I keep coming back to whether compute markets are an investable category or mostly riding temporary excitement around AI. The Blackrock article got me thinking... The bear case is that compute should get much cheaper over time. Early rail travel was expensive because the infrastructure was scarce and expensive to build out. Over time, more railways were built and transportation became much cheaper. Compute may follow a similar path: GPUs are expensive today because demand has grown faster than supply, while power, data centers, networking and chips remain constrained. But hardware keeps improving, models keep getting more efficient and more capacity is being built. That should make a given amount of intelligence much cheaper. Tasks that required an expensive frontier model a few years ago can increasingly be handled by smaller models at a fraction of the cost. If that continues, generic compute starts to look more like a commodity. That creates a pretty credible bear case for many compute markets: - scarcity premiums shrink as supply catches up - GPU hours become easier to source and harder to differentiate - more inference moves onto phones, laptops, cars and other local devices - large cloud providers continue building their own capacity If a marketplace works mainly because someone desperately needs an H100 this week and cannot find one elsewhere, I would be skeptical of the long-term economics. That feels more like monetizing a shortage than owning something durable. The part I am less sure about is whether cheaper compute actually leads to less spending on compute (the reverse should be true via Jevons paradox). Historically, lower computing costs tend to create much more usage. Storage became cheaper and we started saving everything. Bandwidth became cheaper and the internet moved from basic text pages to people streaming video for hours every day. AI could follow the same pattern...if the cost of running an agent falls 90%, a company will probably deploy far more agents and have each one do much more work. This gets especially interesting because software can consume compute at a rate humans cannot. A person might make a few dozen requests to an AI product in a day. An agent could make thousands of model calls while completing a single job, continuously checking information, testing different outcomes and retrying work without anyone sitting there prompting it. So you can have compute getting much cheaper while total demand keeps growing. The price of a unit of compute can fall dramatically while the number of units consumed rises even faster. If that happens, compute markets do not necessarily disappear. The stronger businesses may be the ones solving harder problems around compute: finding large blocks of capacity when they are needed, matching different workloads to the right machines, keeping hardware highly utilized or giving customers reliable performance across different providers. There is also a chance the bottleneck just moves. Today everyone talks about GPU shortages. A few years from now, GPUs could be much easier to get while power becomes the scarce input (h/t Ribbit). Then the valuable asset may be a data center with a large grid connection. Compute has much less obvious limits because cheaper intelligence can create entirely new uses. If compute becomes 100x cheaper, companies can afford simulations that make no economic sense today. Agents can work continuously instead of being used occasionally. Software can make decisions and transact far more frequently than people do. There are still plenty of ways this thesis can be wrong: - better algorithms may reduce compute needs faster than new demand appears - small local models may take a large amount of work away from centralized data centers - some tasks may reach a point where more intelligence has very little economic value - large cloud providers may also capture most of the market themselves. If cheaper compute leads to far more volume and the marketplace has something harder to replicate than access to GPUs, it might be a much more durable business. Until I see that, I remain skeptical. DMs remain open if you want to prove me wrong!