Newsletter

Ever rued the thought: "why should I pay for that? I can do it myself?".

My plan to load Reginald onto his own Raspberry Pi computer, rather than host him in the cloud, has fallen afoul of that DIYer's curse.

First, in a ham-fisted attempt to add a cooling fan to the Pi, I broke it. Only a little bit, but enough to mean my to-do list on this project now includes soldering. Something which I don't claim proficiency at and definitely can't delegate to Claude.

Second, the parts-bin hard drive I was planning to install is incompatible with the Pi. This is a straight-up hardware issue. The only fix: I'll buy a new one.

I had a shock when I looked at the prices. I have been tinkering with hardware for 30+ years; some of you have probably been doing so for longer. A side effect of this AI boom is that it's starting to dampen my DIY spirit, due to rising costs. Just look at the price of memory over time (PC Part Picker Price Trends Graph). Before November 2025, I'd recycled, returned or resold parts that would now be worth multiples of what I first paid for them, due to severe supply constraints (Wikipedia). This is a result of a surge in demand from AI datacentres.

Oh well. Reginald really needs that new home. Today I will get my wallet out and order the part I need.

I am far from the only one to find this building-boom irksome. There's a legitimate concern about impact on the affordability of essential utilities too. For example, New York has put a hold on new AI datacentres (NY Times), citing the excessive thirst for energy and water. These resources are needed for running and cooling those powerful AI chips.

The tech giants are on it, with improvements promised over time. Case in point: Nvidia claims that its new cooling system can use less energy by running 'hotter than a hot-tub' (Tom's Hardware). But the fact remains that the negative externalities of this AI era are making themselves felt right now... macro and micro.

Legal tech & tokenomics

How does this relate to Legal Tech? Cost. Welcome to the world of 'tokenomics'.

The word itself was coined with the development of cryptocurrency, with the idea that intangible tokens (like those stored on a blockchain) can – through simulated scarcity – hold economic value. (Canonical example: Bitcoin). In other words, the tech world has got used to the idea that tokens = value.

Separately, the word token has also been used to describe the mapping between snippets of text (usually words) and the numerical values that large language models (LLMs) store and process. Inputs into LLMs – like your prompt and source material – are broken up into tokens by a 'tokeniser' first. Out of the LLM come different tokens, which are reassembled into the answers you see. Have a play here: OpenAI's Tokenizer demo ... change between Text and Token IDs to see how the mapping works. (Grant Sanderson's YouTube videos are helpful if you want to go deeper into how LLMs work.)

The number of tokens input and output has a somewhat direct relationship to the complexity of the request, the corresponding time taken to process it by the LLM, and therefore the economic impact of using all of that hardware, power, water, etc. As a result, it has become convenient to price AI in terms of tokens used, and - at least in the world of tech – this feels pretty natural.

In the legal world, variable costs can be a bête noire of the harried GC. Legal spend is already hard enough to manage, thanks to the seeming invincibility of the billable hour. Effect? Insightful advice and momentous transactions tainted by surprise bills and frustrated Finance folk. Legal technology like Persuit (Vendor website) has already begun to help manage this, but there's room for improvement.

For a little while, AI Legal Tech pricing gave some respite, with major AI vendors choosing to price per seat on a broadly 'all-you-can-eat' basis. Price-certainty aids adoption of new technology.

But the sands are shifting. For example:

• Frontier vendors: For the first time, Anthropic's most powerful new model Fable 5 – ideal for Claude for Legal – is not included in their standard consumer or enterprise subscription pricing (~ £18 - £180 per month). Now it'll require usage credits, described as 'pay-as-you-go pricing' (Anthropic): $10 per million input tokens and $50 per million output tokens.
• Legal Techs: Legora has introduced consumption-based pricing (Legora) (CBP). Apparently "Three innovation chiefs told The Lawyer that Legora has spoken to their firms in recent months about the legal tech's intention to change its billing model away from seat-based pricing when their contracts come up for renewal." (The Lawyer, via Insta).

Why?

You'd expect compute costs to drop over time, and in some ways they have. But there's a countervailing force: the models are getting more capable, and more capable models are more expensive to run. [...]

Industry analysts have noted for years that consumer AI subscriptions are priced for growth, not profit. The goal was to acquire users, train the market to depend on AI tools, and build the data and feedback loops needed to improve models.

(MindStudio).

It is worth noting that, contrary to their consumer offerings, Anthropic, OpenAI, Google, Mistral etc. all already charge per-token for use of their APIs. Legora and other Legal Techs use these APIs to access frontier models. For Legal Tech vendors, the input cost of AI has always been variable.

The challenge is that the really expensive part of AI – the R&D and the infrastructure – is still in very few hands: these behemoth vendors with the leverage to set, and change, their prices over time. Even where per-token prices stay static, the greater capability of the models and increasingly sophisticated use-cases (plus, clued-up users) can have an inflationary effect (JP Morgan) on token usage. Sentiment is that pay-per-token is not going away any time soon. It could become unsustainable for Legal Techs to resell for a fixed price what they buy at an increasingly variable cost.

Token budgets

For the in-house community, perhaps budgets will need to include an estimate for AI token spend. In its announcement, Legora noted: "With CBP, you only pay for the work Agent Pro delivers, and the cost can be attributed to the project that drove it." Joining the dots ... a new disbursement on external counsel's bill?

Setting parameters isn't necessarily a bad way to manage the models, quite the contrary. Applying the right constraints, Goldilocks style, can help optimise the output. AI agents can even be given a budget directly, and told to spend it 'wisely'. (TryAI; get your Uptown Funk on. Speakers on full volume please.)

So, higher budgets might yield better results, but this is not a given, and limits can help focus the artificial mind. There will be inevitable technical and commercial sweet spots, which may or may not converge.

Could working out the AI budget become an important part of pricing external legal support? Tokenomics might be a new skill we need to master.

Around the web

• Meta lawyer Eric Xiyu Li on the 10X coder v 3X lawyer. (LinkedIn)
• UK Jurisdiction Taskforce's Legal Statement on Liability for AI harms (LawTechUK): an authoritative Legal Statement on liability for non-deliberate AI harms under English private law.
• I'm leading a Virtual Hackathon to write an app for the Society for Computers & Law (LinkedIn). Interested? Drop me a line. If you'd like to be a tester, and are on iPhone, let me know. Breaking news: the first release is now available for private testing.