Everyone has their own definition of art. Mine is: art is anything that elicits an emotional response. Angry, sad, hopeful and happy - it’s art. It follows that anything that doesn’t make anyone feel anything cannot be art.

Two things follow from this:

1. Art cannot be singularly defined, as what is or isn’t art is defined on an individual level. As long as someone thinks it’s art - it’s art.

2. Art is still collective, social and cultural - but to be collective art it needs to elicit a powerful emotional response from many people.

Can AI be used in the creation of art?

I’ve been thinking a lot in the last few weeks about this question: can AI be used in the creation of art?

I think it can. But only in the sense that it’s a tool a human can use to convey the emotional response they desire. AI itself cannot create art - the human is a necessary component of the artistic process.

Think about it this way.

AI is essentially a compression technology with emergent capabilities in producing novel, new content. But it’s still just another technological advancement.

Technological advancements brought about better amplifiers, electric guitars and synths.

In the hands of a technologist those tools remain physics experiments and curiosities at best.

Artists create emotions with new technology

But in the hands of an artist - we have some of the most powerful collective shared art in history. The Beatles. Queen. Phil Collins and the gated reverb drum line. EDM. Mufasa. Bambi. Toy Story.

Artists experiment as much as a technologist, but an artist is experimenting in how to make themselves, and others feel something.

They are finding how to turn this technology - transistors, circuits and software - into something emotional.

In this way, I don’t think AI is that different to previous technological jumps in art history.

Seeing photographs would have blown the minds of people in the early days of that technology.

Movies - just the concept of seeing a moving screen for the first time must have been wild.

Records meant you weren’t limited by what you could play on the piano for a sing along.

Each of these technologies probably felt pretty similar to how we feel today about AI - in terms of change over a lifetime - but one of the great things about being human is that we get to keep inventing new things and feeling like we’re the first ones to ever experience something new.

Artists will create art with AI

The challenge for artists with AI is not “how do I avoid being replaced by this” - this has already been happening for centuries as technology has replaced humans in art. (Even to the point where the DJ that puts on the songs has already been replaced with an algorithm in Spotify).

It is, how do we use this technology to create something that makes people feel.

Personally, I’m really excited to see what people create.

I think we’re going to see amazing movies created in someone’s bedroom. Production company, lighting, video no longer being hurdles to getting a great script on screen.

Are you a great poet, but don’t have a way to record a song - I’m keen to hear your latest fully-produced track with Suno.

The best art is created soon after the technology goes mainstream

Each time a new technology comes along, there’s this window of bohemianism as slower-moving corporations and the capitalist powers that be try to catchup with the sheer inventiveness of people. An explosion of new ideas, new ways of doing things and new creative endeavours that we couldn’t have imagined without the technology. Then the capitalists come in and ruin it for ever by squeezing every last unemotional drop of commercial value out of it.

AI hasn’t had this (yet) - so get out there and create while you still can. Build something. Write something.

Emotion is our most raw and fundamental means to communicate and form relationships. It is something core and central to what it is to be human.

I and everyone close to AI seems to be experiencing some sort of collective psychosis around “human” endeavours. Multiple people I know want to turn it all off and go farm vegetables in a sort of croft. Crofting was objectively pretty shit - but it was human. It was our shit.

We don’t all need to go back to staving to death over winter to recapture humanity - but we DO need to figure out how to find humanity in this new technology, and fast. I’m not keen to read another “here’s the honest reality” in an article online.

I don’t want to be questioning every em-dash in a piece as some sort of gotcha.

And holy shit some of the new AI video production stuff is getting good.

But most of it’s not emotional. It’s a smeared goop of sepia-coloured slop that gives most people the ick.

Changing this is going to require a lot of hard work by very creative people. And I don’t think the job of creatives is going away any time soon.

One other thing. Older people, as a collective, ALWAYS hate whatever the younger generations are experimenting with. As I get older I find myself increasingly intrigued by “kids these days”. This is normal. Doesn’t make getting older any easier. But with age comes wisdom and we should revel in the creativity of youth unencumbered by rose-tinted glasses of the past.

Watt made coal so productive that steam power became worth deploying everywhere, in mills and mines and ships and railways that could never have justified the fuel bill before. Efficiency lowered the cost of an outcome, demand for outcomes exploded, and total coal consumption rose with it.

One hundred and sixty years later, the same paradox is running again. This time the fuel is thinking and the outcomes are almost everything we associate with intellectual work.

What is the Jevons Paradox?

The Jevons Paradox is the observation that when technology makes a resource more efficient to use, total consumption of that resource rises. The mechanism is price. Efficiency lowers the effective cost of achieving an outcome, and when demand for that outcome is elastic, the quantity demanded grows faster than the efficiency gain saves.

The paradox requires an economic environment where there is a large reservoir of demand that is not met at the old, higher price. Britain in 1865 had an almost unlimited appetite for mechanical work it could not yet afford. That is the condition to test AI against.

How fast are AI token prices falling?

In November 2021, running a model of GPT-3 quality cost 60 dollars per million tokens. By late 2024, the same capability cost six cents, a thousandfold decline in three years, a trend Andreessen Horowitz measured at roughly 10x per year and named LLMflation. Epoch AI, measuring across a wider set of tasks, found the median price of a fixed level of capability falling around 50x per year. There is no precedent for this. Compute during the PC era and bandwidth during the dotcom boom both got cheaper at a slower rate.

If intelligence were an ordinary expense, spending on it should be evaporating.

Why is AI consumption rising while prices fall?

Google publishes its own usage stats. In 2024 it processed 9.7 trillion tokens a month across its products. A year later the figure was 480 trillion. By May 2026 it passed 3.2 quadrillion tokens a month, a sevenfold rise in a single year. Enterprise spending tells the same story from the buyer’s side: Menlo Ventures tracked generative AI spend in the enterprise going from 1.7 billion dollars in 2023 to 11.5 billion in 2024 to roughly 37 billion in 2025. The unit price fell by orders of magnitude and the total bill rose by orders of magnitude.

It turns out the latent demand for intellectual work is huge. At 60 dollars per million tokens, you spend tokens carefully, on the few tasks that clearly justify them. At six cents, you spend them on everything. You let a model draft ten versions instead of one. You let it read the entire document set instead of a summary. You let agents burn thousands of tokens checking their own work, because checking is now cheaper than the error it prevents. Whole categories of thinking that were absurd at the old price are routine at the new one, and each new category adds to consumption faster than the price decline subtracts from it.

Does The Jevons Paradox apply to AI and knowledge work?

Jevons was writing about the fuel of physical work, and his paradox explained why an age of efficient engines became the age of maximum coal. Tokens are the fuel of intellectual work, and the same logic now explains why an age of collapsing AI prices has become the age of maximum thinking.

In my next article, I want to tie some of this together. My logic is this: humans performing high-level intellectual and relationship work are complementary to AI. Not only that, but AI promises increased productivity for the humans able to work alongside it.

Both of these point in the direction of rising latent demand AND falling unit costs of human intellectual labour.

I think that where this leads is the biggest boom in intellectual labour demand in history.

That is a complementary good behaving exactly as the textbook says it will. It is also the most useful thing the economics of AI can teach you about your own career.

What is a complementary good?

Earlier in this series we covered the substitute, a good that can be used in place of another. A complementary good is the mirror image. It is a good consumed together with another, so demand for one lifts demand for the other. Just like AI and memory. Formally, complements display negative cross elasticity of demand, which means that when the price of one falls, demand for the other rises.

The substitute and the complement are the two relationships a good can have with another, and they pull in opposite directions. A substitute takes demand away. A complement adds it.

Why AI is making memory more expensive

An AI model runs on memory. Every token a data centre generates is produced on banks of high-bandwidth memory, so tokens and memory are complements in the strictest sense. As the demand for tokens has exploded, the demand for memory has gone with it, and the price has followed. DRAM prices rose by roughly 80 to 90 percent in a single quarter, and the AI hyperscalers, the Microsofts and Googles and Metas, are buying the supply at the front of the queue. That leaves the makers of laptops and phones to compete for what is left, at the price the complementary buyer has set.

This is why your next laptop costs more. The memory in it is the same memory an AI data centre wants, and the data centre got there first.

The squeeze falls on everyone who is not the complement

Here is the part worth sitting with. If you are buying memory because it complements tokens, you can pay the higher price, because the memory is earning its keep inside a far larger machine. If you are buying memory because you want a laptop, you simply pay more, or you go without. The complementary use case sets the price, and the ordinary buyer wears it.

When a good becomes complementary to something in enormous demand, its price is set by that demand, and everyone outside the complement is left competing on terms they did not choose.

The complement you have not yet priced in is your own work

Memory is not the only thing complementary to tokens. Human judgement is too. Every token a firm generates still needs a person to decide what to ask for, to tell whether the answer is any good, and to carry the consequences on execution. The more tokens a business runs, the more it needs that judgement, which makes a certain kind of human work a complement to AI in exactly the way memory is. As demand for tokens rises, demand for that work rises with it.

The big question: Do you feel busier because of AI?

This gives you a simple test. If AI has made you busier than you have ever been, you are most likely operating as a complement, and demand for what you do is rising alongside demand for tokens. If AI has made you feel surplus, the work quietly draining away from you, you are most likely a substitute, and the demand is flowing past you to the people who direct the machine rather than compete with it.

The position is not fixed, and that is the optimistic part. The same technology that takes demand away on the substitute side adds it on the complement side. You have some say over which side of that line your work sits on. Memory had no choice about becoming valuable. You do.

That is a substitute good doing the one thing a substitute does. It is also the cleanest way to understand what just happened to your job.

What is a substitute good?

A substitute good is a good that can be used in place of another. Tea and coffee. Butter and margarine. The bus and the car. Formally, substitutes display positive cross elasticity of demand, which means that when the price of one rises, demand for the other rises with it, because buyers move between them.

The part that matters is the part the bus just demonstrated. Nothing in that shift came from opinion. Nobody’s view of the bus changed, and whether you personally think a bus compares to your own car is irrelevant to the economics. The substitution became attractive enough at the margin, enough people re-decided, and the whole market moved. A substitute does not have to be as good as the thing it replaces. The bus is worse than the car. It still empties the road.

The first substitute for intellectual labour

By intellectual labour I mean the work that can be delivered through a keyboard and a mouse. That is reductive, and it is also most of what white collar work has become.

For all of economic history, that labour had exactly one supplier: a person. If you wanted thinking done, you convinced a human to do it, or the thinking did not happen. Intelligence was never a tradeable good in its own right. It came bundled inside people, with contracts, notice periods, relationships, and management overhead.

LLMs end that arrangement. For the first time, a firm can buy intellectual labour from something other than a person, priced per token. That makes the LLM the first true substitute good for intellectual labour, and the claim does not rest on my opinion of the technology. It rests on how firms now talk. AI is openly named as the reason behind restructures and hiring freezes, and when a buyer explains a decision by pointing at the alternative, the alternative is a substitute.

Does AI have to be better than humans to take work?

This is where the doom sets in, and the doom is half right. The substitute exists, and that fact alone changes the competitive dynamic for labour. What the doom gets wrong is the obsession with quality. Every week someone posts a screenshot of a model doing something ridiculous, as if that settles the matter. The bus is worse than the car. It still empties the road. A poor substitute still moves demand for the thing it stands in for, and each new model quietly improves both the quality and the relative price of this one.

Why this substitute is different from every tool before it

Every previous tool of intellectual work was a complement, not a substitute. The typewriter, the calculator, the spreadsheet, the computer itself. They transformed jobs and compressed whole professions, and through all of it the substitute for a bookkeeper was always a human using a better tool. There was no version of the work with no human in it at all.

That is the line we have just crossed. For the first time, firms are choosing between hiring a person and building an agent.

But a substitute is only half of what AI is. The same technology that can stand in for a person can also make one far more valuable, and that is where this series goes next.

The only substitute for intellectual labour used to be cheaper intellectual labour

For most of the modern era, the sole substitute for expensive intellectual labour was the same labour bought somewhere cheaper. That is the entire logic of offshoring. When a task did not depend on physical presence or close relationships, it could be sent to a lower cost market, and for thirty years a great deal of it was. The substitute for a worker was always another worker.

That is the assumption the last year has broken. The substitute is no longer a cheaper person. For a large class of tasks, the ones that are not built on trust, judgement, or a relationship, the question is no longer where you find the cheapest human. It is whether the task needs a human at all. Work that is being shipped offshore right now may be solving last decade’s problem with last decade’s tool.

Why an option changes a market before anyone uses it

Optionality has a value of its own in economics, separate from whether the option is ever exercised. The moment a buyer gains a credible alternative, their behaviour changes, because every decision is now measured against the thing they could do instead. The alternative does not have to be chosen to matter, it only has to be available.

This is the big event of the last year, perhaps the century.

As a founder, an owner, or a manager, you are now a buyer of intellectual labour with two suppliers instead of one. You can hand a task to a person, or you can do it with an AI agent, a skill, an MCP, or an API. This new second option did not meaningfully exist even twelve months ago. But it does now. If you are still making resourcing decisions on the assumption that intelligent work means a person, you are working from thinking that is already out of date.

Is this a bubble, or a structural change?

If there is one thing to take from this, it is that the ground has moved further than the headlines suggest. This is no longer the territory of crypto promoters who pivoted to AI overnight. The change is structural, and it sits in the resourcing decisions you are making this quarter.

The obvious objection is that this is a bubble. In market terms, that may prove true. We are still working out not only what intelligence is worth, but what the infrastructure behind it costs. The technology is developing so rapidly, and so experimentally, that the price will stay unpredictable for some time.

However, a bubble in the price is a separate question from a fad in the technology. Valuations can fall without the option disappearing, because a capability this revolutionary can never be shelved. It will take many forms, but the option is here to stay.

The way intellectual work gets priced has fundamentally changed, and it will never change back.

A supply shock is an unexpected event or condition that changes the price and quantity of goods or services available for sale in a market.

A positive supply shock is an increase in the quantity of goods or a reduction in price of the good - so there is suddenly way more of a good being produced for a given price.

In the short term, positive supply shocks lower production costs meaning profits for businesses as the market takes time to adjust to the new quantity and price.

In the longer term, businesses and consumers adjust their behaviour, ultimately increasing consumer surplus and driving economic growth.

Short term: lower prices for consumers of the good and higher profits as the increased consumer surplus is eroded by business with market power.

Long term: Structural changes in markets and expansion in production capacity.

We are currently witnessing what is undoubtedly the biggest supply shock to white collar work ever.

The knowledge work revolution

Think about what AI models can do today that required human expertise just a few years ago: write code, analyse data, create marketing copy, draft legal documents, design graphics, translate languages with almost imperceptible accuracy, summarise research papers.

The "good" being produced here is cognitive output - the thinking, writing, analysing and creating that forms the backbone of modern knowledge work. And suddenly, there's an enormous increase in the supply of this cognitive capacity. (Whether LLMs are actually “thinking” is up for debate, but I’m happy with this over-simplification for the sake of this post)

Just as the printing press created a supply shock for written information, or the industrial revolution created one for manufactured goods, AI is creating one for human thinking itself.

The short-term effects are already visible

Right now, we're seeing the classic signs of a positive supply shock playing out in real time:

Businesses are capturing immediate profits by maintaining their pricing while dramatically reducing their production costs. A marketing agency can now produce ten times the content with the same team. A software company can ship features that would have taken months in a matter of weeks.

But consumers haven't yet fully adjusted their expectations or behaviour. They're still paying roughly the same prices for services that now cost a fraction of what they used to produce.

This is the temporary imbalance that always follows a supply shock - the market hasn't yet figured out the new equilibrium.

What the long term looks like

But just like every supply shock before it, the market will adjust. Or more specifically: human behaviour and work culture will adjust. And when it does, the changes will be structural and permanent.

We're already seeing the early signs: new types of businesses that simply couldn't exist before, like AI-first consulting firms that can tackle problems previously reserved for McKinsey-sized teams. Tools that put sophisticated analysis in the hands of small business owners who could never afford it before. Startups with over $1m ARR / staff member.

The real transformation will come when we stop thinking about AI as a replacement for human work, and start thinking about it as an amplifier of human capacity. Just as the industrial revolution didn't eliminate human labor but changed what humans do, the AI revolution will change the nature of knowledge work itself.

In five years, the idea of writing a business plan without AI assistance will seem as antiquated as doing accounting by hand. The humans who thrive will be those who learned to work with these tools, not against them.

And just like farming gave way to factories, and factories to offices, knowledge work as we know it today will give way to something entirely new - something we're only just beginning to see emerge.

The supply shock we're experiencing isn't just about doing the same work cheaper and faster. It's about redefining what knowledge work is when we can produce an infinite amount of output.

This year has been a big one - both professionally and personally - and I have been intentional to take the time out to recharge before what promises to be another season of intense hard work.

As a founder of a business exploring the limits of Large Language Models, I have spent a lot of time trying to understand how this technology will impact work.

I have been reflecting about the changing features of work - how humans work schedules have been defined as much by social norms as by technology:

- Nomadism gave way to farming

- Farming gave way to factories

- Factories to office work

- Office work to a universal basic income and endless leisure ... just kidding.

But something that was lost in the move from farming to factories was seasons.

Human lives used to be governed, on a day to day basis, by the seasons. Both in hunter-gatherer and farming societies, people's daily tasks depended heavily on what season it was - the season effectively defined what was being done, but also, importantly, provided variety throughout the year - you're unlikely to be harvesting much grain when the fields are covered in snow.

If you had a lot of hard, manual work bringing the harvest in in Autumn, you knew that it would be followed by months of cold in winter when you could make tools, repairs and enjoy more sleep due to the shorter days.

This seasonal variety meant that you didn't have to work on the same task for long periods of time. You got to do different things, master different crafts and enjoy more time off depending on the time of year.

The variety this afforded helped to make each period more bearable - rather than looking into the foreseeable future working on the same task, you could instead look forward a much more reasonable amount of time (weeks) to a change (often demarcated by a festival - Easter, Midsummer, Halloween, Christmas).

I believe that we're programmed as humans to work well in 10-15 week blocks of intense work. The changing of the seasons forms a natural cadence to work (regardless of hunter-gather or farming). "Work" and "life" didn't need to have the same distinction - they could blend together into one occupation. Due to the monotony of office work, lacking a work <> life distinction is immensely draining.

Pre-industrialisation, you could have periods of intense work - eg. harvest - followed by periods of rest - winter. You always knew that the work intensity would go up and down, but the predictable cadence reduced the monotony.

So, how does this relate to startups?

Biologically, founders are the same as farmers.

As a startup founder, you need to be across everything: sales, product, marketing, investment.

But working hard consistently, on all tasks, is the path to burnout.

I propose that startup founders need to introduce the concept of seasons to their work.

Have a season of "product", where your mental focus is given as much as possible to product and feature improvements. The season ends once a logical "feature release" has been reached (and set this date at least 12 weeks in advance).

Follow this with a festival - it has to be a proper celebration with your team, partner or other cofounders - and during this time you can take a short holiday break. Then with the change of season, focus on something else: marketing, sales, investment - but it's got to be something different to your previous season. No two adjacent seasons should be on the same focus.

The variety is important - punctuated by periods of intentional rest and celebration.

And it's OK to have periods of incredibly intense work, followed by periods of lower intensity - just make sure that you're disciplined about when those periods start and finish.

I like the seasonal concept because seasons don't have a fixed end *date* per-se - sometimes winter comes early, sometimes autumn is late, summers can be long.

But equally, the changing of the seasons is inevitable. Autumn can be a week or two late, but it's going to come.

Use this concept in your work-life as a founder - you don't need to beat yourself up over missing that deadline: the launch, the sales target etc. But you do need to get the crops in before winter comes. And the prospect of missing the change of seasons should scare you as existentially as it did your farming ancestors.

I think this release highlights a massive issue for GenAI companies like OpenAI - the solution looking for a problem.

With each injection of funding, the biggest challenge for AI companies is rapid adoption - OpenAI's valuation is built off the assumption they can grow revenue fast and it's clear from the outside that a big part of that revenue is going to have to come from direct users (as opposed to API revenue from builders on their platforms).

Many AI companies (OpenAI especially) are busy building new playgrounds for users to play in rather than focussing on integrating into existing workflows.

As a developer, I don't want to have to forgo all the features of my IDE just to access the AI goodness - it's good, but not *that* good. As it stands, Canvas is a good tech demo, but I'm having to give up too much in return.

I want (side note: user's aren't always right :P) a coding assistant the deeply integrates with my existing IDE, not a new place I need to go to edit code - it's duplication I don't need and raises the bar for feature requirements given I now need the new AI thing to also have all the must-have features of my IDE.

A far more useful tool would be an extension to an existing IDE - like VS code or JetBrains. Even better is if I can have an actual IDE that also includes the AI features eg. Cursor.

Focus on customers a fundamental competitive edge that small startups have over the big players. By having fewer resources available, you no option but to to focus on the customer and their actual needs if you want to grow. Focussing on the customer is what ultimately leads to building the billion dollar businesses.

Don't reinvent the wheel - make a tool that brings the magic of AI into user's existing workflow. This doesn't mean it has to be in their existing toolset, just that users shouldn't be expected to give up a whole raft of functionality in order to get access to your AI features.

I think a lot of AI companies have underestimated how entrenched existing workflows are. Even a tool that magically does everything still needs to overcome the incredible inertia in most customer organisations beyond the early adopters.

There are lots of reasons open source exists, but I’m going to look at this solely from an economic perspective.

Open source software exists because the marginal cost per user of software at sufficient scale is (nearly) 0, and as a result in a competitive market firms will produce where marginal cost = marginal revenue -> marginal revenue is $0 and the software is free.

So any competitive market for software will essentially tend to the point where the price of the software is practically $0, and producers will need to make money doing something else (like consulting) in order to stay alive.

This explains why so many libraries, programming languages and databases are open source and free - the switching costs are relatively low, and there is lots of competition with limited barriers to entry. "Firms" (in this case open source developers) are attracted by altruistic value and/or kudos and/or the belief that their software is somehow better and they want to change the world for the better.

If the makers of python suddenly decided to charge $10/user, the user base would drop massively and lots of people would switch to another language.

Software can be charged at non-zero revenue where the product has monopolistic tendencies - for example, where there are agglomeration benefits in everyone using the same product to allow them to collaborate (eg. Microsoft Office), or if the software contains some sort of non-trivial technical or usability innovation that is difficult to replicate (product-led moat).

The middle ground between these two is services - where you have a fundamentally free software product, but the effort required to get it working for is significantly non-zero. This creates an opportunity for a business to package that software into a "service" which is then sold at a huge markup.

If your software product has existing free competitors and you have no way to service-ise the business, you’re going to have a very hard time building a business around it.

The flip side of this is that in software, we see more knowledge sharing and technological catch-up than in other sectors, which in turn fuels even more innovation.

We’re seeing this in the AI space at the moment with tools like HuggingFace drastically reducing transaction costs and improving access to AI models. The most interesting space in AI at the moment is Open Source models - free and easy access to this technology fosters innovation and invention.

My prediction is that the real breakthroughs in AI will come from the open source community (and those participating in it), rather than from closed-source businesses like OpenAI.

Either way, it’s going to be a fascinating space to watch and be a part of.

In this week’s post, I’ll explore what practically $0 marginal costs means in the markets for software products.

Marginal Cost of Production

The Marginal Cost of Production is the cost a producer experiences in order to product one more unit of a good or service.

In traditional industries, the marginal cost is made up of

• The resources needed to make the product

• The labour required to produce and supply the product

• Transport costs required to move the product to where it is needed

• Any taxes / tariffs involved in supplying the product (which idealistically exist to correct any externalities arising from the first 3 points)

Importantly, marginal cost doesn’t include the fixed costs required to be able to produce goods in the first place. Things like having a factory, a distribution network etc are all fixed costs that, while they may impact the final price of the good, don’t impact the marginal cost of the company producing and selling another unit.

Software is different to a traditional good - software can be infinitely replicated with exact copies, there’s no steel needed, nothing needs to be made in a factory and thanks to the internet there are virtually no transport costs (moving electrons is waaay cheaper than moving physical goods). The fixed costs are high, and the public utilities required to supply the goods are immense (all of the infrastructure of the internet, the legal infrastructure required to police licensing), but importantly, the marginal cost of another unit of the good being produced basically boils down to the cost of producing, moving and storing electrons in electric circuits.

This has really interesting implications for markets for software products.

Time for an economics lesson.

The profit maximising condition in a competitive market occurs when marginal costs (MC - the cost of producing and selling an additional unit) are equal to marginal revenue (MR - additional revenue the firm receives from selling an additional unit). That is, where the cost of producing one more good is equal to the benefit a consumer receives from buying that good (and therefore the amount they will pay for the good and the revenue the firm receives).

In a perfectly competitive market, if a producer tries to sell goods at above marginal cost, they will sell nothing as the sale price will exceed the benefit a consumer receives from buying the good.

If marginal cost exceeds marginal revenue, the producer would receive no benefit from selling that good in the market and would therefore be better off producing less (and not producing a good that it could only sell at less than production cost).

If marginal cost is less than marginal revenue, the producer could produce one more unit and sell this unit to increase profit.

Supply and Demand

(While there is an argument in economic circles that there is no such thing as supply curves - there’s a whole deep dive we can do around this and what it means for software … but not for now).

Classic supply and demand curves you might be familiar with: demand slopes down to the right. Supply slopes up to the right. Competitive market equilibrium occurs where Supply = Demand.

But software is different! It’s not a classical good.

Competitive Markets for Software

In a competitive market, the market supply is the sum of the supply of all the firms in the market - meaning that in software markets, the supply curve is downward sloping and asymptotes to $0.

Side note: the supply curve asymptotes to $0 because there will always be some fixed costs of producing and distributing software. Better and cheaper distribution systems mean that the cost can get even closer to $0 (which is why Github et al. helps in the proliferation of open source software).

So for MR = MC (the point where the supply and demand curves intersect) in a competitive market, in the limit, the competitive market price is (practically) $0!

In a competitive market, margins would be expected to be super slim, and the businesses making these products would expect to charge basically $0/user.

This hyper-competitiveness is why open source software exists - it is competitive to the point that the marginal cost per user is basically $0 and the sale price is free.

But most software companies enjoy some level of monopoly power

In a monopolistic market, the profit maximising condition also occurs where marginal costs are equal to marginal revenue, but because there is no competition, the monopolist faces the market demand curve, meaning that increasing quantity sold will require a decrease in the price. This decrease can offset any increase in revenue from the additional product sold, meaning the monopolist stands to make more by selling less (at a higher price).

All software that is paid has some level of monopolistic market tendencies - agglomeration benefits from being able to work with others (eg. Microsoft Office), the incumbent product in a market where training staff is a significant cost (eg. Adobe products, also Microsoft Office) or regulated markets where significant regulatory burden exists (eg. software for hospitals and banking).

There is a middle ground between monopolistic software and free software: services + software. This is where you have a software product that is fundamentally free, but the effort required to get it working / customised for a particularly use case leads to the creation of services businesses providing the setup / customisation for a (often significant) fee.

Sorry for getting a bit technical in this post - I promise it will pay off in future posts where we will explore some of the classic SaaS behaviours through this lens of marginal cost and monopolies.

• Anti-competitive behaviour like supply subsidies

• Marketing to manipulate demand for a particular product

• Great product design to increase value

Stay tuned and thanks for reading 🙂

What is SaaS?

A Software as a Service (SaaS) business provides on-demand computer software as its main product. SaaS is the foundation of the business strategy behind nearly all enterprise software companies.

The service component differentiates the licensing and ownership model - with SaaS products, customers pay for access to a software application hosted on the service providers infrastructure, often through a thin web client.

This business model has been made possible (on an industry wide scale) due to the commoditisation of server infrastructure through cloud computing, drastically reducing the barriers to entry for large-scale hosting.

The key component of a true SaaS business is zero or near-zero marginal costs - we’ll explore this in more detail in a future post, but essentially this means the firm doesn’t need to pay for extra inventory or extra labour in order to serve an additional customer. This is a big deal when we look at the economic model and assumptions for a typical startup business.

85% margins! Holy economic rents batman!

Nothing says economic rents like 85% margins, yet many SaaS businesses boast margins in this order - what’s going on?

While SaaS businesses may enjoy near-zero marginal costs, they face huge upfront development costs.

The development of most SaaS applications requires many months or years of work by highly skilled engineers - these don’t come cheap.

These upfront development costs form a very strong barrier to new firms entering the market and copying an existing SaaS product.

SaaS business customers are often sticky - getting humans to change their behaviour is hard, and getting organisations to change their behaviour is even harder. This represents another barrier to entry as a new entrant needs to show that their product is substantially better than existing firms if they want to attract customers - sometimes even years after a new market entrant, incumbent firms still enjoy customer retention simply due to legacy and habit.

Key disclaimer - “at scale”

The reality for most SaaS businesses is that they are still in the “very expensive years of development” point on their journey, and haven’t got to the “enjoy economic rents” bit (relatively few currently operating SaaS businesses are genuinely hitting 85% margins).

These businesses see very high fixed costs - despite commoditisation, cloud computing is still pretty hard to set up (requiring people who know what they’re doing).

At scale just means that once a company has enough customers, the near-zero marginal costs mean the fixed costs component of goods sold (or services provided) asymptotes to 0.

This is very handy when you have expensive software engineers and executives to pay - get enough customers and who cares what your fixed costs are!

Also very handy when you’re calculating future expected value of a business: no need to worry about those pesky costs - just pure ARR to the moon.

Why do we care?

The ability to charge economic rents (or the theoretical ability to do so at some point in the future) is a huge driver behind company values for SaaS businesses. High revenue multiples flow from the simplification that a high margin SaaS company should be so profitable that revenue is almost as good a predictor of underlying value as profit.

SaaS margins are basically why Venture Capital (at least, those focussed on SaaS startups) exist - it’s a bit circular, but high margins lead to a high valuation which in turn means a good exit for the VC.

Not everything is SaaS, and not everything can have 85% margins. It is nearly impossible to achieve margins this high in anything but software - software is unique because the value created is very detached from the cost of production. Software is ultimately just electrons in logic gates, and the resource cost of using a piece of software can be as cheap as running a lightbulb.

Obligatory AI mention

Like pretty much everything, AI could revolutionise the space - AI doesn’t have practically 0 marginal costs, so it’s possible that AI startups will move away from the “85% margins” focus in the future.

This is going to be a very interesting thing to watch in the coming years.

My feeling is the margins will end up somewhere between full-software and service-based businesses, with scaling difficulties also being somewhere in between. Could AI mean the end of the (effectively) zero marginal cost per user SaaS?

Financial products all the way down.

Founders offer a financial product to a VC - in this case, a very very high-risk, very high reward bet. Basically if you give me $1m, I will return you $100m, but there is a very low chance that I will do so, and most likely I will return you nothing.

Some investors like to buy these highly risky financial products when the odds are very long. They invest relatively small amounts early in startups where they believe in the ability of the founders to build a big business. These investors often will have a specific thesis about founders - they’ll invest in founders that remind them of themselves, come from a certain background or are pursuing a specific industry. This sort of investment enables the entire sector to exist, but a portion of this space is “for the love” - not the usual motivation for buying a financial product!

VCs package up risky financial products (startups) into a “fund” - essentially a bundle of really risky bets that taken together are able to spread the risk enough to offer relatively (compared to investing in an individual startup) reliable returns. These funds are themselves a financial product - though investors in these financial products will invest based on the historical returns of previous products offered by the VC firm and ahead of the fund’s investment in a specific startup.

VCs in turn sell their financial products to limited partners - these could be high-net-worth individuals, family offices and pension funds (super funds in Australia).

So - startups are financial products, packaged into other financial products, which are in turn packaged into other financial products, and ultimately sold to rich people and, perhaps, you.

Subscribe now