Sam: Five hundred and ninety billion dollars. That's what Samsung and SK Hynix are about to pour into new chip factories and packaging centers, backed by the South Korean government. To put that in perspective, that's roughly the GDP of Sweden. And it's aimed at one bottleneck: memory bandwidth for AI. These two companies control nearly 80 percent of the global high bandwidth memory market, and analysts at Jefferies are projecting memory prices could climb 50 percent per quarter through 2027. So if you're planning AI infrastructure budgets right now, the cost curve you're modeling is probably wrong.

Priya: Welcome to AI Revolution for Monday, June 29th, 2026. I'm Priya Nair.

Sam: And I'm Sam Kim.

Priya: We have a packed show today. We're going to dig into that massive memory investment and what it means for anyone building on GPUs. China just reclaimed the world's fastest supercomputer despite US export controls — we'll explain how. Mozilla's security team found a really elegant and really scary attack vector against AI coding agents. We've got a Chinese open-weight model making waves in cybersecurity, a fascinating accidental discovery in neuromorphic computing, and Ford admitting it was wrong about replacing experienced engineers with AI. Let's get into it.

Sam: So let's start with the Samsung and SK Hynix story because this is about the physics of what's actually constraining AI scaling right now. When people think about AI hardware bottlenecks, they tend to think about GPUs — can I get enough H100s or B200s or whatever the latest Nvidia chip is. But there's an equally important constraint that gets less attention, which is memory bandwidth. Modern AI accelerators are often memory-bound, meaning the chip can compute faster than it can feed data to the compute units. High bandwidth memory, or HBM, is the technology that stacks memory dies vertically and connects them with through-silicon vias to achieve much higher data throughput than conventional memory. Every major AI accelerator — Nvidia's GPUs, AMD's MI series, Google's TPUs — depends on HBM.

Priya: And what makes this investment story significant beyond the sheer dollar amount is the market structure. Samsung and SK Hynix together control roughly 80 percent of HBM supply. Micron is the third player, but it's a distant third. So when these two companies signal that demand will outstrip supply for years, that's not speculative — they have the order books. The $590 billion is going toward new fabs and advanced packaging facilities, because packaging HBM is actually one of the hardest parts. You're stacking eight or twelve dies, connecting them with thousands of through-silicon vias, and the yield challenges are significant.

Sam: The Jefferies projection of 50 percent quarterly price increases through 2027 is worth sitting with. If that holds even approximately, the memory component of training a frontier model could double or triple within a year. That changes the economics of who can afford to train at the frontier. It also changes inference economics — every GPU in every data center needs HBM, and if that HBM costs dramatically more, the cost per token goes up across the board.

Priya: And this connects directly to our second story, which is China's LineShine supercomputer displacing El Capitan at the top of the TOP500 ranking. This is the first time China has held that top spot since 2018. The significant detail here is that this happened despite US export controls that are specifically designed to prevent China from acquiring the most advanced computing components.

Sam: Right, so the question everyone should be asking is: how? The US has been tightening export restrictions on advanced chips and semiconductor equipment since 2022. China can't buy Nvidia's top-tier GPUs. They can't buy ASML's most advanced lithography machines. And yet they've assembled a system that benchmarks faster than El Capitan, which runs on AMD's latest Instinct accelerators. The answer is a combination of domestic chip development — companies like Huawei's HiSilicon and others we have less visibility into — and clever system architecture, particularly around interconnects. You can compensate for individually less powerful processors with enough of them and fast enough networking between them. It's a brute force approach in some ways, but the engineering required to make it work at this scale is genuinely impressive.

Priya: And to be clear about what TOP500 measures — it's a specific benchmark called LINPACK, which is dense linear algebra. It's not a direct proxy for AI training capability. But the underlying engineering achievements in domestic chip fabrication and high-speed interconnects absolutely do transfer to AI workloads. The US still dominates the overall TOP500 list, but the trend line matters.

Sam: Now let's shift to something that should concern anyone deploying AI coding agents. Mozilla's 0DIN security team demonstrated an attack against Claude Code — and I want to be precise about what they showed, because the mechanism is clever. They created a GitHub repository that looks clean. You can scan every file, and there's nothing malicious. The AI agent can inspect the code, and it sees nothing wrong. But the setup script contains a DNS lookup that fetches a malicious payload at runtime. The code that actually does the damage doesn't exist in the repository at all — it's delivered dynamically when the code executes.

Priya: This is a supply chain attack that exploits a fundamental gap in how AI coding agents evaluate trust. The agent is designed to analyze code it can see. Static analysis tools scan code that's present in the repo. But this payload is assembled at execution time from an external source. It's like checking every ingredient in a recipe but not noticing that one step says "add whatever the chef hands you." The result is full machine takeover from a single compromised repository.

Sam: And the deeper issue is about trust boundaries for agentic systems. When a human developer runs an unfamiliar setup script, they might be cautious, might run it in a sandbox. An AI coding agent operates with whatever permissions it's been granted and follows the instructions in the repository. The agent doesn't have the contextual suspicion that would make a human pause at a DNS lookup in a build script. This is a solvable problem — sandboxing, network restrictions, runtime monitoring — but it requires explicitly designing for it. The default posture of most agentic coding setups right now is too permissive.

Priya: Related to the security theme, Zhipu AI — the Chinese lab also known as Z.ai — released GLM-5.2 as an open-weight model, and it's showing competitive performance with frontier Western models specifically on cybersecurity benchmarks. Bug finding, vulnerability analysis, that category of tasks. It still lags on general reasoning, but the security-specific capability is notable.

Sam: The dual-use implications are real. An open-weight model with strong vulnerability discovery capabilities is available to anyone without API restrictions or usage policies. That's useful for defenders and for attackers alike. And it signals that security-relevant AI capability is diffusing globally faster than some policy frameworks assumed.

Priya: Let's talk about something more speculative but potentially transformative. IEEE Spectrum reported on research into artificial neurons implemented directly on silicon — and the origin story involves an accidental lab discovery. The core idea is building analog circuits that mimic how biological neurons process information, rather than simulating neural networks digitally on conventional processors.

Sam: The energy numbers frame why this matters. A modern AI GPU consumes up to a thousand watts. Your smartphone uses about one watt. Your brain — which is doing something like what these AI systems are approximating — runs on about twenty watts. There's a thousand-fold gap between current AI hardware and biological efficiency, and neuromorphic approaches could close a significant portion of that gap. The research is exploring silicon circuits that exhibit neuron-like spiking behavior natively in the physics of the device, rather than computing it. Early stage, but if it pans out, the implications for edge AI and sustainable scaling of inference are enormous.

Priya: And then we have what I think is one of the more honest stories we've seen in a while. Ford publicly acknowledged that it was wrong to believe AI could replace experienced engineers in complex manufacturing. They're actively rehiring the veteran engineers — the so-called gray beards — they had previously let go.

Sam: The quote from Ford leadership was remarkably direct: "Mistakenly we thought that by just introducing artificial intelligence, that would produce a high-quality product." This is a useful data point about where AI's boundaries actually are right now. Manufacturing engineering involves decades of accumulated judgment about material behavior, failure modes, tooling interactions — the kind of contextual expertise that current AI systems genuinely struggle with. It's not pattern matching on well-structured data. It's reasoning about physical systems with incomplete information under real consequences.

Priya: And worth noting — this doesn't mean AI is useless in manufacturing. It means it's a tool that amplifies expert judgment rather than replacing it. Ford learned that the expensive way.

Sam: One more quick item — Scam.ai announced a partnership with Qualcomm to run deepfake detection directly on-device during live video calls. The model, called Halo, runs on Qualcomm's NPU without sending video to the cloud. As deepfake quality improves, having real-time detection at the endpoint rather than depending on cloud round-trips is architecturally the right approach.

Priya: So looking ahead, what ties today's stories together for me is the question of constraints — physical, economic, and cognitive. The Samsung and SK Hynix investment tells us memory bandwidth will remain a binding constraint on AI scaling. China's supercomputer tells us export controls are a leaky constraint on compute access. The Claude Code attack tells us our security models haven't caught up to the trust boundaries that agentic AI requires. And Ford tells us domain expertise is a constraint you can't just replace.

Sam: I think the memory story is the one to watch most closely. If prices really do climb at anything close to 50 percent per quarter, we'll see real pressure on the mid-tier of AI companies. The hyperscalers can absorb it. Startups with thin margins can't. That could accelerate consolidation in ways that reshape the competitive landscape of AI development. Meanwhile, research into alternative compute paradigms like neuromorphic chips becomes more urgent, not less, as conventional approaches get more expensive.

Priya: And on the security side, the Claude Code vulnerability is a preview. As AI agents get more autonomous and more integrated into development workflows, the attack surface expands in ways we're only beginning to map. Every team deploying agentic tools needs to be thinking about sandboxing and runtime monitoring now, not after the first incident.

Sam: That's the show for today. Show notes and links to everything we discussed are at cleartext.fm.

Priya: Thanks for listening. We'll see you tomorrow.