The week of August 29 – September 5, 2026 delivered three storylines engineering leads must track: Grok 4.7 is ten days out with 2.1T parameters trained on aerospace telemetry; ChatGPT, Claude, and Grok went dark within 90 minutes on September 3; and Grok Bot for Enterprise crashed the persistent-agent entry price to $20/month.

TL;DR: (1) Grok 4.7 targets September 12 — 2.1T params, 40% larger than 4.6, SpaceX engineering data in training; prepare eval workflows now. (2) Simultaneous AI outage exposed cloud concentration risk — build multi-provider fallback. (3) Grok Bot Enterprise at $20/month (Cursor Pro), isolated VMs, audit controls, two-week free trial. (4) Tesla summer update: Grok controls vehicle functions by voice, FSD supervisor mode in three months. (5) Grok 4.6 won a biosecurity benchmark outright — only model above 50% on both refusing dangerous bio requests and completing legitimate bio work.

Grok 4.7: 2.1 Trillion Parameters, SpaceX Data, September 12 Target

Musk posted September 2 that “Grok 4.7 comes out in 10 days.” The model is 2.1T parameters — 40% above Grok 4.6’s 1.5T. Pre-training completed August 12; the final phase ingests SpaceX engineering records, telemetry, and Starlink satellite data. No other AI lab has aerospace telemetry at this scale.

Known: 40% larger, higher token efficiency, possibly slower inference. Unknown: No model ID, pricing, context window, or benchmark card as of September 5. Timeline slipped twice (August 22 → “3-4 weeks” → “10 days”). Treat September 12 as a target.

Actions: Prepare eval pipelines for a frontier model drop. If you’re on Grok 4.6 via API, Bedrock, Vertex, or Foundry, check rollout timelines. Benchmark against your own eval suite, not xAI’s numbers.

The Great AI Outage of September 3

At 9:30 AM ET, Grok returned errors. By 10:30 AM, Claude followed. By 11:00 AM, ChatGPT failed. All restored by ~3:00 PM ET.

Grok’s outage traced to xAI’s Memphis Colossus data center — hardware, per official statement. OpenAI and Anthropic cited separate causes. No common root cause identified; speculation points to shared Azure East US infrastructure, unconfirmed.

Three independent platforms with different architectures failing within 90 minutes breaks the independent-failure assumption. If your agent infrastructure depends on a single LLM provider — or providers sharing the same cloud backbone — September 3 is your warning shot.

Actions: Build multi-provider routing into agent infrastructure. OpenClaw’s fallback chains and provider-agnostic routing are the right pattern. Test failover before you need it.

Grok Bot for Enterprise: $20/Month Entry, Isolated VMs

xAI launched Grok Bot for Enterprise September 3. Pricing is aggressive:

  • Two-week free trial for Grok and Cursor Enterprise customers
  • Cursor Pro $20/month — lowest paid tier with Grok Bot access
  • SuperGrok $30/month, Cursor Teams Standard $40/seat/month
  • Top tier: SuperGrok Heavy $300/month

Agents run on isolated cloud computers with dedicated VMs — executing commands, browsing the web, interacting with applications, working offline. Governance includes access, network, and audit controls. Invite an entire organization, including people without seats. Use-case framing: “junior teammate” — sales follow-ups, candidate screening, code auditing. Weekly usage allowances per plan; overage billed as on-demand tokens.

Actions: $20/month makes piloting trivial. Start with a non-production workflow — code audit, document summarization, research triage. Evaluate audit logs and isolation before production access.

Grok Bot’s X Integration and Plugin Expansion

On August 29, xAI shipped the first native X connector for Grok Bot — read-only: search posts, read timelines, check mentions, summarize activity, manage bookmarks. xAI auto-creates an X developer account; paid users get free X API credits. No posting or DMs — “the first version.” Bots can still post via external MCP servers (Blotato).

Beyond X, plugins expanded: Microsoft (Sep 1): Outlook, Calendar, OneDrive; Firecrawl: agentic web crawling; Stripe Link: bots can purchase online; webhook triggers, template sharing, Android app (Sep 2), X Ads MCP.

Stripe Link deserves attention — purchasing agents introduce a new threat surface. Evaluate spending limits and approval workflows before enabling.

Tesla’s Summer Update: Grok as Vehicle Controller

Tesla’s 2026.26 update brings Grok Voice Think Fast 2.0 into the car as a vehicle controller — climate, heated seats, Dog Mode, Camp Mode, Bioweapon Defense Mode, glovebox, mirrors, wipers, navigation, tire pressure, Dashcam, Sentry Mode, charge port.

Speech-to-speech architecture: reasons while speaking, not after. Requires AMD infotainment, software 2025.26+, Premium Connectivity or Wi-Fi. No on-device fallback. Still Beta.

Musk indicated voice input reaches FSD’s planning layer in ~3 months — first time speech influences driving behavior. Robotaxi “supervisor” mode coming: “drop me at the entrance, then go park.”

The same models run Optimus humanoid robot’s voice interface. Starlink uses Grok Voice for 15,000+ daily support calls — diagnosing hardware, shipping replacements, 3,000+ orders weekly. Cybercab (no steering wheel) launched September 3 in Austin.

Actions: For voice interface teams, speech-to-speech is the pattern — reasoning during generation eliminates latency that makes agents feel robotic. Tesla proves frontier LLMs can handle real-time control surfaces.

Grok 4.6 Biosecurity Benchmark

On September 1, xAI highlighted LatchBio’s independent BioSecBench-Refusal analysis. Grok 4.6 was the only model above 50% on both metrics: ~59.2% refusal of disguised hazardous bio requests, ~64.8% completion of routine biological work (harmonic mean ~62.1%). On Surveillance (~53.5%), it trailed Opus 5 but beat GPT-5.6 Sol. Refusals are model-reasoning driven, not API filters.

For healthcare, pharma, or biotech, refusal rates without false-positive paralysis determine whether a model is usable in production.

SpaceXAI Corporate Context

SpaceX acquired xAI in February 2026 (~$250B all-stock), then Cursor in August. SpaceX IPO’d June 2026 — largest on record. Grok 4.6 runs on xAI API, Cursor, Copilot, Bedrock, Vertex AI, Foundry, OpenRouter, Vercel, Cloudflare. Pricing: $2/M input, $6/M output (<200K tokens); $4/M input, $12/M output (>200K). Cached: $0.50/M (xAI, OpenRouter), $0.30/M (Vertex AI).

What to Watch

  • September 12 — Grok 4.7 target release. Expect slippage; plan for it.
  • Next 3 months — Tesla FSD voice input reaching planning layer. First speech-to-driving integration.
  • Grok 5 — Targeted before end of 2026. No details beyond monthly cadence promise.
  • Stripe Link — Monitor for spending-control incidents as agents gain purchasing capability.

Looking Ahead

This week’s stories converge on one theme: the boundary between digital and physical AI is dissolving. Grok 4.7 trains on spacecraft telemetry. Grok Voice controls Teslas. Grok Bot spends via Stripe. Starlink’s support line is staffed by Grok. The simultaneous outage proved this convergence creates systemic risk — when your car, coding agent, support line, and research assistant depend on the same model family through the same cloud backbone, one data center failure takes down everything.

Teams that manage this well treat multi-provider routing as infrastructure, not a nice-to-have. They evaluate agent isolation before the next incident, not after. And they pilot Grok Bot at $20/month — the cheapest way to learn what enterprise agents can and can’t do is to run one against a real workflow this week.

Follow the ongoing analysis on X: https://x.com/kkaminsk