AI Daily Dev — July 23, 2026

AMD Launches MI400 Series and Helios Rack With 12 GW Booked by OpenAI and Meta

Day 2 of Advancing AI 2026: MI455X enters volume production — 40 PFLOPs FP4, 20 PFLOPs FP8, 432 GB HBM4 and 19.6 TB/s per chip.
Helios rack packs 72 MI455X GPUs, 31 TB pooled HBM4, 2.9 exaFLOPS FP4 — priced ~$5.25M, 40% above Nvidia's second-gen Rubin.
EPYC 'Venice' launches alongside: first x86 server CPU on TSMC 2nm, 256 cores and 512 threads per socket, ~1 GB of L3.
OpenAI + Meta booked 12 GW of AMD accelerator capacity across generations; Microsoft Azure (ND MI455X v7), Oracle, and Tata are early Helios customers.
MI500-series previewed for 2027; AMD says 8 of the world's top 10 AI companies now run workloads on Instinct GPUs.

industry datacenterdynamics.com

Limited GA on July 22 for eligible enterprises — not a self-serve API, delivered as a deployed product.
Bundles company knowledge, SOPs, approved actions, simulations, evals, guardrails, GPT-Live realtime voice, and a Codex-powered improvement loop.
OpenAI's own support line: 75% of inbound issues resolved without a human; agent-to-human handoffs down 15 points in 10 days.
Early customers include BBVA (Mexico banking), SoftBank (Japanese-language support), and Australian insurer IAG.
HN thread is skeptical — top comments argue an in-house MCP server plus M365 Copilot beats leaning on OpenAI consultants.

tools openai.com

New rates land July 27: critical drops from $20k–$30k+ to a flat $10k.
Medium/high/critical all ≥50% lower against the old ranges; low-severity down about 59%.
Invite-only permanent VIP tier keeps $30k+ payouts, faster triage, and closer access to GitHub's security engineers.
Reports filed before July 27, including everything in the current queue, are grandfathered on the old terms.
Follows a May 2026 policy demanding working PoCs; framing: 'you don't earn more by submitting more.'

industry github.blog

New MCP connector shipped July 22 in the Claude Directory — no install, works in every conversation and every Claude model.
Ask Claude about how AI is used across occupations, regions, and tasks; answers cite Index data and expose the underlying rows.
Full datasets stay free on anthropic.com; Claude flags that the Index reflects Claude usage, not the whole labor market.
Slots into Anthropic's growing connector directory (343+ verified MCP integrations).

tools anthropic.com

Five priority areas: workplace impact, workforce transitions, income-support modernization, worker equity, and public-investment evidence.
Fund was announced June 10 alongside a $150M national fellowship; today's post sets the call-for-proposals shape.
Lands as Anthropic's Q2 2026 federal lobbying rose 26% to $1.97M and the company pushes states on frontier-AI rules.
Companion piece to the same-day Economic Index connector — Anthropic doubling down on being the lab that instruments the AI-labor debate.

industry anthropic.com