Abundance, if we allow it

A report that Goldman Sachs had restricted access to Claude Code made me look more closely at AI pricing. Bloomberg described a review of productivity and risk. The discussion also turned to cost, as people compared personal subscriptions with much more expensive work licences. I don’t know what Goldman pays, and couldn’t verify the figures being shared. But it left me wondering what the access I take for granted would cost at a much larger scale.

On a personal Claude plan, a monthly payment includes a usage allowance. Anthropic’s new Enterprise plans charge for the seat and then for usage from the first token. I pay $200 a month each for Claude, Grok and ChatGPT, and have them running practically around the clock without exhausting my allowances.

I did a rough calculation for Grok, pricing the tokens I get through my subscription at its API rates. I arrived at around $12,000. That is a comparison with another retail price, which may include a substantial margin, rather than a measurement of what xAI spends serving those tokens.

With a flat monthly bill, I barely think about the cost of another task. With metered usage, every task adds to it.

Anthropic’s latest filing implies a gross margin of about 85% for the first nine months of 2026, yet it still reported an operating loss after research and other costs. Those company-wide figures don’t reveal whether my subscription is profitable. Free access is funded elsewhere, and I suspect personal plans substantially subsidise some heavy users too. I don’t know by how much.

The infrastructure behind it is substantial. In April, Anthropic committed more than USD 100 billion over ten years to AWS technologies, securing up to five gigawatts of capacity to train and run Claude. It also linked rising consumer demand to pressure on reliability and performance. The commitment covers training as well as serving users, but it gives some sense of what sits behind the chat window.

The drug dealer analogy came to mind: make something cheap or free, let it become a habit, then change the terms. I don’t know whether that describes anyone’s strategy. Keeping people using these tools every day might be worth more to the companies than raising prices. But if the terms changed once education, work and ordinary life were built around them, how easy would it be to leave?

Switching providers might help with the price. It wouldn’t necessarily change the habit. I see people ask ChatGPT questions they already know the answer to, perhaps for reassurance. A useful second opinion can become the first place someone turns when they are unsure. I wonder what growing up with that reflex does to a person.

I can produce better work with these tools without necessarily becoming better at doing it myself. They can help me learn, but they can also remove the need to try. If a model supplies the first draft, resolves each dead end and handles the debugging before I’ve understood the problem, when do I develop the judgment needed to question its output?

In an Anthropic experiment with 52 software engineers learning an unfamiliar Python library, the group given AI assistance scored lower on a subsequent comprehension quiz. Participants who used it to seek explanations tended to score better than those who delegated the work. A small, short-term study cannot tell us what happens to a generation, but it gives me another reason to think about how I use these tools.

I wonder about the creative side too. Some ideas only become interesting once we’ve spent time with a bad version and worked out what bothers us about it. Models might eventually do most of the inventing, producing more than ever while fewer people learn to invent or evaluate the results. If we also defer to them about what to attempt, we give up some say in what gets made.

Even if these tools create new kinds of work, those jobs have to be realistic options for the people being displaced. Agricultural workers once moved into factories, although that transition was neither easy nor painless. I struggle to imagine the equivalent for a taxi driver whose work is automated. Supervising a hundred agents in a critical system requires technical knowledge and judgment developed over years. For many people, that simply isn’t a realistic next job. Losing a livelihood does not suddenly provide the expertise to do it.

Some people will make that transition, and there may be entirely different jobs I haven’t imagined. But we cannot assume everyone can retrain for those jobs, or that enough will exist. If a handful of people can direct what previously required hundreds, pointing to a new occupation doesn’t answer the question of how many it can employ.

Perhaps a basic income eventually makes the need to find another job less urgent. That still depends on enough real production to support it, and having enough to live on leaves the question of what to get up for. I can imagine having plenty to do without a salary: making things, gardening, getting absorbed in something simply because it interests me. Other people might struggle.

Paid work doesn’t reliably give people a purpose either. Some white-collar work already looks like adult daycare: showing up, filling the hours and following decisions made elsewhere. I wonder how much autonomy would actually disappear with those jobs, and how much has already been handed over.

The habit of leaving decisions to someone else could outlast employment. Being provided for while letting something else make the decisions starts to sound uncomfortably like humans as pets. I would want the freedom to choose what to do with my time, including things a machine could do better. That means keeping some judgment of my own, even when handing another decision to a model would be easier.

Freedom over how to spend our time is harder to hold onto when the tools reshaping work are out of reach. The taxi driver could lose his income while also being priced out of something that might help him find another way to earn it. Access would not solve the whole problem, but making it harder to obtain seems a particularly bad response. That is where a question about subscriptions becomes a political one for me.

I’m comfortable calling myself a capitalist. Libertarian describes part of how I think, but I want a state that provides police, healthcare, education and fire services. I want the law upheld, fewer unnecessary laws, smaller bureaucracies and more freedom to do things that do not harm other people. Living in Denmark, I often find myself at odds with how readily another problem becomes an argument for another layer of government.

I care more about the difference between making and taking than between rich and poor. A billionaire who builds useful things, creates jobs and pays tax can be a maker. So can a small business owner who provides something people need. Wanting more control over your own time is a perfectly good reason to build a business. Teachers and nurses contribute too. Becoming wealthy through creating value is fine with me; wealth alone tells me little about whether someone deserves admiration.

I have much less patience for extracting value without contributing: a company lobbying to keep competitors out, a politician expanding an office without improving the service, or someone with a realistic opportunity to contribute who chooses to leave necessary work to other people. Scale does not make that more respectable. I want building and working to pay better than securing a protected position. We need more makers.

I also want the gains from AI to reach people who never own a technology company. Greater productivity should make a good life easier to afford. There is a difference between leaving necessary work to other people and no longer needing people to do it. Keeping someone occupied just to prove they are working would miss the whole point.

The version of socialist utopia where everyone spends their time painting, drinking coffee and pursuing their interests leaves an obvious question: who does the necessary work? If you oppose the automation that could grow the food, build the homes and keep the electricity running, someone still has to do it. There is no reason to assume you’ll be the one in the coffee shop. You might be the one in the potato field.

I don’t believe communism solves that problem. Changing ownership does not make the work disappear. AI, robotics and cheaper production offer a way to reduce the human labour it requires. For the first time in my life, I can see how a large part of that work might actually be done for us. There is an odd political twist here: the right could end up delivering the material abundance the left has always promised. The left seems to want abundance, as long as it still takes human-intensive labour to produce it. That is a strange condition to put on utopia. Blocking that progress would undermine the very freedom being promised.

Getting there requires investment, engineering, energy and infrastructure. We need more efficient models and hardware as well as more capacity. Competition matters if falling costs are going to reach customers. Wanting abundance while obstructing the means of producing it seems completely backwards to me. We could end up protecting people from the very thing that would let them work less and live better.

The French proposal shows the instinct that worries me. On 30 October, the National Assembly adopted an amendment proposing a charge of EUR 0.055 per gigabyte sent by content providers to French users, subject to revenue-based caps. It never became law: the Assembly rejected the bill’s revenue section on 4 November, taking the amendment with it.

That was a tax on data traffic, rather than AI computation, but it raises the same question for me: who ultimately pays for costs added to digital services? Addressing the bill to a company does not guarantee that its owners absorb it. If an approach like that returns, I expect providers to pass costs on, cut services or leave markets where the economics no longer work. A failed proposal can still reveal an approach I think is dangerous.

Germany has also been debating special taxation of AI-related revenues, although there is no settled scheme and the federal government has said growth should come first. I dislike treating the use of a more productive tool as grounds for an additional tax.

How would we isolate the profit attributable to AI once it is used throughout a business? As ifo’s Clemens Fuest points out, value created through AI is already captured by existing taxes where it appears as higher taxable profits or revenues. Whether foreign companies pay enough tax in a particular country is a separate question; attaching an AI label does not resolve it. I want Europe to build more of its own capability. Making that capability more expensive to develop or use seems a strange place to start.

The same concern applies to the political reaction around data centres. In the US, Democrats have won control of both houses of Congress, while voters have backed restrictions on data-centre development. Opposition to data-centre development crosses party lines. In August, Republican Governor Greg Abbott required an audit before projects could advance through Texas’s grid-connection process.

There are real questions about power, water and who pays for the infrastructure, which Abbott’s directive addresses. I want workable rules that let projects proceed once those questions have been addressed. Turning them into an indefinite halt would create problems of its own.

Blocking new data centres would not make AI go away. Existing facilities would keep running, and the companies that already have capacity would keep using it. What we would be restricting is the ability to serve more people. If demand keeps growing while new capacity is held back, I expect heavy use to become much more expensive than it needs to be. AI could remain available to organisations willing to pay large bills while becoming unaffordable for people who need it just as much.

That is why the difference between the hyperscalers and the developers building their next facilities matters. An established operator with functioning data centres, secured power and substantial cash has options that a developer financing construction does not. A construction delay does not stop financing costs from accruing. If those developers cannot complete their projects, the additional capacity that could ease the shortage never arrives.

Consider the financing involved: Cipher announced USD 2 billion of secured notes to help finance its Black Pearl facility in Texas. Applied Digital announced a USD 300 million bridge facility for construction in North Dakota, with more financing expected to complete the work. I don’t know of either project being stopped, but the financing shows how much can already be committed when a delay arrives.

The relationships overlap too. Nebius has a large infrastructure agreement with Meta alongside its own AI cloud business. New capacity can serve the largest companies while also giving other customers an alternative to established providers. Restricting it can disrupt the large companies’ plans too.

Google, Microsoft and Amazon may face higher costs and delayed projects too. But if there are fewer viable competitors and scarcer access, I expect the advantage of already owning productive infrastructure to grow. A policy can hurt an established company while leaving it in a stronger position relative to everyone trying to catch up. My concern is what that means for everyone buying access: fewer alternatives, less pressure to lower prices, and more of the available capacity going to whoever can afford the largest bill. The divide could become one between wealthy people who can afford to put AI to work continuously and everyone else rationing their use.

It would also affect who gets to train the next models. We already have labs releasing open-weight models and specialised GPU cloud providers such as CoreWeave offering infrastructure for training and inference. That gives developers alternatives to buying everything from a hyperscaler or relying entirely on a closed model. If compute becomes scarcer and compliance becomes more expensive, I expect independent labs to find it harder to experiment, train and compete. Open weights alone do not pay the compute bill.

Both OpenAI and Anthropic actively advocate binding rules for frontier AI. They frame these as safety measures, and OpenAI explicitly says smaller developers and open models should be protected. But exempting small labs does not guarantee they can afford to grow into serious competitors. If rules make it prohibitively expensive for others to reach the frontier while the largest labs can absorb the cost, that would be regulatory capture in practice. It would let incumbents shape who can realistically challenge them, even when the stated aim is safety.

Nor do I expect China to wait for us. Its government has set out plans to expand its national computing network. A pause in Texas or a new European levy does not oblige China to slow down. I think restricting our own ability to build will leave us more dependent on whichever companies and countries carry on. Even within the US, organisations with enough money and existing capacity will keep finding ways to use the technology.

Everyone can win from a sufficiently large increase in productivity. I don’t expect that to happen automatically, and I don’t trust a future where either a handful of companies or the government decides who gets access to these tools. Open competition, more capacity and lower costs seem much more likely to give people a chance.

The people I expect to suffer most from getting this wrong are precisely those the restrictions are supposed to protect. The taxi driver, the artist and the small business owner would still face competition from organisations using AI, but would have less affordable access to it themselves. We would have preserved the disruption while limiting the opportunity to respond.

I also need to keep the noise in perspective. The French amendment has already fallen away. A tax debate is not an enacted tax, and an election does not tell me exactly which projects will be built. Sometimes “nothing ever happens” is a useful corrective to the feed. But the consequences of making capacity persistently harder to add are worth thinking through before that becomes the policy.

And I remain excited about what these tools make possible. Building with a small team gives me new reasons every day. There are things we can attempt now that would have been out of reach a few years ago. I want to keep finding out what else we can do.

The difference between my subscription allowance and paying for each task on a company account started this train of thought. I can afford to keep models working around the clock today. I want that kind of access to become cheaper and more widely available, rather than something we look back on as an unusually generous introductory offer. That requires the capacity to serve far more people and enough competition for lower costs to reach customers. A subsidy can help people get started. I don’t want lasting affordability to depend on one.

I want to remain capable of questioning the tools I use. I don’t want us to get as far as making AI indispensable, then leave most people unable to afford it.