The "dogshit economics" of AI according to Cory Doctorow

The "dogshit economics" of AI according to Cory Doctorow   ai philosophy

This is a response to this post. I'll respond to the post point by point, because I think that in his rush to discount something he finds personally distasteful, Doctorow gets his economics and arguments very wrong.

First, Doctorow argues that the current excitement around AI is a massive economic bubble, larger than previous bubbles like the dot-com boom or the Worldcom fraud. A significant portion of the stock market is tied up in a few AI companies that are not profitable and have no clear path to profitability.

I agree that this is a bubble, and I also agree that it's larger than any previous bubble we've seen. I'm not blindly bought into the AI hype.

However, I think Doctorow is overdoing it on the doomsaying a little. The end isn't nigh. I don't think that a significant portion of the stock market is actually tied up with specifically AI companies: the vast majority of these "AI companies" are not pure AI ventures whatsoever, but diversified tech giants like Google, Meta, and Microsoft, possessing substantial existing businesses and thus very much able to weather the impact of the AI bubble popping. That's part of the entire reason they're investing so much in it in the first place: they have a ton of money laying around that they want to put to use. Should AI not deliver as expected, these companies will likely remain solvent. This fact strongly suggests Doctorow's concerns regarding the severity of the AI bubble's impact are almost certainly overblown. However, they're spreading like wildfire on progressive/leftist/anti-AI social media, so that makes them right!

I also disagree that any of these companies "have no path to profitability" — this is a distortion of the truth, as we will see.

Second, Doctorow echoes Ed Zitron's claim that LLMs have "dogshit unit economics" and that frontier models are too costly to produce for anyone to do them in a finacially viable, economically profitable way.

This is just wrong. Contrary to his assertions, costs are low when users pay per token, while still allowing for substantial profit margins that cover inference, past training, and operational upkeep. Here is a very good, in depth breakdown, backed by significant market evidence, of this fact. To quote:

The LLM API prices must be subsidized to grab market share – i.e. the prices might be low, but the costs are high - I don’t think they are, for a few reasons. I’d instead assume APIs are typically profitable on a unit basis. I have not found any credible analysis suggesting otherwise.

First, there’s not that much motive to gain API market share with unsustainably cheap prices. Any gains would be temporary, since there’s no long-term lock-in, and better models are released weekly. Data from paid API queries will also typically not be used for training or tuning the models, so getting access to more data wouldn’t explain it. Note that it’s not just that you’d be losing money on each of these queries for no benefit, you’re losing the compute that could be spent on training, research, or more useful types of inference.

Second, some of those models have been released with open weights and API access is also available from third-party providers who would have no motive to subsidize inference. (Or the number in the table isn’t even first party hosting – I sure can’t figure out what the Vertex AI pricing for Gemma 3 is). The pricing of those third-party hosted APIs appears competitive with first-party hosted APIs. For example, the Artificial Analysis summary on Deepseek R1 hosting.

Third, Deepseek released actual numbers on their inference efficiency in February. Those numbers suggest that their normal R1 API pricing has about 80% margins when considering the GPU costs, though not any other serving costs.

Fourth, there are a bunch of first-principles analyses on the cost structure of models with various architectures should be. Those are of course mathematical models, but those costs line up pretty well with the observed end-user pricing of models whose architecture is known. See…

…

The reason he doesn't see this is that Zitron, and subsequently Corey Doctorow — who relies on his reporting — are trying to determine the necessary profitability potential of AI, as an inherent property of the technology itself, by looking at the contingent finances of current companies doing AI, and more than that, by looking at a hyper-abstracted version of their financials. They're assessing profitability by comparing Open AI, Google, Microsoft, Meta, etc.'s expansive user base to their income from AI services, as well as fixed-price subscriptions from companies like Open AI and Anthropic.

But as the same blog post I quoted above points out in response to a similar objection:

…

But OpenAI made a loss, and they don’t expect to make profit for years! - That’s because a huge proportion of their usage is not monetized at all, despite the usage pattern being ideal for it.

Essentially, the problem with their analysis is that it doesn't take into account that most users are not paying at all, and many subscription models are loss-leaders — i.e., of course they're not making money, or even losing money, on most users on average. To interpret these as indicators of LLM inference's true economic viability is flawed; instead, they reflect a strategic choice to delay profitability. The underlying economics will shift once these companies prioritize profit. So Zitron and Doctorow are pretending that profitable AI inference isn't possible, but there's no reason to think that's the case at all.

The predictable rebuttal, of course, would be that charging even those "relatively cheap" market rates for LLM usage would unavoidably make LLM inference too expensive, deterring the users that are currently enjoying free service. This is false for several reasons. First, again as the blog post I am quoting says,

OpenAI reportedly made a loss of $5B in 2024. They also reportedly have 500M MAUs. To reach break-even, they’d just need to monetize (e.g. with ads) those free users for an average of $10/year, or $1/month. A $1 ARPU for a service like this would be pitifully low.

Thus, users could be monetized in a profitable way without forcing them to pay directly and thus losing them. We can discuss the economics, ethics, politics, and societal impacts of surveillance capitalism — and I for one do agree they are bad — but this is purely a question of whether they can be made profitable, and I think they clearly can.

The same article I've been quoting directly compares the unit prices of AI inference to a similar API service — web search — and found it to be significantly cheaper, meaning that it'll likely be even easier to monetize, and get b2b API business for:

What is the price of a web search?

Here’s the public API pricing for some companies operating their own web search infrastructure, retrieved on 2025-05-02:

  • The Gemini API pricing lists a “Grounding with Google Search” feature at $35/1k queries. I believe that’s the best number we can get for Google, they don’t publish prices for a “raw” search result API.
  • The Bing Search API is priced at $15/1k queries at the cheapest tier.
  • Brave has a price of $5/1k searches at the cheapest tier. Though there’s something very strange about their pricing structure, with the unit pricing increasing as the quota increases, which is the opposite of what you’d expect. The tier with real quota is priced at $9/1k searches.

So there’s a range of prices, but not a horribly wide one, and with the engines you’d expect to be of higher quality also having higher prices.

What is the price of LLMs in a similar domain?

The purpose is just to get rough numbers for how large typical responses are. A 500-1000 token range seems like a reasonable estimate. …

What’s the price of a token? The pricing is sometimes different for input and output tokens. Input tokens tend to be cheaper, and our inputs are very short compared to the outputs, so for simplicity let’s consider all the tokens to be outputs. Here’s the pricing of some relevant models, retrieved on 2025-05-02:

Model Price / 1M tokens
Gemma 3 27B $0.20 (source)
Qwen3 30B A3B $0.30 (source)
Gemini 2.0 Flash $0.40 (source)
GPT-4.1 nano $0.40 (source)
Gemini 2.5 Flash Preview $0.60 (source)
Deepseek V3 $1.10 (source)
GPT-4.1 mini $1.60 (source)
Deepseek R1 $2.19 (source)
Claude 3.5 Haiku $4.00 (source)
GPT-4.1 $8.00 (source)
Gemini 2.5 Pro Preview $10.00 (source)
Claude 3.7 Sonnet $15.00 (source)
o3 $40.00 (source)

If we assume the average query uses 1k tokens, these prices would be directly comparable to the prices per 1k search queries. That’s convenient.

The low end of that spectrum is at least an order of magnitude cheaper than even the cheapest search API, and even the models at the low end are pretty capable. The high end is about on par with the highest end of search pricing. To compare a midrange pair on quality, the Bing Search vs. a Gemini 2.5 Flash comparison shows the LLM being 1/25th the price.

Note that many of the above models have cheaper pricing in exchange for more flexible scheduling (Anthropic, Google and OpenAI give a 50% discount for batch requests, Deepseek is 50%-75% cheaper during off-peak hours). I’ve not included those cheaper options in the table to keep things comparable, but the presence of those cheaper tiers is worth keeping in mind when thinking about the next section…

Objection!

I know some people are going to have objections to this back-of-the-envelope calculation, and a lot of them will be totally legit concerns. I’ll try to address some of them preemptively. …

Surely the typical LLM response is longer than that - I already picked the upper end of what the (very light) testing suggested as a reasonable range for the type of question that I’d use web search for. There’s a lot of use cases where the inputs and outputs are going to be much longer (e.g. coding), but then you’d need to also switch the comparison to something in that same domain as well.

Even if we want to talk about longer output use-cases, like AI coding agents, the longest session I've ever had with qwen-code was 40,000,000 input tokens (98% of which was cached, so didn't need to be sent repeatedly) and 300,000 output tokens. This was a session that lasted multiple days. Using the average prices for Qwen3 Coder 480B A35B on OpenRouter, that would end up being about…. drumroll… $8. I'm an extremely heavy agentic coding user, and assuming I'm using it every day, multiple times a day, based on a reasonable-length session, I'm looking at maybe 100 million tokens. So $16/mo. That's peanuts for the amount of use it gives me, and that's a pretty high end level of AI usage. I think most people would be willing to pay $16/mo for productivity and entertainment software — we already do it a ton with things like streaming services, Microsoft Office, etc. Not to mention that there are cheaper AI models to switch to if price was really a concern for me.

And sure enough, if you construct a first-principles account of how much running a frontier AI model should cost, based on DeepSeek R1 and H100 hardware at market rates, what you find that it's not just the unit economics on token inference that are good — the margins on AI subscriptions are also pretty good:

So to summarise, I suspect the following is the case based on trying to reverse engineer the costs (and again, keep in mind this is retail rental prices for H100s):

  • Input processing is essentially free (~$0.001 per million tokens)
  • Output generation has real costs (~$3 per million tokens)

These costs map to what DeepInfra charges for R1 hosting, with the exception there is a much higher markup on input tokens.

[…]

A. Consumer Plans

  • $20/month ChatGPT Pro user: Heavy daily usage but token-limited

    • 100k toks/day
    • Assuming 70% input/30% output: actual cost ~$3/month
    • 5-6x markup for OpenAI

This is your typical power user who’s using the model daily for writing, coding, and general queries. The economics here are solid.

B. Developer Usage

  • Claude Code Max 5 user ($100/month): 2 hours/day heavy coding

    • ~2M input tokens, ~30k output tokens/day
    • Heavy input token usage (cheap parallel processing) + minimal output
    • Actual cost: ~$4.92/month → 20.3x markup

  • Claude Code Max 10 user ($200/month): 6 hours/day very heavy usage

    • ~10M input tokens, ~100k output tokens/day
    • Huge number of input tokens but relatively few generated tokens
    • Actual cost: ~$16.89/month → 11.8x markup

The developer use case is where the economics really shine. Coding agents like Claude Code naturally have a hugely asymmetric usage pattern - they input entire codebases, documentation, stack traces, multiple files, and extensive context (cheap input tokens) but only need relatively small outputs like code snippets or explanations. This plays perfectly into the cost structure where input is nearly free but output is expensive.

  1. API Profit Margins
  • Current API pricing: $3/15 per million tokens vs ~$0.01/3 actual costs
  • Margins: 80-95%+ gross margins

The API business is essentially a money printer. The gross margins here are software-like, not infrastructure-like.

While going hog wild on AI may incur high costs, considerable utility remains accessible at a sensible price, and the whole point of market economies is to inherently incentivize efficiency (not going hog wild) through pricing. As the article I just quoted says in the conclusion, not all usages of AI are alike in terms of compute cost, and as a result, some applications of it will be economically viable and others won't:

The key insight that most people miss is just how dramatically cheaper input processing is compared to output generation. We’re talking about a thousand-fold cost difference - input tokens at roughly $0.005 per million versus output tokens at $3+ per million.

This cost asymmetry explains why certain use cases are incredibly profitable while others might struggle. Heavy readers - applications that consume massive amounts of context but generate minimal output - operate in an almost free tier for compute costs. Conversational agents, coding assistants processing entire codebases, document analysis tools, and research applications all benefit enormously from this dynamic.

All told, then, I doubt the market would collapse if appropriately priced. Most individuals would likely maintain subscriptions, akin to popular streaming services, or be sold ads, or just use products that use AI under the hood — perhaps without them ever really needing to know or think about it — that are monetized in other ways, which then pay for inference on the back end.

Moreover, AI models are getting cheaper. There have been numerous innovations in cost-effective training and inference, from the massive breakthrough that was mixture of experts, to very important ones for agentic and chat use like prompt caching, to quantization-aware training, and even custom hardware like Cerebras and Google use.

Finally, the assumption of requiring astronomically massive capital outlays for frontier models is misguided. Projects by DeepSeek, Alibaba's Qwen, Moonshot, and various open-source initiatives prove otherwise: according to the first major critic of DeepSeek, SemiAnalysis, they spent $1.6 billion over the life of the company to acquire the GPUs they used to train DeepSeek. Mid-tier streaming and enterprise companies like Netflix and Salesforce typically spent about $0.6 to $0.8 billion dollars (adjusted for inflation) annually prior to 2020, meaning that, worst case scenario, a mid tier tech company could achieve that level of capital in about 3 years, training intermediate models in the meantime, selling inference, or doing one of the very many other things you can do with GPU training time.

While significant upfront investment is necessary, it is neither prohibitive nor unusually large, then: definitely a few times larger than a pre-2020 mid tier tech company, but much closer to that than to the capex of large tech companies like Amazon ($33.8 billion in 2020). to that of a mid-to-small-sized tech company. Thus, while this is likely to be a somewhat consolidated market, the economics of training frontier models are not as truly absurd and economically impossible as anti-AI advocates like Zitron and Doctorow want to make it seem.

Ultimately, the first point stands: LLM unit economics enable profitable sales at competitive prices. The opposing view stems from Zitron's moralistic, distorted, and biased reporting. He disregards actual unit economics, instead focusing on high-level figures easily manipulated when disengaged from the granular details of cost structure, especially regarding user segments where OpenAI might intentionally incur losses.

The third claim Doctorow makes is that the AI industry is fraudulent, because it has circular investments — companies giving credits that are then used to purchase services back from them. This is a bubble signature, yes, but I don't think it substantially indicates that there is no possible market or profitability to AI.

The fourth claim that Doctorow makes is that AI is fundamentally useless, economically speaking — a passing fad or psychosis. He makes this claim by citing a University of Chicago study showing "no significant impact on workers’ earnings, recorded hours, or wages."

This is a really economically illiterate — or disingenuous, depending on how charitable you are — way of looking at things, though, as Cory should himself know: we've long known that it's extemely difficult to measure the productivity gains of even having computers at all, or the internet itself in the workplace — that's the Solow paradox at work — for one thing. But more relevant to his study's specific findings, we also know that due to modern neoliberal capitalism, worker earnings, wages, and working hours usually won't actually reflect productivity increases in the modern economy (this has long been known as the Productivity-Pay Gap) because it'll all be soaked up by either the corporation's bottom line, or its bureaucracy. Cory should know this. Thus the Univesity of Chicago study Cory cites (which is somehow "well known" but not actually cited anywhere, by anyone, according to OpenCitations, lmfao) is likely just not very useful in determining how useful AI actually is.

Similarly, the oft-cited-by-AI-haters Salesforce study which showed a 35% success rate for long-running agents only pertained to B2C and Customer Service use-cases, and when they did expand it to some Workflow Execution related tasks, the success rate for single-turn tasks skyrocketed from 58% to 85%. Additionally, although they don't explicitly say or show the code of the agents they used, they appear to have used home-brewed, likely simplistic agents — not cutting edge agentic systems with automatic type checking, linting, error handling, or anything like that, like e.g. Claude Code and other agentic coding agents do — and assumed that the best use case for agents was running completely unattended, which it simply isn't, but that doesn't mean they're actually useless. Not only that, but they had the agents generating actions/queries in their own custom, obscure query language — two of them, in fact — instead of something AIs are trained well on, like MCP function calling format, JSON, Python, JavaScript, Bash, SQL, and so on.

The MIT study claiming 95% of business AI initiatives failed also doesn't actually say what AI haters think it does: it's not talking about applications of existing, first-party AI systems ike those from OpenAI, Microsoft, or Google, nor even about the application of a business's custom-made tools to their own affairs. Instead, it's talking about the number of custom B2B AI systems that are finding profit by being bought by other businesses — in explicit contrast to how successful first-party AI systems are:

Tools like ChatGPT and Copilot are widely adopted. Over 80 percent of organizations have explored or piloted them, and nearly 40 percent report deployment. But these tools primarily enhance individual productivity, not P&L performance. Meanwhile, enterprise-grade systems, custom or vendor-sold, are being quietly rejected. Sixty percent of organizations evaluated such tools, but only 20 percent reached pilot stage and just 5 percent reached production. Most fail due to brittle workflows, lack of contextual learning, and misalignment with day-to-day operations.

And indeed the main complaint of the study is primarily that, while individual productivity is being increased, AI-integrated products aren't being produced for other businesses and consuemrs and becoming profitable in most industries — not that AI isn't useful within corporations. It crucially also found that 90% of employees within those companies regularly and substantially leveraged AI to boost productivity. Far from demonstrating that AI is useless, all this demonstrates is that businesses can't build products exclusively around it and don't know how to leverage it effectively. That's a very different thing.

Lastly, his "AI as arsenic" concern — that managers, falsely convinced of AI's capabilities, will fire workers, replacing them with AI, leading to long lasting labor force attrition that might damage our economy permanently when AI inevitably falls short — is completely baseless. This perspective is divorced from reality. The assumed mechanism of permanent labor attrition — people forgetting the jobs they were trained for, being re-trained for other jobs and leaving that job market, or being permanently "discouraged" and leaving all job markets forever — also assume that it will take corporations years, if not decades, to realize what Cory considers blindingly obvious even with respect to corporate bottom lines, and completely backed by all available financial, investor, and scientific wisdom that these very same companies are releasing, namely, that AI can't actually replace people.

Not only is this assuming corporate incompetence to a degree even I, an anti-capitalist, can't quite take seriously because of how legitimately self-contradictory it is, since the studies and financial reports he's quoting are, as I said, from these very same circles, recent history shows companies quickly realize AI's limitations, often within a few months to a year, not decades. This isn't just a corporate/market rationality argument: this is literally what we already know happens literally from watching corporations actually do their thing. Gary Marcus has dubbed this "The Klarna Effect" after the first occurance, but it's happened with several more companies (a few examples are listed in his blog post, and Duolingo is another example).

The notion that entire job categories become "lost arts," damaging our economies for decades, also ignores our capacity to train new workers or re-employ those laid off once the market corrects. Labor markets adapt rapidly: substantial openings in a specialty quickly attract new talent. Moreover, a massive resurgence of jobs, initially mismanaged by AI, would stimulate a significant economic recovery as consumer spending rebounds, thus literally counteracting the economic bubble he fears so much!

Corey Doctorow is correct in a broader sense, even if his reasoning is flawed, though. As I said at the start, I do believe there is an impending crash. However, in my opinion, it won't be due to poor unit economics or prohibitive costs for frontier models. Instead, it's driven by ludicrous capital expenditure fueled by the pursuit of currently-impossible Artificial General Intelligence (AGI) or superintelligence — a hype bubble that's parasitic on the actually functional AI technology we possess in the here-and-now. The colossal data center investments (e.g., "Project Stargate") that they're going to struggle to pay back the investments on in the short amount of time they have before they burn through their GPUs and/or the investment depricates irrevocably as they upgrade to new GPUs are not intrinsic to advancing current state-of-the-art AI, and it is this overextension, rather than AI's core utility, renders the bubble unsustainable given the rapid depreciation of GPUs meaning that these investments cannot yield sufficient returns within the short operational lifespan of the hardware.

And honestly, if my analysis is correct — that LLM inference unit economics are robust and new foundation models are feasible, as well as that jobs will be more recoverable, and the stock market less crashed, than he expects — the post-bubble market dynamic shifts profoundly. We would see an abundance of powerful, inexpensive Nvidia GPUs, available for trivial sums for business in AI that can be made immediately profitable on the back of existing open source frontier models (thus doubly discounting the up front costs) just to provide inference, which can then fund more frontier model development as needed down the road. This would catalyze the emergence of numerous new AI companies requiring minimal initial capital for used hardware to build cutting-edge frontier models or run inference at healthy profits, which looks like a healthy and lasting market after the bubble to me, the exact same kind of thing we saw with the dotcom bubble (a hype driven bubble leaving something useful and profitable behind) or the Worldcom fiber optic situation (a bubble leaving behind — in this case, granted, more temporary — useful infrastructure, unless you count all the physical data centers, power stations, etc, that are buing built) to me.

Therefore, the dismissive belief that this AI "bubble" is uniquely fragile, lacking in economic value, incapable of profitability, and destined for instant collapse, seems more borne out of dislike of AI than a rational analysis of the economic conditions, to me.