large number of potential targets. Moderna’s computational system then ranks them according to factors such as whether the mutated gene is active, whether the resulting protein fragment is likely to be displayed by the patient’s immune cells, and whether it is sufficiently different from healthy human proteins.

This is where most of the "AI" sits. The work runs on a specialised bioinformatics and machine-learning pipeline rather than on LLMs. Moderna has discussed using neural networks in vaccine and mRNA design before, and has done so here too. It has not published the architecture, the training data or the current version of the production models behind intismeran.

The software selects up to 34 tumour-specific targets called neoantigens. These are encoded into a single mRNA sequence, manufactured and injected into the patient. Their cells read the mRNA and temporarily produce the selected antigens. This trains the immune system to recognise cancer cells carrying the same mutations. Keytruda then removes one of the controls that can prevent immune cells from attacking the cancer.

The central breakthrough is the programmability of mRNA. Once the delivery and manufacturing platform exists, the informational payload can be changed without developing an entirely new production method. For intismeran, the same platform produces a different molecular instruction for every patient.

The exciting progress here is that molecule design may be starting to resemble software engineering. Researchers define a required function, generate a sequence, predict how it will behave, compile it into a physical product and test the result. The comparison is useful, although biology is much less predictable than a computer. Cells are variable, molecular interactions are difficult to observe, and a sequence that works in one biological environment may fail in another.

A new Anthropic experiment shows where other forms of AI could contribute. Claude was given access to specialist protein-design models, scientific literature and computing resources. It planned design campaigns, selected target sites, generated candidate proteins, evaluated them and chose which should be tested physically. Across the reported work, it produced 1,320 designs and laboratory testing confirmed 354 as binders.

Claude was coordinating the scientific process rather than simulating proteins itself, much as an agentic coding system can plan a software change, use specialist tools, run tests and revise its work. Moderna style predictive models can rank molecular candidates. Generative models can propose new sequences. Reasoning models can increasingly connect those tools into an experimental workflow.

Takeaways: Intismeran should not be reduced to an AI vaccine. Its success combines tumour sequencing, machine learning, mRNA engineering, automated manufacturing and immunotherapy. The deeper development is that medicines are becoming programmable at the level of individual patients. AI will help choose, generate, test and refine those molecular instructions, but mRNA is the wonder technology that allows the programme to be executed inside the body.

This week’s chart shows what happened when 26,811 Chinese pupils aged 12 to 18 were assessed on their educational AI usage. Homework scores rose by 18%, while average completion time fell from 64 minutes to 45. Six months later, exam results were 20% lower than those of pupils who had not used it.

The obvious conclusion is that AI damages learning. We think there is a more subtle one. Homework helps to develop cognitive skills through self-directed effort. Its purpose is not the finished answer alone. The difficulty, repetition and occasional frustration are part of how knowledge develops. When ChatGPT produces the answer, the result will improve while the pupil loses the exercise.

Katherine Rundell recently made a powerful case against AI in education. But these arguments completely miss the opportunity of using AI in so many other ways to support educational exploration. The technology can already be used to create flashcards, simulations and educational games. Even with a small amount of innovation it could start to be used to identify gaps in understanding, adjust the difficulty of questions and encourage a pupil to practise more often.

And what the commentators often forget is that this opportunity matters ever more beyond wealthy countries with good schools, universities and private tutors. Around 70% of children in low and middle-income countries cannot read by age ten. Another 44 million teachers will be needed globally by 2030. Affordable AI could provide explanations, practice and individual support where none is currently available.