Diego Oppenheimer

These three show up in every enterprise deployment I've seen. Structured outputs break first. Validation fails silently. The pipeline looks fine until it doesn't. The 8-component framing is useful. More useful signal: this layer is consolidating. Teams that built it internally 18 months ago to ship their own product are now the vendors.

Most retail coverage of tariffs and AI stops at the surface layer: Nvidia costs go up, hyperscalers absorb it, smaller players get squeezed. That's accurate and almost entirely priced in. Here's where I'd actually look for signal. Power infrastructure is the binding constraint nobody writes about: The bottleneck isn't GPUs. It's power. The US imports more than 80% of large power transformers, lead times are already 18-24 months, and 55% of in-service units are over 33 years old. Tariffs hit foreign competitors here but not domestic manufacturers. Companies that build and install power infrastructure for data centers can't be offshored. That's a structural position, not a cyclical one. This category doesn't show up on AI coverage. It shows up on grid coverage. The efficiency premium goes up when compute gets expensive: When frontier compute costs more, the relative value of doing more with less increases. Models optimized for specific tasks, faster inference, edge deployment. These gain ground. The companies solving for efficiency aren't the ones with the most GPUs. Worth paying attention to which direction the architecture is moving. The colocation consolidation play: Larger operators can absorb tariff-driven cost increases. Smaller competitors can't. The sorting mechanism here runs toward consolidation. The players with the most existing capacity and customer relationships benefit from competitors exiting. The middle squeeze in frontier model training: The companies training at frontier scale but not owning their infrastructure are in an unusual position. Cloud-dependent, which insulates them in the short term. But no custom silicon, no owned data centers, no geographic routing. When the tariff costs eventually pass through to cloud pricing, they face pressure with no offset. That lag is months to a couple of years. It's not fully in the current discussion. Don't ask who makes AI work. Ask who makes AI work when compute is expensive. And ask who has a hedge against the cost curve, and who's exposed to it with no structural offset. The power infrastructure story, the efficiency architecture story, the frontier lab middle squeeze. Those don't seem to be priced in.

Noam Brown (OpenAI): single-number AI benchmarks no longer make sense. Right. Performance now depends on how much reasoning time the model gets. A benchmark captures one compute budget. Production runs another. Enterprise buyers are purchasing a number that doesn't describe what they're deploying.

Enterprise data leaders report AI accuracy anywhere from 30% to 85%. That variance isn't about which model they're using. Most of them don't know their actual number. They've never built the feedback loop to find out. The model sounds equally confident either way.

All 5 of my direct reports at Algorithmia are now CEOs. Either I was a great manager or a terrible one. They're too polite to say which.

The founders who are deepest in AI adoption have shifted to a different operating mode. @garrytan, @rwaliany, @mvanhorn, @jheitzeb aren't doing things sequentially. They're running multiple projects at real depth, simultaneously. Serial entrepreneurship exists because execution was the bottleneck. Capital and ideas scaled; the actual work of pushing things forward didn't. You had fixed bandwidth. Serial was the workaround: do one deeply, then the next. What AI removes isn't time. It's the execution overhead between decisions: the research loops, the drafting, the ops work, the synthesis. Offload that and what remains is judgment. Judgment actually scales across multiple projects at once. I'm running it this way myself: investing, advising multiple companies, teaching, sitting on boards, all simultaneously. The "serial" frame assumes sequencing was a choice. It was a constraint. The founders at the forefront of AI adoption are proving it. Parallel entrepreneur is the operating mode that emerges when the constraint goes away. 2026, the year the AI Parallel Entrepreneur was born.

.@ryancarson runs one of the most interesting patterns in agent first company building I have seen out there. Well worth checking out...

In traditional ML, we learned the hard way that error compounds in pipelines. Testing each model in isolation didn't predict what happened when you chained them together. One early paper suggests we're hitting the same wall with LLM agents. Researchers tested 10 frontier models on long-horizon agentic tasks. Memory scaffolds (the standard approach for extending context across steps) degraded performance on all 10. Meltdown rates on complex tasks reached 19%. Model rankings from standard benchmarks inverted as tasks got longer. One study, not a conclusion. But the shape is familiar. If this generalizes, it's the same problem as ML pipelines: evals designed for individual models don't predict system behavior. Worth testing your agents at the actual length and complexity you're deploying against, rather than waiting to find out.

Capital allocation is a cleaner signal than product announcements. Three weeks ago Cisco bought Galileo for agent observability. Now in talks to acquire Astrix for $250-350M. Astrix discovers every agent, MCP server, and shadow identity in your network and scans each for vulnerabilities. Two acquisitions, same thesis, three weeks apart.

Stanford 2026 AI Index: top models are so close on benchmarks nobody's choosing on that anymore. Real differentiators: reliability, consistent instruction-following, cost at scale. We moved from "which model is smarter" to "which model breaks least." Time to pay attention to uptimes.

Honored to have played a small part in this. What Jenni, Abby, Celeste, Alfredo and the AI@GSB team built from a single guest lecture is remarkable. Now the largest student org at Stanford's MBA program. The next generation of leaders learning to build with AI, not just about it.

PwC 2026: top 20% capture 75% of AI gains. What separates them: they measured revenue growth, not cost reduction. Sounds like 2021. It's not. Cost reduction projects are easy to start and easy to abandon. Revenue growth means you knew what you were building before you started.

People's expectations are wild to me sometimes. "heavily guarded military base, wasn't allowed to come in and take pictures," one star.