Karthik Senthil

Mar 5, 2026

Here are the secular themes I’m most interested in that I expect to play out over the next several years with ideas on how to express bets against them:

AI Infra & Frontier Labs

1. Tokens will eat the world. Yes finance will move to being crypto rails, but more importantly and significantly, inference is absolutely going to cook. The amount of tokens used today will look like child’s play 5 years from now. The % of APIs that invoke models (current penetration % is low) and demand for “only possible with AI” things (e.g agents, generative UX + content, robots) will be both up only. While I suspect there will be compute + energy supply shocks along the way, all indications suggest this is a solvable problem. The world will need:
   1. Highly liquid, price-efficient compute markets that maximize utilization of GPU clusters (especially important as agents proliferate)
   2. Cost per token optimizations at the chip, model (eg quantization/distillation), and app layers (memory). Cost per token has largely been subsidized by labs and AI companies in a race to acquire users and pump ARR (similar to consumer software companies in the ZIRP era). This won’t last forever.
2. The leading AI supply chains (model + compute + chips) become nationalized and censored in the name of “national security”. We’ve already begun to see this with Anthropic vs. OpenAI vs. US government. China will eventually have purview over TSMC/Alibaba/Deepseek/etc even if they don’t nationalize it formally.
3. AI backlash will skyrocket. Datacenters will be vandalized, AI policy will be heavily politicized, Sam/Dario/Elon will become public enemies (10X worse than Zuck at his tech bro peak). People are going to get sick of hearing how AI will replace their jobs, and then even worse when they experience it as employees from these companies become insanely wealthy.
4. Public distrust of AI labs is building the same way it did with Wall Street pre-2008. Labs are getting nationalized, Sam and Dario have become household names for the wrong reasons, employees becoming obscenely wealthy while everyone else gets told AI will eat jobs. I don't know if the “Lehman moment” in this analogy looks like a model causing mass harm, a lab getting caught lying about safety/AGI, circular financial deals coming home to roost, etc.

Investment Opportunities:

• Credibly neutral, “always on” AI - there’s an opportunity for a genesis-like moment for a decentralized trained open source / open weights model that users can opt into and acts as the counter weight to frontier labs (similar to how Bitcoin became the counterweight to the financial system). In the ideal, this is built and designed by a group of Satoshi-like Internet anons with a similar mining-like algo that incentivizes people to provide compute/data/RL/etc to continually to fine-tune this model over time. The hard part is getting to reasonable parity with closed-source models but the tech to do distributed pre-training via model parallelism is getting close-ish. Im particularly bullish Plularis because they possess the fundamental building blocks to enable a platform to create decentralized OS models via model parallelism (trading) and model tokenization (durable economics where inference is only possible here).
• Liquid compute markets - Agents will need dynamic and persistent access to compute at the best price execution similar to how humans require electricity. They wont be signing long duration contracts for compute the same way that frontier labs do with datacenters today. This will require a marketplace between those who need compute and those who have compute. On the supply side, compute will be aggregated from existing clouds and long-tail hardware (miners, over-provisioned clusters, etc), priced based on demand, and continuously be routed to the best available demand. To bootstrap supply before organic demand hits, it can borrow the DePIN playbook (early suppliers get token upside to take on early risk), but unlike DePIN, the incentives must have a mechanism to wind down quickly and run off actual demand. On the demand side, compute will be metered per token or per second this market becomes the utility broker for compute and takes a spread similar to an exchange. Whatever gets built will need to be agent-native (both on demand and supply sides). Projects like DoubleZero provide a good architecture design pattern for how the compute marketplace’s routing might behave. Prime Intellect is a proj worth monitoring.

AI Apps

1. I’ve become more convinced that Jevon’s paradox is the right long-term mental model for AI usage. There’s probably a bunch of unmet demand today that humans + software simply can’t meet, but can be met by AI. This being said, the ride will be bumpy and we’ll probably experience significant short-term job displacement (not massive layoffs but slowed hiring). The 2nd order impacts here are interesting to explore:
   1. Property value in tier 2 cities bolstered by tech wages that has seen meteoric rises (Seattle, Charlotte, Austin, Phoenix, etc) have a strong probability of round-tripping. Demand will be destroyed by loss of discretionary income, and this will ironically address housing affordability crisis.
   2. NYC will look very different. Young employees making $100K+ have historically driven the cost of rent, food and liquor. What happens if/when that gets decimated? Similar to (a), affordability crisis in NYC might get solved by AI too!
2. The app layer will unbundle. The same way that desktop apps in the 90s unbundled to the browser, mobile apps will unbundle to an AI agent that runs on your phone. Why bother opening up individual apps vs. telling your agent your intent for it to execute? The apps that have real physical network effects will be the big winners.
3. I remain bullish specialized models but less so than I was 6 months ago.I’ve been surprised/impressed by the continued improved capability of base models (eg on METIS) that it’s hard to know what niches/areas that specialized models can durably outperform. It’s still in the best interest of AI apps to build + fine-tune their own specialized models, but specialized models only win if they assume base models 2x in capability every 18 months and compete accordingly
4. I’m most psyched about the unlock AI will provide specifically in biotech, healthcare and education. My thesis is that software was the wrong modality for these sectors because people who work in these fields are obsessed about purpose, not about making $$. Software enables them to be more efficient but not deliver better drugs, health outcomes or learning experiences. AI offers a fundamentally different modality that can deliver on purpose in these fields, which opens up the design space + market for startups to go direct to consumer vs. only being able to sell into the industry against entrenched players
5. Agents reach parity with humans as economic actors within the next decade. Early signals of increased [x402](https://panteracapital.com/financial-rails-of-agentic-commerce/](https://panteracapital.com/financial-rails-of-agentic-commerce/)) usage and protocols like Giza enabling $20M+ to be managed by autonomous agents are an early signal, but what gives me the most confidence here is the activity on OpenClaw’s GH repo. There’s a ton of latent demand for agents.

Investment Opportunities: