Sam: OpenAI and a coalition of major hardware and software companies just published an open-source networking protocol called MRC. It lets a cluster of over 100,000 GPUs communicate using only two switch layers instead of the three or four that conventional fat-tree architectures require. That's not a minor optimization — the number of switch layers directly determines network depth, latency, cost, and power draw at scale. It's already running on OpenAI's Stargate supercomputer.

Priya: Welcome to AI Revolution for Thursday, May 7th, 2026. I'm Priya Nair, here with Sam Kim. Today we're deep in infrastructure — two stories that show just how aggressively compute capacity is being built out right now. We've also got Gemma 4 doing something genuinely interesting with inference speed, a new learning mechanism for Claude agents, some alignment research worth paying attention to, and a handful of industry and policy developments. Let's get into it.

Sam: So let's start with MRC, because the technical idea here is worth unpacking. Traditional GPU cluster networking uses a fat-tree topology — you have spine switches, aggregation switches, and top-of-rack switches. Each hop adds latency and cost, and the more GPUs you're connecting, the more switch layers you need to avoid bottlenecks. The standard answer has been three or four layers for anything approaching 100K GPUs.

Priya: And the reason that matters is that GPU utilization in training runs is heavily dependent on how fast gradients and activations can move between accelerators. If your network is the bottleneck, your GPUs are waiting. And at 100K+ GPUs, even small inefficiencies compound.

Sam: Right. What MRC does differently is multi-path routing — it sends data across hundreds of paths simultaneously rather than routing packets down a single path through the hierarchy. By spreading traffic that broadly, you can flatten the topology. Two switch layers becomes sufficient because you're not creating hot spots on any single path. The bandwidth aggregates across all those parallel routes.

Priya: And the fact that this is open-source, with AMD, Broadcom, Intel, Microsoft, and NVIDIA all involved, matters a lot. This isn't one vendor's proprietary fabric. If this gets adopted broadly, it could affect how every large AI cluster gets built — and the power savings from eliminating a whole switch layer at scale are meaningful. We're talking about facilities drawing hundreds of megawatts.

Sam: Which brings us directly to the second infrastructure story. Anthropic is taking over the entire Colossus-1 data center — the one SpaceX built for xAI in Memphis. We're talking 220,000 NVIDIA GPUs, over 300 megawatts of capacity, expected online within a month.

Priya: The sheer scale of this is worth sitting with. 300 megawatts is roughly the output of a mid-sized power plant, dedicated to running one lab's models. And Anthropic says they're immediately using this to double rate limits for Claude Code and significantly raise API limits on Opus. So this isn't a build-for-the-future story — it's addressing a current capacity crunch.

Sam: The cross-company dimension is also notable. Anthropic leasing xAI's facility is a pragmatic compute deal that cuts across what you'd normally think of as competitive boundaries in this space. It tells you something about how tight GPU supply still is — you take capacity where you can get it.

Priya: Let's move to Gemma 4, because the speculative decoding story is one I've been waiting to see land in open-weight models at this scale. Sam, can you walk through what speculative decoding actually is?

Sam: Sure. Standard autoregressive inference generates one token at a time — the model does a full forward pass to produce each token sequentially. Speculative decoding changes this by using a smaller, faster draft model to predict several tokens ahead, then using the full model to verify those predictions in parallel. If the predictions are right, you've effectively generated multiple tokens per full model pass. If they're wrong, you fall back. The math works out to significant speedups when the draft model's accuracy is high enough.

Priya: And the key word there is "native" — Google baked this into Gemma 4's architecture rather than bolting it on afterward. That's what makes the three-times speedup credible. The draft model and the verification model were co-designed.

Sam: Three times is at the high end of what you'd expect — you typically see one-and-a-half to two-and-a-half in practice depending on the task. The claim of no quality loss is plausible if the draft model is well-calibrated, but that's something developers will need to validate on their specific workloads. The broader point is that this substantially changes the economics of running open-weight models locally or at the edge. You can now get throughput that was previously only achievable with larger serving infrastructure.

Priya: And open-weight means developers can actually inspect and customize this, which matters for production deployments where you want predictability.

Sam: Shifting to the Claude agent updates — Anthropic announced something they're calling Dreaming, which is an asynchronous process that runs after agent sessions end. It reviews what happened, cleans up redundant or outdated memory entries, and distills new insights that carry forward into future sessions.

Priya: The analogy to biological consolidation is obvious and probably intentional. The idea is that agents currently forget everything between sessions, or they accumulate memory stores that get noisy over time. Dreaming is a maintenance pass that curates that memory. It shipped alongside Outcomes — which lets agents evaluate their own task performance — and Multiagent Orchestration, now in public beta.

Sam: The combination of those three is interesting architecturally. Outcomes gives the agent a signal about whether it succeeded. Dreaming uses that signal to update what it remembers and how it weights future decisions. Orchestration lets multiple agents coordinate. That's a fairly complete loop for persistent, improving agentic systems — without needing to retrain the underlying model.

Priya: Though we should be clear that this is learned behavior within the memory layer, not weight updates. The base model isn't changing. The question for practitioners is how well the distillation step actually generalizes and whether the memory consolidation introduces new failure modes.

Sam: There's also a piece of alignment research out of the Anthropic Fellows Program that connects to this. The finding is that training models on texts explaining the rationale behind intended values — before training them on specific behaviors — leads to significantly better adherence to those values in novel situations. Not just in distribution, but in cases the model was never explicitly trained on.

Priya: This is a meaningful result for alignment methodology. The conventional approach has been behavioral: you show the model examples of good behavior, it learns the pattern. What this suggests is that understanding the "why" behind a value gives the model something it can actually reason from when it encounters an edge case.

Sam: It's analogous to how you'd want to train a human engineer. Teaching them principles produces more reliable behavior than just teaching them rules, because rules don't cover every case. The question is whether this scales and whether the effect holds under adversarial pressure — but it's a concrete, empirical finding that safety teams can act on.

Priya: Quick note on DeepSeek: the lab is reportedly targeting a $45 billion valuation in its first external funding round. To put that in context, they came to prominence in early 2025 training competitive frontier models at a fraction of the compute cost of US counterparts. A $45 billion first-round valuation would put them in the same tier as established US frontier labs. The efficiency-first approach has real market validation now.

Sam: On the policy front, two things worth flagging. The Pentagon has added Microsoft, Amazon, NVIDIA, and Reflection AI to its approved supplier list for classified operations — they join OpenAI, xAI, and Google. Reflection AI is notable because they haven't released a publicly available model yet, but they've apparently cleared the bar for classified use. Anthropic's government role is described as being reconsidered, which is a conspicuous data point given the Colossus deal and rate limit expansions they're focused on.

Priya: And separately, the US and China are reportedly exploring formal bilateral talks specifically on AI, per the Wall Street Journal. This would be the first structured diplomatic engagement between the two countries focused specifically on AI governance. Given where export controls and chip restrictions stand, even getting to a formal table would be significant. Nothing is confirmed yet, but it's worth watching.

Sam: One more technical item before we look ahead — Google announced at Cloud Next that GKE is getting an Agent Sandbox built on gVisor kernel isolation. It can spin up 300 sandboxes per second, which is the right order of magnitude for dynamic agent workloads. The companion announcement is Hypercluster, which manages up to a million chips from a single Kubernetes control plane.

Priya: The sandbox piece matters for anyone thinking seriously about deploying agents in production. Agents that execute code need real isolation — not container-level, but kernel-level. gVisor gives you that with much lower overhead than a full VM. The fact that this is open-source and being proposed as a Kubernetes SIG subproject means it could become infrastructure that the whole ecosystem builds on.

Sam: So stepping back — today feels like a day where infrastructure is doing a lot of the talking. MRC flattens the network layer for massive GPU clusters. Anthropic absorbs 220,000 GPUs to address an immediate capacity constraint. Google is building out both the compute management layer and the secure execution layer for agents.

Priya: The thread I keep pulling on is: what does this level of infrastructure investment unlock that wasn't possible six months ago? The answer seems to be longer-running, more persistent agents with more reliable memory — and inference that's fast and cheap enough to use those agents interactively at scale.

Sam: The questions I'm watching: does MRC get adopted outside of Stargate, or does it stay proprietary in practice despite being open-source? How quickly does the Colossus capacity translate into observable improvements in Claude's capabilities and availability? And on the alignment research — does the "teach the why first" finding replicate across different model families?

Priya: And on the geopolitical side, whether those US-China AI talks actually materialize into formal talks is worth following closely. The structural tensions around compute access and safety standards haven't gone away.

Sam: That's going to do it for today. Thanks for listening to AI Revolution. Show notes and links to everything we covered today are at cleartext.fm. We'll be back tomorrow.

Priya: See you then.