Quick Hits — September 3, 2026
A G20 consensus on lighter-touch AI rules, an MCP spec rewrite that quietly deletes the protocol's handshake, and a phone-side zero-knowledge identity kit from World. Plus: Google recycling DDR4 from retired servers and a fresh look at China's grip on U.S. data-center power gear.
The G20 just signed on to a U.S.-shaped AI rulebook — and the biggest lab CEOs were in the room. All G20 nations endorsed the non-binding "Carolina Principles" at the summit, a U.S.-proposed framework that pushes for lighter, sector-specific AI rules over horizontal statutes. Separately, the Wall Street Journal reports Jensen Huang, Mark Zuckerberg, Sam Altman and Elon Musk were on hand lobbying policymakers against heavy regulation alongside Trump officials. The endorsement is voluntary, but unanimous: every member government signed. The take: the political alignment is real but the teeth aren't — non-binding + sector-specific is a policy that lets everyone in the room say they did something, and the labs already know it.
MCP just deleted its own handshake. The 2026-07-28 revision of Anthropic's Model Context Protocol removes the initialize handshake and the session ID that came with it, pushing capability signaling onto per-request metadata and a new server/discover call so any instance behind a load balancer can answer any request without stickiness. Roots, Sampling and Logging are formally deprecated (earliest removal after July 2027); the old HTTP+SSE transport is on a faster three-month clock; and the new "Multi Round-Trip Requests" replaces the old server-push elicitation with an input_required result the client retries. The spec also gets a real lifecycle policy and an extension registry so new features (MCP Apps, Tasks, a Skills working group) graduate through their own repos instead of fighting into the core. The take: this is the boring kind of important — stateless, serverless-friendly, with security duties that have to move into the application layer whether anyone is ready or not.
World has open-sourced ProveKit, a zero-knowledge proving toolkit that runs entirely on a phone. Released under MIT, the toolkit lets a device generate ZK proofs of age, nationality or document validity without uploading the underlying ID — an answer to last week's leak of 170M identity documents on a dark-web service called Nexus. The cryptography leans on the WHIR IOP and hash-based commitments (no pairings), and ProveKit's engineering win is fitting it inside phone-class memory: a few seconds on a typical device, under 30 on a low-end test handset, with offline support. For on-chain use it currently exports a Groth16 wrapper; v2, in development, drops the wrapper. The take: the cryptography is the conservative part — the hard part was making it fit on a handset, and the proof this works as a verification channel comes when a regulated verifier outside World's orbit accepts it.
Google is pulling DDR4 modules out of retired servers and reusing them in new AI racks. Google's supply-chain senior director Nikhil Cherian disclosed at a summit that the industry has flipped from compute-constrained to memory-constrained, with high-performance memory now around 75% of an AI server's bill of materials. The fix is unglamorous: an internal recycling loop that dismantles decommissioned servers, salvages the DDR4 sticks, and redeploys them alongside the new TPUs while software work (KV-cache compression, model architecture) tries to cut memory per unit of compute. The take: when a hyperscaler is tearing down old boxes to keep new ones fed, the DRAM shortage has clearly passed the "soft" line — and it confirms the CXMT / Micron thesis that memory is now the real lever on AI capacity, not silicon.
Trump's "national emergency" on bulk-power equipment lands in the middle of a multibillion-dollar AI data-center buildout. A late-August executive order lets the Energy Department prohibit or condition transactions involving foreign-made transformers, switchgear, batteries and similar grid components — and analysts say the U.S. data-center build is exposed in exactly those categories. Chinese suppliers hold around 30% of certain transformer and switchgear segments, over 40% of battery imports, and roughly two-thirds of data-center optical-transceiver unit supply. U.S. capacity is still forecast to climb from 62 GW in March 2026 to 152 GW by 2030, and Western alternatives aren't ready to absorb that demand. The take: Washington has been hammering chip and compute supply chains for two years — the power stack is the next front, and unlike chips there isn't a domestic substitute waiting to ramp.
