đ Hidden Economics of the AI Boom
AI infrastructure bottlenecks, economic risks, and the real limits of the current AI boom.
Iâm Ivan. This is a weekly founder-first market intelligence brief. A faster way for founders & investors to understand startup markets.
This weekâs sponsor is AI CRM Attio:
Attio is the AI-native CRM for the next era of companies: Connect your email and calendar, and Attio instantly builds a CRM that matches your business model.
Instantly prospect and route leads with research agents
Get real-time insights from AI during customer conversations
Build powerful AI automations for your most complex workflows
Hello there!
It has been a massive quarter for AI, with a record breaking number of >$100M deals (the list of 40+ deals + 3 dominant categories here). So massive that I keep thinking about historical parallels to what might be (really) going on in our economy.
So, on this very topic, I ran into this must listen interview between Paul Krugman (Nobel-winning economist, NYT) and Paul Kedrosky (investor, MIT research fellow, long time tech + markets nerd).
It is one of the most contrarian and interesting conversations Iâve come across.
Here are the top 10 insights + my 2 cents:
1. The AI boom looks a lot like Dutch disease
What happened:
Krugman and Kedrosky frame AI as a classic âDutch diseaseâ story.
The details:
Dutch disease is what happens when one sector gets so hot that it quietly starves everything else of capital and attention. In the 70s, Dutch gas exports pushed wages up, strengthened the currency, and made every other industry less competitive. The gas boom was real but it hollowed out manufacturing because capital, talent, and policy chased the âeasy returnsâ.
We see capital, talent, and political focus are being sucked into this one AI vertical today. I see it every time I talk to founders, LPs, VCs, banks, you name it (to the detriment of those building âoutsideâ of being a âpureâ ai company, for now).
So what:
When an economy gets Dutch disease founders donât feel it as a âmacro phenomenonâ but they do feel it as harder fundraising, more hype expectations (as youâve probably listened to in a most vc podcastsâŚ) and less (perceived) room to build outside the narrative.
2. We already burned through the Saudi Arabia of data
What happened:
Transformers worked so well because we discovered a giant free data reservoir: the open internet.
The details:
Large language models got good by training on public text at internet scale. Once we found a way to turn raw text into predictions, scaling laws kicked in: more data, more compute, better models. That trick got us from âthis is a cat?â to GPT 5 level systems today. The problem is that most of the high quality public data is now used and new gains need way more work for smaller jumps. Thereâs a must watch interview on this topic by Ilya (co-founder of OpenAI) on this topic which I summarised here.
So what:
The easy part of the curve is behind us and it looks like the future progress is less âjust add more dataâ and more âwe need new sources, new architectures, and smarter trainingâ. Which (might be?) slower and potentially more expensive.
3. LLMs are great at code, less great at the rest
What happened:
Kedrosky is very skeptical that current LLMs are on a straight road to AGI.
The details:
Models are incredibly good at software because code gives a sharp learning signal. Natural language is fuzzy: you and I can disagree on which sentence is better. So the model learns much faster from code than from prose. Benchmarks that show models âcrushing coding tasksâ can mislead people into thinking we are close to general intelligence. Kedrosky argues that we are not and that we are close to âvery good stochastic programmerâ.
So what:
Betting on AI to keep eating software workflows (and in some sectors, parts of labour) is far more realistic than betting on near-term âgod-level intelligence.â LLMs have clear strengths (i.e. code, text, structured tasks) and clear ceilings. The architecture is biased toward certain domains and plateaus in others.
4. GPUs age like fruit, not factories
What happened:
The chips inside AI data centers do not behave like normal industrial capital.
The details:
Training runs keep GPUs at full throttle, 24/7, under heavy thermal stress. That is not like your laptop but more like running a race car flat out all season. What ends up happening is high early failure rates, slowdowns, constant replacement etc etc.
Which means that in a 10,000 GPU data center you should expect regular failures every few hours. Which means that long before âthe next generation chipâ arrives, a lot of your fleet has already worn out.
So what:
AI infra is not laying railroads you can use for 50 years (thereâs been a lot of analogies on this lately). It is closer to a warehouse full of bananas, which matters for depreciation, returns, and how fragile the whole buildout might be.
5. Power (not compute) is the real bottleneck
What happened:
We are hitting the limits of the electrical grid faster than the limits of model size.
The details:
Data centers want hundreds of megawatts and even gigawatts of power. These need the kind of power normally reserved for aluminum smelters or small cities. The problem is that the grid was never built for this. Substations are full, transmission lines take years to approve and local communities push back when they hear âyour bills might go up.â So utilities do strange things:
speculate on power for future AI sites
cancel projects when the demand slips
push hyperscalers into âbring your own powerâ deals
slow-roll approvals because the grid canât move at startup speed
Which ties into the broader idea of fragility in this boom, because the physical world moves slower than the hype cycle.
So what:
You can raise a billion dollars in a quarter but you canât build a gigawatt in a quarter and if your AI forecast assumes infinite, instant electricity, you are not modelling reality. It is basically modelling a wish, and this makes the economy more fragile.
6. The financing stack looks like the âsum of all bubblesâ
