Dispatches 002: AI Economies & Future of Work

Ah, AI.

There is a joke doing rounds in tech circles right now that Anthropic’s next releases will be Claude HR, Claude Sales, and — at long last — Claude Customer.

(I say this, as Anthropic literally pilots an agentic marketplace…)

Funny as it is, it does raise a lot of questions. Our entire economic system is built on the assumption that it runs on human labor. People receive income for their work, which is in turn taxed to fund everything from their social security to public services. Usually, work does not simply arrive from the digital aether to hand down your quarterly financial report.

So what happens when we have autonomous systems that do not sleep, running execution 24/7? When 5% of the knowledge economy runs on AI? Or 25%? What if it cuts through to other sectors?

Sometimes this all sounds farfetched, but as a recent grad myself, I would very much like to know the answers to these questions. And the discussion is just frustratingly binary at times. We’ve all seen that one graph with runaway AI and Singularity extinction on the same scale. The extremes are real and to some degree possible, but it’s the staggered, messy, chaotic middle that I want to understand.

So, in an effort to avoid peddling street GPUs to AI agents as my next career move, I’ve been thinking about what the future of work in an age of AI looks like. And what that really means is what AI Economies will resemble. Because this is far larger a trend then just movements in sectoral employment.

Frankly, this is a think tank’s worth of work. I am in fact working on a proper report with scenario modelling and economic analysis, but for now, I thought it would be helpful to jot down some thoughts. My journals on the subject are lengthy.

(And yes, just the journals ended up 5x my standard essay size)

What I know for sure is that we have to start working through these questions now, so that we understand our response options later. The shifts will arrive in waves, but they will do so quickly, and we probably won’t have a decade to deliberate policy and institutional responses once it all lands.

Known Unknowns

The entire AI Economies conversation hinges on a couple fundamental assumptions. Many have to do with AI itself. Some of them are organisational or questions of efficiency. Others are geopolitical.

Right now, we don’t really have good answers to these questions, but we do know what to ask — and there is a lot.

Understanding the trajectory and the factors at play is right now more important than what the end outcome will be. That’s unknowable. It also isn’t something that just emerges out of the woodwork; that outcome should be a collective project.

Here is my non-exhaustive list, compiled from my journals:

On AI & Tech:

- Will the rate of advancement in AI capabilities continue to hold, accelerate, or decelerate?

- Can the current paradigm (LLMs) deliver significant economic transformation, or will the next step require a transition to a different technology? (World Models, something else, etc)

- What are the next waves? (Chatbots → Agents → Swarms → ?)

- Which capabilities are missing to enable significant economic impact?

- When & how will the next set of barriers be overcome? (I personally hold these to be memory, agent coordination, real-world interfacing, and a couple others)

On Economics:

- Will AI actually end up being cost-effective? Will that be the case market-wide, per sector, something else?

- Can organisations effectively adopt AI? Which ones, and how? How will that adoption end up distributed across sectors?

- Will compute remain a major bottleneck?

- Who owns AI? Who owns the resulting productivity and income from AI?

- What do we do with the layoffs and job market impact?

- What do we do with the employment-GDP divergence? (Aka revenue being generated by non-humans)

On Geopolitics:

- What will happen with the question of sovereignty? Does the benefit of AI advancement just accrue to the 2-3 major players, with everyone else left out? (Currently those are a US-China duopoly, with EU trailing)

- What will the picture look like on global AI adoption?

- Will energy costs & data center investment become a major barrier? Will that affect which states accrue the most benefit?

- ...

On the list goes. Were I to list everything, we would be here forever, but hopefully I have picked up all the major ones. Even if I didn’t, logic still holds.

I think of these questions as inflection points. Some combination of events will produce one outcome — say, large-scale unemployment. Another will result in major economic growth with people perhaps transitioning to a 3-4 day work week. On it goes. There is no binary answer here, and I suspect the outcomes will be deeply unequal across sectors, countries, and regions.

The real challenge isn’t so much predicting specific impacts as it is the second-order effects. Something this major will have knock-on consequences from work to education to policy to welfare. It’s a lot closer to ripples on an ocean emerging from several waves than it is to a straight line trajectory.

What I will do now is go by layer: tech, work & organisations, and economies. Keep in mind that these may as well be working notes, so don’t bash me for forgetting to quote Keynes or something.

Baseline: What’s the Tech Trajectory?

Technology first. Half this conversation hinges on capability.

As Ethan Mollick often says, even is AI development virtually halted, we would spend the next five to ten years and then some figuring out all the way to harness just what we have today.

Except that it’s not stopping. It’s very much accelerating.

Frontier models are being released roughly every 2-3 months. Task time horizon for AI is consistently doubling roughly every 7 months, per METR. At current rates, frontier models will reach day-length execution capability sometime around 2028. If current trends hold, they will cross a week in early 2030s. Beyond that, the question becomes whether what we humans do is a continuous month-long+ process, or is a day / week / etc of continuous work a sufficient unit.

Time Horizons for AI, METR

A similar picture is playing out across tasks & jobs.

Anthropic Economic Index, Jan 2026

The interesting thing to note above is the shape of automation. Anthropic naturally has a bit of a conflict of interest here, but they’re also closest to the data. What they’re seeing is uneven, jagged augmentation; AI’s impact on a role fundamentally depends on how well it can do their highest-value, highest-time-spent tasks. A kind of most of the way and last mile situations.

What this leads to is, well, complex. There are mainly two outcomes:

- Suppose AI replaces ≈80% of a job at task level. The person could then transition into overseeing “last mile“ tasks, or whatever requires direct decision liability, across several streams. That would drive redundancy, or at the very least some sort of reshuffle.

- Let’s say AI in fact replaces a lot of the procedural tasks (Which is what is often happening across coding, data analysis, etc). The person could then shift much of their time to higher-value tasks and projects which simply did not make the cut into their ‘time budget‘ before. That could drive massive increases in creativity in innovation.

A binary, I know. The real dynamic is likely somewhere in between the two.

There has also been a lot of evolution at interface level. Claude Cowork, standalone apps for Codex and Lovable, now Claude Design, etc. All of it is essentially scaffolding, knobs and control panels that are easy to understand and manipulate.

And that’s just the big, generalist labs. Chatbots & Agents is what gets the most use on consumer and enterprise levels, these tools being conveniently adaptable and general. The recent crop of scientific models has been extensive, though: NVIDIA Modulus for physics, Berkley’s OPAL, new AlphaFold variants, and more. Everything from bio modelling tools to physics-first simulations.

So, AI is accelerating massively. All rosy, buy me some Anthropic secondaries.

What are the current limitations?

Most of it is in implementation, adoption, and economics.

On tech, my view is that current barriers are in memory, coordination, and reliability. These are the core roadblocks for swarm-grade AI to be effective. We’ll come back to that particular subject in a second.

There is also the questions of LLMs as a technology. Whether an LLM is just a predictive engine or not may be somewhat of a moot point if it is incredibly effective regardless, and the current crop of models is very good at much of regular knowledge work. But the question remains. My guess is that we will in fact need a paradigm shift if AI is to become reliably embedded into manufacturing, biology, scientific research, and other areas of the economy.

AI x Work: Where’s the Threshold?

Frankly, one doesn’t need to look at the economic data to see that even current generation models change the game entirely. I have been continuously rewiring and adapting my workflows every two months.

The specifics change, but what I’m personally seeing is three parallel enabling vectors:

- Breadth: AI lets me reach into other domains to assist my work, which before would have been too time consuming (Build custom tools, create some quick exploratory designs, etc)

- Acceleration & Depth: because I can now call up a pretty solid 15 page research report at will, I can spend far more of my time cutting deeper into topics, synthesising material, looking for interesting patterns, and so on. This is the case in other areas, too.

- Connectors: combining the two above means I work more like a consistent, singular flow state, shifting between all my instruments. I can quickly review, stress test ideas, run planning, etc. Effectively, AI is plugging a lot of gaps in my workflow just on an admin / org level.

This is just anecdotal, but I am doing something on the order of 3-5x the scope of work I have before. And the thing is, I also spend more of my time doing what I find interesting and valuable – experimental projects like Meridian. I can say for certain that it will take months to years to rewire my personal workflows even with the current technology. At the very least it’s made me consider reskilling into other areas like product, since much of my analytical work happens on far more compressed timescales.

Similar things are happening in other knowledge work areas. Depth and priority reallocation. Coders can be more intentional about what they build and iterate faster. Designers can plug the gaps on code for their projects. Analysts of various shades can do research faster, and shift focus on synthesis and deeper analysis.

Is this what is actually happening? Well, hard to tell.

Ballpark, we’re currently looking at a mix of AI-driven job displacement and accelerated productivity.

There has been plenty of headlines about Big Tech AI-led layoffs, but that may be in part PR cover. What I would really pay attention to is per-employee revenue — and that has shifted. There are of course the mad numbers at AI-native hypergrowth startups (Cursor / Anysphere at $3M+ per employee, Midjourney’s $2M PE, etc), but the bar is also rising in other tech firms (Salesforce, Crowdstrike, Workday, all in the ≈$400k PE ranges).

Outside of tech, this trend largely hasn’t materialised. The very idea of per employee rev is coming from startups right now. Likely a mix of coding & product being the frontier of AI development right now, and leading startups & tech companies pouring far more resources and effort into AI adoption and implementation.

I’ll dig into this specifically in future issues. What I want to note here is that the picture is quite uneven, both geographically, across organisations, and across different jobs. As I’ll go over in a bit, worldwide AI adoption in particular is incredibly spread out. Coupled with what we’re seeing on per-employee data, that suggests a massive split between leading adopters in top-adoption countries (Mostly ‘Global North‘), and lagging adopters in low-adoption countries. If that is persistent and systemic, this will show up in GDP data, which will in turn show up in quality of life and prosperity, whatever numbers you use.

Bottom line: very, very jagged adoption and automation curves all over.

How exactly this will play out for individual jobs is tough to say. Here’s the two broad scenarios:

1. Displacement: AI cuts down knowledge workforces across the board, contracting organisations. That will slow down hiring, in turn slowing down the education streams in the affected fields. The job market itself will contract in those sectors, necessitating reskilling.
2. Reallocation: as AI optimises workflows, people can shift their effort onto higher-value or more experimental parts of their work. Meaning projects which wouldn’t have been cost-competitive now get done, “inessential“ but delightful product features get built, and there is all around more focus on creativity and innovation.

My guess is that these will be co-occurring, effectively rolling in waves.

One sector gets streamlined to some sort of optimum. That talent shifts elsewhere. The same thing occurs again. Perhaps this time, people shift geographies — in turn bringing the know-how with them. Propagation will speed up. On it goes.

The real question here is just how extensive will this be. If it’s 5% of the economy, things will likely reorganise by themselves. If we start creeping into 10, 20 per cent numbers, however, that will likely necessitate social reform.

My predictions here start getting very speculative. Frankly, it’s impossible to answer where will this many people even go.

If it’s a bifurcation between employment and GDP, that can be fixed up with redistribution of revenues (UBI, AI Taxes, all very dicy and divisive in of themselves). If it’s closer to reallocation, perhaps people will reskill with state & public support, and we will plug gaps in other segments of the economy. A rebalancing away from knowledge work, perhaps.

The larger the scale of the shift, however, the more disruptive it will become. If what we’re looking at here is swarm AI adopted by 20% of organisations in the leading 20 adopter countries circa 2035, these will become test cases for new hybrid organisational forms, state reskilling programs, and UBI pilots.

Speaking of…

AI Economies: Waves and Waves

The larger the scale, the murkier the picture.

So far, I have been very focused on AI capabilities and progress. Now is a good time to touch on the limits, though, because this is where they will really start to bite. Elevated token costs, energy prices, and compute limits will all show up at the economic level more than they do in AI-first tech companies that can work around them or fund them away,

And so at last, we have reached the compute conversation.

Let’s say we want to more or less fully automate 20% of knowledge work in top 20 AI adopter countries. Here’s some back of the envelope math for you.

- Average employee does around 2,000 hours of work per year.

- Token counts vary widely, but let’s say ballpark 500k per knowledge worker (Approximated from METR, Claude Code usage, etc, basically a low-end middle. These numbers will be 10-100x at hyperscalers and Big tech).

- 2,000 x 500k = 1B token per year per employee

- Top-20 AI adopters (US, China, UK, Germany, France, Japan, S. Korea, Canada, Israel, Singapore, Netherlands, Sweden, Switzerland, Australia, Finland, Denmark, India, Italy, Spain, Belgium). Combined labor force ≈ 1.3B. Knowledge workers as a share of employment ≈ 35%, ≈460M people. "Automate 20%" = ≈90M knowledge-worker jobs.

- 90M workers × 1B = ≈9 × 10¹⁶ tokens/year (90 quadrillion) (Yes, really)

- $5/1M tokens for a frontier-tier 2026 model

- 9 × 10¹⁶ / 10⁶ × $5 = ≈$450B/year in pure inference spend.

- 90M jobs cost ≈$6.3T/year in total comp. Token cost ≈ 7% of the labor cost being displaced. Ouch.

Let’s keep going.

- Compute fleet. Per-GPU throughput in a 2026 B200/H200 stack: ≈3,000 tok/s sustained → ≈95B tok/year at full util, derate to 70% → ≈66B/year/GPU.

- 9 × 10¹⁶ / 6.6 × 10¹⁰ = ≈1.4M frontier GPUs dedicated to inference.

- That number is something like NVIDIA’s entire GPU shipment order for 2024-2025.

- Capex at ≈$80K all-in per deployed GPU (silicon + server + datacenter share + networking + interconnect) = ≈$110B for the inference fleet.

- Add ≈$150B for redundancy, training pipeline, and 1–2 model-generation refreshes → ≈$250–300B total compute capex to stand it up over 3–4 years.

- Electricity. 1.4M GPUs × ≈1.5 kW facility load (PUE-included) = 2.1 GW continuous = ≈18 TWh/year.

- Global data-center electricity (2024): ≈415 TWh → this adds ≈4%.

- US total electricity: ≈4,000 TWh → ≈0.4%.

TLDR: $450B/year token cost · $250–300B compute capex · 18 TWh.

Alright, done. A lot of numbers.

The shortest version: the compute and electricity to do this exist or are being built. It is, however, a colossal investment, and extremely sensitive to all sorts of factors.

There are three numbers here: token usage, token cost, and electricity cost. Double the token usage and you’ll need something like twice the infrastructure. Double the cost, and that will cut down on adoption / token usage. If this colossal data center buildup overloads the grid, that will put a hard brake on the scaleup.

Now , look. I am not an economist. I am not qualified to tell you what this means. I’m not sure the actual economists know either, if the Singularity hypergrowth / extinction graphs I occasionally see are anything to go by.

What I do see here is a sort of silver lining: AI will not magically take over entire slices of the economy. It just cannot happen. The servers aren’t running on some secret AI installation on the Moon, they’re right here, using whatever grid capacity we have available and virtually the entire planet’s GPU supply.

Full-scale automation of certain jobs is possible, and I will venture as far as saying that it will eventually happen. But it will happen in waves. It will slowly creep up across growing task horizons, additional data center capacity, and the grid scaling up to match that. We will in effect be building synthesis cognition clusters. Odd thought, but here we are.

All of this will, however, be incredibly uneven. For starters, developing countries do not have the resources to deploy anything close to such data center and grid installations. Unless we start shipping nuclear reactors to Nigeria, forgive me the quip, AI infra will be owned by US, China, Europe, and perhaps some other rich economies. That presents a challenge.

The gradually rolling waves of automation also present a challenge to reform our education. Labor policy. Taxation. Much, much more.

This is perhaps the Grand Project of the early 21st century.

There are literally hundreds of points I couldn’t touch upon here. As I have hopefully emphasised, this is not a definitive analysis. Even just this one essay raises dozens of questions, though.

The bottom line is that the impacts of AI as a technology and transformation force will have ripple effects across all of society. The second- and third-order effects are going to be a minor nightmare. All of this will offer a lot of opportunity for positive change, but we have to handle it proactively, or the situation will get very dicy. This is a whole-of-society challenge demanding a whole-of-society response.

I will continue working on this for the foreseeable future. A far more comprehensive version of these Dispatches is also in the works.

We have a lot of work to do, that’s for sure.