Sam: OpenAI just announced they've secured 10 gigawatts of AI compute capacity in the United States — years ahead of their original timeline. To put that in physical terms: 10 gigawatts is roughly the output of 10 nuclear reactors. This isn't a projection or a roadmap item. They're saying it's done. And on the same day, the UK AI Security Institute confirmed that GPT-5.5 can autonomously solve full network attack simulations. Those two facts together tell you something important about where we actually are.

Priya: Welcome to AI Revolution, I'm Priya Nair, joined as always by Sam Kim. It's Friday, May 1st, 2026, and we have a genuinely dense news week to unpack. Compute infrastructure hitting milestones years early. A government body confirming autonomous cyber offense capability in a publicly available model. The White House blocking a frontier lab from expanding access to its most powerful system. And Anthropic potentially closing a round at a 900 billion dollar valuation within two weeks. We'll also get into a real interpretability tool that's moving from research curiosity to production, Elon Musk testifying under oath about how Grok was trained, and where agentic infrastructure is actually landing. Let's get into it.

Sam: So the 10 gigawatt compute story — let's actually think through what this number means technically. Compute capacity for AI is ultimately a power story. The bottleneck on how many GPUs or TPUs you can run, and how fast you can train or serve models, comes down to how much power you can deliver and cool. 10 gigawatts is an enormous figure. For context, large hyperscale data centers today might draw 100 to 500 megawatts. OpenAI is talking about an aggregate capacity that's orders of magnitude beyond that. What this enables is simultaneous large-scale training runs and inference at a scale we haven't seen. When you have this much compute headroom, you can run multiple frontier training runs in parallel, you can afford to do extensive post-training work — RLHF, Constitutional AI-style fine-tuning, red-teaming loops — without those activities competing for the same resources as your main training job.

Priya: And reaching this years ahead of schedule tells you something about how capital and political will are aligning right now. Building data center capacity at this scale isn't just a money problem — it's permitting, grid interconnection agreements, cooling infrastructure, physical construction. The fact that OpenAI pulled this forward suggests either the regulatory pathway got smoother, or the capital deployment was so aggressive it compressed timelines, or both. What I keep coming back to is the strategic implication: this kind of moat is cumulative. More compute means better models faster, which means more revenue, which funds more compute. The feedback loop is real.

Sam: Right. And the timing connects directly to the next story, which is the UK AI Security Institute's findings on GPT-5.5. The UKAISI ran GPT-5.5 through autonomous network attack simulations — these are structured environments where a model has to identify vulnerabilities, chain exploits together, and compromise systems without human guidance at each step. GPT-5.5 performed at a level nearly matching Anthropic's Claude Mythos. That's a real, autonomous offensive cyber capability confirmed by an independent government body — a meaningful step beyond benchmark results on curated datasets.

Priya: And the asymmetry here is important. Mythos is restricted to a small group of approved users. GPT-5.5 is shipping in ChatGPT and through the public API right now. So you have two models at roughly the same capability tier for autonomous cyber offense, and one of them is broadly available. The defensive side of the security community needs to be thinking seriously about what that access curve looks like. Attackers can iterate on a capability that's now publicly available. Defenders need to move accordingly.

Sam: Which flows directly into the White House blocking Anthropic's plan to expand Mythos access. Anthropic wanted to bring roughly 70 additional companies onto the platform. The White House said no, and the reported reasoning is a combination of compute capacity constraints and national security concerns. The compute angle is interesting — it suggests the administration views Mythos inference capacity as a finite strategic resource, not just a commercial product. There's an implicit argument that running Mythos at scale for 70 more enterprise customers could strain the capacity available for priority use cases.

Priya: This is genuinely new governance territory. A government directly intervening in a private company's customer expansion decision for a software product — and the justification being partly about compute scarcity — that's a precedent. It signals that the most capable AI systems are being treated more like controlled dual-use technology than like cloud software. Whether that framing holds as more models reach similar capability levels is the open question. Because if GPT-5.5 is publicly available and matches Mythos in these tests, the access control strategy starts to look difficult to sustain.

Sam: Quick beat on the Anthropic valuation: sources are saying a 900 billion dollar-plus round could close within two weeks, with investors asked to submit allocations inside 48 hours. For comparison, that's larger than the market cap of most Fortune 50 companies. Investors are pricing in the expectation that frontier AI capability translates to durable, defensible revenue at extraordinary scale. Whether that's right is a separate question, but the capital conviction is unambiguous.

Priya: Now let's talk about something that I think deserves more attention than it's getting: Goodfire's Silico tool. This is a mechanistic interpretability tool that lets engineers inspect and adjust model parameters during training, not just after the fact. So to explain what mechanistic interpretability actually is — the core idea is that instead of treating a neural network as a black box and only observing its inputs and outputs, you try to reverse-engineer what specific circuits, attention heads, and feature directions inside the model are actually computing. You're asking: which components activate when the model processes a particular concept, and can we trace causally how that leads to behavior?

Sam: Previous interpretability work was largely post-hoc and observational. You'd train a model, then probe it to understand what had happened. What Silico is claiming to enable is intervention during training — you can identify that a particular cluster of features is encoding something undesirable, and adjust the training process in response while it's still happening. That's the jump. It's moving interpretability from forensics to active feedback in the training loop. For production model builders, this matters because right now a lot of safety and alignment work happens through RLHF and preference data, which shapes behavior indirectly. Mechanistic tools that let you locate and adjust specific internal representations more directly could give substantially finer-grained control.

Priya: The honest caveat is that this is still early. Mechanistic interpretability at scale — on models with hundreds of billions of parameters — remains genuinely hard. The techniques that work cleanly on smaller models don't always transfer. But the direction is right, and having production tooling rather than just research code is a meaningful step.

Sam: Elon Musk testified under oath this week that xAI used OpenAI model outputs to train Grok. He characterized distillation — training on outputs from another model — as standard practice across the industry. And he's not wrong that it's widespread. But the legal question is whether using a specific competitor's model outputs, potentially in violation of terms of service, constitutes IP infringement or fair use. This is the first time someone at the CEO level has testified to this under oath, which makes it a landmark moment for how distillation gets adjudicated legally. The outcome could set precedent that affects how every frontier lab approaches data sourcing.

Priya: Briefly on two applications stories worth flagging: Stripe's Link wallet now supports AI agent transactions with user-defined approval flows. This is payment infrastructure built natively for agentic workflows — agents can connect to financial accounts and transact within boundaries the user sets. The authorization architecture matters here. Clear human-defined approval thresholds before an agent spends money is exactly the kind of constraint that makes agentic commerce viable rather than chaotic. And Meta deployed unified AI agents to autonomously detect and resolve performance issues across its global infrastructure. Self-optimizing systems at hyperscale, in production. That's a concrete proof point for agentic AI in operations.

Sam: And one quick note on Anthropic's BioMysteryBench: they published a benchmark claiming Claude matches human expert performance on real bioinformatics problems. The benchmark design looks thoughtful — problems drawn from actual research contexts rather than curated textbook cases. But it's a self-published benchmark from the model provider, so independent validation matters before drawing strong conclusions. Worth watching for third-party replication.

Priya: So stepping back to what this week actually points toward — the compute and governance stories are converging on something important. We're entering a period where the most capable AI systems are being treated as strategic infrastructure, not consumer software. Compute capacity, model access, and national security framing are all getting entangled in ways that will shape who builds what and who can use it.

Sam: The cyber capability story is the one I'd be watching most closely. When an independent government body confirms that a publicly available model can autonomously execute network attack simulations, the question isn't whether this capability exists anymore. The question is how fast defensive tooling, detection systems, and organizational security practices can adapt to a world where that capability is broadly accessible. That adaptation race is happening right now.

Priya: And the interpretability thread — Goodfire's Silico — is the technical story I think has the longest tail. If mechanistic tools mature to the point where you can reliably locate and adjust specific behaviors during training, that changes the fundamental architecture of how safety and alignment work gets done. Worth keeping close tabs on as the techniques scale.

Sam: That's AI Revolution for Friday, May 1st. If you want to go deeper on any of these stories, the links are in the show notes. Have a good weekend, stay curious, and we'll see you Monday.

Priya: See you then.