Sam: Ten gigawatts. Let that land for a second. OpenAI has hit 10 gigawatts of AI compute capacity in the United States — several years ahead of schedule. To put that in physical terms, you're talking about the output equivalent of roughly 10 nuclear reactors, dedicated entirely to training and running AI models. The original timeline assumed this would take until the end of the decade. It's April 2026. The compute scarcity story just got a lot more complicated.

Priya: Welcome to AI Revolution, Thursday April 30th, 2026. I'm Priya Nair, here with Sam Kim. Today is one of those days where the infrastructure story and the policy story are actually the same story told from two different directions — and we need to pull that thread. We've also got OpenAI launching a cybersecurity-specific frontier model that won't be publicly available, Anthropic reportedly heading toward a $900 billion valuation, a 440 megabyte translation model that runs entirely on your phone, and a quiet but important signal that AI evaluation itself is becoming a bottleneck. A lot to get through.

Sam: Let's start with the compute story because it reframes everything else today. When OpenAI announced the 10 gigawatt goal, the assumption in the industry was that physical infrastructure — power, cooling, land, grid interconnects — would be the binding constraint on frontier AI development through at least 2028. The logic was: you can design better chips, you can improve training efficiency, but you can't just conjure gigawatts of reliable power. Utilities move slowly. Permitting takes years. That assumption is now clearly wrong, or at least much less true than it was.

Priya: And what's interesting technically is what 10 gigawatts actually buys you at the inference layer, not just training. We're past the era where the headline number is about training a single large model. This capacity is supporting continuous inference at scale — millions of concurrent requests across consumer products, API customers, and enterprise deployments. The architectural implication is that OpenAI can now run larger, more capable models at serving time without the latency and cost tradeoffs that have historically forced them to deploy smaller distilled versions.

Sam: Which brings us directly to story two, because here's where it gets genuinely interesting. The White House has blocked Anthropic from expanding access to its frontier model Mythos to roughly 70 additional companies. The stated reason is compute capacity constraints within the US. And you have to hold both of those things in your head simultaneously — OpenAI just announced it hit 10 gigawatts ahead of schedule, and the federal government is simultaneously arguing there isn't enough compute to let a competitor's model reach more customers.

Priya: These aren't necessarily contradictory, but the tension is real. OpenAI's 10 gigawatts is OpenAI's capacity — it doesn't transfer to Anthropic. What the White House intervention signals is something more structural: the federal government is now actively making resource allocation decisions about who gets to deploy frontier AI to whom, and at what scale. That's a governance posture that goes well beyond chip export controls. This is domestic gatekeeping of model access based on compute as a scarce national resource.

Sam: The policy mechanism here matters. This isn't a regulatory framework with notice and comment periods. This is direct executive intervention in a commercial distribution decision. And the justification — that we need to preserve compute headroom — implies the government has a view on what that compute should be used for. That's a pretty significant precedent.

Priya: And then you put Anthropic's valuation story next to that and the picture gets even more interesting. Multiple pre-emptive offers at $850 to $900 billion. A potential $50 billion raise. For context, that would make Anthropic one of the most valuable private companies in history, and it's happening right as the federal government is throttling their distribution. The capital markets and the regulatory environment are pointing in completely different directions.

Sam: The revenue signal underneath the valuation is the part I keep coming back to. Anthropic reportedly surpassed OpenAI in LLM revenue. That's not a rounding error in a funding announcement — that's institutional investors reading the enterprise adoption curve and concluding the frontier AI market has real room for two dominant players. The compute constraint might actually be helping Anthropic's revenue story by making their existing capacity more valuable.

Priya: Okay, let's talk about GPT-5.5-Cyber, because this is a different kind of model release and the deployment model is as interesting as the capability. OpenAI is launching a frontier cybersecurity-specific model, and it won't be generally available. It goes first to a vetted group of cyber defenders — the explicit logic being: let institutions harden their defenses before the same capability reaches potential adversaries.

Sam: The underlying technical question is what makes a cybersecurity-specific frontier model different from GPT-5 with a system prompt. The honest answer is probably a combination of things: specialized fine-tuning on security-relevant corpora — CVE databases, exploit code, threat intelligence reports, network forensics — combined with RLHF or preference tuning that rewards responses useful to defenders and penalizes outputs that primarily benefit attackers. Getting that balance right is genuinely hard.

Priya: The controlled rollout is a real departure from how frontier models have typically been released. The standard pattern has been broad access, then iterate on safety based on observed misuse. This inverts that — restricted access first, let defenders build institutional capability, then expand. Whether that actually works as dual-use risk management depends entirely on how tight the vetting is and how long the gap between defender access and general availability actually is.

Sam: Shifting to research — Anthropic released BioMysteryBench, a benchmark built from real bioinformatics problems, and Claude reportedly matches human expert performance on those tasks. I want to be honest about what that means and what it doesn't. The benchmark is built on actual scientific problems rather than synthetic ones, which is a meaningful improvement over a lot of existing evals. But expert-level performance on a curated benchmark is not the same as expert-level performance in an open-ended research context.

Priya: The caveat about generalizability is doing a lot of work in this story. Bioinformatics is a domain with well-defined problem structures — sequence alignment, protein function prediction, genomic data interpretation — and those map reasonably well to what large language models are good at. Whether Claude can handle the messier, less well-specified problems that actual researchers face day to day is a different question. Still, this is a real signal that domain-specific capability is advancing.

Sam: On the hardware side — Google is going to start selling TPUs directly to select customers rather than requiring cloud access. This is architecturally significant. TPUs have always been Google's internal competitive moat, accessible externally only through Google Cloud. Selling them directly opens a new competitive dynamic against NVIDIA in the accelerator market. For organizations building their own AI infrastructure, having a Google-sourced TPU option changes the vendor calculus.

Priya: I want to quickly flag two stories that are easy to underweight. Tencent released a 440 megabyte translation model that runs fully offline on smartphones, supports 33 languages, and reportedly outperforms Google Translate. That's a model compression and efficient inference story — the techniques that get frontier-level performance into 440 megabytes are the same techniques that will eventually put much more capable models on edge devices. Watch that research direction.

Sam: And the eval cost bottleneck piece from Hugging Face deserves attention from anyone who thinks seriously about AI reliability. As models become more capable, properly evaluating them requires more compute, more expert time, and more sophisticated benchmark design. The cost of doing evals rigorously is scaling faster than many organizations anticipated. The practical consequence is that benchmarks you see published today may reflect what an organization could afford to evaluate, not what was actually sufficient to evaluate. That's a real quality signal problem.

Priya: One more quick story — drone strikes have damaged data center infrastructure in the Middle East, and the fact that war damage is uninsurable is forcing major tech companies to pause expansion plans in the region. The geographic concentration risk implication is real: if conflict zones become no-go areas for infrastructure, compute capacity concentrates in fewer, more politically stable regions. That changes who has access to AI capacity and on what terms.

Sam: So looking ahead — the thread connecting today is that physical infrastructure and political control are becoming the actual binding constraints on frontier AI development, not just model capability or training algorithms. We have OpenAI hitting 10 gigawatts early, we have the White House rationing competitor access based on compute scarcity, we have geopolitical conflict literally destroying data centers. The question I'm watching is how quickly compute capacity decouples from any single company or government's control.

Priya: For me, the GPT-5.5-Cyber rollout model is the thing to watch most carefully. If restricted-access deployment for dual-use models becomes the norm rather than the exception, that reshapes the entire competitive landscape for frontier AI — not just in cybersecurity. The companies with existing trusted relationships with government and enterprise customers have a structural advantage in a world where model access is gated.

Sam: A lot of moving pieces today. As always, we'll link everything in the show notes.

Priya: Thanks for listening to AI Revolution. We're back tomorrow.