Sam: Google just split its AI chip strategy in two — and that architectural decision tells you more about where the industry is heading than any benchmark result.

Priya: Welcome to AI Revolution for Wednesday, April 22, 2026. I'm Priya Nair.

Sam: And I'm Sam Kim. Today we've got a lot of ground to cover — Google's TPU 8 dual-chip architecture, Anthropic's Mythos model leaking to unauthorized users after finding 271 vulnerabilities in Firefox, a $5 billion Amazon deepening of the Anthropic relationship, and an AI agent that designed a complete RISC-V CPU core from scratch. Plus some fast-moving industry deals that are reshaping the competitive map.

Priya: Let's start with the hardware story, because I think it's genuinely underappreciated. Google announced two TPU 8 variants at Cloud Next — one optimized for training, one for inference. Sam, walk us through why splitting the chip matters.

Sam: The underlying insight is that training and inference have fundamentally different computational profiles. Training is about throughput — you're doing massive, sustained matrix multiplications across enormous batches of data, and you need high memory bandwidth and deep parallelism over long runs. Inference is about latency and cost efficiency — you're serving individual requests, often with smaller batch sizes, and you care about how fast you can generate a token and how cheaply you can do it at scale. Nvidia's GPU approach has always been a generalist solution. A single architecture with enough headroom to do both reasonably well. Google is saying: we're operating at a scale where the inefficiencies of that compromise are costing us real money and real performance. So they built two specialized chips.

Priya: And the Arm move is significant too. Google is swapping out x86 host CPUs for their own Axion cores across this infrastructure stack.

Sam: Right. The host CPU in an AI accelerator cluster is handling orchestration, data loading, preprocessing, communication between accelerators — it's not doing the core matrix math, but it's everything around that math. And x86 carries a lot of legacy overhead. Arm-based designs can be more power-efficient and more tightly integrated with custom silicon. Apple proved this at the chip level, and now Google is applying the same logic at the data center level.

Priya: What this signals to me architecturally is that the era of general-purpose accelerators as the default assumption is ending. The competitive moat going forward is vertical integration — your training chip, your inference chip, your host CPU, your interconnect fabric, all co-designed. Nvidia has the software ecosystem advantage with CUDA, but Google has the volume and the incentive to build around it.

Sam: Now let's talk Mythos, because there are two separate stories here and they're deeply connected. Start with the Firefox result, because that's the capability demonstration.

Priya: Mozilla ran Anthropic's Mythos cybersecurity model against Firefox 150 and it found 271 security vulnerabilities. Their CTO described it as every bit as capable as the world's best security researchers. To be clear — these aren't vulnerabilities that are beyond human capability to find. But the speed and the coverage are the point.

Sam: This is what AI-powered vulnerability discovery looks like when it's actually working. The traditional model for finding bugs at this depth is expensive, slow, and bottlenecked by the availability of top-tier security researchers. What Mythos apparently does is combine deep static analysis, semantic understanding of code, and the ability to reason about multi-step exploit chains — and run that process continuously across a massive codebase in a way no human team can match in the same time window.

Priya: Which leads directly to the second story. Mythos was accessed by a small group of unauthorized users via a third-party contractor and a private online forum. Anthropic is investigating. They say no core systems were impacted. But the security implications here are serious.

Sam: A model capable of finding 271 real vulnerabilities in a major browser, in unauthorized hands, is a qualitatively different threat than a jailbroken general-purpose model. The cybersecurity capabilities that make Mythos valuable for defense are structurally the same capabilities that make it dangerous for offense. Anthropic knew this — they had flagged Mythos internally as potentially dangerous and restricted access for exactly this reason. The contractor pathway bypassed those controls.

Priya: And this is a structural problem for the industry, not just an Anthropic problem. As frontier labs build specialized models for high-stakes domains — cybersecurity, biology, critical infrastructure analysis — the access control perimeter becomes as important as the model itself. A model is only as safely contained as its least-secure integration point.

Sam: On to the capital story. Amazon put another five billion dollars into Anthropic, and Anthropic is using a substantial chunk of it to secure five gigawatts of Amazon's custom Trainium silicon capacity. The headline reads like circular finance, but the actual dynamic is more interesting.

Priya: It's compute commitment locking. Anthropic is reserving a massive slice of Amazon's silicon roadmap, which gives them supply chain certainty in an environment where accelerator availability is still genuinely constrained. And Amazon gets a major anchor tenant for Trainium — which helps justify continued investment in a chip that needs to compete with Nvidia and Google's in-house silicon.

Sam: Claude demand is apparently driving this. Which makes sense — Claude has become the default choice for enterprise AI integrations in a way that even a year ago wasn't clear. The question is whether Trainium can get close enough to Nvidia performance that the ecosystem tradeoff is worth it.

Priya: Now let's talk about the RISC-V story, because I find this one genuinely fascinating. A startup called Verkor.io claims their agentic AI system designed a complete RISC-V CPU core — called VerCore — entirely from scratch. It runs at 1.5 gigahertz with roughly 2011-era laptop performance.

Sam: Let me give some context for why this is hard. Chip design involves multiple interdependent stages — RTL design, functional verification, synthesis, place and route, timing closure. Each of these has specialized tools and requires domain expertise. Prior work with AI in chip design has largely been about improving individual stages — AI-assisted floorplanning, AI-optimized synthesis. What Verkor.io is claiming is end-to-end agentic design, where a unified system handles the full pipeline.

Priya: The 2011 performance benchmark is actually the honest framing here. This isn't competitive with modern CPUs. But the claim that matters is that a unified agentic approach outperformed specialized AI tools for individual tasks. If that holds up to scrutiny, it suggests something important about how agency and context-awareness across a long workflow can compensate for raw per-task performance.

Sam: It's early. The RISC-V ISA is intentionally simple compared to x86 or ARM, which makes this a more tractable design target. But the trajectory from 2020 GPT-2 generating logic fragments, to 2023 GPT-4-assisted 8-bit processors, to a full functioning CPU core in 2026 — that's a real progression.

Priya: Two quick industry notes. Mira Murati's Thinking Machines Lab signed a multi-billion-dollar Google Cloud deal for infrastructure powered by Nvidia's GB300 chips. Murati founded the lab after leaving OpenAI, and landing this kind of compute commitment this fast says something about how seriously Google Cloud is taking the strategy of hosting frontier AI startups as anchor tenants.

Sam: And SpaceX apparently has an option to acquire Cursor — the AI coding assistant — for sixty billion dollars. The strategic logic: Cursor has the developer platform and workflow integration, xAI has models, and together they'd have a more complete stack to compete with Anthropic and OpenAI in the developer tools market. The number is eyebrow-raising. But the underlying gap it's trying to close is real — neither Cursor nor xAI currently has models competitive with Claude or GPT-5 for coding tasks.

Priya: And one more. Wired reviewed permits for gas-powered data center projects linked to OpenAI, Meta, Microsoft, and xAI. Combined projected emissions: over 129 million tons of CO2 per year. That would exceed the annual emissions of entire nations. These aren't hypothetical numbers — these are permits that have been filed.

Sam: The regulatory and political surface area for AI infrastructure just got a lot more concrete. Hard numbers attached to named companies make this a different kind of conversation than general concern about AI's carbon footprint.

Priya: So looking ahead — what does today's news actually point toward?

Sam: The TPU 8 split is the one I'll be watching most closely. If Google's dual-track architecture delivers meaningful cost reductions on inference, it creates real pressure on the economics of running large models. That affects every lab's pricing strategy and every enterprise's build-versus-buy calculation.

Priya: For me it's Mythos and what it represents for the broader category. We now have clear evidence that specialized AI security models are operating at expert-researcher level. The questions that opens up: what does responsible deployment of these models actually look like at scale, who bears liability when access controls fail, and how fast are other labs building in this direction?

Sam: And the agentic chip design story is one to keep tracking — not because VerCore is competitive today, but because the question of whether AI can compress hardware design timelines is enormously consequential for the pace of the whole AI development cycle.

Priya: If AI can design better AI chips faster, that feedback loop becomes very interesting very quickly.

Sam: That's it for today. Thanks for listening to AI Revolution. We're back tomorrow with whatever the next 24 hours bring — which lately has been quite a lot.

Priya: Find us wherever you get podcasts. See you then.